An essay by Dario Amodei, published at darioamodei.com/post/on-deepseek-and-export-controls around January 28, 2025, shortly after the release of DeepSeek-R1. Written as a direct response to the claim that DeepSeek's releases undermine the case for US export controls on chips to China, the essay argues the opposite: that DeepSeek's performance falls on the normal cost-reduction curve, does not represent a breakthrough, and therefore leaves the strategic situation — who will have millions of chips first — unchanged, making export controls more important rather than less.
Three-dynamics framing
Amodei organizes the analysis around three dynamics he describes as foundational to how AI development proceeds:
- Scaling laws. More compute yields smoothly better AI. The gains at each level are valuable enough that efficiency improvements do not reduce overall spending; instead they are reinvested in bigger models, so companies "traverse the curve faster."
- Shifting the curve. Engineering innovations and new architectures shift the cost curve but do not change the underlying dynamic. As of 2025, algorithmic progress shifts the curve roughly 4× per year, and efficiency gains are immediately reinvested into more capable models.
- Shifting the paradigm. Occasional new training approaches, such as RL-based reasoning and chain-of-thought, add a new dimension. Amodei argues RL for reasoning was at a "crossover point" when DeepSeek released — temporarily cheap enough that multiple companies could produce comparable outputs.
DeepSeek analysis
Amodei identifies DeepSeek-V3, released about a month before R1's viral moment, as the real innovation, citing engineering improvements in KV cache management and mixture-of-experts scaling. He characterizes the result as modest relative to the public framing: "a fair statement is that DeepSeek produced a model close to the performance of US models 7–10 months older, for a good deal less cost — but not anywhere near the ratios people have suggested."
The essay rejects the claim that DeepSeek trained a frontier model for $6M versus the billions spent by US labs. Amodei estimates company-level spend at roughly 50,000 Hopper-generation chips and about $1B total, putting DeepSeek within 2–3× of major US labs. He characterizes DeepSeek-R1, the viral release, as adding RL reasoning on top of V3, replicating what OpenAI did with o1 and being less innovative than V3 itself.
On export controls, the essay argues they did not fail. DeepSeek's chip fleet, by Amodei's account, includes H800s shipped before the October 2023 ban, H100s that were likely smuggled, and H20s that were not yet banned. He frames the controls as working and the remaining loopholes as in the process of being closed.
Unipolar versus bipolar world argument
Amodei frames the strategic stakes as a choice between two outcomes. If China obtains millions of chips, the world becomes bipolar, with both the US and China holding powerful AI and China, in his account, likely directing more of it toward military applications. If China cannot obtain millions of chips, the US and its democratic allies retain at least a temporary unipolar lead which, because AI can help build better AI, could become durable. He concludes: "Well-enforced export controls are the only thing that can prevent China from getting millions of chips, and are therefore the most important determinant of whether we end up in a unipolar or bipolar world."
The essay presents the case as a geopolitical rather than technical one. It acknowledges DeepSeek's researchers as "smart, curious researchers who just want to make useful technology" while arguing against their government's access to powerful AI. It is Amodei's detailed policy argument for export controls, written in response to the most prominent challenge to that position, and provides the technical scaffolding — the three dynamics and curve-shifting analysis — for the broader export-controls debate. It explicitly rejects the argument that DeepSeek makes AI cheaper for everyone and therefore renders chips less relevant.
The essay pairs with Dario Amodei — The Urgency of Interpretability, which treats export controls as a "security buffer" giving interpretability time to develop before AGI. The CSIS — DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race (Allen, March 2025) analysis by Gregory Allen offers additional institutional corroboration.
Relationships
- supports: Export Controls (AI) — the central policy argument; Amodei's most detailed defense
- supports: Scaling Laws — the three-dynamics framing builds directly on scaling law concepts
- related: DeepSeek-V3 Technical Report, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — the releases this essay responds to
- related: (Source: Raw Sources/SemiAnalysis - Huawei Ascend Production Ramp.md) — chip supply analysis corroborating the "50,000 Hoppers" figure
- related: Compute Governance — export controls as compute governance
- related: Dario Amodei — author
- related: DeepSeek — the company analyzed