Compute is necessary, not sufficient.

A research note on the DeepSeek precedent and the limits of FLOPs-supply restriction as AI strategy. The export-control regime was not wrong. It was wrongly scaled.

Prologue. The morning the export-control thesis broke.

On 26 December 2024 a Chinese laboratory called DeepSeek released the technical report for V3, a Mixture-of-Experts language model with 671 billion parameters total and 37 billion active per token. The reported headline was the cost of the final pre-training run: approximately 2.78 million H800 GPU-hours, which the lab translated into a dollar figure of around 5.6 million USD at internal rental rates. The model achieved performance broadly competitive with the contemporaneous frontier of Western proprietary systems on standard benchmarks. Five weeks later, on 20 January 2025, the lab released R1, a reasoning model derived from the V3 base, with comparable headline numbers and chain-of-thought capability that closed the remaining gap on hard reasoning benchmarks.¹²

Several Western analysts subsequently argued that the headline cost figure understated the true cost of the programme by roughly an order of magnitude once the cluster amortisation, the failed runs preceding the successful one, the salaries and the prior research investment were factored in. The argument is correct on its own terms and beside the point on the architectural one. SemiAnalysis's 2025 cost decomposition put the all-in cost in the high tens of millions, not the high single digits. That number is still an order of magnitude below the reported all-in cost of comparable Western frontier programmes.⁴

The export-control thesis assumed that frontier-model capability was a function of compute supply, and that constraining compute supply to the People's Republic of China would constrain the frontier in PRC laboratories by a corresponding factor. The thesis was not silly. It was tractable and measurable, with the additional virtue of being politically saleable in a Washington that has rarely passed up that combination. It was also, as the DeepSeek release demonstrated, structurally incomplete.

The thesis.

Compute is necessary. Compute is not sufficient. The cost of training a frontier model is the product of the FLOPs required by the algorithm and the cost per FLOP of the available hardware. Restricting the second factor while the first factor is improving by 30 to 50 percent per year, as Epoch AI's 2025 trends data documents, is restricting an asset whose value is depreciating relative to its complement.³

The DeepSeek precedent is not the assertion that compute restriction is pointless. It is the assertion that the restriction's binding magnitude is smaller than its political magnitude, and that the restricted party will buy time and engineering ingenuity in lieu of the restricted hardware. Whether the time and ingenuity prove sufficient depends on the curve of algorithmic efficiency improvement and on the depth of the export control. As of April 2026, both vectors favour the restricted party more than the restricting policy assumed in 2022.

The hardware ladder, in concrete.

The current US export control regime designates GPU classes by interconnect bandwidth, peak compute, and total chip-level performance metrics. The PRC-allowed lineage has migrated through several variants of NVIDIA's Hopper and Blackwell architectures with progressively tighter bandwidth restrictions.

The H800 was the workhorse on which DeepSeek V3 was reportedly trained. Compared to an H100, the H800 has the same peak FP16 throughput and a deliberately reduced NVLink bandwidth that targets the all-reduce step of distributed training. The reduction is real; the engineering response, as documented in DeepSeek's V3 paper, is to schedule communication around the bandwidth bottleneck and to use FP8 mixed precision to reduce the volume that must move. The bottleneck was a constraint, not a stop. The ingenuity premium for working with the bottleneck is real and bounded.¹

Performance against cost.

The cost-versus-capability frontier as of early 2026 places DeepSeek V3 and its derivatives substantially below the cost line of Western frontier proprietary models, at performance levels close to (though not at) the absolute frontier. The shape of the chart matters more than the position of any single point. The Pareto frontier moved.

The dashed line traces the empirical cost-performance frontier as reported. The DeepSeek points sit on or just inside that frontier at the low end, which is exactly the position the export-control thesis predicted no PRC laboratory could occupy. The Western proprietary points cluster at the top right at much higher reported cost, having paid for absolute frontier performance at a premium that, in the post-DeepSeek market, is going to be difficult to justify on inference economics alone.

What the export-control regime gets right, and where it stops.

The regime gets the direction right. The PRC laboratory ecosystem is materially less productive at the absolute frontier than it would be with unrestricted access to Blackwell-class hardware. Training a 1-trillion-parameter dense model is harder on H800 clusters than on H100 clusters by a meaningful factor. The frontier of what is theoretically feasible inside the PRC has been pushed back, and the gap between PRC and Western frontiers has not closed at the highest tier. Both points are true and both are politically important.

Where the regime stops is at the substitutability of compute for algorithm. The DeepSeek V3 paper documents a series of architectural and training-recipe choices (auxiliary-loss-free MoE balancing, multi-token prediction, FP8 mixed-precision, communication-aware all-reduce scheduling) that collectively reduce the FLOPs needed for a given level of capability by a factor that, on conservative reading, is in the 3 to 5 range. That factor is a multiplicative discount on the binding compute constraint. Regulators do not have an instrument that addresses this layer.¹

The libertarian read.

The libertarian reading is that physics, distributed knowledge production and incentive gradients are not regulated artefacts, and the policy that pretends otherwise will be circumvented by them at low cost. The American export-control regime correctly identified compute as the strategic chokepoint of mid-2020s frontier AI. It correctly applied policy instruments to that chokepoint. It under-estimated the rate at which the chokepoint would be partially routed-around by algorithmic efficiency, and it over-estimated the political authority of unilateral US export law over the rest of the world.

This is not a recommendation against export control. The policy will, on net, slow the PRC frontier and the slowing has strategic value. It is a recommendation against treating export control as if it were a stop-gap. The chokepoint is partially routable; the politics that produced the routing have run faster than the politics that produced the original control. Subsequent rounds of policy will need to address algorithmic-efficiency leakage as a category, which is materially harder, both technically and diplomatically.

What it means for builders.

If you are building or buying frontier-class AI systems, the operational lessons of the DeepSeek precedent are the following.

Frontier model price is collapsing toward inference-economics floor. A $100M reported training cost, amortised across the inference base of a Western frontier-API customer, converges on the per-token economics of an open-weights model whose training was reported at $5M to $20M. The pricing power of a closed-weights frontier model erodes as the gap closes; the closing is happening regardless of geography.

Open-weights frontier-class is a credible procurement target. An open-weights model that lands within 2 to 3 percentage points of the absolute frontier on relevant benchmarks, at a fraction of the inference cost, is a procurement decision a competent CTO has to take seriously. The political optics around using PRC-origin weights are non-trivial. The economic optics around not using them are also non-trivial.

Algorithmic efficiency outranks raw FLOPs. Building or licensing a model whose training was efficient is more strategically valuable than building or licensing a model whose training was expensive. The expensive one is less likely to be cost-competitive at inference and is more likely to be obsolete sooner.

The hardware tier that actually matters is the one you can keep paid for. The H100 cluster you cannot afford to leave running 24x7 has worse production economics than the H800 cluster you can. Capability per running dollar matters more than peak capability per nominal dollar.

Closer.

The DeepSeek precedent will be remembered as the moment the AI policy debate stopped being about FLOPs and started being about FLOPs-times-multiplier. The multiplier is algorithm. Algorithm is harder to constrain than silicon. The Chronicle's view is that this has been obvious to a small subset of the AI policy community since 2022, and that the broader community is now catching up under the pressure of an empirical example. The pressure is healthy. The catching-up is overdue.

Tobias Hager

Founder · CEO · Sole engineer · mi media intelligence SL

Tobias designs and operates every system inside the house. He builds AI-native products and the offensive/defensive cyber capability stack the company runs alongside its research thread toward agentic generative reasoning. He writes the Chronicle from inside the operational stack rather than from outside of it.

tobias-hager.com LinkedIn contact@mi.network

mi media intelligence SL · Palma de Mallorca · April 2026 · SIGIL CHR-0005