The Allocation Turn: When Compute Is Rationed by Capacity, Not Price

There is a sentence buried in the reporting about Google and Meta that should stop anyone who still thinks of cloud compute as something you simply buy. Around March of this year, Google told Meta that it could not supply as much Gemini capacity as Meta wanted, and Meta — a company with a market capitalization north of a trillion dollars and more cash than most national treasuries — responded by telling its own engineers to start conserving tokens. Not to negotiate a better rate. Not to sign a bigger contract. To conserve. To use less of a thing they were willing and able to pay full price for, because the supplier had run out of the thing to sell.

On May 17, Google formalized what had been an informal squeeze, imposing compute-based usage limits on Gemini access broadly. The operative phrase in the new regime is the one worth memorizing: access now scales with available capacity, not with how much you are willing to spend. For most of the history of cloud computing, that sentence would have been incoherent. The entire promise of the cloud was elastic supply — the idea that capacity was effectively infinite and the only question was how much you wanted to pay for it. Money bought compute. That was the deal. It is not the deal anymore.

This is what I want to call the allocation turn: the moment the market for AI compute stopped clearing on price and started clearing on rationing. In a normal market, when demand exceeds supply, the price rises until enough buyers drop out to match the available quantity. The market clears, painfully but cleanly, through price. What happened between Google and Meta is different in kind. The price did not rise until Meta walked away. The supplier simply said no — allocated the scarce capacity by priority, relationship, and strategic interest rather than by the highest bid — and the richest buyer on earth got rationed anyway. When price stops being the mechanism that clears a market, something fundamental about that market has changed, and everything downstream of it has to change too.

The tell: what actually happened between Google and Meta

Strip the story to its mechanics and the significance becomes hard to miss. Meta, in the course of building out its internal AI systems, had been buying access to Google's Gemini models — using a competitor's frontier model to power its own workloads, which is itself a sign of how commoditized the model layer has become. Meta wanted more. Google, despite spending on the order of 180 billion dollars on AI infrastructure this year, could not give Meta all the capacity it requested, because that capacity did not exist to be sold. The result was delay: multiple internal Meta AI projects slipped, and Meta instructed employees to use tokens more sparingly and to improve their efficiency per call.

Look closely at the asymmetry. Google is not a marginal supplier scraping to serve a whale. Google is one of the three or four best-capitalized compute providers on the planet, in the middle of the largest infrastructure buildout in its history, and it still had to ration a strategic customer. And Meta is not a startup that ran out of runway. It is the buyer that, in the old world, would have been able to summon any quantity of anything by writing a large enough check. Both ends of this transaction are among the most powerful economic actors alive, and the transaction still failed to clear on price. That is the tell. When the constraint binds this hard this high up the food chain, it is not a temporary imbalance. It is the new shape of the market.

There is a second detail that matters as much as the cap itself. While Google was rationing Gemini to Meta, it was simultaneously securing additional GPU capacity from SpaceX — buying compute from an unlikely source to feed its own demand. A company spending 180 billion dollars a year on infrastructure was, at the same moment, both a rationer of compute to its customers and a scavenger of compute for itself. If that does not convince you that capacity, not capital, is the binding constraint, nothing will.

From price to allocation: the mechanism actually changed

Economists have a precise way to describe what has happened, and it is worth being precise, because the imprecise version — "there is a GPU shortage" — misses the structural point. A shortage in the ordinary sense is a price signal: the good is underpriced relative to demand, the price rises, and the shortage resolves. What is happening in AI compute is a rationing regime, where the supplier holds price roughly fixed relative to contract and instead allocates a fixed quantity among competing claims by criteria other than price. The two look similar from a distance — in both cases you cannot get as much as you want — but they behave completely differently, and the difference determines strategy.

The chart is a schematic, not a measurement, but it captures the shape of the argument. In 2023, if you wanted more compute, you paid more and you got it; price was the lever. By 2026, the lever that actually moves capacity is allocation — your place in a queue determined by contract size, strategic relationship, and how far in advance you committed. The willingness-to-pay curve has flattened at the top, because above a certain point there is simply nothing more to buy at any price, and the thing that distinguishes buyers is no longer their budget but their position.

Why does this matter beyond the vocabulary? Because a price-clearing market and an allocation-clearing market reward completely different behavior. In a price-clearing market, the winning move is to have money and be willing to spend it. In an allocation-clearing market, the winning move is to secure position early, lock in reserved capacity, build relationships with suppliers, and — the endgame — remove yourself from the allocation queue entirely by owning the supply. Money still matters, but it matters as a means to position rather than as a substitute for it. This is why the wealthiest AI companies are behaving less like buyers shopping for the best price and more like nations securing strategic reserves. They understand that the market no longer clears on price, so being rich is necessary but no longer sufficient.

Why price stopped clearing the market

For price to clear a market, supply has to be able to respond to price. Charge more, and more of the good appears. The reason AI compute has slipped out of the price-clearing regime is that the supply chain underneath it cannot respond to price on any timescale that matters, because every link in it is gated by physical lead times measured in years rather than weeks.

Start at the top. When Google, Microsoft, Amazon, and Meta placed multi-billion dollar forward orders for Nvidia's Blackwell generation in 2025, those orders consumed most of the available allocation through the end of 2026 and into 2027. That is not a market that can respond to a new buyer's higher bid; the output is already spoken for years ahead. Go one level down. The advanced packaging that bonds high-bandwidth memory onto the accelerator substrate — TSMC's CoWoS process — is fully allocated through at least mid-2027, and you cannot conjure a new packaging line with a purchase order; it takes years and billions to build. Go down again, to the high-bandwidth memory itself, which is dominated by a handful of manufacturers running flat out. Down once more, to the leading-edge fabrication capacity, and finally to the data center power and the grid interconnects, which I have written about at length as the deepest constraint of all.

When every link in the chain has a multi-year lead time, the aggregate supply curve for AI compute is nearly vertical in the short run. A vertical supply curve means price can rise without quantity responding at all — you can pay double and still get nothing more this year, because the physical capacity to make more does not exist yet and cannot be summoned by a bid. In that regime, a rational supplier stops using price as the rationing device, because raising price on a fixed quantity just transfers surplus without solving anyone's capacity problem and alienates the strategic customers you most want to keep. Instead the supplier rations by hand, allocating the fixed pie to the claims it values most. That is exactly what Google did to Meta, and it is what every capacity-constrained provider is now doing, whether or not they describe it that way.

This is the direct downstream consequence of the shift I described in the physical layer turn, where the binding constraint on AI moved past the processor to the memory that feeds it and the power that runs it. Once the constraint lives in physical assets with multi-year lead times, the market that sells access to those assets cannot clear on price, because price cannot summon atoms fast enough. The allocation turn is the physical layer turn observed from the demand side: if the substrate is scarce and slow to build, then access to the substrate must be rationed, and rationing is what we are now watching happen in public.

Allocation as a weapon

Here is where the story stops being about economics and starts being about power. When a supplier rations a fixed quantity by discretion rather than by price, the act of allocation becomes a strategic instrument. Who you serve first, who you throttle, and who you cut off entirely are now decisions with competitive consequences, and the companies making those decisions are often the same companies competing with the customers they are rationing.

Consider the Google-Meta relationship from this angle. Google sells Gemini capacity to Meta, and Google also competes with Meta across advertising, social, consumer AI, and the race for frontier models. When Google decides how much of its scarce capacity to allocate to Meta versus to its own products versus to other customers, it is making a decision that shapes a competitor's ability to ship. I am not alleging that Google throttled Meta out of malice; the simpler explanation, that Google genuinely ran short and served its own needs first, is almost certainly the true one. But that is precisely the point. Even the benign version of allocation is a competitive act, because a supplier that is also a rival will, entirely rationally, serve itself before it serves you. The moment your compute supplier is also your competitor, your growth runs on their sufferance.

The allocation hierarchy is not a conspiracy; it is a rational ordering that any capacity-constrained supplier converges on. A provider will allocate reserved capacity to a hundred-million-dollar committed customer before it releases on-demand inventory to a startup, because the committed customer has paid for priority and the startup has not. A provider will serve its own products before either, because those products are the reason the infrastructure exists. What this produces, in aggregate, is a market where your access to compute is a function of your position in someone else's priority stack — and that position is something the supplier controls and can change. Compute has become a lever that suppliers can pull on the companies that depend on them, and the largest AI companies have noticed.

The vertical-integration reflex

The rational response to being rationed by a supplier who is also a rival is to stop depending on that supplier. This is why the deepest consequence of the allocation turn is a wave of vertical integration — companies racing to own the layers they were previously happy to rent, precisely so that no one can ration them.

Watch what Meta did in response to the Gemini cap. The reporting is explicit that the restriction accelerated a transition Meta was already pursuing: a shift away from reliance on external frontier models toward internal alternatives capable of handling critical workloads at scale. Meta began moving workloads to its own Muse Spark model to reduce external dependence. The cap did not just delay some projects; it taught Meta a lesson it will not forget, which is that depending on a competitor's rationed capacity for a critical workload is a strategic vulnerability. The cheapest insurance against being told no is to own the thing you would otherwise have to ask for.

The same reflex is visible everywhere you look. Anthropic is reportedly in talks with Samsung to build a custom AI chip, moving to own its silicon rather than compete for someone else's allocation. The hyperscalers have spent years building their own accelerators — Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia — specifically so that their most important workloads do not sit in Nvidia's allocation queue behind everyone else. The pattern is consistent: at sufficient scale, the answer to an allocation-clearing market is to exit the market by building the supply yourself. You do not want to be a buyer in a market that rations its buyers; you want to be your own supplier.

This is the same logic I traced in the training decoupling, where a Chinese lab built its stack on domestic chips to escape dependence on a supply it could not control. Whether the constraint is an export control or a capacity cap, the strategic response is identical: reduce your dependence on a supply someone else allocates. The allocation turn generalizes that logic from geopolitics to ordinary commerce. You do not need a trade war to be cut off from compute. You just need a supplier who runs short and serves someone else first.

The mid-market gets crowded out

Every allocation regime has winners and losers, and the losers here are predictable: the companies too small to command priority and too large to run on scraps. The hyperscalers can build their own chips. The Metas and Anthropics can sign reserved-capacity deals and pursue custom silicon. But the vast middle of the market — the mid-sized enterprises and the well-funded-but-not-giant startups — has neither the scale to own supply nor the leverage to command a priority tier, and they are the ones getting squeezed out of the queue.

The mechanism is straightforward. When hyperscalers place enormous forward orders that consume most of a chip generation's allocation, they crowd out the mid-market customers who previously bought through standard channels and resellers. When a provider allocates reserved capacity to its hundred-million- dollar customers before releasing on-demand inventory, the startup that runs on on-demand finds the inventory gone. The result is a bifurcation: a small number of giants with secured supply, and a long tail of everyone else competing for whatever is left over, at prices and availability the giants do not have to tolerate.

The chart shows the structural reason the squeeze persists: deliverable capacity keeps growing, but demand grows faster, so the gap that has to be allocated by rationing rather than filled by price does not close. As long as that gap exists, someone has to be told no, and the someone is whoever sits lowest in the priority stack. For the mid-market, this changes the entire calculus of building on AI. You can no longer assume that the compute you need will be available when you need it at a price you can plan around. You have to treat compute the way a manufacturer treats a scarce raw material — something you secure in advance, diversify your sources of, and design your product to consume as little of as possible.

This is the demand-side companion to the repricing I described in the efficiency turn, where the industry stopped paying for maximal token consumption and started paying for value delivered per token. Efficiency is not merely a cost discipline anymore; it is an allocation discipline. Every token you do not need to spend is a token you do not have to secure a place in a queue for. When capacity is rationed, the most efficient consumer of compute has a genuine strategic advantage over a wasteful competitor, because it needs less of the scarce thing to deliver the same product.

This has happened before, and it always ends the same way

The allocation turn feels unprecedented if AI is the only industry you have watched, but it is one of the most familiar patterns in economic history. Every time a critical input goes structurally scarce faster than its supply can be built, the market for that input abandons price-clearing and adopts rationing, and the downstream industry reorganizes itself around securing the input rather than buying it. The specifics change; the shape never does.

The oil industry learned it in the 1970s, when supply shocks turned crude from a commodity you bought at a posted price into a strategic resource nations stockpiled, hedged, and went to considerable lengths to secure at the source. Airlines learned it with landing slots at capacity-constrained airports, which are not sold to the highest bidder in real time but allocated administratively, grandfathered, and traded as scarce rights — because the runway cannot be made larger by paying more for a takeoff. The semiconductor industry itself learned it repeatedly, most recently in the pandemic-era chip shortage, when automakers who had treated chips as a just-in-time commodity discovered they were at the back of an allocation queue behind consumer electronics buyers who had committed earlier and larger, and lost billions in unbuilt vehicles as a result. In every case the lesson was the same: when supply cannot flex, the firms that secured position early and diversified their sourcing survived, and the firms that assumed the input would simply be there when they needed it got rationed at the worst possible moment.

AI compute is now running the same play, only faster and with higher stakes, because the input in question is not a raw material that goes into a product but the substrate of the product itself. An automaker rationed on chips builds fewer cars; an AI company rationed on compute cannot run its product at all. That is why the vertical-integration reflex is so much more aggressive here than in earlier scarcity episodes — the dependency is more total, so the incentive to escape it is more total. But the underlying dynamic is the oldest one in industrial economics. A scarce, slow-to-build input reorganizes the industry that depends on it around control of supply, and the winners are the ones who saw it early and moved to secure their position before the queue formed. The companies still treating AI compute as an elastic commodity are the automakers of 2021, and the ceiling is coming for them the same way it came for everyone who made that mistake before.

What buyers should actually do

If you build on AI and you are not a hyperscaler, the allocation turn has concrete implications that are worth stating plainly, because the instinct carried over from the elastic-cloud era — assume capacity is always there, buy it on demand, optimize later — is now actively dangerous.

The first move is to secure position rather than assume availability. In an allocation-clearing market, the reserved-capacity contract is not a way to save money; it is a way to guarantee you get served at all. The companies that will be caught short are the ones running critical workloads on on-demand inventory, because on-demand is the first thing a rationing supplier cuts. The second move is portability. If your compute supplier is also a potential competitor, or simply a single point of failure that can run short, you want your workloads to be able to move — across providers, across models, across accelerators — so that being throttled by one supplier is an inconvenience rather than an outage. I have argued before for building a resilient multi-provider posture rather than betting the company on a single model vendor, and the allocation turn raises the stakes on that argument from cost optimization to survival.

The third move is efficiency, for the reason above: in a rationed market, using less of the scarce input is a competitive advantage, not just a line-item saving. And the fourth move, available only at sufficient scale, is to own the critical path — your own model for the workloads you cannot afford to have rationed, and eventually your own silicon or reserved power. Most companies will never reach the scale where owning silicon makes sense, but every company should know where its critical-path dependencies sit and whether any of them run through a supplier who could, on a bad quarter, decide to serve someone else first.

What actually breaks

If this analysis is right, the failure modes of the AI industry shift with it, and they are worth naming because they are not the failure modes people are watching for.

The first thing that breaks is the assumption of elastic supply baked into a generation of business plans. A great many AI products were designed on the premise that compute would always be available in whatever quantity, at a price that would fall over time. In an allocation-clearing market, that premise is false: the compute might not be available at any price you can pay, and the companies that planned around infinite elastic supply will discover the ceiling at the worst possible moment, when a growth spurt collides with a supplier's capacity cap. Meta discovered its ceiling and had the resources to build around it. A smaller company that discovers the same ceiling may simply stall.

The second thing that breaks is the neutrality of the compute supplier. As long as capacity was elastic, it did not much matter that your cloud provider also competed with you in some markets, because the provider had no reason and no ability to disadvantage you — there was always more capacity to sell. Once capacity is rationed, that neutrality is gone. Your provider now makes daily decisions about who gets served, and if you compete with them, you are on the wrong side of those decisions. The comfortable fiction that infrastructure is a neutral utility does not survive contact with scarcity.

The third thing that breaks is the mid-market's access on ordinary terms. If the gap between demand and deliverable capacity stays open for years — and every lead time in the supply chain says it will — then the bifurcation between the supply-secured giants and the rationed everyone-else hardens into the permanent structure of the industry. That has consequences for competition, for where innovation can happen, and for who gets to build at the frontier. An industry where only the companies that own their own compute can be sure of running is an industry with a very high wall around its center.

None of this means the buildout fails. On the contrary — TSMC just posted record revenue on exactly this demand, which tells you the atoms are being made and sold as fast as they can be. The buildout is working. It is simply not working fast enough to restore the elastic-supply world, and until it does, the market will keep clearing on allocation rather than price, and the companies that understand the difference will keep pulling away from the ones that do not.

The turn, stated plainly

For fifteen years, the defining promise of cloud computing was that capacity was something you summoned rather than something you secured. You did not build a data center; you rented one by the hour, in whatever quantity you needed, whenever you needed it. Supply was elastic, the market cleared on price, and no serious buyer was ever told no. That world produced a generation of companies that treated compute the way earlier generations treated electricity from the wall — abundant, metered, and never in doubt.

The allocation turn ends that world for AI. Compute is no longer summoned; it is secured, rationed, and fought over. The market clears on your position in a queue rather than on the size of your check, and the queue is controlled by suppliers who are increasingly your competitors. The proof is that the two richest and most capable actors in the entire industry — Google as supplier, Meta as buyer — ran into the ceiling in public, and the buyer, unable to pay its way past the limit, started building its own model instead. When the wealthiest buyer on earth responds to a capacity cap by conserving tokens and going vertical, the era of elastic compute is over, and the era of allocated compute has begun.

The strategic lesson is old and unglamorous, the same one every industry learns when a critical input goes scarce: control your supply, or someone else controls your growth. The companies that internalize it early — securing reserved capacity, staying portable, ruthlessly efficient, and where they can, owning the substrate — will keep building at the frontier. The companies that keep assuming the compute will simply be there will find, at the moment they most need to scale, that it is not, and that no amount of money will change the answer, because the answer was never about money. It was about who got in line first.

If you want the near-term, falsifiable version of this thesis — a specific, dated claim about compute rationing becoming an openly disclosed, formalized feature of how large AI providers sell capacity — I have written it up as a standalone prediction on capacity-tiered compute allocation becoming a disclosed commercial norm. For the reported detail behind the deals and disputes that anchor this piece, see the companion news analysis of the Google-Meta Gemini cap and the compute-rationing regime.