Dario v Jensen

There’s been a bit of a UFC vibe to Anthropic CEO Dario Amodei, and Nvidia’s CEO Jensen Huang’s recent blog posts and podcasts. The debate revolves around whether the U.S. should give Chinese AI researchers access to Nvidia’s GPUs for training AI models. Jensen says yes; Dario says no.

In a contentious interview on the Dwarkesh podcast, Jensen Huang challenged the case for restricting Chinese researchers’ access to Nvidia GPUs. He noted that there are five layers to the AI environment. At the bottom sits energy, then comes chips, above that is infrastructure (data centres), then models (the AI labs) and finally the applications for AI. All five layers are critical, and America should lead all of them.

If the U.S. constrains AI researchers from Nvidia’s GPUs, China will then have to build its own tech stack that does not involve U.S. technology. He notes that Huawei can make 7 nanometre chips. Although these are not as efficient as Nvidia’s, China’s renewables investments and grid connections mean they have significant energy resources for data centres. China could cluster the larger GPUs and utilize the cheap energy for model training. Additionally, most of the efficiency gains do not come from the semiconductors, but from the algorithms that run at the chip layer. Nvidia has spent two decades cultivating an ecosystem and optimizing for their use cases. When Chinese AI lab researchers try to use Huawei, they encounter enormous frustrations. As a result, they utilize Nvidia and its CUDA ecosystem and are reluctant to switch, even under Chinese government pressure to develop its domestic chip industry.

China now graduates 50% of the world’s AI lab researchers. They utilize an open-source architecture in conformity with the Chinese governments edict that AI should be a utility for the whole world. If the open-source movement eventually clusters around a Chinese tech stack, then Jensen believes it will be detrimental to U.S. interests.

Jensen envisions a world where China will end up with the lowest cost energy, powered by Chinese chips and data centres, building world class models and the whole Global South (and maybe the world) building on top of it.

Dario takes a diametrically opposite view. In a recent blog post, designed to help Congress campaign against President Trump giving concessions on chips to Beijing, Dario cited a study that the U.S. will have access to 11x more compute than China. This allows American AI labs to create indomitable leads in AI models and be the backbone for the global economy.

In addition, Chinese AI lab researchers note that they are compute constrained and there aren’t enough chips, even Chinese. One Chinese podcast noted that Chinese AI researchers could absorb 100x-1000x times the compute capacity than what they currently have. Many Chinese companies are innovating, but for them the iterations are not fast enough.

On Jensen’s point that constraining access will force AI labs in China to innovate on other axes, and that algorithmic improvements matter more than smaller chips, Dario says algorithmic improvements are a function and multiplier of compute, not a substitute for it and discovering advances is a compute-intensive process: more compute allows labs to run more experiments and then make algorithmic improvements.

Anthropic thinks if you restrict chips and prevent Chinese labs from using distillation techniques, America’s will have a compute edge, American AI models will stay 12-24 months ahead of China’s, American AI will be the backbone of the global economy, the West will enjoy cyber advantages and the U.S. will be a more attractive partner for AI applications and compound its leadership.

On the other hand, if China can get the chips, Chinese models will stay near the frontier, there is rapid commercial and state adoption due to China’s AI policies, AI-enabled cybersecurity becomes a serious threat, and Beijing wins global adoption on cost and on-premises flexibility (i.e. – open-source).

I fed Dario’s blog and Jensen’s podcast transcript into Claude to evaluate the debate. Claude said its “a fascinating strategic dilemma with compelling arguments on both sides.” But it said Jensen’s position is stronger. First in terms of historical precedent, when the U.S. restricted GPS access to satellites, China built its own system. When the U.S. denied China access to the International Space Station, China built its own space station. The pattern suggests restrictions accelerate rather than prevent indigenous developments, as Jensen argues. Claude also notes that cutting off such a large talent pool from the U.S. ecosystem could rapidly accelerate independent innovation. China also already possesses indigenous capability. China is planning comprehensive AI integration with state coordination. Combined with cheap energy and manufacturing advantages, they already possess strong capability. The Financial Times reported this weekend that Chinese labs seem to have better image generation models trained as the Chinese ‘shorts’ ecosystem for training data is multiples of the U.S. Claude also says Jensen is likely right that algorithms and ecosystems matter more than chips. The CUDA moat has proved remarkably durable even as hardware became more commoditized.

Claude says Dario did raise valid concerns. The Xinjiang surveillance state and military AI development show genuine risks of China Communist Party AI leadership. Dario also cited a Chinese cybersecurity researcher who compared Anthropic’s Mythos model to a gatling gun versus China sharpening its swords. That suggests the gap may be quite wide. And China is indeed doing massive distillation attacks (although note Anthropic does not exclude that from their annual recurring revenue).

Claude recommends that the U.S. maintain software and security restrictions to prevent distillations but allow hardware sales. The U.S. should copy China’s strategy to accelerate AI adoption. The U.S. could also insist that Chinese researchers conform to safety protocols to have access to CUDA. The CUDA dependency would also allow the U.S. to inhibit China as a future chokepoint in the event of a conflict.

The U.S. government allowed Jensen to join their China trip, so it seems to have sympathy for Jensen’s views (maybe they use Claude too), but China seems determined to create a separate tech stack so it may be too late. In the short-term (or at least this year), the U.S. AI ecosystem and stocks should enjoy robust demand, but as China grows its domestic capability there is the potential for pricing erosion and a correction.

Bond Vigilantes Talk to Burnham. With Labour’s abysmal results in local elections, the knives are out and the rank and file want a change. Andy Burnham has put up his hand and will run for a by-election to become a member of parliament and join the fray in Westminster. It will be an important referendum; in recent local elections Reform received around 50% of the vote in the area Burnham is running. compared to Labour’s 25%. If he can beat Reform’s attack, it could reset Labour’s popularity. According to Polymarket, Burnham has a 66% chance of winning the seat.

The problem is the bond market. A Burnham reset is likely to involve more spending. And there is no more money in the pot. The UK has the highest bond yields amongst major developed economies. Last year, I talked about my tail risk fears that the U.S., France, Japan or UK would suffer from a bond revolt, leading to contagion risk in all bond markets that do not have control of their fiscal spending. We seem to be getting closer to that point in the UK. But first Burnham must win, then he needs to win a leadership challenge (likely) and then he needs to ignore the bond market (less likely).

Its Takes Two to TACO. With China seemingly not putting much pressure on Iran to end the war, oil prices pushed up on Friday. Trump still has no solution to open the Strait. Higher energy prices are feeding into global inflation expectations. The bond moves this week were significant with the U.S. 30-year Treasury hitting 5.1%, the highest level since 2007. The 10-year approached 4.6%. The bond market traded stable but stressed.

I Was Hoping for a Bit More. It seemed that President Trump and his coterie had a nice holiday. They saw the Temple of Heaven, the Great Hall of the People and some nice gardens. Children greeted Trump with an up and down dance that seemed a bit North Korean in style. Xi Jinping promised to send some rose seeds.

China bought a few planes from Boeing, but lower than expected. They agreed the war in Iran should end and that Iran shouldn’t have nuclear weapons. However, how that opens the Strait is a mystery. China made clear it would not be happy if Trump sold $14 billion of arms to Taiwan. Trump said he would consider allowing China to buy Iranian oil without sanctions. China does not seem to have requested any access to Nvidia chips and continues to restrict access domestically.

They talked philosophy, specifically the Thucydides Trap, that a rising power often ends up in conflict with the incumbent hegemon. China also brought up the concept of “strategic stability” which is defined as keeping “competition within proper limits” for stability and peace. This concept was used in the Cold War.

The lack of results is negative for inflation expectations. Overall, it seemed like a lot of time and expense to get a few planes orders, some agriculture orders and some rose seeds. Optimists may say there will be more talks, and this is only the first step.

Blackrock in the Principal’s Office. The Manhattan U.S. Attorney’s office is investigating Blackrock TCP Capital Corp’s valuation practices. This is a BDC, or business development corporation for private credit. In addition, federal prosecutors brought executives in for questioning. TCP reported a 19% drop in NAV. Several investors have filed a class action suit that TCP had not marked its loans correctly.

This validates what I have written in prior posts about private credit NAV opacity and that the marks are not necessarily correct.

End Note

I listen to TBPM, a podcast that OpenAI bought. It used to primarily focus on AI news. But since the takeover announcement, the amount of time spent on luxury items has expanded exponentially. One episode devoted half its time to expensive watches after over 600 OpenAI employees cashed out of $6.6 billion worth of stock. This signals that AI labs culture may shifting from cutting edge AI research to who owns the nicest watch.

Omar Sayed