DeepSeek was inevitable: what lawyers and policymakers don't get about export controls

Constraints breed creativity and lawyers are not systems thinkers

Introduction

For decades, governments and regulators have tried to control technological advancement by restricting access to high-end hardware. The logic is simple but flawed: if cutting-edge chips are kept out of the hands of certain entities, those entities won’t be able to build competitive technology systems. But this mindset fundamentally misunderstands software and systems thinking.

DeepSeek, the advanced AI model developed by Chinese researchers, exemplifies why this approach is doomed. Even without access to NVIDIA’s top-tier GPUs, developers in restricted regions make up for it with sheer ingenuity. The secret? Software-driven optimization.

The Regulatory Assumption: Hardware as a Bottleneck

Governments, especially in the US, view hardware as the ultimate gatekeeper of AI capabilities. Export controls on NVIDIA’s A100 and H100 GPUs were imposed specifically to slow down China’s AI progres. The assumption was that training and running state-of-the-art AI models require massive computational resources that only the best chips can provide.

However, this view is myopic for several reasons.

1. AI Progress is More About Algorithmic Efficiency than Raw Compute. While high-end GPUs accelerate training, software optimizations can reduce reliance on them.

2. Distributed Computing Sidesteps Single-Chip Bottlenecks. Large clusters of weaker hardware can approximate the performance of a few high-end chips.

3. Optimization Techniques Close the Performance Gap. Advances like model quantization, weight pruning, and retrieval-augmented generation (RAG) enable efficiency gains that make even model hardware capable of high-level AI tasks.

Software Engineers Always Find a Way

The defining characteristic of software development is its adaptability. Unlike hardware, which has physical limitations, software can be improved indefinitely through smarter design. John Carmack, the famous video game programmer who optimized graphics processing in the early ‘90s, has this to say about the importance of software optimizations:

In the 90s, there were a dozen companies making graphics accelerators, and Nvidia wasn’t initially a clear winner. Their first product was terrible, and 3DFX, 3DLabs, Rendition, and others all had important pieces of the puzzle earlier. However, they relentlessly improved and avoided missteps until they were clearly at or vying for the top spot on hardware capabilities.

The real differentiator was taking software so much more seriously than competitors, which allowed them to weather periods when AMD slipped ahead in raw hardware performance, and had them building the CUDA ecosystem that underlies so much of modern AI work.

Software is, to put a not too fine point on it, both important and extremely powerful. Its optimization frequently obviates hardware limitations. DeepSeek’s existence demonstrates this principle. Even without unrestricted access to NVIDIA’s most powerful chips, Chinese researchers managed to develop a competitive LLM. How? By leveraging:

• More Efficient Model Architectures. Instead of relying on brute-force parameter scaling, they optimize transformer layers, use Mixture of Experts (MoE), and tweak activation functions.

• Better Training Paradigms. Techniques like reinforcement learning from human feedback (RLHF) and transfer learning improve output quality while reducing compute requirements.

• Compression and Quantization. Reducing model precision from FP32 to FP16 or even INT8 cuts memory requirements dramatically with minimal accuracy loss.

• Distributed and Parallel Computing. Instead of needing a single super-powerful GPU, developers break tasks into smaller pieces and distribute them across many weaker processors.

These methods mean that a country restricted from cutting-edge hardware can still build world-class AI models simply by being smarter about how they use their available resources.

Why Legal Minds Don’t Get It

The failure of regulatory bodies to grasp this inevitability comes down to a lack of systems thinking. Lawyers and policymakers tend to see the world in terms of static rules and linear cause-and-effect relationships. But technology operates in feedback loops, where constraints spur innovation.

A classic historical example of this phenomenon was during the Cold War, when the Soviet Union, lacking Western semiconductor technology, became extremely advanced in software-based optimizations. (They also invented the classic video game Tetris.) Their programmers learned to squeeze performance out of primitive hardware in ways that Western engineers never had to consider. The same dynamic is playing out today in AI.

Export restrictions on hardware will never stop innovation. They only change its direction. If you make the straight road impassable, the builders will carve a path through the mountains.

The Future: Where This Is Heading

The real battle in AI won’t be fought over who has the best chips. It will be about who has the best minds solving the efficiency problem. Companies and nations that invest in software-first AI approaches will future-proof themselves against hardware constraints. The next wave of breakthroughs will likely come from:

• Neural architectures that are fundamentally more efficient than transformers.

• AI training techniques that require exponentially less compute.

• Hybrid models that integrate symbolic reasoning with deep learning to reduce brute-force computation needs.

At this point, the only way export controls would meaningfully slow down AI progress is if they restricted not just hardware but also the free exchange of ideas. But even that would be a losing battle. Knowledge always finds a way to flow.

Final Thought

The legalistic approach to technology control is reactive and static, while innovation is proactive and dynamic. The more restrictions imposed, the more workarounds are developed. DeepSeek is just one example of this process in action, and it won’t be the last. The real story here isn’t just about Ai. It’s about how every imposed constraint becomes the seed of the next breakthrough.

---

Addendum: After I published the post above, I found a recent post by Dario Amodei, in which he argues:

There is an ongoing trend where companies spend more and more on training powerful AI models, even as the curve is periodically shifted and the cost of training a given level of model intelligence declines rapidly. It's just that the economic value of training more and more intelligent models is so great that any cost gains are more than eaten up almost immediately — they're poured back into making even smarter models for the same huge cost we were originally planning to spend. To the extent that US labs haven't already discovered them, the efficiency innovations DeepSeek developed will soon be applied by both US and Chinese labs to train multi-billion dollar models. These will perform better than the multi-billion models they were previously planning to train — but they'll still spend multi-billions. That number will continue going up, until we reach AI that is smarter than almost all humans at almost all things.

Making AI that is smarter than almost all humans at almost all things will require millions of chips, tens of billions of dollars (at least), and is most likely to happen in 2026-2027. DeepSeek's releases don't change this, because they're roughly on the expected cost reduction curve that has always been factored into these calculations.

This means that in 2026-2027 we could end up in one of two starkly different worlds. In the US, multiple companies will definitely have the required millions of chips (at the cost of tens of billions of dollars). The question is whether China will also be able to get millions of chips⁹.

These are fair points, and Dario Amodei is certainly much more informed about AI than I am. That said, I will reiterate the thrust of my argument: software lends itself to optimizations in a way that hardware does not. While Dario is likely correct that he who controls millions of state of the art GPUs will rule the world, it is still true that software optimizations can do a hell of a lot for China’s AI initiatives.