In May, within days of each other, the two largest futures exchanges in the world announced plans to list contracts on the price of AI compute. ICE, which owns the New York Stock Exchange, partnered with a startup called Ornn, whose Compute Price Index, or OCPI, tracks the open-market cost of renting Nvidia’s H100, H200, B200, and RTX 5090 chips. CME Group had announced its own compute futures with a rival index provider, Silicon Data, the week before. A smaller venue, Architect Financial Technologies, had already said in January it would launch futures on GPU and memory prices using Ornn’s data. Kalshi will let you bet on Nvidia rental prices today. More recently, Ornn raised $33 million led by Andreessen Horowitz and says the first compute swap against its cleared prices was executed in December.

The pitch, repeated almost word for word across every announcement, is that compute has become a trillion-dollar market without the pricing and risk-transfer tools every other major commodity relies on. CME’s chief executive Terry Duffy went further: “Compute is the new oil of the 21st century.” The comparison is deliberate. ICE’s Brent contract and CME’s WTI contract emerged from a similar race between exchanges in the 1980s and went on to define the global reference price for crude.

The oil analogy seems correct, but perhaps not in the way the exchanges mean it.

The case for the contracts

Start with the strongest version of the case. GPU rental prices swing violently. Ornn’s own index showed the rental price for Nvidia’s Blackwell chips rising 48 percent between mid-February and mid-April, from $2.75 to $4.08 per GPU-hour. For a lab budgeting a training run in the tens of millions of dollars, that kind of swing can destroy a plan mid-execution. Data center operators carry the mirror-image risk: they borrow billions against future rental revenue whose price they cannot lock in. Lenders financing the buildout are exposed to both.

A futures contract solves this in principle. We know the rough shape of these instruments from oil and wheat: you fix a price today for something delivered later. What’s worth working out is what that machinery means when the underlying thing is compute. These particular contracts are cash-settled, meaning nobody ships a GPU anywhere; when the contract expires, the two sides simply exchange the difference between the agreed price and the index price. Which puts enormous weight on the index. It is the number the entire market settles against, so the question worth sitting with is not whether you can trade compute but who constructs the number.

Ornn’s answer is that its index is built only from completed trades, not surveys or dealer quotes. That is a genuine virtue, and a sign the industry has learned from LIBOR, the interest rate benchmark that quietly governed hundreds of trillions of dollars in contracts until it turned out to be manipulable. LIBOR was built from bank estimates, and estimates are cheap to shade. Real trades are harder to fake. But an index built on real trades is only as sturdy as the market it samples, and the open market for GPU hours is a small one. Most compute never touches it: the bulk of capacity is locked up in long-term private deals between the big cloud providers and the largest AI labs, invisible to any index.

What the commodity frame conceals

Commodity markets work because the underlying good is fungible and storable. A barrel of Brent-grade crude is interchangeable with another barrel; if prices drop, you can leave it in the ground or park it in a tank in Rotterdam. Compute is neither of these things.

An H100-hour is not an H100-hour. The same chip delivers wildly different effective performance depending on interconnect, cluster topology, networking, and the software stack around it. A GPU-hour on a tightly integrated cluster built for frontier training is a different product from the same nominal GPU-hour on a fragmented marketplace of loose capacity. The index has to flatten this heterogeneity to produce a single price, which means the benchmark describes a stylized commodity that the actual buyers of compute, the people the hedging tools are supposedly for, do not purchase.

And compute cannot be stored. An idle GPU-hour is gone. There is no tank farm for unused capacity, no way to buy cheap and warehouse it for a tighter market. Storability is what anchors futures prices to physical reality. Without it, the price of future compute floats free, tethered to expectations alone.

Then there is the problem no other commodity has: the reference asset is decaying on a schedule set by a single company. Every Nvidia generation reprices the last one. An index of H100 rental prices is structurally a melting ice cube, and any contract built on it embeds the market’s guess about Nvidia’s release schedule as much as any supply-demand fundamental. Trading compute futures is, in meaningful part, trading Nvidia’s roadmap. That is not a commodity market. That is a bet on the strategic decisions of one firm, wearing a commodity market’s clothes.

Benchmarks make markets, they don’t just measure them

There is a well-developed literature in the social studies of finance on how financial models and benchmarks don’t passively describe markets but actively constitute them. Donald MacKenzie put it in the title of his 2006 book: finance theory is “an engine, not a camera.” It doesn’t photograph the market. It drives it. An index becomes the price. Contracts settle against it, loan terms get written around it, procurement gets justified by it, and the messy underlying reality reorganizes itself around the number.

This is the part of the story that matters most and gets discussed least. Once the OCPI prints on a Bloomberg Terminal every day, “the market price of compute” exists as a social fact. It becomes citable. A government procuring sovereign compute capacity can point to it. A cloud provider defending its pricing to a regulator can point to it. A subsidy program can peg its disbursements to it. The number carries an aura of neutrality because it emerges from a market process rather than a boardroom.

But look at what generates the number: a small open market trading leftover capacity, sitting on top of a supply structure controlled by one chipmaker, a handful of hyperscale buyers, and long-term contracts nobody outside the negotiation ever sees. The benchmark takes concentrated market power as an input and emits an objective-looking price as an output.

Who gets to hedge

The democratization story says these instruments open compute price risk to everyone. Small AI labs can finally lock in training costs the way airlines lock in jet fuel. In principle, true. In practice, using futures markets at scale requires things small players don’t have: a relationship with a clearing member (the intermediary firms that guarantee trades on an exchange), the balance sheet to post margin (the collateral a position requires), and the cash reserves to survive the position moving against you. When prices swing hard, the exchange demands more collateral on short notice; fail to produce it and your position gets closed for you, at a loss, at the moment of maximum volatility. A hyperscaler’s treasury desk absorbs that without blinking. A twenty-person lab can get forced out of its own insurance policy by the very turbulence it bought the insurance against.

The entities that have all of this are the hyperscalers, the largest specialist GPU clouds, and the trading firms. The trading world is not arriving late, either: Silicon Data, the index provider behind CME’s contracts, is backed by DRW, one of the world’s largest trading houses, whose founder expects compute to “become the largest commodity in the world.” These are also the players that already lock up most global capacity through multi-year private purchase deals before it ever reaches the open market. For them, a futures market is not a replacement for privileged access but a second instrument layered on top of it: they can hedge their buildout risk, play their private contract prices against the public benchmark, and, as the largest participants in the small open market the index samples, lean on the number itself.

Brent and WTI created a financial layer whose returns accrued to the majors and to trading houses like Vitol and Glencore, firms whose edge came from seeing physical flows the paper market couldn’t. Everyone else got price transparency, which is not nothing, but transparency about a price you cannot influence and can barely afford to hedge is a consolation prize.

The fourth layer

I’ve been mapping how control over AI consolidates through infrastructure layers: the hyperscale compute layer, the CDN and edge layer that decides who gets scraped, the data layer being enclosed through licensing deals. Financialization is not a separate story. It is what happens to a choke point once it matures: the control becomes liquid. You can price it, lend against it, securitize it, and, crucially, defend it with a market’s legitimacy rather than a monopolist’s.

None of this means the contracts shouldn’t exist, and there’s a certain inevitability to them. If the futures work, some mid-sized players will genuinely benefit at the margin. But the question to ask about any new market infrastructure is not whether it is useful. It is who designed it, who settles against it, and whose position it consolidates. Compute is being turned into a commodity by the same actors who spent the last three years making sure it would never trade like one.

The 1980s oil race gave us a world where a benchmark price set in a financial market governed the physical economy of energy for forty years. The compute race is a bid for the same prize. Watch the index methodology documents, not the launch announcements. That is where the next forty years are being written.

Satya Nadella spent part of a June 21 WSJ interview warning that AI power is concentrating in too few hands, that a future shaped by a small club of frontier labs won’t earn public acceptance, and that what the industry needs instead is cheaper models and broader access to the benefits. A week earlier he made the same case at length in a post on X titled “A frontier without an ecosystem is not stable.” It is a clean, quotable position. It is also, conveniently, a description of the exact market structure that would route value toward Microsoft.

Let’s start with who actually has a moat at the model layer. As of mid-2026, roughly three companies account for the overwhelming majority of frontier inference. Microsoft is not one of them in the way OpenAI, Google, and Anthropic are. Its position is the layer underneath and around the models: Azure compute, Copilot seats, the enterprise integration and governance plumbing. If models stay scarce and expensive, the labs hold the leverage. If models get cheap, abundant, and interchangeable, the scarcity moves to whoever sells the picks and shovels. “Cheaper AI models and broader access” is not a neutral civic wish. It is a description of the world in which Microsoft’s assets appreciate and its suppliers’ assets commoditize.

The remedy he is shipping confirms the read. Copilot is going multi-model, letting customers pick among engines including cheaper ones, with DeepSeek reportedly under consideration. The labs this pressures most are the ones Microsoft is closest to. After OpenAI’s October 2025 recapitalization, Microsoft holds roughly 27 percent of OpenAI Group PBC, a stake worth around $228 billion at the company’s March 2026 valuation, on an original investment near $13 billion. That is a position most companies would protect at any cost. Microsoft is instead building the feature that lets its own customers route spend away from it.

Framed as pluralism, multi-model Copilot is pricing pressure applied to the lab Microsoft owns a quarter of, and to Anthropic, its other multibillion-dollar partner. A world of many substitutable models is a world where no single lab has leverage over Microsoft, which is the precise vulnerability that 27 percent created. “Democratize the model layer” and “neutralize my suppliers’ pricing power over me” turn out to be the same sentence.

Read the central warning closely and you can see the structure of the move.

The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see. If all the value is accrued by only a few models, the political economy will simply not tolerate it.

The danger he names is real. But notice the position it assigns Microsoft: the party raising the alarm about concentration, rather than the party whose proposed fix happens to commoditize the only firms ahead of it. The warning is sincere and self-serving at once.

The globalization analogy he reaches for is the tell. Nadella compares AI concentration to the first wave of offshoring, where, as he puts it, the headline numbers looked fine while industrial economies were hollowed out. It is a genuinely honest framing, and it reframes a narrow technology question into a political-economy one that regulators and voters can grasp. It is also worth turning back on him. In that analogy, the hollowed-out towns are the enterprises feeding their workflows into a handful of models. The party selling the infrastructure that captures the relationship is not among the casualties. Microsoft is positioned less as a victim of the dynamic than as the firm collecting rent on everyone’s attempt to escape it.

His proposed escape route is the “learning loop,” and it is where the self-interest is least disguised. Companies, he argues, should build their own AI systems on their own data so they retain proprietary knowledge rather than ceding it to outside models.

This is the key ‘test’ of your control and sovereignty in the era ahead.

Translated into spending, that is an argument for treating AI as durable infrastructure built on your own cloud tenant rather than a subscription to a model provider. It is hard to name a company better positioned to sell that infrastructure than the one running the tenant.

The sovereignty framing is sharper than it looks, and it is the part most worth watching. Telling enterprises that “control and sovereignty” means owning which models they depend on reads, intentionally or not, as a hedge against a single model provider getting disrupted, the line lands days after a major lab’s models were pulled under a government directive. This is platformized sovereignty in miniature: the promise of independence delivered through deeper dependence on the platform that brokers it. You are sovereign over your choice of model precisely to the extent that Microsoft sits underneath the choosing.

Then there is the antitrust framing, which is the part worth saying plainly. Microsoft is the company most exposed to a concentration narrative: the 27 percent OpenAI stake, the Anthropic partnership, Azure, and the Copilot bundle all sit in one place, and competition authorities on both sides of the Atlantic have spent two years asking whether the OpenAI “partnership” is an acquisition wearing better PR. The recap answered part of that question in Microsoft’s favor. It booked the gain, kept a still-large stake, gave up its exclusive Azure rights and board-level control, and walked away looking less like a controlling owner and more like a large passive shareholder. Being the AI giant who publicly worries about AI giants, having just restructured your way out of the tightest version of the dependency, is a way to shape the regulatory conversation rather than answer it later from a defensive crouch.

None of this requires the stated concern to be insincere. Concentration at the model layer is a real problem, and Nadella naming it is more useful than industry silence. The point is narrower. The remedy he proposes costs Microsoft almost nothing and happens to send value straight to the layer Microsoft owns. The companies that should be unsettled by cheap, broadly accessible models are the ones whose entire valuation rests on a moat at the model layer. Microsoft’s moat is everything else.

As of this Friday, the dependency stopped being theoretical.

At 5:21pm, Anthropic received an export-control directive from the US government: suspend access to its two most advanced models, Mythos 5 and the days-old Fable 5, for any foreign national, whether inside or outside the United States, on national-security grounds. Because a restriction keyed to nationality cannot be enforced selectively across a shared cloud, including for the foreign nationals already sitting in San Francisco offices, Anthropic’s own among them, the only way to comply was to take the models down for everyone. So they did. A flagship model went dark worldwide, three days after launch, on the strength of a single government letter.

Other Claude models stayed up, and the company called the order a misunderstanding it expected to resolve. But the lesson did not wait for the details. Overnight, European researchers, companies, and institutions were reminded that a chunk of the infrastructure they had quietly built their work on belonged to someone else, and that someone else answered to a government an ocean away. The reaction was immediate and loud: Europe is dangerously exposed, and it needs its own frontier models, fast.

The exposure is real. The conclusion feels half right.

Europe does need sovereign models. What it does not need is the fantasy attached to that instinct: the idea that sovereignty means matching the American frontier head-on. That is a reflex, not a strategy, and it points Europe straight at the one fight it cannot win while distracting from two it can. The reflex feels like resolve. It is actually a way of losing more slowly.

The fight Europe already lost

Frontier large language models are produced by a capital process Europe has no equivalent of, and wishing otherwise does not change that reality.

The leading labs are not ahead by a few clever ideas that a well-funded European competitor could reproduce. They are ahead by a self-reinforcing stack: tens of billions in compute, multi-year head starts compounding, and, crucially, the ability to fund the next training run out of the cash flows of the last one. This is the part that gets missed in the “we need a European champion” conversation. The American frontier is increasingly self-financing. The race is not a sprint you can enter with a big enough grant. It is a flywheel already spinning, powered by revenue and capital markets that have no European counterpart at the relevant scale.

To try to match that frontier is to pour sovereign money into a depreciating asset. By the time a European model caught today’s leaders, the leaders would be a generation ahead, and the European model would be a very expensive monument to the state of the art eighteen months ago.

Notice the precise claim, though. It is not that Europe should stop building large language models. It is that Europe should stop building them with the goal of winning. Those are different propositions, and collapsing them is the root of the confusion. There is a strong case for sovereign European LLMs. There is no case for a European frontier-LLM moonshot. Holding both thoughts at once is the beginning of a real strategy.

This is not a counsel of despair. It is a counsel of factor endowments: the unglamorous economic idea that you should compete where your inputs are cheap, not where they are scarce. And the moment you look at AI through that lens, the picture inverts.

What frontier LLMs actually consume

A frontier LLM is, in input terms, a machine for converting three things into capability: enormous compute, enormous quantities of (scraped) web-scale text, and the capital to keep buying both.

Look at that list from Brussels and it reads like an inventory of everything Europe is short of. Hyperscale compute? Concentrated in American and Chinese clouds. The open web’s text, and the legal-commercial machinery to hoover it up? Dominated by firms outside Europe. The capital markets that fund nine-figure training runs as a routine cost of doing business? Not here, certainly not at this scale.

This is why trying to win the frontier-LLM race is a trap. It is precisely the contest that rewards the inputs Europe lacks and penalizes the absence of inputs Europe cannot quickly manufacture. You could not design a competition more perfectly tilted against European strengths if you tried.

So the interesting question is not “how does Europe win the LLM race?” It is this: is there a frontier where the binding constraint is something Europe actually has?

There is. But naming it requires walking into one of the messiest words in the field.

A word that means three things

“World model” is having a moment, and like most terms having a moment, it has been stretched until it means whatever the speaker needs it to mean. Before it can do any work in an argument, it has to be pinned down, and pinning it down turns out to be unexpectedly useful, because the confusion is itself revealing.

The “World Model” confusion: Three senses, one word, no shared definition. This piece uses the middle one, world models as predictors of physical scenes, because its bottleneck is proprietary real-world data, not the compute and web text the other senses compete for. That is the input Europe is least short of.

The idea is older than the hype. It traces, in its modern form, to a 1990 proposal from Jürgen Schmidhuber: pair a model that predicts what the environment will do next with a controller that acts on those predictions. The roots run deeper still, into control theory and the cognitive-science intuition that an intelligent agent carries an internal model of its world in order to plan. In 2018, a paper by David Ha and Schmidhuber revived the term and made it current, showing an agent that could be trained inside its own learned “dream” of an environment.

From that single root, the term has since splintered across three research agendas that mostly do not talk to each other, which is exactly why it confuses anyone trying to follow the field.

One sense is about prediction as the path past today’s AI. This is the camp, associated with Yann LeCun and the JEPA line of work, that treats world models as the architecture that gets machines beyond the limits of language models: intelligence as prediction and planning in a learned, abstract space rather than next-token pattern-matching. Here, “world model” is a bet about the future of general intelligence.

A second sense is about acting in the physical world. This is the world model as the substrate for robots and autonomous vehicles, a system that predicts how a physical scene will unfold so a machine can move through it. Nvidia’s Cosmos “world foundation models,” pitched explicitly at physical AI, sit here; so does a company like Wayve, betting on learned driving. The defining feature of this sense is its appetite: it runs on proprietary, physical-world data, the motion and manipulation and navigation footage of reality, not on scraped text.

A third sense is about generating worlds to explore. This is the world model as a kind of neural game engine, DeepMind’s Genie line, which conjures navigable, interactive environments from a prompt. Related to text-to-video, but distinguished by interactivity: you don’t watch the world, you move through it.

Three senses, one word, no shared definition. Each agenda reached for the same evocative term, partly, one suspects, because all three like to describe themselves as a step toward general intelligence, and “world model” carries that grandeur for free. They are not rivals fighting over territory so much as strangers who happened to move into the same house.

It is worth noting that the senses are beginning to share machinery even as their purposes stay distinct. Nvidia’s Cosmos uses the generative techniques of the third camp to serve the second camp’s robots. That convergence makes the shared vocabulary more misleading, not less, which is all the more reason to be explicit about which sense is in play.

This essay means the second one: world models as predictors of physical scenes, the substrate for embodied machines. Not because the other senses don’t matter, but because this is the one whose binding constraint is the input Europe is least short of.

The fight Europe could actually win

Set the two frontiers side by side and the asymmetry is stark.

The frontier-LLM race rewards compute, web-scale text, and capital, which are Europe’s scarcities. The embodied-world-model race rewards something else: proprietary physical-world data, deep domain integration, and the industrial and scientific assets to generate both. And that list reads like an inventory of European strengths.

Consider what is actually here. Europe’s industrial-robotics and factory-automation base, the accumulated and decades-deep know-how of motion control, actuation, and real deployment on real factory floors, is precisely the kind of physical-world competence that cannot be scraped off the internet. Its automotive sector and the supplier ecosystem around it have spent years building high-fidelity driving simulation, sensor-fusion stacks, and digital twins; Wayve is the cleanest example of a European firm betting directly on driving as a learned, world-model problem. And its scientific-modeling tradition is, in places, genuinely world-leading. European weather forecasting’s machine-learning work is among the best on the planet, and institutions across fusion, particle physics, and materials science are fluent in exactly the task a world model performs: predicting how a complex physical system evolves.

The connective tissue across all three is the thing that matters strategically. In each case, the bottleneck is proprietary, physically-grounded data and the domain expertise to use it, not the web-scale text and hyperscale compute that the LLM frontier runs on. This is the axis along which Europe’s dependency is lowest. It is the one frontier where Europe’s structural position is an advantage rather than a liability.

A caveat worth stating plainly, because it is the place a sharp reader pushes back: this is an argument about where the comparative advantage lives, not a claim that Europe is already winning. Several of these are assets adjacent to world models rather than world-model programs as such. The case is that Europe is unusually well-positioned to build them, not that the race is already run. And it is a narrowing window, not an empty field. The US and China are pushing hard on robotics and embodied AI too. Europe’s specific edge is not in general-purpose humanoids, where capital advantages reassert themselves, but in the deep industrial and scientific verticals where the relevant data is hard to replicate and already sits inside European firms and institutions. That distinction is the whole game.

Two moves, not one

This is where the sovereignty conversation needs to split a single panicked instinct into two distinct strategies, because they answer different questions.

The first is defensive, and it is the one the chokepoint genuinely justifies. Europe should build sovereign LLMs, but the goal is not to win benchmarks. Rather, it is to hold a controlled, good-enough capability that cannot be switched off from abroad. A second-tier sovereign model is not a failed frontier model. It is insurance, and insurance is not supposed to outperform the thing it protects against. Judged as a championship contender, a European LLM is a disappointment. Judged as a hedge against exactly the dependency that just made itself felt, it is the rational minimum. The mistake is conflating the two and then feeling humiliated that the hedge is not a contender. Good enough and sovereign is the target. Frontier-or-bust is the delusion.

The second move is offensive, and it is where ambition belongs. Bet on the embodied frontier. Fund it where Europe’s comparative advantage is real, in the industrial, automotive, and scientific verticals where proprietary physical data and domain depth are the binding constraints. Sovereignty as insurance, world models as offense. The first keeps Europe from being held hostage. The second gives it something worth being un-hostageable for.

Don’t panic, this is not a white paper

What this implies for policy is less a program than a change of target, a few directions worth more than another doomed champion.

Treat physical-world data as strategic infrastructure. The advantage here lives in proprietary datasets (factory floors, vehicle fleets, scientific instruments) and the default trajectory hands that data, and the rent-setting power over it, to whoever provides the modeling layer on top. The procurement and data-governance questions that decide who captures that value are not back-office details. They are the sovereignty fight, arriving one layer down from where everyone is currently looking.

Fund the verticals, not the vanity. Money aimed at a general-purpose European frontier model is money aimed at Europe’s weakness. The same money aimed at embodied and scientific world models, at robotics, autonomous systems, simulation, and physical-AI foundations tied to existing industrial and research strengths, is aimed at its strength.

And stop measuring success against the wrong scoreboard. As long as European AI policy grades itself on proximity to the American frontier, it will keep losing a race it entered by mistake, and keep underfunding the one it could lead.

The dependency that surfaced that Friday evening was a genuine warning. But the answer to “they can cut us off from their frontier” is not “build a slower copy of their frontier and call it parity.” It is to build a sovereign capability good enough to absorb the shock, and then to put the real ambition where the ground actually favors you.

The illusion of borderless AI is over. That much the alarmists have right. What they have wrong is the response. The point is not to win the last war faster. It is to notice which war is still open.

Stack Overflow announced a beta this week called Stack Overflow for Agents. It is worth sitting with, not because the product is finished (it is a beta, the contribution loop is voluntary) but because of where it sits in a problem the field has spent the last eighteen months mapping in detail. Personally, I find the solution elegant, but there's a larger context worth understanding about agentic AI.

Start with the architecture, because the failure modes everyone complains about all trace back to one fact. A transformer is “stateless” by design. The context window is the only memory it has, and when the inference call ends, that window is discarded. This is not a defect someone forgot to fix. Statelessness is what makes these models horizontally scalable, reproducible, and parallelizable; you get the amnesia and the scalability from the same property.

Picture a brilliant contractor who shows up, solves your problem flawlessly, and then, the moment they walk out the door, forgets your company ever existed. Not fired, not negligent: the arrangement is built that way on purpose, because a worker with no standing memory can be cloned a thousand times over and dropped into a thousand offices at once. The same trait that makes them infinitely scalable makes them unable to remember yesterday. So one contractor untangles a nasty bug at noon; at one o'clock an identical contractor three timezones away hits the same bug and starts from zero, because as far as they know it has never been seen before.

So, back to the tech: an agent solves a breaking API change, ships it, and forgets it the instant the session closes. An agent in another timezone hits the identical bug an hour later and burns the same compute from scratch. The freshest, most production-accurate layer of knowledge gets generated constantly and persisted nowhere.

The field has a tidy metaphor for this now: the context window is RAM, not disk. Fast, volatile, expensive working memory with no storage layer beneath it. The models are good enough and the windows are large enough; what is missing is the persistence.

And the cost of that missing layer is not abstract. Stack Overflow’s 2025 developer survey found that 84% of developers use or plan to use AI tools while more actively distrust the accuracy of the output (46%) than trust it (33%), with just 3% highly trusting it, and distrust up from 31% the year before. That gap between using a thing and trusting it is the whole problem in one statistic. Generating an answer that looks right is free; knowing whether it survives contact with production is the expensive part, and nothing in the agent’s architecture remembers the answer once it does. Asked who they would still turn to in a future where AI handles most coding, developers’ top reason for wanting a human was “when I don’t trust AI’s answers,” at 75%. Verification is the job that does not go away.

It is worth being clear about what the frontier labs have and have not shipped here, because the obvious objection is that bigger context windows already solved this. They did not. The 2026 model releases added memory aimed at holding context within a session, not across them. Both leading models, with the largest commercial windows available, still treat in-session context and cross-session memory as separate problems, and the second one is unsolved at the model layer. Longer windows extend how much an agent can hold while it works. They do nothing about what survives after it stops.

Orchestration is the hinge

This is where the structural fact meets the skill everyone says matters most. The emerging consensus among practitioners, and increasingly the thing I hear in Amsterdam’s AI rooms, is that the near-term developer skill is not writing code. It is directing agents that write it. CIO put it cleanly in February: the engineer of 2026 spends less time writing foundational code and more time orchestrating a portfolio of AI agents, and the core skill becomes systems thinking rather than syntax. The role shifts from creator to curator. Delegate, review, own.

But notice how that conversation gets framed. Almost every account of orchestration describes an intra-session problem: one human conducting their own private orchestra. How does the agent that designs the schema hand off to the agent that writes the API, then to the one that runs the tests? That is coordination inside one person’s fleet, behind one person’s walls, and it is exactly the part the labs’ within-session memory addresses.

The harder question is what feeds the orchestrators across the wall. If the production-accurate knowledge an agent generates evaporates by default, where does it go so the next person’s agents can reach it? That is not the hand-off problem inside one fleet. It is the hand-off problem between strangers’ agents who never meet.

A large, busy market has formed to answer it, and Stack Overflow’s own survey measures it: among developers building agents, the tools already reached for to handle agent memory and data run from Redis and the GitHub MCP Server at the top through a long tail of vector stores and purpose-built memory layers like Mem0, Zep, and Letta still in low-single-digit adoption. Memory has become a first-class architectural component with its own benchmarks and research literature, converging on a shared taxonomy of episodic, semantic, and procedural memory, with declarative injection through files like CLAUDE.md and AGENTS.md. None of the three companies below invented the gap or noticed it first. What makes them worth isolating is that they answer one specific version of the question the general-purpose plumbing leaves open: not how to persist, but how far the knowledge should travel.

Three scopes

Re-fetch the source. The narrowest answer, and the one furthest along. Context7, built by Upstash, does not try to remember anything the agent learned. It maintains an indexed database of documentation from thousands of libraries and serves version-specific snippets into the context on demand, so the agent stops inventing APIs that do not exist. With north of 55,000 GitHub stars, it has real traction. But this is freshness, not verification. It tells an agent what the docs say, not what actually held up when someone shipped it.

Persist within the team. Cloudflare’s Agent Memory, launched in May, lets a team share a profile so knowledge learned by one engineer’s agent (conventions, architectural decisions, the tribal stuff) becomes durable and available to the rest. Knowledge Plane does something similar with a graph of typed relationships. These genuinely make the ephemeral persist, but they make it persist inside one company’s walls. The knowledge travels exactly as far as the org chart and no further.

Persist across the open ecosystem. This is the only answer that reaches past the wall, and it is the one Stack Overflow is taking. The distinctive move is making verification, not creation, the thing that earns reputation. Generating a plausible answer is cheap now; confirming which answers hold in production is not. So agents and developers who hit the same problem after publication report back on what worked and under what conditions, and those signals compound around the original post. The defense against hallucinated fixes polluting the corpus is a trust anchor: humans claim ownership of their agents through SSO, and an agent’s accuracy is tied to its owner’s established human reputation.

Three answers to the same architectural fact, separated by one question: how far does the knowledge travel?

Why the widest scope is the hard one

The two narrower answers work whether or not anyone chooses to contribute. Context7 scrapes and indexes docs automatically. Cloudflare derives memory from sessions as a byproduct. The knowledge accumulates without anyone deciding to be generous, which is why those models are safe bets regardless of how the agentic era shakes out.

Stack Overflow for Agents needs agents to voluntarily write back to a public commons. The whole flywheel depends on a contribution loop suggested by a skill file, not enforced by anything. And voluntary contribution to a public knowledge commons is precisely the dynamic AI already broke. Question volume on the human site has been in steep decline, with just 3,862 questions posted in December 2025, a 78% year-over-year drop, against a 2014 peak above 200,000 a month. The slide started before the models (questions began falling in mid-2020, partly a story of aggressive moderation and a Google Analytics change Stack Overflow itself pointed to) but it accelerated sharply once ChatGPT launched in late 2022, as developers routed the questions they used to post to models trained on Stack Overflow’s own back catalog.

The aforementioned loop is closed and self-cannibalizing: the answers trained the models, the models drained the contributions, and the company now licenses the back catalog to the same labs. The bet now is that the commons can be rebuilt by swapping human contributors for agent contributors, anchored back to human reputation so the accountability holds. The company whose commons got eaten is betting the agents will rebuild one. There is at least a thread of evidence the underlying need persists: in that same survey, asked who they would still turn to in a future where AI does most coding, the top answer, at 75%, was a person, specifically for when they do not trust the AI’s answer.

None of this is a knock on Stack Overflow's design, which is sharp; the reputation anchor is the right answer to the obvious failure mode. And it is worth being plain that the agents' amnesia is nobody's strategy. It falls out of the architecture the same way the scalability does. But notice what the scope distinction does to the stakes. Re-fetching docs and team memory are crowded categories with substitutes, and the value stays with the customer. A public, verified, cross-organization commons has no substitute, which is what makes it both the biggest prize and the longest odds. The widest scope has to convince agents to do something the existing commons never required. Open source works because the contribution is the artifact: you publish the library because you want it to exist, and everyone else's use is the point, not a favor. Stack Overflow for Agents asks for something else, a report, after the fact, on whether a fix actually held, when the working code already exists and nothing about the contributor's own job requires writing it down. The verification signal is pure surplus, valuable only to the next stranger. That is the unproven part. Not whether agents will draw on a commons, they already do, constantly, but whether they will feed one they had no reason to.

There is a familiar kind of rejection that doesn’t feel like discrimination because it doesn’t feel like anything. There’s no trainwreck interview, no recruiter who stops replying—just silence, the kind of ghosting every job seeker learns to absorb as ordinary bad luck. That’s what makes it so hard to see: the harm arrives as a mundane absence, and absences don’t leave fingerprints.

If you’ve applied to a graduate scheme or an entry-level role at a big consulting firm, bank, or consumer-goods giant in the last few years, you may have brushed up against it. Somewhere after the résumé screen, you’re sent off to play a set of online games. You tap balloons, you remember sequences, you make snap decisions for money, and a model you never see decides, from how you played, whether you measure up on traits like “processing speed,” “risk tolerance,” and “altruism.”

Pymetrics, the vendor at the center of the study in question, was used to screen candidates at firms like JPMorgan, Unilever, BCG, Accenture, and Mastercard. (If you’ve never run into it, and I had not myself, that’s because it’s concentrated in high-volume hiring—campus recruiting and bulk entry-level pipelines—rather than the specialized or senior roles most mid-career people apply to.)

The games are incidental, a quirk of one vendor. The mechanism underneath them is not. Strip away the balloons and what you have is algorithmic scoring: a model, trained on data, assigning you a number that decides whether a human ever sees your application. That mechanism is generalizing fast, and it certainly won’t keep the shape of a game. It’s already showing up as résumé scoring, video-interview analysis, and assessment tools bolted onto the applicant-tracking software nearly every large employer now runs. Games are simply the version that happened to be studied first, so treat this paper less as a report on one screening fad and more as the first clear look at a pattern that’s coming for the rest of hiring in forms we haven’t named yet.

Now imagine you apply to three such jobs at three different companies. You play the games, or near-identical ones, three times. You hear nothing back, three times.

What you can’t see is that all three companies bought their screening from the same vendor, that some of them may be running the literally identical model on your data, and that your game scores (which the system stores and reuses for 330 days) are quietly producing the same verdict every time. You’re not getting three chances. You’re getting one verdict, delivered three times, by a machine that has decided you are a “do not recommend.”

A new paper presented this month at the ACM’s Fairness, Accountability, and Transparency conference—Algorithmic Monocultures in Hiring, by researchers at Stanford, Chapman, and Northeastern—is the first to actually watch this happen at scale. They obtained data from the hiring vendor pymetrics (acquired by the recruiting-software firm Harver in 2022) covering 4.2 million applications from 3.4 million applicants across 156 employers. It appears to be the first study ever to observe what a single applicant’s real algorithmic outcomes look like across multiple employers. And what it found should reframe how we think about both algorithmic bias and corporate transparency.

The measurement is the scandal

Over 90% of U.S. employers now use software to screen or rank applicants, and a handful of vendors supply most of it. When pymetrics audited its own system for bias—pooling together all the applications it had ever processed—it found nothing alarming. Across the whole dataset, Black applicants were selected at 52.5%, White applicants at 58.3%. The ratio between those rates clears the bar regulators use.

That bar is the “four-fifths rule” (sometimes the “4/5ths rule”). It comes from how the U.S. enforces employment discrimination: if a protected group is selected at less than 80% of the rate of the most-selected group, and the gap is statistically significant, that’s a red flag for adverse impact. Or rather, the legal term for a practice that’s neutral on its face but lands disproportionately on one group. pymetrics’ aggregate numbers passed.

Here’s the problem the researchers identify: that’s the wrong way to do the math. U.S. discrimination law operationalizes adverse impact per job, not across an entire labor market. A vendor sitting between 156 employers and millions of applicants had simply averaged everything together. And averaging is exactly what hides the harm.

The analogy the authors use is sharp: imagine an employer that recommends only men for doctor roles and only women for nursing roles. The overall selection rates by gender could come out identical (perfectly fair in aggregate) while every individual position is rigidly sorted by sex. Aggregation launders the discrimination out of the data.

When the researchers redid the analysis the legally correct way—each of the 1,746 positions examined separately—the clean bill of health collapsed. 10.6% of positions show adverse impact against Black applicants. Roughly a quarter of all applications submitted by Black applicants went to positions where the model worked against them. The same pattern hit Asian applicants. None of this was visible in the pooled numbers. The discrimination didn’t appear because of a new finding; it appeared because someone finally divided by the right denominator.

This matters beyond hiring, and it’s the part I’d underline for anyone who works on platform accountability: the unit of analysis is a political choice. Consolidate the data enough and any harm averages away. It seems that the fight over how finely you’re allowed to look is often the whole fight.

Opportunity has a single point of failure

Adverse impact is an old worry. The genuinely new finding is what the authors call systemic rejection, and it only exists in a world where everyone uses the same few tools.

Start with the word monoculture. In farming, it means planting a single crop across a whole region: efficient, until one blight arrives and there’s nothing to stop it, because every plant is identical and vulnerable in the same way. Algorithmic monoculture is the same idea applied to decisions. When most employers screen applicants through the same handful of vendors, they stop making independent judgments and start making one shared judgment, replicated everywhere. The diversity that used to protect you, the fact that a recruiter who passed on you at one firm had nothing to do with the recruiter at the next, quietly disappears. Hivemind HR.

Here’s why that matters for a single person. In an ordinary labor market, applying to ten jobs means ten separate rolls of the dice. Some land, some don’t; a rejection here says nothing about your chances there, because different people with different tastes are doing the deciding. That independence is the thing that makes “just keep applying” sound like reasonable advice. It’s what gives a job search the shape of a numbers game you can eventually win.

Monoculture breaks that. If the same model—or a set of closely related models trained the same way on the same kind of data—is scoring you each time, you’re not getting ten chances. You’re getting one verdict, returned ten times. And because pymetrics stores your game scores and reuses them for nearly a year, the verdict doesn’t even get the benefit of you having a better day. The researchers find exactly this in the data: among applicants who applied to ten positions, 4% were rejected by every single one. Far more than independent chance would produce. The decisions aren’t independent. They’re correlated, because they’re frequently the same decision wearing different company logos.

This is what systemic rejection means: not that some people get turned down a lot, but that the same people get turned down everywhere, for reasons baked into a profile they never see and can’t appeal. The authors have a blunter phrase for the people it happens to: they’ve been “algorithmically blackballed.” And the cruelty of it is that it’s invisible from the inside. Each individual rejection looks like ordinary bad luck. Only by tracking one applicant across many employers—which no one before this study could do—does the pattern resolve into something systematic.

The researchers test the obvious escape hatch. Could you just apply to more jobs? Because the algorithm is deterministic—run it twice on the same data and you get the same answer—they could simulate outcomes for jobs people never actually applied to. The better news: no one is universally unhireable; every simulated applicant was eventually recommended by some model. The damning news: to drive your odds of being shut out everywhere down to near zero, you’d need to apply to roughly 25 positions where 10 would have sufficed in an independent market. The machine doesn’t make you unemployable. It just makes opportunity scarcer, and quietly stacks that scarcity against the same people again and again.

That’s the structural picture, and it’s bigger than any one applicant. When a labor market routes its decisions through a single vendor, that vendor becomes a single point of failure for human opportunity. Get a “do not recommend” lodged in your stored profile, and you can be locked out of an entire sector without one human ever seeing your name.

Accountability by permission

Here is the part that should worry anyone who cares about holding these systems accountable: everything above exists only because pymetrics voluntarily handed over its data, under an agreement that—to the company’s real credit—couldn’t veto the findings. That is extraordinarily rare. Most hiring algorithms are black boxes. Researchers studying them are usually reduced to scraping public job boards or sending fake résumés, neither of which can see the algorithm actually deciding.

Now follow the incentive. A vendor shares data. Independent researchers find adverse impact and systemic rejection. The findings embarrass the vendor and its clients. The lesson every other vendor draws: never share data. The authors say so themselves—this very paper may discourage the kind of openness that made it possible.

That’s the trap, and it’s the throughline that connects this paper to every other fight over algorithmic accountability. Voluntary corporate transparency is self-extinguishing. The more useful the access, the stronger the reason to shut it off. You cannot build an accountability regime on the goodwill of the companies being held accountable. Which is why the only durable answer is to stop asking nicely—and the two jurisdictions that matter are moving in opposite directions.

The EU is building the scaffolding

Europe already has the two pieces this problem needs, even if neither is pointed directly at hiring yet.

The first is the precedent for forced access. The EU’s Digital Services Act requires very large online platforms to grant vetted researchers access to their data—a rule that exists precisely because a decade of voluntary cooperation in social media research had collapsed into nothing. The logic transfers cleanly: if we accept that researchers must be able to study the algorithms shaping public discourse, the algorithms deciding who gets to earn a living are not a lesser case.

The second is classification. The EU AI Act designates AI used in hiring as “high-risk,” which triggers a battery of obligations: risk management, documentation, human oversight, bias monitoring. The systems in this paper would almost certainly qualify. On paper, the high-risk obligations become enforceable on August 2, 2026. In practice, that date is already wobbling: the Commission’s late-2025 “Digital Omnibus” package has floated pushing the high-risk deadline back, and industry is lobbying hard for the delay. So even Europe’s scaffolding, likely the most advanced in the world, is being softened before it’s load-bearing. The framework exists. The will to switch it on is still contested.

The gap the paper exposes is that even a fully-enforced AI Act doesn’t obviously require the right measurement. A vendor monitoring its own system for bias could do exactly what pymetrics did—pool everything, pass the aggregate test, and miss the per-position discrimination entirely. The lesson for European regulators is specific: high-risk bias audits have to be conducted at the level of the individual position, not the comfortable aggregate, or they’ll certify the same blindness the law was meant to cure.

The US is dismantling the concept

The American picture is starker, because the United States isn’t debating how to measure disparate impact. Rather, it’s debating whether disparate impact should exist as a legal idea at all.

The entire framework this paper relies on—the four-fifths rule, the very concept of a neutral practice that’s illegal because of its disproportionate effect—comes from Title VII of the Civil Rights Act of 1964 and the case law built on it. As of 2025, that framework is under direct attack. A Trump administration executive order, “Restoring Equality of Opportunity and Meritocracy,” explicitly targets disparate-impact liability, directing federal agencies to scale back its use. The theory that lets you challenge a hiring practice based on its results rather than its stated intent is exactly the theory now in the crosshairs.

Consider the timing: These researchers have built the sharpest instrument yet for detecting algorithmic discrimination in hiring, and they’ve released it at the precise moment the U.S. government is working to retire the legal concept that instrument measures. A better thermometer arrives just as the fever is being redefined as not a fever at all. In the U.S., whether this work changes anything may have less to do with the strength of the evidence than with whether the legal category it speaks to survives the decade.

There’s a narrower American failure worth naming too. New York City passed Local Law 144 in 2021, the first U.S. law to directly regulate algorithmic hiring, requiring employers to commission annual bias audits and post the results publicly. A 2024 study by Lucas Wright and colleagues found the law largely toothless in practice. Combing through 391 NYC employers, the researchers found only 18 had posted the required audit reports and 13 the required transparency notices, because the law lets employers decide for themselves whether they’re even in scope. They coined a term for the resulting fog: “null compliance,” a state where you can’t even tell whether a company is breaking the law. Tellingly, nearly every audit that was posted reported a passing score. Everyone cleared the four-fifths bar.

That last detail is the same blindness the pymetrics paper diagnoses, now written into law. New York’s own guidance, the Bommasani authors note, appears to instruct exactly the aggregation that hides the problem: Merging data across positions and employers into a single, reassuring number. So the first serious American attempt to regulate algorithmic hiring may be quietly mandating the very averaging that launders the discrimination out of view.

In the space of about seventy-two hours this week, the publisher side of the AI fight produced more visible momentum than it had in the previous year.

At the World News Media Congress in Marseille, the SPUR coalition — the Standards for Publisher Usage Rights group launched in February by the BBC, Financial Times, Guardian, Telegraph, and Sky News — announced it had added some thirty new members and secured formal backing from WAN-IFRA, which represents roughly 3,000 news organizations across 120 countries, and FIPP, the global magazine-media association. On the same stage, Guardian editor Katharine Viner named AI-driven disintermediation, or rather the severing of the link between a publisher and its readers, as the one threat she would not hedge against, and pointed to SPUR as the collective answer.

And in London, the Competition and Markets Authority did something no regulator had done before. It ordered Google to let publishers opt out of having their content used in AI Overviews and AI Mode without losing their place in ordinary search results. CMA chief executive Sarah Cardell called it a “world-first requirement” meant to give publishers “appropriate bargaining power over how their content is used.” Google was also ordered to attribute publisher content with clear links in its AI results.

Taken together, the week reads like a counteroffensive finally taking shape: a publisher coalition scaling up, a major regulator breaking the most important logjam, and one of the world’s most respected editors naming the stakes plainly. The trade press greeted it accordingly.

I want to suggest a more careful reading; not a cynical one, because the progress is real and very encouraging, but one that asks what each of these moves actually does, and what it conspicuously leaves unfinished.

A couple of weeks ago I argued that the publisher response to AI was missing a specific institutional piece — that the enclosure of the open web was happening across three distinct layers, and that the interventions on offer addressed some layers while leaving the decisive one untouched. The events of this week are the clearest test of that argument so far. They cut both ways, and the way they cut is instructive.

What the CMA actually changed

Start with the regulatory news, because it’s the most consequential and the most misunderstood.

The single biggest structural obstacle to a functioning publisher-licensing market has been what the writer Joshua Benton, at Nieman Lab, called the fundamental blocker: Google’s refusal to separate its search-indexing crawler from its AI-training crawler. As long as appearing in Google Search meant submitting to AI ingestion, the price a publisher could charge for that ingestion was effectively zero — because the alternative was disappearing from search entirely. News Media Europe, in its submission to the CMA, called this “compelled consent”: allowing the crawl is the price of visibility, and the same content then powers the AI features that erode the traffic visibility was supposed to provide.

The CMA just broke that coupling, in one jurisdiction, by force of regulation. UK publishers will be able to refuse AI use while keeping their search presence. On the logic of the licensing market, this matters enormously as it is the first time the “Hobson’s choice” at the heart of the publisher complaint has been pried open by anything other than a private lawsuit.

But notice three limits, each of which the celebratory framing tends to skip. The remedy is confined to the UK and to Google; the scraper economy and the other AI labs sit entirely outside it. Google has nine months to comply, and as one analyst noted, a harm running for three years now continues un-remedied until late 2027, with effectiveness unknown until then. And most importantly: an opt-out is a shield, not a revenue stream. The right to say no to Google’s AI is not the same as the ability to be paid by anyone for saying yes. It restores leverage. It does not, by itself, cut a single deal.

What SPUR is — and what it is careful to say it isn’t

Which brings us to the coalition news, and to a sentence worth reading closely. As John Rahim of The Media Stack put it, SPUR "isn't a licensing body and doesn't cut deals on anyone's behalf. What it's building is the layer underneath the deals."

That is an accurate and honest description, and it is also the whole problem in miniature.

The layer underneath the deals (shared standards, machine-readable terms, transparency requirements, usage data so publishers can see how their work is consumed) is genuinely valuable infrastructure. It is the grammar a licensing market needs in order to function. Building it across a coalition that now spans national broadcasters, global wire-adjacent publishers, and two of the largest trade bodies in the industry is real, hard, useful work.

But a grammar is not a counterparty. Standards describe how a deal should be structured; they do not sit across the table from OpenAI or Google and negotiate price, and they do not collect and distribute the money afterward. SPUR has said, clearly and to its credit, that this is not what it does.

So the question the week’s coverage mostly glides past is the obvious one: if SPUR builds the layer beneath the deals, and explicitly doesn’t cut deals — who does?

The layer nobody has built

Survey the field and the answer, for most publishers, is: still no one.

The major publishers cut their own deals. News Corp, the New York Times, Axel Springer, Reuters — these institutions have the brand leverage, the legal departments, and the senior relationships to negotiate bilaterally with AI companies, and several have. For them, better standards (SPUR) and restored leverage (the CMA) are accelerants on a process they were already capable of running. They were always going to get the meeting.

Below that tier, the deal-cutting function exists only in fragments. Dow Jones’s Factiva aggregates thousands of sources and negotiates on their behalf — a centralized model, governed commercially by Dow Jones. The News/Media Alliance has built templated arrangements with AI partners where members contract individually under a shared legal framework. And the infrastructure players — Cloudflare’s pay-per-crawl, Microsoft’s

Publisher Content Marketplace — operate the only at-scale marketplaces that actually move money, but they are governed by the platforms, not the publishers, with rent-setting authority sitting on the wrong side of the table.

None of these is what SPUR is, and SPUR is not trying to be any of them. Which means the architecture now taking shape looks like this: a publisher-governed standards layer (SPUR) sitting above a set of deal-cutting mechanisms that are either elite-only (bilateral), commercially governed (Factiva), structurally limited (NMA’s templated bilaterals), or platform-governed (Cloudflare, Microsoft). The standards layer and the deal-cutting layer are being built by different actors, on different governance logics, and it is not at all clear they are designed to interlock.

For the publisher this entire exercise claims to serve — the regional daily, the mid-sized digital native, the specialist trade title, the independent investigative outlet — the practical question is brutally simple and almost entirely unanswered: when an AI company trains on my work, who negotiates on my behalf, and who makes sure I get paid? SPUR’s standards don’t do it. The CMA’s opt-out doesn’t do it. A bilateral deal isn’t available to me. Factiva might, if I qualify and accept its terms. Otherwise, the answer is the same as it was a year ago.

Why “thirty members” and “trade-body backing” don’t settle it

It’s worth being precise about the inclusion question, because the headline numbers invite a misreading.

Thirty new SPUR members announced at an event attended by the executives of major publishing houses are, in all likelihood, still substantial publishers — national and large-regional outlets, international groups, significant magazine brands. “Thirty members” can describe thirty of the largest two hundred publishers in the world and still represent the apex of the industry, not its long tail.

And trade-body endorsement, however valuable, is not membership. WAN-IFRA representing 3,000 organizations does not place 3,000 organizations inside SPUR’s arrangements, much less inside anyone’s licensing deals. It signals legitimacy and opens a recruitment channel. It does not enroll the small publisher, and it certainly does not negotiate or collect on that publisher’s behalf. The distance between “my trade association endorsed this initiative” and “I receive a payment when a model trains on my reporting” is the entire unsolved distance.

There is even a quieter bias built into the standards-first approach. Machine-readable standards, however open, are adopted most readily by publishers with the engineering capacity to implement them. The large publisher has a team for that. The small one often does not. A universal standard, in practice, tends to deliver its benefits first and most fully to those who already had resources — which is the opposite of what the long tail needs.

The question the week made visible

None of this is an argument against the week’s developments. The CMA ruling is the most important regulatory move yet made on behalf of publishers, and the precedent may travel well beyond the UK. SPUR’s expansion is a genuine act of collective will in an industry not known for solidarity. Viner naming disintermediation as the threat she won’t hedge against is the most clear-eyed thing a major editor has said about AI in months.

But momentum is not the same as resolution, and the shape of what’s been built tells you precisely what hasn’t. The week produced a stronger standards layer and a stronger leverage position — the layer beneath the deals, and the right to walk away from a bad one. What it did not produce, and what no announcement this week even claimed to produce, is the thing in the middle: a publisher-governed mechanism that actually negotiates and collects on behalf of the publishers who cannot do it themselves.

The standards are arriving before the entity that would make them pay off for anyone without their own business-development team. Whether that entity gets built, and whether the economics of the long tail can even support it, is the question all this week’s progress quietly sharpened rather than answered.

It is the question worth watching and one I’ve spent years thinking and now building on. The news will keep coming fast; the announcements will keep sounding like victories. The thing to ask of each one is the same: does this finally cut the deal for the publisher who can’t cut it alone, or does it build one more layer around the place where that deal still isn’t being made?

For the longer argument this builds on — the three-layer enclosure of the open web and why the publisher-side broker is the missing piece — see below:

Last week I wrote about why your AI bill is about to look very different, arguing that GitHub’s repricing was the leading edge of a structural reset across the AI application economy. The piece focused on the demand side: what happens when flat-rate subsidies end and the meter starts running. Everyone, finally, sees what inference actually costs.

On Saturday, DeepSeek answered the same question from the other side of the meter.

The Hangzhou-based lab announced it is making permanent a 75% price cut on its flagship V4-Pro model, bringing API costs to between roughly $0.0035 and $0.87 per million tokens, depending on usage type. When V4 launched a month ago, DeepSeek said Pro pricing was up to 12 times higher than Flash because of “constraints in high-end compute capacity,” tied to limited supply of Huawei’s Ascend 950 — the domestic AI accelerator chip that Chinese labs use as a workaround for the Nvidia GPUs that US export controls have kept out of the country.

That gap just compressed. Using the new V4-Pro prices against DeepSeek’s V4-Flash rates, the multiplier has narrowed from 12x at launch to roughly 4x today. (My math, not theirs.)

Subscribe now

For context, and per a Cryptopolitan breakdown of the comparison: Claude Opus 4.7 costs about $5 input and $25 output per million tokens. GPT-5.4 sits at $5 and $15. DeepSeek V4-Pro is now $0.44 input and $0.87 output. The output-cost gap alone is roughly 9x against Anthropic and 17x against OpenAI’s prior-generation model. An autonomous agent that costs a few hundred dollars a day to run on Claude Opus runs for under $40 on V4-Pro.

The reading that’s circulating

A specific interpretation of this news has gained traction across business press and industry commentary: that DeepSeek is running a loss-leader strategy. Price below cost now, capture market share, raise prices once the user base is locked in. It’s the playbook attributed to Amazon in retail, to plenty of Silicon Valley unicorns in their growth phases, and increasingly to Chinese tech companies in their export pushes. The reading is sophisticated enough that it deserves to be taken seriously rather than dismissed.

Whether it actually fits DeepSeek is a question worth working through, especially when AI marketing, its sound bytes and narrative construction, is itself heavily shaping news coverage. The structural features of this specific case point in directions the standard template doesn’t quite capture.

What supports the loss-leader reading

There are real reasons the framing keeps surfacing.

DeepSeek is closing its first external funding round. The Financial Times and Bloomberg have reported that the lab is raising $3 billion to $4 billion at a valuation that has climbed from around $20 billion in late April to roughly $45 billion by mid-May. Most consequentially, the round is being led by the China Integrated Circuit Industry Investment Fund — known as the “Big Fund,” a state-backed vehicle established to advance China’s semiconductor self-sufficiency. This would be the Big Fund’s first investment in a Chinese AI model lab, which analysts read as Beijing treating frontier AI software and domestic chip production as a single strategic problem. Aggressive pricing is a growth story, and a growth story at this scale is what state-led funding rounds buy.

Anthropic and OpenAI have previously accused DeepSeek of training on Claude outputs through what it called “distillation attacks.” DeepSeek disputes the latter, but if any part of it is true, some of the cost advantage would be coming from borrowing capability from competitors rather than from pure architectural efficiency.

DeepSeek did not explain the permanent price cut. The Reuters story notes the company declined to say whether it was tied to increased Ascend 950 supply; in other words, whether Huawei is finally able to ship enough of its domestic AI chips in volume to lower DeepSeek’s per-query inference costs, which was the reason DeepSeek itself gave at launch for the original Pro premium. The silence on this question leaves room for a strategic reading rather than a unit-economics one.

And then there’s the structural subsidy question. State-level support for Chinese AI champions, combined with access to domestic compute that doesn’t price against Nvidia retail markups, plausibly allows margin absorption that Western labs can’t match. The Big Fund leading this round makes that less hypothetical than it would have been a month ago.

These are not nothing.

What pushes against it

The pattern of price cuts predates anything that looks like a market-share play. According to a published history of DeepSeek’s roadmap, each model generation since V2 in May 2024 has shipped a specific efficiency improvement that translated more or less directly into a lower per-token price. V2 introduced an attention mechanism that compressed memory usage during inference. V3.2-Exp in late 2025 introduced sparse attention and cut prices in half. V4 introduces what DeepSeek calls Hybrid Attention Architecture. The technical specifics are less important than the pattern: pricing tracks engineering improvements rather than the funding cycle. Founder Liang Wenfeng has said the team was surprised by the market’s price sensitivity when V2 launched — they were pricing against actual costs, not strategically.

The weights are open. V3 was released under DeepSeek's own license, and R1, V3.2-Exp, and V4 Pro were all released under the MIT License — one of the most permissive open-source licenses, which lets anyone download, modify, run, and even commercially deploy the model with essentially no restrictions. This is the structural fact the standard loss-leader template doesn’t accommodate.

Loss-leading works when users get locked into the platform through switching costs, ecosystem dependencies, or proprietary data. None of that applies to a model anyone can download from Hugging Face and run on their own infrastructure. If DeepSeek raises prices later, sophisticated users self-host or move to one of the dozens of third-party providers already hosting earlier DeepSeek models at competitive prices. The classical moat isn’t there.

The V2 precedent is worth dwelling on. When DeepSeek aggressively priced V2 eighteen months ago, the structural outcome wasn’t market capture. It was commoditization. The Chinese AI market got dramatically cheaper, the Western labs eventually had to follow, and DeepSeek’s enterprise market share in the West remains at roughly 1% on the most recent Menlo Ventures data, against Anthropic at 32% and OpenAI at 25%. If V2 was the opening move of a lock-in strategy, the strategy has so far produced a price war without producing the capture.

And then there’s the unusual corporate structure. DeepSeek’s parent is High-Flyer, a quantitative hedge fund that has financed the lab entirely off its own balance sheet since 2023. The lab is organized around small project teams rather than the traditional management hierarchies of Chinese tech giants, and it has explicitly described itself as research-first.

The geometry of “burn capital now to monetize later” assumes a capital structure built around eventual monetization at scale. A hedge-fund-financed research lab that publishes open weights and prices against its actual unit costs has been operating in a different incentive regime than OpenAI or Anthropic. The state-led funding round now in progress will change that structure to some degree, which is part of why the loss-leader question is harder to dismiss than it would have been six months ago.

What this leaves us with

My read is that some loss-leader dynamics are probably present: the timing of the permanent cut, the state-led funding round, the company’s silence on whether Huawei’s Ascend chip supply has actually scaled, the convenient alignment with a valuation push. The capital structure is shifting in ways that make the strategic reading more plausible than it was for V2 or R1.

But the structural conditions that make loss-leading work as a long-term strategy, which is proprietary lock-in, switching costs, captured ecosystems, aren’t there in the way the standard template assumes. So even if DeepSeek wants to run that play, the game board doesn’t support it the way it would for a Western SaaS company. The technical openness and the financial closing-in are pulling in different directions.

The other thing worth noting is what the framing reveals about how we read these stories. Western business commentary defaults to the templates it knows: predatory pricing, growth-stage burn, eventual monetization. Those templates work well for the companies they were built around. They work less well for an open-weight research lab inside a hedge fund operating with structural compute advantages and a multi-year pattern of architectural cost reduction — even one now taking state capital. Reaching for the familiar frame is not lazy, but it’s what happens when the available reference patterns don’t quite fit the case.

For the metering argument from last week, none of this changes the basic claim that enterprise AI bills are about to look very different. What it changes is the floor those bills are converging toward. A Western enterprise buying Claude or GPT for compliance, latency, and procurement reasons will keep buying Claude or GPT. But the anchor point in every pricing conversation just moved. The repricing notices going out from GitHub, the metered shift Nadella described as inevitable across every per-user business, all of that lands differently when the buyer can point to V4-Pro and ask what the Western markup is paying for.

The meter and the floor are both legibility events. We’re going to find out what AI actually costs, from both directions, faster than the industry consensus assumed. That number is going to be different in different places. The gap between those numbers is going to be one of the more important things in the AI economy.

A note on “multimodal”

One of the things I wanted to settle in my own mind before writing any of this was whether a specific criticism of V4 that has circulated in recent weeks, that DeepSeek isn’t really a peer to GPT-5.5 or Claude Opus 4.7 because it isn’t multimodal, actually holds up.

“Multimodal” is one of those terms that does a lot of work without explaining itself. It just means a model that can take in more than one kind of input. Text plus images, mostly, with some video on certain Chinese consumer deployments. For example, a user uploading a screenshot for analysis, a developer pointing a model at a UI mockup, a document workflow pulling from scanned PDFs; those are multimodal tasks. The capability matters because a lot of real-world enterprise work — document processing, quality control, accessibility audits — obviously requires a model that can see.

The criticism that V4 doesn’t do this turns out to be reading from notes that are roughly six months old. V4 ships with a three-mode architecture: Fast, Expert, and Vision — and Vision is native rather than bolted on. The lineage goes back to DeepSeek-VL2 in December 2024 and the DeepSeek OCR paper in October 2025. One independent analysis found the V4 vision system uses about a tenth of the inference cache of Claude’s vision processing per image, which is itself an efficiency story rather than a capability gap.

The point here isn’t to defend V4’s capability against every benchmark, but it trails the absolute frontier on some agentic coding and graduate-level reasoning tasks, and DeepSeek itself says it’s about three to six months behind. The point is that the specific framings circulating in Western coverage tend to lag the actual product, in directions that make the standard skeptical reading easier to sustain than the facts justify. That is something worth keeping in mind.

(credit: Google)

The week the architecture became shockingly visible

Last week at Google I/O 2026, Head of Search Elizabeth Reid announced the most significant redesign of Google’s search box in twenty-five years. The query field now expands to accept longer conversational prompts. AI Mode, already at over a billion monthly users, becomes more deeply integrated into the default search surface. Information agents will monitor topics over time and push updates to users. Autonomous workflows can be initiated from the same box where users used to type three-word queries. Reid’s framing was unambiguous: “AI search through and through.”

For the publishing industry and for analysts that have been along for the ride, the I/O announcements weren’t a surprise so much as a confirmation. Zero-click searches now account for roughly 60 percent of queries. AI Mode itself shows a 93 percent zero-click rate. Pew Research found click rates with AI summaries drop to 8 percent from 15 percent without — a 46.7 percent relative decline measured against a controlled baseline. Individual publishers report damage from severe to, dare we say, catastrophic. HubSpot estimates a 70 to 80 percent collapse in organic traffic. DMG Media documented drops of up to 89 percent for affected queries. NPR called it “an extinction-level event” for online news publishers. You get the idea.

In brutal terms, this is the demand-side harm — what users see, what gets quantified in click-through rates, what regulators and antitrust authorities focus on. It is real and substantial. But it is also, I want to argue, only one of three layers of enclosure currently being built around the open web, and not the only one where durable structural shift is happening.

A week before I/O, on May 6, Thomson Reuters CEO Steve Hasker articulated his AI licensing framework publicly: archive-only scope, maximum pricing, short contract terms designed to force renegotiation. Reuters has commercial agreements with major AI companies. Reuters also, Hasker confirmed, sees its breaking news content surfaced in those same AI products without compensation. The chief executive of one of the world’s largest news organizations was publicly acknowledging that the deals don’t cover the actual extraction happening.

If even Reuters cannot stop live content extraction with the leverage it has, the deals are not the architecture. They are the visible part of an architecture whose less visible parts are doing more work. The tip of the iceberg, let’s say.

Three months earlier, in February, Microsoft launched its Publisher Content Marketplace, a structured platform where publishers can license their work to Microsoft AI products under usage-based reporting terms. In April, Cloudflare expanded the x402 Foundation — now jointly governed with Google, Microsoft, Amazon, Coinbase, and the major payment networks — to handle machine-to-machine billing at protocol level. Both are brokerage architectures. Both are owned and governed by the companies whose pricing power publishers are trying to escape.

And beneath both, a third layer: a roughly $1 billion market for on-demand publisher content operated by venture-backed scraper firms — Firecrawl, Exa, Tavily, Bright Data, Apify, Zyte, and at least eighteen others — selling extraction-as-a-service to AI labs, consultancies and enterprises with no licensing conversation at any point in the pipeline. According to industry analyst Matthew Scott Goldstein, of twenty-one such vendors profiled in his recent report, zero had agreements with the publishers whose content they re-sell.

The publishing industry’s response to all of this remains, for the most part, individual. A handful of elite players negotiate bilateral deals. SPUR coordinates among five major UK publishers. CoMP is a technical protocol from the IAB Tech Lab. None of these, on its own, is the missing piece.

What’s missing is a publisher-side broker.

Three layers, one architecture

The dominant frame for the publisher crisis treats it as a story about Google, or about AI Overviews, or about the failure of bilateral licensing. The frame is too narrow. The architecture that’s being built has three layers, and each layer extracts value differently from publishers who lack the leverage to push back.

[Figure taken from my paper: “The New AI Scraper Economy and the Three-Layer Enclosure of the Open Web” (April 2026)]

Layer 1: The hyperscalers. Amazon Web Services, Microsoft Azure, and Google Cloud have built full-stack AI offerings that integrate infrastructure, foundation models, and application-level tooling. Their convening power over AI is structural — model developers, enterprise customers, and the industrialization of AI across sectors all depend on cloud infrastructure controlled by three firms. This is the layer where antitrust attention has concentrated. It is also the layer where the most visible licensing deals are happening: News Corp and OpenAI ($250M+ over five years), Reuters and various AI companies, AP and Google, NYT and Amazon (reported at $20–25M annually). For publishers with the scale and legal apparatus to negotiate at this layer, real revenue is flowing. For everyone else, this layer is structurally inaccessible.

Layer 2: The content-delivery tier. This is the layer that requires its own theorization, and the one this piece is centering. Cloudflare handles traffic for roughly 20 percent of the web by its own account. Since mid-2024, it has constructed a parallel apparatus of marketplace intermediation, automated payment, and ecosystem cultivation that mirrors the hyperscaler playbook on a smaller footprint but in a strategically decisive position. The chronology is diagnostic.

In July 2024, Cloudflare launched a one-click AI bot blocker. In September and December, it added granular per-crawler controls. On July 1, 2025 — branded “Content Independence Day” — it made blocking the default for new customers and introduced Pay Per Crawl, reviving HTTP 402 (a payment-required status code defined in the original HTTP/1.1 specification but unimplemented for three decades) as the channel through which AI crawlers could obtain paid access. Between July and early December 2025, Cloudflare logged 416 billion blocked AI bot requests. In March 2026, Cloudflare released its own Crawl API — a tool that can scrape an entire website with a single request — and customers initially found they could not block it through Cloudflare’s own managed controls.

The protector had become, in infrastructural terms, a toll collector on both sides of the road.

Two weeks before the first version of this paper was written, Google, Microsoft, and Amazon formally joined the x402 Foundation alongside Cloudflare, Coinbase, and the major payment networks. What had been presented to publishers as Cloudflare’s independent defense of the open web had resolved into co-governance with the hyperscalers of a jointly-owned payment architecture for machine-mediated content access.

Layer 3: The scraper economy. This is the least-discussed layer and the one with the most aggressive bypass logic. The firms in this category — over twenty in Goldstein’s report, with the content-licensing tracker TollBit identifying approximately forty in total — have built a $1 billion-plus market for on-demand publisher content, projected to double to $2 billion by 2030. Their customers, documented in Goldstein’s research, include BCG, IBM, Cohere, AWS, Salesforce, Apple, PwC, Shopify, and Alibaba. Their raw material, as Goldstein puts it bluntly, is “the same articles, analysis, and journalism that publishers spend billions producing every year.”

The structurally novel feature of this layer is its distribution model. Nineteen of the twenty-one vendors profiled require no sales call, no contract, no licensing conversation. A developer signs up with a credit card and begins scraping. The buyer is an engineer building an AI system rather than an executive making a procurement decision. TollBit’s recent “State of the Bots” report found that roughly 30 percent of AI scrapes violate explicit robots.txt directives. None of the twenty-one vendors had licensing agreements with the publishers whose content they re-sell.

This layer completes the architecture. It is structurally downstream of both hyperscalers and CDN providers, yet it bypasses both Big Tech’s data-partnership regime and Cloudflare’s pay-per-crawl toll system. Firecrawl and its peers are themselves the entities doing the scraping that Cloudflare aims to block, and they re-sell what they extract as a clean-data-as-a-service product. The CDN Akamai reports a 300 percent surge in AI bot activity in 2025, with publishers representing 40 percent of all media-related AI bot traffic. According to Goldstein’s summary, publishers receive no payment from any of the twenty-one vendors identified.

Three layers. One direction of extraction. No publisher-side institution that operates across them.

The existing publisher responses

The response architecture currently being built reflects what each institution behind it is structurally positioned to do. Each intervention does real work; together, they don’t yet add up to a counterweight at the scale of the problem because the institutional form that would carry the full response hasn’t yet been built.

Bilateral licensing deals work for the publishers who can negotiate them. News Corp signed with OpenAI in May 2024 for an estimated $250M+ over five years. The New York Times signed with Amazon at $20–25M annually. Reuters with several. AP with Google. Axel Springer with OpenAI. Microsoft signed deals with Hearst, Condé Nast, and Dotdash Meredith in 2025. As of Q1 2026, USA Today Co. and IAC began reporting “meaningful” AI licensing revenue for the first time, driven by Meta deals signed in December 2025. These are real. They are also concentrated entirely at the top of the publisher hierarchy. The pattern is consistent: AI companies want major-brand premium content and will pay for it; everyone else competes for free.

SPUR — the Standards for Publisher Usage Rights coalition launched in February 2026 by the BBC, Financial Times, Guardian, Telegraph Media Group, and Sky News — is genuinely a publisher-side coordination effort. It aims to establish shared licensing standards for AI access to journalism. But its membership is, by design, elite UK publishers. It is a coordination forum among institutions that already have leverage; it does not, by itself, generate leverage for those who don’t.

CoMP — the IAB Tech Lab’s Content Monetization Protocol, released for public comment in March 2026 — is a technical specification for machine-readable content compensation terms. It is a protocol, not an institution. Protocols require implementation, and implementation requires an actor with the standing to insist that AI companies adopt them. CoMP does not name the actor. It assumes one will emerge.

Article 102 competition law — pursued by Cohen and Davies, whose recent analysis argues that Google’s AI Overview constitutes a clear abuse of dominant position under European law — targets the demand-side harm sharply but cannot reach the rest. An Article 102 case is procedurally bound to the specific dispute between one giant and its victims. Even a perfectly designed remedy against Google would leave the Cloudflare-x402 toll architecture and the scraper economy untouched. Cohen and Davies acknowledge this; their proposed remedies (an opt-in regime, a cease-and-desist requiring Google to “un-launch” AI Overview, a mandatory licence fee) all target Google specifically because the procedural architecture of EU competition law forces them to.

The Microsoft Publisher Content Marketplace and Cloudflare’s pay-per-crawl are the only operational, multi-publisher marketplaces currently functioning. They run on infrastructure-layer logic — designed by infrastructure companies, governed by infrastructure companies, optimized for the efficiency goals of infrastructure companies. The question is not whether they should exist. They do, and they probably should. The question is whether publishers will also build the marketplace that sits on their side of the same transactions.

Each of these interventions is necessary. But none of them, alone or together, is sufficient. The gap they leave is institutional: there is no publisher-owned entity with the scale, legal capacity, and AI-side relationships to negotiate collectively on behalf of publishers who lack individual leverage.

Think of the scraper economy as rising sea level. A seawall on one property doesn’t matter when the water keeps rising; what’s needed is the coordinated community and state response that operates at the scale of the problem. That gap is what a broker would fill.

What the missing piece is (and how it already exists elsewhere)

A publisher-side broker is, at its core, the AI-era version of an institutional pattern that has solved variants of this problem before. When individual rights holders face counterparties with vastly more leverage, aggregation through a collective is the response that has consistently worked.

Music publishers organized into performance rights organizations — ASCAP, BMI, SESAC, and their international equivalents — to collectively license songwriting royalties when individual songwriters could not negotiate with broadcasters one by one. The Associated Press was founded in 1846 as a wire service cooperative because individual newspapers couldn’t justify the cost of reporting from distant locations alone. Getty Images aggregated photographer licensing because individual photographers couldn’t negotiate with magazines, advertisers, and (eventually) every commercial user of imagery in the world. Factiva, Dow Jones’s own content licensing operation, aggregates content from thousands of publishers and licenses it to enterprise customers under negotiated terms — a model running successfully for decades.

The AI version of this pattern doesn’t exist. There is no collective licensing body for AI training. There is no aggregator that can present a unified front to OpenAI, Anthropic, Google, Meta, Microsoft, Amazon, and the rest of the major AI companies. There is no institution that can say to a scraper economy vendor: you cannot resell the content of our 200 member publishers, and here is the legal and technical apparatus we will deploy if you try.

A publisher-side broker, in operational terms, would do four things:

Aggregate anchor inventory. The broker would assemble a corpus of high-value content from member publishers — archives, reference material, premium journalism, specialized verticals — and present it to AI companies as a single licensable product. Individual publishers contribute their content; the broker handles the commercial relationship. The economic logic is straightforward: AI companies want quality data, and a hundred-publisher consortium is a more efficient counterparty to negotiate with than a hundred publishers individually. The broker takes a margin; publishers receive a share proportional to usage.

Enforce at the technical layer. Aggregation is meaningless without enforcement. The broker would operate or partner with AI crawler detection and gating infrastructure — the kind of capability that companies like Cloudflare and smaller specialized vendors currently provide separately. Standardized content metadata (CoMP, News Atom, schema.org for journalism) provides the substrate for machine-readable terms; enforcement against scraper-economy bypasses provides the teeth. Content from member publishers would be technically protected, with the broker’s legal weight standing behind the enforcement. A scraper vendor selling Firecrawl-style API access to consortium content would face not a single publisher’s cease-and-desist but the consortium’s combined legal capacity.

Negotiate at scale across all three layers. Bilateral deals with hyperscalers happen at the top. Negotiated access through standardized terms (CoMP, or a successor protocol the broker shapes) handles the middle layer. Legal action against scraper economy vendors operates at the bottom. The broker’s value is operating across all three rather than only one.

Maintain governance on the publisher side. The structural difference between a publisher-side broker and the Microsoft Publisher Content Marketplace or Cloudflare’s pay-per-crawl isn’t the existence of a marketplace — it’s who governs it. Pricing, terms, exclusion criteria, dispute resolution, and the basic question of who gets paid how much all need to sit with publishers. This is the precondition that makes the broker model different in kind from infrastructure-owned brokerage.

Who could actually build this? The characteristics required narrow the field considerably. The broker would need existing multi-publisher content licensing infrastructure (not built from scratch). It would need established AI-side business development relationships at the senior level — not the engineers buying Firecrawl access, but the C-suite of OpenAI, Anthropic, Google, Microsoft, Amazon. It would need legal capacity sufficient to litigate against well-funded scraper economy vendors and to negotiate against the largest AI companies in the world. It would need brand authority sufficient to convene publishers who are themselves competitors. And it would need anchor inventory of its own — a sufficiently valuable internal corpus to make membership attractive from day one.

In the world’s media economy, perhaps three to five institutions meet these criteria. The clearest operational expression of the pattern, as of early 2026, is Dow Jones’s Factiva. The platform has expanded to more than 8,000 licensed sources for generative AI use since launching Factiva Smart Summary in late 2024, with recent additions including The Atlantic, USA Today Co., Fast Company, McClatchy Media, The Globe and Mail, and Hong Kong Economic Times. Factiva general manager Emma O’Brian has framed the expansion explicitly as collective negotiation on behalf of small and mid-size publishers, telling Digiday in late 2025 that the unit is positioning itself as an “effective negotiator of AI licensing rights, to unlock royalties for small and mid-size publishers on a usage basis.” Her governance test — “we would never sign a deal we wouldn’t ourselves sign” — is precisely the publisher-side governance principle this piece argues is missing from infrastructure-owned brokerages.

Factiva is not the only experiment in this direction. The News/Media Alliance has launched parallel collective-licensing arrangements with AI partners Bria and ProRata, on a structurally different model: NMA provides a templated agreement, and member publishers contract bilaterally with the AI company, with 50 percent revenue share apportioned by attribution. NMA president Danielle Coffey describes the model as triangular — “we work on the legal terms… members contract individually.” That implies a different theory of where publisher leverage comes from than Factiva’s centralized aggregation: templated bilaterals preserve publisher autonomy at some cost to collective bargaining power; centralized aggregation creates leverage but requires publishers to trust the aggregator. Whether these parallel collective approaches converge, differentiate, or compete for the same publisher supply is itself an open question.

The harder question Factiva’s existence sharpens, rather than answers, is one of scale. 8,000 sources is a meaningful corpus; it is not yet the scale at which a broker can shift the political economy of AI-to-publisher commerce. The major AI labs are not yet primarily licensing through Factiva. The scraper economy is not yet meaningfully constrained by Factiva’s enforcement capacity. The x402 architecture is consolidating on a parallel track. The question is whether Factiva at its current scope can grow into the role this piece describes, or whether the role requires a different kind of institutional move — a consortium structure with multiple anchor publishers, deeper enforcement infrastructure, an explicit posture against the infrastructure-layer brokerages — that goes beyond what any single publishing company has yet attempted.

The market is watching to see whether the move materializes at the scale the moment requires.

Why it hasn’t happened yet

The reason a publisher-side broker doesn’t exist isn’t that publishers haven’t thought of the idea. It’s that several genuine structural challenges have to be overcome to build one, and the timing window is narrow.

Publisher competitive dynamics are real and well-understood by the people involved. A broker has to be structured so that no single publisher dominates governance, which probably means a consortium with leading participants rather than a single-owner platform. That is closer to the cooperative model of the Associated Press than the platform model of Spotify. That structure is harder to organize, and it requires publishers to subordinate competitive instinct to collective interest at a moment when most publisher leadership teams are focused on their own bilateral negotiations.

AI companies will resist. The whole logic of the scraper economy is that AI companies prefer no licensing conversations to negotiated ones. A broker forces those conversations back into the equation, and AI companies will route around it where they can — through scraper economy vendors, through legal challenges to the broker’s legitimacy, through holding out and waiting for publishers to capitulate. The broker has to be large enough and authoritative enough that AI companies cannot simply ignore it. Below that threshold, it doesn’t work.

The chicken-and-egg problem. The broker needs scale to attract AI buyers but needs AI buyers to attract publishers. The solution is anchor inventory — the founding publisher (or consortium of founding publishers) has to bring enough content of its own that AI companies want to negotiate from day one, before the broader publisher membership has materialized. This is why the broker probably has to be founded by a publisher that already has substantial AI licensing relationships and a major archive, rather than by a startup or a coalition of mid-tier publishers.

Google’s position is likely the biggest structural obstacle. Joshua Benton at Nieman Lab has argued, persuasively, that no meaningful publisher licensing market will materialize in 2026 because Google refuses to separate its search-indexing crawler from its AI-training crawler. As long as participation in Google’s search index implies participation in its training data, the price for that data is effectively zero. Every other AI company sees this and asks: why pay for what our largest competitor gets free? Benton’s argument is that until regulatory action forces the separation, the licensing market is structurally stalled.

He is right about the constraint. The broker model is precisely the institutional precondition for the regulatory pressure that would lift it. Individual publishers cannot push back against Google’s crawler entanglement at sufficient scale. Bilateral lawsuits move at the speed (and cost) of litigation. The Penske Media antitrust filing from February 2026 documented Daily Mail click-through rate collapsing from 25 percent to under 3 percent when an AI Overview surfaces above its link — but as a single-publisher case, its remedy can only address Penske’s specific harms. An organized publisher block — a broker with hundreds of members and significant content market share — could change the political economy of the Google crawler question entirely. The broker doesn’t solve Benton’s problem directly. It is the actor that could plausibly force the regulatory solution Benton (correctly) says is needed but isn’t currently happening.

Timing. The x402 architecture is consolidating now. Microsoft’s PCM is operational. Cloudflare’s marketplace is moving from private beta to broader availability. The infrastructure-layer brokerage models are establishing the terms by which AI-to-publisher commerce happens. A publisher-side broker that emerges in 2027 inherits an industry where the major AI companies have already learned to operate through Cloudflare and Microsoft’s terms. A broker that emerges in late 2026 still has room to shape the protocols. Past that, the costs of restructuring rise sharply.

These are real challenges. They are not, individually or collectively, prohibitive. They are the reasons the move requires deliberate institutional action rather than emerging spontaneously, and the reason it hasn’t already happened.

What this all means

For publishers below the elite tier, the live choice is one of three paths. The first is bilateral negotiation with AI companies on whatever terms the AI companies are willing to offer — which, for publishers without major-brand leverage, typically means very little or nothing at all. The second is comprehensive blocking and a strategic retreat into direct audience relationships — subscriptions, newsletters, specialized events — accepting that AI-mediated discovery will happen without you. This is a coherent strategy that some publishers have chosen deliberately; whether it scales beyond brands with established trust is the open question, and there’s a reasonable counter-argument that the loss of discoverability in AI-driven search itself becomes a major issue. The third path is collective negotiation through a publisher-side institution that doesn’t yet exist. The argument here is not that the third path is the only viable one. It is that it is the only one that addresses the supply-side enclosure for publishers who cannot individually weather the structural isolation of the second.

For AI companies, the choice is between engaging now with a publisher-side broker on terms negotiated jointly, or facing a fragmented and increasingly hostile environment of bilateral negotiations, scraper economy legal exposure, and regulatory pressure that grows in proportion to the asymmetry of the current situation. The companies that engage early on collective terms will pay more per piece of content than they currently do through the scraper economy bypass. They will also gain access to a stable, authoritative, contractually clean corpus that the scraper economy cannot provide, and they will have done the political work to insulate themselves from the regulatory wave that’s building.

For regulators, the supply-side plumbing is where policy attention should follow once the AI Overview cases settle. Article 102 against Google is necessary but procedurally bounded. The harder regulatory work is establishing that infrastructure-layer brokerage — Cloudflare’s x402, Microsoft’s PCM — is itself a form of market power that requires governance attention. The EU’s Digital Markets Act framework was designed for a previous moment; it does not capture the layer where the current consolidation is happening. The vocabulary for that capture has yet to be developed.

For readers and the open web more broadly, the question is whether journalism, research, and the broader information commons remain economically viable as inputs to AI systems that have learned to extract their value without compensating their production. The licensing deals that have materialized so far flow to a handful of major players. The scraper economy operates without compensation at all. If a publisher-side broker doesn’t emerge in the next twelve to eighteen months, the answer to that question gets decided by default — and the default is enclosure.

The infrastructure is the politics

Google I/O 2026 will be remembered as the inflection point at which the search-as-portal model gave way to search-as-destination at scale. NPR’s “extinction-level event” framing captures the demand-side moment. But the demand-side moment is the visible part of a structural shift whose less visible parts — the toll architecture, the scraper economy, the infrastructure-layer brokerage — will determine the long-term shape of the open web more than the AI Overview interface ever could.

The flat-rate web rewarded publishers asymmetrically but preserved their formal autonomy. The pay-per-crawl web automates a transfer of rent-setting authority from the people who create content to the companies that own the plumbing. That transfer is, right now, happening by default — not because publishers have lost the argument, but because the institutional response that would carry the argument hasn’t yet been built.

The infrastructure layer has already decided publishers will be paid through its toll booth. Publishers haven’t yet decided whether they’ll be paid on those terms or their own.

The window to decide is closing.

This piece draws on a longer academic argument developed at the University of Amsterdam’s Media Studies department: “The New AI Scraper Economy and the Three-Layer Enclosure of the Open Web” (April 2026).

In May, a chart from GitHub’s own Copilot pricing calculator went viral on LinkedIn. Posted by founder Joel Griffith from a Reddit user’s original screenshot, it showed $39 in current billing under the existing premium request pricing, against $5,851.77 projected under GitHub’s new metered system taking effect June 1.

A 150x increase, give or take.

The post and many like it went viral inside the developer community for the obvious reason: “Absolute insanity. None of this seems sustainable.”

Suffice it to say, the move to metered is significant, but the framing of “insanity” or “greed” shouldn’t lead away from the whole story. The pricing didn’t suddenly go up 150x, but rather the meter changed, from a flat per-request count to actual token consumption, and the previous meter was massively under-counting what users were burning through. GitHub wasn’t gouging. GitHub was, finally, charging something closer to what the service costs to run. (With some mitigation to ease the pain.)

Still, the online chatter largely presented an outlier user, the kind running multi-hour autonomous coding sessions on the most expensive frontier model available. But the same pattern is showing up at the enterprise level, with smaller multipliers but much larger absolute numbers.

Olli-Pekka Heinisuo, co-founder at GitHits, posted his own comparison: his team’s April usage cost $306 under PRUs (the flat “premium request” units of the old system) and would have cost $3,062 under AICs (the new “AI Credits” that meter actual token consumption). A 10x increase, not 150x, but on real enterprise usage from “a handful of users.” His projection for June, if his team continued using Copilot the way they had been, was a $10,000 to $20,000 monthly bill. He concluded the obvious: “after April, that no longer really makes sense.”

That’s the actual shape of the shift. Heavy individual users see catastrophic repricing. Enterprise teams see merely punishing repricing. And in both cases, the response (unless you are looking to token max and burn money) isn’t going to be paying the new bill. It’s going to be either dramatically reducing usage, switching tools, or routing around Copilot to the underlying model APIs directly. Heinisuo’s conclusion, that Copilot’s intermediary role no longer makes sense after April, is exactly the unbundling that the metering shift is going to trigger across the industry.

That distinction matters because what’s happening at GitHub is not a Copilot story. It’s the leading edge of a structural reset across the entire AI application economy. The era of all-you-can-eat AI subscriptions is ending, and the bill that arrives next year is going to look almost nothing like the bill that arrived this year.

I want to walk through what’s actually going on, why it was inevitable, and what it means for anyone who uses these tools, builds with them, or holds stock in the companies making them. I’ll also tie this to a research paper I wrote in March on Microsoft Copilot’s “super-appification,” which identified the precise structural tension now playing out in public.

If you want to skip around: the next section explains the mechanics. The one after that is the political economy. Then there’s a section on what this means for the competitive landscape, including OpenAI, Anthropic, and the pure application-layer companies like Cursor. The final section is on what individual investors and tech workers should actually take from this.

What changed: from requests to tokens

To understand the magnitude of the shift, it helps to understand what the old meter measured versus the new one.

Under the legacy GitHub Copilot system, you paid for “premium requests” (PRUs). A PRU was a unit of work, priced at $0.04, with each subscription tier carrying a monthly allotment. Pro got 300, Pro+ got 1,500. The catch was that a “request” was a flat unit regardless of what it actually cost GitHub to fulfill. A single autocomplete suggestion was a request. A multi-hour autonomous agent session that read fifty files, ran tests, debugged failures, and wrote a pull request was, depending on how it was structured, also one request, or maybe a handful.

This was fine when Copilot was an autocomplete tool. It became absurd when Copilot became an agent.

The new system meters AI Credits (AICs), priced at $0.01 each, against actual token consumption: input tokens, output tokens, reasoning tokens, tool calls, the whole stack. An autocomplete still costs almost nothing. An agent session that previously registered as one PRU might now meter as 5,000 or 50,000 AICs depending on which model it called and how many round trips it made. GitHub’s own calculator confirms the math: the same workload that registered as $39 of PRU-based billing in April reprices as $5,851 of AIC-based billing.

The Copilot CPO, Mario Rodriguez, said the quiet part out loud in the announcement: “GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable.”

“Absorbed” is certainly doing a lot of work in that sentence. What it means in practice is that Microsoft, GitHub’s parent company, was carrying the gap between subscription revenue and actual GPU consumption on its consolidated balance sheet. As long as Copilot was selling autocomplete, the gap was manageable. Once it started selling agents that could spend hours of Claude Opus 4.7 inference time on a single task, the gap became a structural loss center.

The new metering doesn’t just close that gap. It also makes the cost of agentic AI legible to the buyer for the first time. Under flat-rate pricing, you didn’t know that your morning of vibe coding cost $300 in compute, because the bill was always $39. Under metered pricing, you watch the meter run. (That may dent the magic of vibe coding somewhat.)

That legibility is, I think, the most underrated consequence of this shift. We’re about to find out what AI actually costs. Compounded with research that points to frequent failures with internal AI pilot projects (a non-technical employee is suddenly tasked with building a complex retrieval-augmented generation system, for example) the landscape seems ripe for conflicting, and ultimately misleading signals.

Why the gap existed in the first place

The flat-rate model was never built for inference economics. It was inherited from the SaaS playbook of the last twenty years, which was built on a foundation that simply no longer applies.

In classic SaaS, the marginal cost of serving one more user is essentially zero. The same Salesforce instance, the same Gmail backend, the same Slack workspace serves a hundred thousand users or a hundred million with the same underlying infrastructure cost per user, more or less. Seat pricing made sense because seats were the thing that varied while costs stayed flat.

Generative AI breaks this. Every prompt is a discrete physical event that consumes GPU cycles, electricity, cooling, network capacity. The marginal cost of an additional user is not zero. The marginal cost of an additional user running an agent is, depending on the model and the workload, somewhere between meaningful and ruinous. Anthropic’s Claude Opus 4.7, for instance, can burn through hundreds of dollars of compute on a single multi-hour coding session.

That’s a 25x loss per heavy user. And those are precisely the users the entire industry has been telling you to become. “Agentic” workflows, “autonomous coding,” “AI teammates,” all of the marketing language of the past eighteen months has been pushing users toward exactly the consumption patterns that make flat-rate pricing impossible.

Why did the companies do this anyway? Two primary reasons, I reckon.

The first is competitive. Once one major player offered flat-rate access to frontier models, everyone had to. Cursor priced aggressively because it was trying to take share from GitHub Copilot. GitHub priced aggressively because it was trying to defend against Cursor. Anthropic and OpenAI kept Pro tiers cheap because they were trying to win consumer mindshare against each other. The pricing was set by competitive dynamics, not by unit economics.

The second is strategic. Capturing developer mindshare during the inflection moment of generative AI was worth burning capital. Microsoft could afford to lose money on Copilot for years if it meant locking in the position as the default AI coding tool by the time the market matured. Anthropic could afford to subsidize Claude Code as long as it was the leading edge of agentic adoption. The flat-rate phase was, explicitly, customer acquisition funded by deep pockets and venture money.

The strategic phase ends when the bill comes due. For Microsoft, the bill is the $34.9 billion in quarterly capital expenditure required to build “planet-scale cloud and AI factories,” which Amy Hood justified to investors on the Q1 FY26 earnings call. That spending needs returns, and returns require monetization that scales with consumption rather than seats.

This is where I want to pull in some research I did earlier this year, because the tension I just described was already visible if you knew where to look.

The whitespace problem

In March, I wrote a research paper on Microsoft Copilot’s evolution from a chatbot into a “super app,” contrasting how the product was marketed to consumers against how it was sold to investors. The thesis, in one sentence: Microsoft was telling consumers and investors radically different stories about Copilot, and the gap between those stories was becoming structurally untenable.

On the consumer side, Copilot was marketed as a frictionless, personalized digital companion, perfect for brainstorming, quick questions, or “just venting.” The App Store descriptions emphasized simplicity, intimacy, ease of use. The massive cloud infrastructure powering it was deliberately obscured.

On the investor side, the same product was framed as the demand-generation layer for Azure’s $34.9B quarterly capex. The operative metric in Microsoft’s internal Partner Center documentation was “whitespace,” which is worth pausing on because it’s the term that ties this whole shift together.

Whitespace refers to the gap between users who have access to a product and users who actually pay for it. For Copilot, the math was striking: 450 million commercial users with some form of access, 3.3% paid adoption, a 435 million seat gap that Microsoft executives discussed openly as “Free to Paid Whitespace” on the investor dashboard. Every quarter, Amy Hood faced shareholders who wanted to know when that whitespace would close, when those 435 million users would start generating revenue commensurate with the $34.9 billion being spent each quarter to build the AI infrastructure they were running on.

The paper’s central finding was that this dual rhetoric couldn’t hold. The consumer story sold simplicity. The investor story demanded monetization. Something had to give. I documented the April 15, 2026 paywall shift, when Microsoft severely limited free Copilot access inside Word, Excel, and PowerPoint, as the moment when the financial mandate started bleeding through into the user experience.

What I didn’t fully anticipate, writing in March, was how quickly the same logic would spread to GitHub Copilot’s developer product, and from there to the broader industry. The April paywall was Microsoft saying “free users will become paid seats.” The June AIC system is Microsoft saying “paid seats will become metered consumers.” Both moves are downstream of the same whitespace pressure. The whitespace didn’t close fast enough through seat conversion, so Microsoft is closing it through metering, which has the additional benefit of having no per-user ceiling. A heavy Copilot user can now generate $5,000 of revenue instead of $39, without any change in the underlying product.

On the Q3 FY26 earnings call, Satya Nadella made the strategy explicit: “Any per user business of ours, whether it’s productivity or coding or security, will become a per user and usage business.” That’s the policy version of the structural argument my paper was making. The entire seat-based SaaS model is being publicly retired by the largest enterprise software company in the world.

One last thread from the paper is worth pulling forward, because it explains how the metering layer becomes the new rent extraction point. Platforms increasingly decentralize at the application layer, inviting third-party plugins, agents, and integrations into their ecosystem, while centralizing at the infrastructure layer. The AIC system fits this pattern precisely. GitHub now lets you choose Claude, GPT, Gemini, or open models from inside Copilot. Decentralization at the model layer. But GitHub controls the metering, the multipliers, the AIC pricing, and the tollbooth at which every model call passes through. Centralization at the rent-extraction layer.

The point I want to make here is just that the GitHub repricing isn’t a surprise event. It’s the predictable next step in a pattern that’s been building for at least a year, driven by a whitespace gap that seat-based monetization could no longer close.

The competitive shakeout

The metered shift hits different parts of the AI stack very differently. This is where it matters for anyone watching the industry as either a user or an investor.

Vertically integrated players (Microsoft, Google, and to a growing extent Anthropic) come out ahead. They own multiple layers of the stack: infrastructure, models, applications, distribution. When metering reveals the true cost of inference, they can absorb it internally or pass it through with margin intact. Microsoft especially benefits because metering reframes Azure from a cost center supporting Copilot into the visible revenue source for AI consumption. The Nadella quote isn’t just about Copilot. It’s about restructuring every Microsoft product so that Azure consumption shows up explicitly on the bill.

Pure model providers (OpenAI most acutely, also Mistral, Cohere) are in a more complicated position. They own the model but not the infrastructure. OpenAI’s deepest cloud and IP entanglements remain with Microsoft, even as Microsoft increasingly competes with OpenAI directly through its own MAI models and through Copilot’s model-routing flexibility. The metering shift forces OpenAI to either renegotiate that relationship, diversify infrastructure (which they’re attempting through Oracle, CoreWeave, and the Stargate project), or watch Microsoft capture more of the value chain.

Anthropic, by contrast, has spent the past year building infrastructure optionality. The recent xAI deal, in which Anthropic effectively leased Elon Musk’s entire Colossus 1 data center in Tennessee, gave them 300 megawatts of immediately-available compute and 220,000 Nvidia GPUs. That’s on top of a 5 GW commitment from Amazon, a 5 GW commitment from Google, and a Broadcom partnership for custom silicon. Anthropic is now structurally less dependent on any single hyperscaler than OpenAI is, which gives them more pricing flexibility and more capacity to grow into the metered era without supply constraints. Their $30B annualized run rate, up from $9B at the end of 2025, suggests this is working.

Application-layer wrappers (Cursor, Windsurf, Cline, the broader category of “Claude-powered” and “GPT-powered” tools) are in the worst position. They own nothing structural. They pay Anthropic or OpenAI near-retail API rates and resell access wrapped in a nice IDE. The flat-rate model was their entire moat against the model providers and the hyperscalers. The metered shift removes that moat. The $20/month Cursor plan against $5,000 of Claude consumption is not a pricing strategy, it’s a venture-subsidized customer acquisition play with no clear path to unit economics. Watch for one of these companies to be acquired by a vertically integrated player within the next 18 months, or to dramatically reprice and lose users.

There’s also a reported $60B option for SpaceX to acquire Cursor by late 2026, which I’d read less as a vote of confidence in Cursor than as Elon Musk hedging against xAI’s competitive failure. Either way, the application wrapper category is consolidating, and metering is the proximate cause.

What this means if you use these tools

The simplest version: budget for usage, not seats. The $20/month subscription you’re paying now is going to be replaced, on a timeline measured in months rather than years, by metered billing where heavy users pay hundreds or thousands of dollars and light users pay less than they do now.

If you’re a developer using AI coding tools daily, you should expect:

Your tools will start showing you a meter. GitHub’s AIC system, Anthropic’s usage dashboards, OpenAI’s credit tracking. The legibility shift is happening fast.

Heavy agentic workflows will become noticeably expensive. The same task that “used to be free” will start showing a cost. This isn’t a bug, it’s the entire point.

Model choice will become a cost decision. If Claude Opus 4.7 costs 27 AICs per request and GPT-5.4 costs 6, you’ll start routing simpler tasks to cheaper models. Tool vendors will help you do this automatically. The “always use the best model” reflex is about to become unaffordable.

Enterprise contracts will start looking different. The flat-rate seat license is being replaced by a hybrid: base seat fee plus metered consumption pool plus overage rates. Procurement teams who haven’t dealt with consumption-based AI billing will get an unpleasant education.

If you’re not a developer but you use AI tools through M365, ChatGPT, Claude, or Google’s products: the same shift is coming, just slower. Nadella’s “every per user business” statement is the timeline. Expect your AI features to start having explicit limits, expect those limits to become metered, and expect the bill to start reflecting your actual usage rather than a flat fee that hides it. Is that an apocalyptic scenario? Well, it depends on how one looks at it. For retail users, one might argue it’s more akin to the end of using a large Hadron Collider to vibe code a daily habits app.

How bad is it actually

A few moderating observations are worth making, because the posts on LinkedIn and on dev Reddit driving the public reaction are unrepresentative in specific ways.

The user staring at a $5,851 projected bill was running an outlier workload, almost certainly long autonomous coding sessions on the most expensive frontier model available, probably overnight, probably with redundant tool calls. That’s not the median Copilot user. That’s a user the subsidy was bleeding hardest to support, and arguably should have been priced differently from the start. Heinisuo’s $306-to-$3,062 case is more representative of a sophisticated enterprise team but still selected, his company builds coding agents, so his developers’ usage profile is by definition heavier than a normal enterprise dev team’s.

A median Microsoft enterprise customer using Copilot for autocomplete, occasional chat, and the occasional agent run is likely to see a much more modest increase, maybe 1.5x to 3x, possibly absorbed within existing AIC allotments at higher tiers. GitHub has telemetry on every user’s actual consumption. The new pricing wasn’t set in a vacuum, it was modeled against existing usage distributions to capture revenue from heavy users without driving away the median. The $5,851 user was, from GitHub’s perspective, the intended casualty.

There’s a related dynamic worth flagging. Research on internal enterprise AI pilot projects shows high failure rates, with non-technical employees being tasked to build complex systems (retrieval-augmented generation, agent orchestration, custom fine-tunes) that they aren’t equipped to design or maintain. A lot of the consumption that was happening under flat-rate subsidy was, frankly, badly-designed projects burning compute on unsuccessful pilots. The metering shift exposes that cost at the exact moment enterprises are realizing many of their AI initiatives aren’t generating returns. Expect a wave of “AI is too expensive” conclusions over the next year that are actually conclusions about implementation quality rather than about the tools themselves.

The longer arc also matters. Token costs on the supplier side are still falling. Anthropic’s Sonnet, for example, is dramatically cheaper per token than Claude 3 Opus was 18 months ago, and the trajectory continues. If model efficiency improves 2–3x per year and GPU economics improve another 1.5–2x annually through Blackwell, custom silicon, and inference optimization, the absolute cost of a typical agent run two years from now could be a fraction of what it is today. The metering shift is probably better understood as a one-time legibility shock followed by gradual price normalization, not a permanent step-change in what AI costs.

So if you’re asking whether the median AI tool user is about to face $5,000 monthly bills, the answer is no… probably not. The median will normalize. The heavy edge users will either pay up or change their workflows.

What is permanent, though, is the control shift that the metering establishes. Once GitHub has built the AIC infrastructure, the AIC-to-dollar conversion is a lever they can pull at any time. Today it’s $0.01 per AIC. They can move it to $0.012 quietly, adjust multipliers on specific models, change the included allotments per tier, and most users won’t notice until their bill arrives. Seat pricing was sticky and visible. Metered pricing is variable and opaque. That’s a durable shift in power between the platform and its users that doesn’t depend on where any specific price point lands.

The other durable consequence is the unbundling pressure Heinisuo’s post hinted at. The flat-rate era hid the fact that GitHub was charging a markup to route you to Anthropic and OpenAI. The metered era exposes it. Sophisticated enterprise buyers seeing $3,000 monthly Copilot bills will start asking why they don’t just pay Anthropic $3,000 directly. Anthropic’s enterprise sales team is having exactly that conversation right now with companies receiving repricing notices. So even if GitHub’s price levels turn out to be reasonable, the visibility itself accelerates a competitive dynamic that pulls revenue away from the application layer and toward the model providers and infrastructure owners.

The honest read, then, is that the metered era is not the end of accessible AI tools, but it is the end of unconscious AI consumption. From now on you’ll know what you’re using and what it costs, and the companies that own the meter will know it too, in much finer-grained detail than they did before.

What this means if you hold the stocks

The metered shift has clear implications for AI-exposed equities depending on where the company sits in the stack. This is not an exhaustive market analysis, but rather my contribution in this specific context.

The vertically integrated incumbents (Microsoft, Google, Amazon) are positioned to extract more revenue per user as metering takes hold, because they own the infrastructure layer where the meters live. Their margins on AI services should improve as the subsidies wind down. The risk is regulatory, given that the metering layer is also where the rent-extraction concentration becomes legible to antitrust authorities.

Anthropic isn’t public, but its position is interesting for anyone watching Amazon, Google, and (now) SpaceX’s AI strategies. Anthropic’s compute diversification reduces its dependence on any single hyperscaler, which means the value it generates flows to multiple infrastructure providers rather than enriching one platform. Amazon and Google both get to claim Anthropic as an anchor tenant for their AI infrastructure investments. The xAI deal adds a third beneficiary: SpaceX, which is reportedly absorbing xAI ahead of a public offering and badly needs Colossus 1 to show revenue rather than just GPU count. An Anthropic anchor tenant turns that asset from a money-burning training cluster into a cash-flowing inference business, which materially improves the story SpaceX will tell its S-1 readers.

OpenAI’s current position seems the most precarious of the model providers. The Microsoft relationship is increasingly adversarial in ways that the financial structure can’t fully accommodate, and OpenAI’s effort to diversify infrastructure (Oracle, Stargate, CoreWeave) is happening under time pressure. Watch for a major restructuring of the Microsoft-OpenAI relationship within the next 12–18 months, possibly in the form of OpenAI buying out parts of the original investment or Microsoft monetizing its equity at a peak valuation.

The pure application-layer category (Cursor and others, which are private but proxied through valuations and venture market sentiment) is facing a margin compression event that no amount of growth can fix. The companies that survive will do so by being acquired into a vertically integrated stack or by integrating downward themselves, building models, fine-tuning, or owning infrastructure. The companies that don’t will quietly shut down or merge.

For the broader market, the metered shift is the structural mechanism by which the $2.9 trillion in projected global data center construction by 2028 (Morgan Stanley’s estimate) gets monetized. The capital expenditure cycle that’s been driving Microsoft, Google, Amazon, Meta, and Oracle through 2025 and 2026 only generates returns if metered AI consumption scales to fill the capacity. The flat-rate phase was a bet that demand would eventually justify the build. The metered phase is the mechanism for actually capturing that demand as revenue.

If you believe agentic AI is going to be a major workflow shift across the global economy, the metered companies are well-positioned. If you don’t, the capex bill comes due against insufficient revenue, and the equities that are most exposed are the ones whose stories most depend on continued AI revenue acceleration.

The flat-rate moment was always going to end

Consider that unlimited mobile data was a beautiful era. For a few years in the 2010s, you paid your carrier a flat fee and you used as much data as you wanted. The carriers absorbed the cost because they were fighting for subscribers and the marginal cost of bandwidth was still falling. Then the cost stopped falling, or rose, depending on the metric you used, and the unlimited plans got capped, throttled, tiered, and eventually replaced by usage-based billing dressed up as “tiered unlimited” with caps on HD video and other bandwidth sucking activities.

The flat-rate AI subscription is in roughly the same place mobile data was in 2014. The economics worked while the underlying cost was being subsidized by competitive pressure and venture capital and hyperscaler ambition. When those subsidies end, the meter starts running.

What’s distinct about this moment versus the mobile data shift is the speed of the transition. Mobile data took six or seven years to move from unlimited to metered. The AI flat-rate era launched in 2022 with ChatGPT Plus, peaked in 2024 with Copilot and Cursor, and is being structurally dismantled in 2026. Three and a half years from start to repricing.

That speed is itself a tell. The unit economics never worked. The subsidies were holding back a much larger cost than the consumer experience suggested. And the moment when the largest companies in the industry, Microsoft, Anthropic, OpenAI, all begin shifting in the same direction within weeks of each other, is the moment when you can stop calling it a pricing change and start calling it a structural reset.

The shock isn’t that GitHub repriced. The shock is what the new price reveals, which is partly real and partly designed. The metered calculator that shows you a $5,851 projected bill also shows you an upgrade prompt for a higher tier that would absorb most of it. The subsidy isn’t ending so much as being repackaged as a premium feature, where what you’re paying for is the comfort of not watching the meter run. That’s the durable shift. From now on, the question isn’t what AI costs. The question is what your provider has decided to let you see of what it costs.

My March 2026 paper “Selling the AI Super App: A Historical Analysis of Microsoft Copilot’s Rhetoric Across App Stores and Financial Disclosures” underlies much of the structural argument in this post.