DeepSeek and the Compute Thesis: How a Chinese Lab Broke the Model

In January 2025, a Chinese AI lab called DeepSeek released R1. A frontier reasoning model competitive with OpenAI’s o1. It was open-weight. Anyone could download it. It hit the top of Apple’s App Store rankings within days. On January 27, Nvidia lost $589 billion in market capitalization in a single trading day. The largest single-day loss in stock market history.

The lab had done this on restricted hardware, for reported training cost of approximately $5.6 million, using a novel architecture that nobody outside the lab had built before.

The entire AI compute governance thesis took the most expensive afternoon in semiconductor history to absorb.

The Setup

The October 2022 US Bureau of Industry and Security semiconductor export controls were the most consequential AI governance action any government had taken to that point. Updated in 2023 and again in January 2025, the controls targeted the physical layer. The chips that make frontier AI possible.

The geography was stark. TSMC, a single Taiwanese company, fabricated approximately ninety percent of the world’s most advanced semiconductors. ASML, a single Dutch company, manufactured every extreme ultraviolet lithography machine on the planet. The tool required to print transistors at the scale modern chips demand. Two companies. Two countries. Two chokepoints that controlled the entire frontier of computational capability.

The export controls divided the world into tiers. Unrestricted allies got full access. A middle tier faced licensing requirements and volume caps. Embargoed nations, primarily China, got nothing above a specified performance threshold. The Great Firewall, applied to hardware.

The CHIPS Act reinforced the architecture with $52.7 billion in subsidies. The catch: recipients were prohibited from expanding semiconductor manufacturing capacity in China for ten years. The subsidy was a leash. Take the money, accept the geopolitical alignment.

The thesis was clear. Controlling who has the chips controls who can build frontier AI. Restrict the chips, restrict the AI. The thesis was the foundation of the entire emerging AI governance regime.

What DeepSeek Did

DeepSeek operated under the export controls. Its engineers trained R1 on H800 chips. The hobbled variant that Nvidia had designed specifically to comply with the US restrictions, slower than the H100s available to American labs. They used a novel architecture (Multi-head Latent Attention, MLA) that extracted more performance from less hardware. Memory bandwidth was the binding constraint on the H800. MLA reframed attention computation to ease the constraint.

They also used a mixture-of-experts design. The full model is 671 billion parameters, but only a small fraction of those parameters activates for any given query. The compute footprint per inference is far smaller than the parameter count would suggest. The training footprint, distributed across the hobbled hardware, came in at approximately $5.6 million according to DeepSeek’s own technical report.

Five point six million dollars.

OpenAI was reportedly spending hundreds of millions per model. Microsoft had committed $100 billion to data center construction. Anthropic had raised multiple rounds at valuations premised on the compute-as-moat thesis. The entire argument for compute-as-moat, that frontier AI requires so much computational power that only the richest companies and best-equipped nations can compete, had been challenged by a lab operating under the restrictions designed to prevent exactly this outcome.

The model was released open-weight. Anyone could download it, inspect it, fine-tune it, run it on their own hardware. The mobile app hit the top of Apple’s App Store rankings within days.

The market reaction was the largest single-day loss for a semiconductor company in history. Not because Nvidia made bad chips. Because a Chinese lab with restricted chips had demonstrated that the restriction didn’t work as intended.

What It Challenged

DeepSeek challenged three pillars of the compute-governance thesis simultaneously.

The export controls thesis: if restricted hardware produces frontier AI, the controls don’t prevent what they claim to prevent. The H800 was the chip Nvidia had designed specifically to comply with US restrictions. The chip was supposed to be deliberately slow enough that nobody could use it for frontier training. The frontier training happened anyway. The deliberate slowdown turned out to be a constraint, not a barrier. Constraints have always inspired better architecture.

The compute moat thesis: if $5.6 million trains a competitive model, the hundred-billion-dollar data center plans look less like infrastructure investment and more like capital misallocation. The thesis underpinning the AI capex super-cycle was that the frontier capability gap widens with compute, so the labs with the most compute win. DeepSeek demonstrated that the gap is closeable, at least intermittently, by algorithmic innovation rather than by raw compute. The investors who had priced AI labs at multi-hundred-billion valuations on the moat thesis re-priced quickly.

The safety argument for restricting model access: DeepSeek open-sourced a frontier reasoning model, and the world did not end. The capabilities that the safety community had argued required careful release-management were now distributed to anyone with a GPU and an internet connection. Whatever the consequences turned out to be, they would be the empirical answer to a question that had previously been argued in the abstract.

The compute governance thesis, that controlling hardware controls AI development, was not disproven. It was complicated. The chokepoints still exist. TSMC and ASML still have no competitors. But the thesis that restricting chip access would maintain a durable capability gap between American and Chinese AI took a $589 billion hit in a single afternoon.

The Novel Architecture

Multi-head Latent Attention is the innovation worth understanding because it explains why the result is harder to dismiss than the safety-community first response wanted it to be.

Standard transformer attention requires the model to store key and value vectors for every prior token in the context window during inference. The memory bandwidth required to read and write that key-value cache is the dominant inference cost on long contexts. On a memory-bandwidth-constrained chip, the H800, this is the binding limit.

MLA compresses the key-value cache by projecting it through a low-rank latent space. The compression is lossy in theory and turns out to be substantially lossless in practice for reasoning workloads. The model fits in the available memory bandwidth. The model trains and infers at a footprint the export-controlled hardware can support. The innovation is algorithmic. It works on any GPU. It works particularly well on the GPU the export controls were supposed to bottleneck.

The mixture-of-experts design layered on top of MLA gives the activation sparsity that lets the 671-billion-parameter model run with the compute cost of a much smaller dense model. The combination is what produced the price point.

Neither technique was invented at DeepSeek. Mixture-of-experts has been in the literature since the 1990s. Low-rank attention compressions have been explored for years. The DeepSeek contribution was assembling them into a working frontier system, under hardware constraints that forced the assembly. The constraint produced the innovation. This is the boring historical pattern of every engineering field. It is also exactly the pattern the export controls were supposed to prevent from happening on the other side of the chip embargo.

The Broader Context

DeepSeek was the dramatic moment. The broader context made the dramatic moment legible.

Mistral, the French open-source AI lab, has released competitive open-weight models throughout 2023 and 2024. The European open-source ecosystem is real. Meta’s Llama series is open-weight. The frontier-capability-distribution question was not entirely about China. China made the question impossible to dismiss because the chip embargo was supposed to specifically prevent the Chinese case.

California SB 1047, which would have imposed liability on developers of open-source frontier AI, was vetoed by Governor Newsom in September 2024. The veto was the last meaningful regulatory action before DeepSeek dropped. Had SB 1047 passed, the regulatory regime in the United States would have been incompatible with the kind of release DeepSeek then performed from a jurisdiction that did not have such a regime.

The EU AI Act’s open-source exemption is instructive. Small models get the exemption. Frontier models, defined as those trained with more than 10^25 floating-point operations, do not. The threshold is the moat. Below it, you can open-source freely. Above it, you need the compliance apparatus of a major corporation. The regulation makes it economically irrational to open-source frontier AI into the European market. DeepSeek’s open-weight release does not face this constraint in its home jurisdiction.

The Bletchley Declaration of November 2023, signed by twenty-eight countries and the EU including China, established the consensus that frontier AI poses catastrophic risk and that international coordination is required. Bletchley imposed no binding commitments. It established the problem definition. DeepSeek demonstrated that the problem definition does not match the operational landscape.

The Ratchet Implication

The hardware control layer was supposed to be the most durable layer of the AI governance ratchet. Physical, trackable, concentrated supply chain, two chokepoint companies in two friendly jurisdictions. Other governance layers are slower or softer. The hardware layer was supposed to be the hard constraint.

DeepSeek proved that algorithmic innovation can route around hardware restrictions. The chokepoints still exist. The compute floor still matters. But the relationship between chip access and frontier capability is now empirically discontinuous in at least one direction. The fastest path from restricted hardware to frontier capability turned out to be algorithmic. The fastest path from algorithmic capability to global distribution turned out to be open-weight release. Neither path is controlled by the hardware governance regime.

John Gilmore said the famous version of this for the first generation of internet censorship: “The net interprets censorship as damage and routes around it.” The Chinese chip embargo is not the first time someone has discovered that the actually-restrictive thing was not as restrictive as the people imposing the restriction believed. It will not be the last.

Lennart Heim and colleagues at the Centre for the Governance of AI published a proposal in February 2024 describing the endgame of hardware governance: on-chip verification. Chips that cryptographically attest to the workloads they are running. Compute usage caps burned into firmware. Tamper-resistant usage logs. Remote attestation by regulators verifying chips are being used as authorized. Critics called it “a kill switch in every GPU.” The proposal is academically serious. It has not been implemented. It addresses the case where the chip is the binding constraint. DeepSeek is the data point that the chip is not always the binding constraint.

The hardware governance apparatus is being built anyway. The export controls were expanded in January 2025. The CHIPS Act incentives are being disbursed. The TSMC and ASML chokepoints are being reinforced. The on-chip-governance proposal is in the policy literature where the next generation of regulation begins. The infrastructure persists because the infrastructure serves purposes beyond the stated thesis. The compute-moat justification took a $589 billion hit. The compute-moat infrastructure did not get unwound.

The compute governance thesis was that controlling the chips controls the AI. DeepSeek did not disprove the thesis. It complicated the thesis on the most expensive afternoon in semiconductor history, by demonstrating that the relationship between chips and capability is not as smooth as the thesis required.

The chokepoints are still real. The export controls still bite. The frontier still requires compute. And a lab in China, working under the embargo, released a model that the embargo was specifically designed to prevent. On the chip the embargo had specifically engineered to be too slow for the purpose. The chip was too slow. The lab was not.

The receipts (free, on this site): the AI-governance tracker · nation-state ratchets

This research appears in The Ratchet, Chapter 20.