AI Export Controls Decide the Compute You Can Buy

Export controls get filed under policy news, which is why most founders skip them. That is a mistake, because they sit upstream of your inference bill. The rules decide which chips and which memory legally exist in which markets. That decides what the clouds and the inference providers can buy. That decides what you pay per token, how long you wait for reserved capacity, and how often your provider reprices.

Y Combinator's own request for startups on semiconductor supply chains compresses the operating reality into four words: "Export controls change quarterly." The same clip notes that "A single advanced AI chip goes through about 1400 process steps of manufacturing, crosses a dozen countries, and takes 5 months to build." A policy change at one point in that chain takes months to show up in your bill, and by then it is not negotiable.

The most useful founder-facing read I have heard on what the controls have actually done came from Aravind Srinivas, co-founder and CEO of Perplexity, in a long interview with Harry Stebbings on 20VC published in June 2026. Asked straight out whether the controls helped or hurt, he did not take either available side. "Jury's still out."

Here is what the rules say, what his argument is, where it needs a correction, and what to do about it.

What AI export controls actually restrict

The structure most people carry in their heads, that the United States banned selling advanced GPUs to China, is roughly half of it.

The current shape comes from the Bureau of Industry and Security interim final rule that took effect on 2 December 2024 and was published in the Federal Register three days later. It tightened three things at once: advanced computing integrated circuits, the equipment used to manufacture semiconductors, and high bandwidth memory. That third one is the part the coverage usually skips and the part Srinivas keeps coming back to. HBM sits under its own classification, ECCN 3A090.c, controlled by bandwidth density, which is package bandwidth in gigabytes per second divided by package area in square millimeters, above a threshold of 2 GB/s/mm2. There is a narrow license exception for shipments to packaging sites under tight conditions, which tells you how carefully the memory path was closed.

Two details matter for planning. First, the controls are not limited to goods made in America: a foreign direct product rule reaches some chips produced abroad using US technology or software. Second, the rules move. On 15 January 2026 the Bureau shifted its review policy for a band of less advanced parts, those under 21,000 total processing performance and under 6,500 GB/s of total DRAM bandwidth, a band that covers the Nvidia H200 and the AMD MI325X, from a presumption of denial to case by case review for direct exports from the United States. That came with conditions: certification that domestic supply is not being diverted, independent third party testing in the United States, and screening obligations on the buyer. Reexports and in country transfers stayed under presumption of denial. It is a loosened review standard, not an open door, and the point for a founder is simply that the line moved once and can move again.

The argument that the controls worked, and then backfired

Srinivas gives the short term case first, and it is the strongest version of the pro control argument. His claim is that export controls are the only reason a development gap exists at all between open source and the frontier. He also names who pushed for them: "companies like Anthropic lobbied very hard for it." That is a speaker's claim about a third party, so it is worth checking rather than repeating, and it checks out. Anthropic has published its position openly, arguing in its April 2025 response on the export control framework that preserving the American compute advantage is a national security requirement.

Then he turns it over. Restricted from the American stack, Chinese labs are not stuck. They are being pushed to build their own, and the thing they are being pushed into is the thing the United States is worst at.

"Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem."

His conclusion, in his own words: "by forcing them to go out there and build all this, you're converting them into a far more like potent competitor." Asked whether America still underestimates Chinese capability, he says it does, and specifically on the physical side, where you have to build fabs, robots, chips and energy rather than ship software. "I think they have a lot more advantages than America."

This is a thesis, not a finding, and it is being argued by a company that buys from, and increasingly wants to replace, the labs that lobbied for the rules. It is worth setting against someone scoring the same question from the policy side. Writing for the Council on Foreign Relations in April 2026, senior fellow Chris McGuire read DeepSeek's V4 release as evidence that the controls are working and the American lead is holding, noting that DeepSeek's own technical paper places V4 roughly three to six months behind state of the art frontier models, broadly in line with the Epoch AI estimate he cites of a seven month US lead.

The two readings are not actually in conflict, which is the useful part. One is scoring the gap as it stands. The other is predicting what happens to that gap once a restricted competitor has finished rebuilding its stack around the restriction. A founder needs both numbers, because they point at different risks: the gap is real today, and the thing that would close it is physical build out rather than model research.

A restricted stack becomes a different stack, not a slower one

The mechanism underneath the argument is the part worth internalizing, because it generalizes well beyond chip policy.

Srinivas's reading is that the memory controls, on top of the GPU controls, are what pushed the architecture somewhere different. Cut off from high bandwidth memory and from the most advanced storage, a lab cannot build the architecture everyone else is building. So it builds a different one: innovations in the KV cache small enough to sit on SSDs, an inference design that does not assume HBM, changes to the attention layer and to the training algorithm that reduce how much interconnect capacity a run consumes. "They're building with the Huawei stack," he says, and the whole thing, hardware, chips and fabs included, gets vertically integrated around that constraint. "That's a very different bet from what America's making."

One correction, because his phrasing goes further than the public record does. Srinivas describes this as a clean break from Nvidia. Reporting since suggests a split rather than a swap. DeepSeek did launch its V4 model on Huawei silicon, and later partnered with Huawei on chip programming tools to ease migration away from Nvidia's CUDA. Bloomberg was blunt about what did not move with it, reporting that the Huawei order still leaves Nvidia doing the training. Serving has shifted toward Ascend. The training side has not followed, at least not on the public record, and DeepSeek has not said otherwise. The direction supports his argument. The absolutism does not.

The founder lesson is the same either way, and it is one you can apply to your own stack this quarter. A constraint applied to a competitor does not slow them down proportionally. It pushes them onto a different curve, and sometimes the different curve is better. Srinivas puts the odds that a radically more efficient architecture arrives and strands a lot of expensive capacity at "it's probably like 20%, 30% chance." If you are building on the assumption that today's cost structure is the permanent one, that is the scenario that breaks you.

The constraint you will feel is power and memory, not GPUs

Ask Srinivas what actually gates the buildout and he does not say chips. "I think power will remain the bottleneck." He puts a number on how much of that is public resistance: "I think right now 40 out of 100 are not being developed because of public resistance." That figure is his own read rather than a published statistic, and I could not source it independently, so treat it as an estimate from someone close to the market and not as data.

The pricing rule he draws from this is the sentence to keep: "Whatever is the bottleneck will command the price."

That rule is why he expects memory suppliers to be repriced upward, and why he argues enterprise CPUs are quietly becoming scarce too, since agent loops and harnesses run on CPUs even when the tokens come off GPUs. It is also the cleanest explanation of your own bill. You are not paying for intelligence. You are paying for whichever physical input is currently scarcest in the chain that produced your tokens, and export controls are one of the things that decide which input that is. The same logic runs through data center power as the binding AI constraint and through why inference chips are built for the wrong job: the scarce layer moves, and whichever layer is scarce sets your price.

Asked what he would do with unlimited money, his answer is consistent: "I would build data centers." He frames the whole period as "physical infrastructure build-outs is like the return of the industrial age again." Y Combinator's supply chain clip is the same observation from the other end, calling TSMC advanced packaging "the single biggest bottleneck in AI compute right now, and Nvidia has locked up over 60% of it."

Where chip policy becomes a security question

This is the part that lands in your lap rather than a policymaker's, and it is where the export control story stops being macro.

The second order effect of the controls is a competitive, cheap, openly licensed model ecosystem coming out of a jurisdiction with different data rules. That is genuinely useful. It is also a decision with two separate halves that are routinely collapsed into one.

Running open weights on infrastructure you control is a software supply chain question. You are importing an artifact you cannot audit, trained on data you cannot inspect, and the right response is the same one you would apply to any third party dependency: your own evaluations, your own red teaming, and a named owner. Calling a hosted endpoint in another jurisdiction is a different question, because now the prompt leaves. The prompt is usually the sensitive part, not the response.

Srinivas hits the same distinction from the routing side when he explains why teams use a model router at all, and one reason is simply that "you don't want your API tokens going to China." If you serve regulated customers, that sentence is a contract term waiting to happen. The practical move is to decide which model classes are approved for which data classes, write it down, and make the routing layer enforce it. A policy that depends on each engineer remembering it is not a policy. If you have not built that seam yet, the AI capex picture and choosing an open source LLM are the two adjacent decisions it touches.

The architectural hedge Srinivas argues for, moving some inference to local or owned infrastructure so you are not metered on every token, happens to be the same hedge that gives you a jurisdiction story. That is rare and worth taking.

What to do this week

1. Write down which physical input prices your tokens. For your main workload, name the scarce layer: memory, power, packaging or CPU. If nobody on the team can name it, your cost forecast is a straight line drawn through someone else's supply chain.
2. Ask your provider the repricing question. Not what you pay today. What their contractual right to reprice is, with how much notice, and what happened to that number the last two times supply tightened.
3. Classify your models against your data. One table: model, where it runs, which jurisdiction, which data classes it is approved for. Hosted Chinese endpoints and self hosted open weights go on separate rows, because they are separate risks.
4. Enforce it in the router, not the wiki. If the routing layer cannot refuse a call that violates the table, you have documentation rather than a control.
5. Price the 20 to 30 percent scenario. Run your 24 month model on the assumption that a far cheaper architecture arrives and your provider's pricing power collapses, and separately on the assumption that memory stays scarce and prices rise. If only one of those two leaves you solvent, you have a concentration problem, not a forecast.
6. Put export controls in the risk register, once. One line, one owner, one review per quarter. The rules changed in December 2024, changed again across 2025 when a whole diffusion framework was issued and then rescinded, and changed again in January 2026. They will change again, and the cost of noticing late is paid in procurement, not in legal.

The course is the structured version of this: AI Operating System for Startups covers how to budget compute, instrument the spend, and build the seams that make a provider, model or jurisdiction swap cheap instead of existential. The wider strategic map these decisions sit inside is AI for startups.

Sources

• Aravind Srinivas on 20VC with Harry Stebbings, published June 2026, the interview this article distills. All quotes attributed to Srinivas are from this episode.
• Y Combinator, Supply Chain 2.0 for Semiconductors, the source of the chip manufacturing and advanced packaging figures.
• US Bureau of Industry and Security, the interim final rule effective 2 December 2024 and the High Bandwidth Memory license exception, for what is controlled and how HBM is classified.
• Revision to License Review Policy for Advanced Computing Commodities, 15 January 2026, and the accompanying Commerce Department announcement, for the case by case review change and its conditions.
• Congressional Research Service, US Export Controls and China: Advanced Semiconductors, for background on how the control regime developed.
• Anthropic's position on the export control framework, the company's own statement of the advocacy Srinivas refers to.
• Council on Foreign Relations, DeepSeek V4 Signals a New Phase in the U.S.-China AI Rivalry, April 2026, for the counter-reading that the controls are working, including the three to six month gap DeepSeek concedes in its own technical paper.
• On the Chinese stack: Tom's Hardware on DeepSeek launching V4 on Huawei chips, Reuters on the DeepSeek and Huawei chip tooling partnership, and Bloomberg on Nvidia still doing the training.