Nvidia: The Arms Dealer That Picked a Side

Jensen Huang does not want to be a geopolitical actor. He has said so repeatedly, in earnings calls, in conference keynotes, in interviews where the journalist asks about China and Huang gives an answer that is simultaneously candid about the business reality and conspicuously light on political interpretation. He is not being evasive. He is being a CEO of a company that would, in a rational world, sell to everyone who can pay.

The world Nvidia finds itself in is not rational by that standard. It is a world where the most valuable product Nvidia makes — the H100, the H200, the Blackwell B100, and whatever comes next — is simultaneously a commercial product and a strategic asset that the US government has decided belongs in the category of instruments of national power. Huang has spent the past four years navigating the gap between those two realities with the controlled frustration of someone who built a company by ignoring exactly this kind of constraint.

Nvidia’s pre-eminence in AI compute is not an accident of the current moment. It is the result of a twenty-year accumulation of exactly the kind of unglamorous technical investments that large companies tend to deprioritize: CUDA, the software platform that Nvidia launched in 2006 when GPU computing was a niche interest of video game developers; the interconnect technology (NVLink, then NVSwitch) that allows multiple GPUs to work together as if they were a single large processor; the memory architecture that feeds data to arithmetic cores fast enough to keep them busy. None of these was obvious at the time. Collectively, they constitute a moat that has proven extraordinarily difficult to cross.

AMD has tried. Intel tried with its Gaudi accelerators. Google built the TPU. Cerebras designed a processor the size of a dinner plate. Amazon built Trainium. Every major cloud provider now has its own AI chip program. None has seriously disrupted Nvidia’s position in the market that matters most — training frontier models, where Nvidia’s software ecosystem is as important as its hardware.

CUDA, specifically, is the thing that makes the moat real. Switching from Nvidia to a competitor is not a hardware decision. It is an organizational decision about how much engineering effort to invest in rewriting or porting the libraries, frameworks, and custom kernels that every AI system depends on. Most organizations, most of the time, conclude that the cost exceeds the benefit. This gives Nvidia pricing power that no normal market would sustain — H100s were selling for well above list price at $30,000 to $40,000 per unit during the 2023-2024 shortage — and it gives the company strategic importance to the AI industry that its revenue barely captures.

The export controls put Nvidia in a position that the arms dealer analogy captures imperfectly but usefully. Arms dealers have historically operated in the space between states, selling to multiple parties in conflicts while maintaining enough distance from the politics to preserve their commercial relationships. The implicit claim — “we sell weapons, not sides” — works as long as the weapons are genuinely fungible and the states involved are content to let the transactions proceed. When a superpower decides that the weapons are too important to sell to its adversary, the arms dealer’s neutrality becomes untenable.

Nvidia’s initial response to the October 2022 restrictions was to design neutered versions of its chips that were technically compliant with the export rules — the A800 and H800, which had reduced memory bandwidth and modified interconnect performance specifically calibrated to fall below the control thresholds. The US government’s response, in the subsequent rounds of export control tightening, was to close the loophole by setting the thresholds based on overall chip performance rather than specific technical parameters. It was an escalating game in which Nvidia’s product engineering and the Commerce Department’s rule-making were responding to each other with the dynamic that characterizes any adversarial technical specification process.

The company lost that game by mid-2024, when the rules became precise enough that no realistic H100 variant could be sold to Chinese customers under the restrictions. Huang’s public response was characteristic: he acknowledged the revenue loss, pointed out that the restrictions were benefiting Chinese competitors (Huawei specifically) by forcing Chinese companies to invest in domestic alternatives they would not have built otherwise, and committed to complying with the law while continuing to advocate for its revision. He was diplomatically honest about disagreeing with the policy while making clear that Nvidia would not operate outside the law.

The financial impact has been significant but not catastrophic for Nvidia, which is itself a geopolitical data point. The company’s revenue growth from 2022 to 2026 was extraordinary, driven by demand for AI training infrastructure from American hyperscalers who were spending enormous capital on data center build-outs. The Chinese revenue that was lost was substantial in absolute terms but modest as a share of total revenue in the years when US demand was growing fastest. Nvidia’s stock performance during this period was driven primarily by US and European data center demand, which meant that the export controls damaged a revenue stream that was, in relative terms, declining in importance anyway.

This created a perverse dynamic in corporate incentives. The restrictions that Huang publicly criticized hurt Nvidia less than they would have if the domestic US AI boom had not been so aggressive. If American cloud providers had been spending less on AI infrastructure, the Chinese revenue loss would have been more painful and Nvidia’s advocacy against the controls would have been more commercially motivated. Instead, the company was able to absorb the loss while maintaining hypergrowth, which made its public opposition to the controls less credible as a market signal and more like the principled-but-not-mobilized advocacy of a company that disagreed with a policy that was not, ultimately, an existential threat.

The Chinese AI chip ecosystem that developed in Nvidia’s absence is Huawei’s Ascend program plus a constellation of smaller chip designers — Cambricon, Biren, Enflame — who benefit from the forced import substitution environment that the export controls created. None of these are yet competitive with Nvidia’s full ecosystem. But they represent a market that did not exist in 2021 and has real commercial traction in 2027. Whether Nvidia could ever re-enter that market — if the political environment changed — is a question the company has to consider against the backdrop of an ecosystem that its customers have now invested years in building.

The deeper transformation that the export controls have forced is in how Nvidia understands its own strategic position. The company spent twenty years building an ecosystem that was deliberately agnostic about political alignment — CUDA works for physics simulations, medical imaging, cryptocurrency mining, game rendering, and AI training. The idea that any particular use case was off-limits, or that the company’s products were weapons in a geopolitical competition, was not in the original product philosophy.

That product philosophy is no longer available. The US government has decided that Nvidia’s products are strategic assets whose distribution is a matter of national policy. Nvidia is now, whether it wants to be or not, an instrument of American industrial strategy — constrained in who it can sell to, required to police its supply chain for potential re-export violations, and watched by both governments in the competition as a proxy for the effectiveness of export restrictions.

Jensen Huang’s preferred world — where a great product company sells its great products to anyone who will pay for them — is not available in the AI age. The technology is too important, and the competition for it is too serious, for any company at the frontier to maintain genuine neutrality. The arms dealer analogy is imperfect. But the underlying dynamic — that the most powerful weapons cannot be sold without choosing a side — is exactly right.