AI ARMS RACE 02: Let’s Bring Back “Targeted and Precise”—And Mean It
In April of this year, I published what may be seen as the AI Imperative 2030’s manifesto: “The U.S. ‘winning’ the global technology race requires gaining dominance over a range of strategic investments that pay dividends in both conflict and peacetime and then using those investments for public good at home and to promote U.S. influence abroad.”
To accomplish these goals, I outlined four objectives for U.S. artificial intelligence policy: (1) maximize national return on investment, (2) align the ensuing benefits with the public interest, (3) mitigate the risk of exploitation by adversaries, and (4) further foreign policy goals.
This is also Beijing’s tech policy framework—and that of the European Union, and India, and North Korea, and any other nation aiming to benefit from AI. The perceived relative importance of each objective differs by country, as do the tactics used to achieve them. However, the underlying blueprint is the same: to pursue technological advancement on a national level and then allocate the spoils on national priorities. This is why the race for AI is a race and not a war, dance, stage play, or some other metaphor. Everyone’s competing within their own lanes, but they’re barreling towards the same destination.
To win, the U.S. needs to stay in the lane that allowed it to lead for nearly a century; namely, promoting maximum possible free markets and a thriving innovation ecosystem, only applying targeted and precise guardrails to avert clear threats to national security. Assuming no such threats are present, the U.S. AI industry must be empowered to succeed in the 95% of the world outside our borders.
That brings me to the news of the week: Nvidia will be allowed to sell its prior-generation (soon to be two generations old) GPU chips in China, given express permission from the Department of Commerce within yet-unknown national security constraints.
If these constraints are implemented effectively, lifting the ban on Nvidia H200s will displace some Chinese chip sales in the commercial arena, replace millions spent on export control oversight and enforcement with a 25% government kickback, and strengthen U.S. AI influence in the Chinese AI stack. These are fair tradeoffs, so long as H200 sales don’t enable China’s military and intelligence agencies to accomplish objectives that they wouldn’t be able to otherwise. But it’s important to know for sure how China’s access to an increased volume of Nvidia H200s will change their national security calculus before deciding whether lifting the ban is a threat to U.S. national security or a win for AI leadership.
1. The first factor to evaluate is technical sophistication; is the H200 chip sufficiently advanced as to grant novel advantages to China’s military and intelligence ecosystem?
The answer appears to be no. The H200 has strong memory bandwidth and inference, but it does not confer exclusive capabilities, break known scaling constraints, or unlock military-relevant applications for the People’s Liberation Army. It has advantages for commercial AI workloads, but it does not change what China’s defense or intelligence services can plausibly accomplish with their existing U.S. technology stockpiles, indigenous A100- or H100-class substitutes, and clustered legacy systems.
2. The second factor to evaluate is volume; will an increased number of non-leading imported chips overcome Chinese structural barriers to AI supremacy or military intelligentization?
Assuming chip imports are fungible between China’s private and defense sectors and Xi’s military objectives supersede the interests of commercial AI developers, it is unlikely that China’s national security establishment is lacking for non-bleeding edge chips of any quantity. While domestic memory die production was the major bottleneck that U.S. chip imports would have alleviated if this ban was lifted last year, research from SemiAnalysis finds that there is no longer a limit as to the number of dies Chinese chipmaker SMIC can produce. There is certainly more than enough to support domestic commercial demand, meaning China’s intelligence and defense needs have been long met.
3. The final consideration is practical: will Beijing sacrifice key national priorities like chip indigenization and economic self-reliance (自主可控) for the benefits above?
That does not appear to be the case; Beijing is planning to limit domestic access to H200s and has ordered its data centers to remove all foreign chips and cancel future purchases. The idea of U.S. chips flooding Chinese markets is a fallacy; any short-term gains will be ringfenced by both U.S. Department of Commerce export and CCP import restrictions, and Chinese long-term policy will continue to prioritize indigenous supply chains, domestic champions, and insulation from external (particularly U.S.) leverage. They reiterated as much just last month.
My position has always been that broad, sweeping industrial policies carry real and growing costs for the United States. They demand expansive bureaucratic oversight, constant investigation and closure of evasive workarounds, and sustained public spending to police global supply chains that are inherently porous. Our government was not designed for them, and forcing our enforcement capacities and those of our allies to maintain them constrains the ecosystem that drives U.S. technological leadership. If we don’t allow American innovators to compete, scale, and lead globally, we’re sacrificing long-term strength for the illusion of control.
