The Allocation Ratchet

Methodology: AI-assisted evidence infrastructure · Human-directed thesis · Primary-source verified

In February, Ford's chief financial officer told investors to expect roughly a billion dollars of extra cost this year, driven partly by the price of memory chips.

That is an unusual statement from an automaker. Ford has bought memory chips without drama for thirty years. They are a rounding error in a new vehicle. This year they are a nine-figure line item. Ford attributes it to memory pricing and input inflation together and has not split the two. But memory is the variable that changed, because data centers now outbid Ford for the same silicon wafers.

Nobody planned this. There was no policy, no embargo, no cartel. The AI buildout became hungry enough, fast enough, that its appetite began showing up on adjacent industries' income statements. What happens when the machine everyone is building starts taxing the people who aren't building it?

The answer takes us well beyond 2026. What matters isn't the shortage. It's what the shortage causes us to build, and who pays for it long after the shortage is over. This is the next layer of the Grid-Silicon Order, the regime in which the electrical grid and compute have fused into a single binding system. The new layer: its costs now land on adjacent industries.

When a scarce input starts hurting stakeholders with political weight, governments acquire the power to decide who gets access. When the scarcity passes, government holds onto that power. The shortage is the event; the power to allocate is what's left behind.

Between June 3rd and July 14th, five things happened that were unremarkable alone and hard to ignore together.

Three letters landed on desks in Washington, all concerning memory chips. On June 3rd, nine industry associations (telecom, automakers, medical devices, retail among them) wrote to Treasury and Commerce asking for help with the shortage. They asked for nothing specific; the point was that a coalition of the taxed now exists.

On July 1st, the chip industry's own trade group wrote back to the same officials: whatever you do, don't allocate. In other words: no price controls, no capacity mandates. They asked for tax incentives instead.

On July 14th, the House Select Committee on China asked Commerce to ban U.S. purchases of Chinese memory chips for AI, federal, and critical-infrastructure use. This ask is different in kind: it doesn't ask Washington to build anything new, only to point an existing weapon at a new target. It covers the strategic tier (AI, federal, critical infrastructure), not Ford's supply chain. But strategic-tier instruments have a history of widening, and that history is this letter's subject.

In the same six weeks, two states acted on power. New Jersey created a separate billing class for very large electricity users, so data-center costs land on data centers instead of everyone's bill. New York's governor ordered a one-year pause on new large data centers. Letter #9, The Grid Bill Comes Due, traced the cost-assignment doctrine both instruments extend.

Three letters, two state actions, six weeks. Every one is a group squeezed by the AI buildout asking the government to do something about it.

The AI buildout is consuming three inputs it does not manufacture: chips, equipment, and people. Each is scarce for a different reason. All three produce the same political result.

Three companies (Samsung, SK Hynix, and Micron) make more than ninety-five percent of the world's qualified memory. AI accelerators need a specialized, stacked version of that memory that earns far more revenue per wafer than the ordinary kind. So all three are doing the obvious thing at once: shifting production toward the profitable product. There is a fourth supplier at scale, China's CXMT, and the July 14th letter is a proposal to disqualify it from the strategic tier.

Two things make this worse than a swap. The AI-grade product is stacked, which means it consumes more wafer capacity per usable bit and yields worse, so a modest shift of wafers takes a disproportionate bite from ordinary supply. And memory has done this before. Suppliers swung capacity toward the premium tier in 2017 and again in 2021; it is well worn behavior in the industry. What is different this time is the response, not the reallocation. In 2017 nobody wrote to Commerce about it.

The shortage travels sideways, out of the AI sector entirely, into cars and appliances and factory equipment that run on older, cheaper memory nobody makes enough of anymore. Micron has said its entire 2026 supply of the AI-grade product is already sold, and can fill only half to two-thirds of demand for everything else.

As discussed in The Equipment Wall, the delivery of a large power transformer now runs from roughly eighteen months to nearly three years depending on voltage class and backlog, against a pre-buildout norm of under a year. The queue was already clogged; solar, wind, and storage interconnection has been stretching the same supply chain for years, which is why the buildout's marginal demand lands so hard. The specialty steel every transformer core requires has exactly one American producer, a single mill behind every transformer in the country.

The labor squeeze is the least well-measured of the three, and its numbers come mostly from the trades themselves. The IBEW puts the incremental electrician gap above 300,000, against a construction workforce already short by nearly half a million; estimates of electrical work as a share of data-center construction cost run from forty-five to seventy percent. Wages are running up to thirty percent above trend. Which means every school retrofit, hospital expansion, and grid-hardening project in the country is now bidding against a hyperscaler's general contractor for the same crews.

A margin decision, a materials bottleneck, and a labor squeeze — structurally unrelated, each squeezing a politically organized constituency. Automakers have lobbyists. Ratepayers have public utility commissions. Electricians have a union.

In the AI infrastructure buildout, compute was the first shortage. Power was the second. Memory is the third, and moving faster than both shortages before it.

Advanced chips went from a predictable market to a managed input in about four years: export licenses, revenue-sharing arrangements, government-brokered access deals. Electricity for data centers went from a utility service to a political allocation: the ratepayer classes and moratoria states are currently writing.

Memory is running that same arc: scarcity, spillover into unrelated sectors, a coalition of the squeezed, a counter-coalition of suppliers, a letter from Congress. That three-sided fight (one side asking the state to act, one asking it to abstain, one redirecting the whole question to China) is what a market looks like just before the government steps in as its referee.

The machinery is general-purpose, and none of it is new. New Jersey's billing class and New York's pause authority were written with data centers in mind, but nothing about either instrument is specific to data centers. Once a state can separate out a class of large customers and assign them their own costs, the precedent waits for whatever fight arrives next. Washington is no different: the China letter asks Commerce to extend a restriction it has already imposed, and the industry coalition asked for a review of a law that already exists. This is how allocation power accumulates: existing agencies quietly acquiring one more commodity.

The chip industry's response is that markets correct shortages, and on the narrow question of price, the numbers currently support the premise. Contract-price increases are already decelerating sharply with more than a trillion dollars of new memory capacity announced or committed. Credible forecasters expect oversupply by 2028 or 2029, a predictable overcorrection.

They may be right about prices. Prices were never the whole exposure.

One thread this letter leaves open: who finances that trillion dollars (supplier balance sheets, hyperscaler prepayments, securitized paper) determines how the overcorrection lands, and on whom. Announced capacity is not shipped capacity, a memory fab runs two to three years from final investment decision to first wafer, and the price-deceleration case assumes that capital converts on schedule. A tightening in the financing stack would slow the correction and widen the political clock's head start.

Industrial economies run a well-worn three-act response to strategic supply shortages.

Act one: the shortage arrives, prices spike, and the losers can name what it cost them.

Act two: the losers organize, and the government builds machinery: priority ratings, set-asides, cost assignments, export restrictions.

Act three, the act everyone forgets: the shortage resolves, and the machinery sticks around.

The Defense Production Act was written for the Korean War's materials crunch. The war ended in 1953. The priority system it created is what Washington reached for to allocate COVID vaccines in 2021 and baby formula in 2022, still the first tool off the shelf. The 1973 oil shock produced the Strategic Petroleum Reserve and a ban on exporting American crude. The embargo lasted five months. The export ban lasted forty years. The reserve is still there and has been doing work with the prolonged closure of the Strait of Hormuz.

Memory matches on the diagnostics: a strategic input, a visible price shock, organized losers, and a government that builds capacity to allocate rather than waiting out the cycle.

Timing cuts both ways. Memory supply responds in years, not decades: a shorter window for machinery-building than oil offered. But the United States no longer manufactures ordinary memory at scale, which takes the classic resolution (building your way out of the dependency) off the table. A country that cannot build its way out regulates instead.

And then the difference that decides it. In 1973 the Strategic Petroleum Reserve and the export ban had to be stood up from nothing. In 2026 the machinery is already on the shelf: an export-control apparatus assembled for AI chips, field-tested, waiting for a new commodity. When the machinery doesn't have to be built, expect it to get used.

By the end of 2027, I put roughly 60 percent odds on Washington taking at least one formal allocation-side action on memory that reaches commercial buyers, not just the strategic tier.

Two instruments, not one, and both sit inside Commerce. A procurement restriction governs only what the government itself buys, and it needs no vote. An export-control licensing rule governs what anyone can ship, and it needs no vote either. The first is faster and narrower. The second is slower, and it is the one that would reach Ford. New institutions need political capital to stand up. Old institutions need only a new reason to come off the shelf.

A second-order consequence carries through the decade: allocation machinery pointed at memory redraws the map for the three firms that make it and the allies that host them. If U.S. rules begin to decide who gets advanced memory first, the hosts must choose between subordinate access and stacks of their own. Seoul has already signaled its answer with a $520 billion national memory plan, even as SK Hynix and Samsung run major fabs inside China, within reach of either side's rules. Whether that plan is a hedge against American rules or an independent bet, Seoul has not said. Tokyo's answer is the one still outstanding. Beijing's answer runs through standing instruments: a domestic memory push and the export restrictions it has already imposed on critical minerals. The apparatus built to manage the shortage becomes the instrument that fragments the market it manages.

To be clear about what I am not forecasting: I do not expect memory to stay scarce. It will likely be oversupplied by 2028. I am forecasting that Washington acquires the power to allocate it first — and keeps that power after the shortage is gone.

This is a race, and worth saying which side I think wins. The supply response is fast: the overbuild is announced and prices are already decelerating. But the political clock has a head start: the machinery doesn't need to be built, only pointed at its next target. A licensing widening moves at rulemaking speed where a fab moves at construction speed; the slower political instrument still lands before the supply side does. That asymmetry is why I expect Washington to move first.

Three ways this forecast fails.

For forty years the default assumption about any input a business buys — chips, steel, power, labor — was that price was the only variable that mattered. That assumption is being quietly withdrawn, one distortion at a time, and 2026 is the year the withdrawal became visible outside tech. You do not have to be in AI to feel the effects downstream. You only have to be in the business of buying something AI also buys.

More consequentially, U.S. economic interests are increasingly entangled with security interests. The past several months supplied the evidence: the Pentagon negotiating terms directly with frontier labs, strategic relationships being rewired around AI sovereignty, export and import action to control the AI supply chain, and national control over who has access to the world’s most powerful models. That is where this is heading.

For executives and boards, this changes the planning question. The 2027 budget question isn't only what will memory cost. It's who could decide what we're allowed to buy, and through which instrument. A price problem is a hedging and contracting problem. An allocation problem is a compliance and relationship problem, and most companies that aren't defense contractors have little muscle for it. If your products contain memory chips, your facilities draw serious power, or your capital projects need electricians, you now have a policy exposure that wasn't on your last risk register.

The useful question isn't how to hedge the next quarter; it's which capabilities appreciate over the decade this regime runs. I see three, each with a first move today.

Policy exposure needs an owner. Defense contractors have entire departments for that question. The first move costs nothing: put allocation exposure on the risk register, by name, and assign it: to the trade-compliance function if you have one, to enterprise risk or the CFO's office if you don't. What gets owned gets watched.

Supply optionality becomes a standing posture, not a procurement project. The exposure runs across every input that shares a buyer with the AI buildout. The first move is a map of what you can see: which of your inputs (chips, power, transformers, trades labor) sit in that overlap, and which tier-one suppliers are already allocating. Visibility past tier one is a multi-quarter build, not a first move.

Contracts need to survive allocation, not just price. Hedges and escalators handle a price shock. They say nothing about a priority order that moves you to the back of the line. The first move is an audit of the force-majeure and allocation clauses in your supply agreements. These decide who gets served when a supplier can only fill two-thirds of demand.

None of these pay off this quarter. All three compound over the ten years the machinery sticks around.

The stake here isn't whether the memory shortage ends. It will. It's that the industrial economy on the other side of it will have acquired allocation institutions it didn't have going in. And institutions, unlike shortages, tend to stick around. That's the ratchet. It only turns one way.

About Aroko: Aroko provides strategic advisory and capital allocation intelligence at the intersection of the energy transition, AI infrastructure, and geopolitical risk. Our analytical process combines proprietary evidence infrastructure with human-directed thesis formation. Every keystone claim is verified against primary sources, and all editorial judgment and capital allocation framing is conducted by Aroko’s team. The Letter is published biweekly.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment, legal, or tax advice. The opinions expressed regarding macro trends and infrastructure investments are solely those of the authors. Past performance does not guarantee future results. Readers should consult with a qualified financial professional before making any investment decisions.