The US-China AI Gap Widens in 2026

Nvidia Blackwell moves the "chip war" to the "cloud war"

Every other week, others and I who follow the AI space in both the US and China get asked this question: who is ahead? Depending on the news of the day, the answers range from two years, to one year, to 6 months. Generally speaking, the consensus is that the US is leading, but the race is neck and neck. 

Things are about to change. This somewhat small but persistent gap will widen, I believe, in 2026. 

That’s because Nvidia’s next generation GPU system, Blackwell, is about to hit the market in volume starting Q4 of this year (right now) and throughout 2025. Nvidia revealed last week to Morgan Stanley that its Blackwell GPU products are “booked out 12 months”. And no Chinese company or organization can place a legal purchase order on them, thanks to US export control.

Blackwell is not a chip, it is a large system. How large? The most popular version of this system, GB200 NV72, is a coherent rack that contains 72 Blackwell GPUs with 36 Grace CPUs (thus the “GB”), all connected with a ton of NVLink networking cables (thus the “NV”) with liquid cooling pipes running through it all. It stands roughly 6 foot 1 inches tall, or 186 cm, or roughly the height of Allen Iverson or Chris Paul, if you are an NBA fan. That’s one single product. 

GB200 NV72

The Blackwell line is the embodiment of the vision that “data center is the new unit of computing” – something Jensen Huang has primed the market for many times, though the physical manifestation of it is finally arriving in volume, after overcoming some production and design challenges. Putting it plainly, the competition over chips is evolving into a competition over large systems installed in bulk into cloud data centers – from chip war to cloud war.

Data centers in China are about to be shut out of this evolution and, consequently, its frontier AI progress will slow compared to US counterparts. 

Export Control is Working

Why accessing the Blackwell system, not just one but hundreds and thousands of this system, is so important deserves some explanation in the context of US export control.

There has been a lot of debate about whether US export control on advanced semiconductors is working or not in slowing down China’s AI progress. Because the state of play has been that the competition is neck to neck, the consensus conclusion is that it’s not working effectively enough. Every time Huawei shows some progress in developing its Ascend GPUs, many in DC fret about the policy’s ineffectiveness.

I don’t think making Huawei or any Chinese company give up on AI or building its own advanced semiconductor was ever the goal of US export control policies, though some may have unrealistically hoped that that would happen. But the policy’s effectiveness needs time to play out, when leading American companies, like Nvidia, roll out new products that are multiple generations ahead and well above the restricted compute power line that the policy has drawn. 

To visualize this effect, ChinaTalk made a helpful graphic to illustrate where Nvidia’s Blackwell (next gen), Hopper (current gen), and Ampere (previous gen)’s computing power sit in relation to Huawei’s Ascend GPU series, contextualized in US export control restrictions. The purple shaded region is what’s offlimit to Chinese companies. And you don’t need to be an expert in the FLOPS number of each of these GPUs to see the exponential difference in performance that the Blackwell system is capable of over the other chips. 

Source: ChinaTalk

To overlay some simple numbers over this graphic, Blackwell is supposedly 4-times more performant on training workloads over Hopper, and 30-times more performant on inference workloads over Hopper. Meanwhile, GPT-4, the model that many companies in both the US and China are still benchmarking against as the “model to beat” was rumored to be trained on roughly 25,000 Ampere 100 GPUs, or two generations behind the Blackwell series.

So it is no surprise that Chinese companies, who had a few windows to stockpile the A100s, H100s, and the other export control-compliant versions of Nvidia products (A800s, H800s and H20s), as well as access to Huawei’s Ascend GPUs (limited only by SMIC’s chip fabrication capabilities), have had no compute barrier to developing GPT-4 level or slightly better capabilities. Chinese distributed system engineers are also some of the best in the world, battle-tested by the country’s rapid Internet sector growth pre-2020, so there is a lot of distilled know-how to optimize a limited supply of compute power to squeeze the most out of these precious GPUs.

But when the Blackwell systems flood into the cloud data centers of OpenAI, Microsoft, Amazon, Google, Oracle, Meta, and xAI next year, all the hard won progress that Chinese technologists made will pale in comparison to their American counterparts. The only reason this gap is widening in 2026, and not 2025, is because it takes about a year to rack up, install, and get these GPU systems ready to train (and infer) new AI models. 

So if you are a DC policymaker who played a role in architecting US export control policies intended to widen the AI competition gap with China, you should celebrate. Export control is working; don’t mess with it. 

GB20 and GPU Smuggling

But what about all the Nvidia GPUs that China has been smuggling to circumvent export control? And what about the rumored GB20 product – Nvidia’s effort to sell a compliant version of its GB200 systems to China?

  In a “cloud war” world, both are red herrings and won’t matter much in the long run.

On GPU smuggling, the rumor is that there are about 100,000 smuggled H100s in China. That may sound like a lot, and certainly would make a crappy headline for Commerce Department officials, but when you consider that each of the leading American AI companies either does, or will soon, have their own 100,000 H100 GPU clusters with 200,000 and 300,000 GPU-scale clusters on the way, then the smuggled GPUs to China would be small potatoes. Not to mention that these GPUs only work best when strung together in a coherent cluster under one roof, not scattered across a big country. 

After all, Allen Iverson-height server racks each containing 72 GPUs is the new unit of computing, not a bag of book-sized, easy to carry chips. 

As for the modified GB20 system, assuming it is real and won’t be struck down by a new round of export control, there are not enough details to determine whether its server-level redesign will make up for the inevitable performance downgrade of the rumored B20 chip (similar to downgraded H20 to the H100). However, time to market, not relative performance, is what makes all the difference. 

Given that the backlog of Blackwell orders are already stretching to a 12-month wait time with “insane”, Nvidia will prioritize its entire supply chain to fulfill the unfettered orders from the entire global-ex-China market, before turning to meet the restricted, highly scrutinized, though no less enthusiastic demand of its Chinese customers. I would not be surprised if the GB20 system does not hit volume rollout until 2026 or 2027. 

This slower time to market means Chinese companies are at least one year behind in getting what the best available Nvidia product is available to them, and that product is nowhere near as powerful. Meanwhile, all the American companies are highly motivated to race ahead to rack up as many Nvidia systems as they can get their hands on to compete with each other. At the extreme, this motivation has pushed Elon Musk’s xAI to go from installation to conducting the first training run of its 100,000 Nvidia GPU supercomputer in Memphis, Tennessee in 19 days, as Jensen recently revealed.

While I have no doubt that any highly motivated Chinese tech company can also pull off a similar timeline (China speed is a real thing after all), if those GPUs only arrive in bits and pieces via smuggling routes or in small shipments due to de-prioritization from Nvidia, then no amount of motivation can make up for that lost time. 

For people working in China’s AI ecosystem or rooting for its success, my conclusion can be upsetting. For people using China’s AI ambition as a threat to advocate for certain policy or business outcomes, my conclusion can be inconvenient. But what I’m presenting is an impassioned analysis of what I think is most likely to happen – no biases nor invested interest for one side or the other.

Can things change and I’m proven wrong? Certainly. 

If we have a change in administration after the US election, existing export control could be gutted or lobbied away. If distributed computing for AI training across multiple data centers becomes more mainstream, Chinese companies may be able to rent all the Blackwell systems they need from cloud hyperscalers outside of China, if the “cloud loophole” doesn’t get closed.

I will of course update my thinking when material changes happen.  But until then, as an investor, I must live in the world that is, not the world that someone wants it to be.