Today we're talking about the latest developments and how the global AI race has changed. Although I don't even know if 'global' is anymore the best description.
Today we are doing a wide-ranging update on the state of the global AI race. And before we dive in, let me give you my pitch for why you should care about this, even if you're mostly here for 'how does AI impact me in my job.' The question of the global AI race is about much more than political machinations and geopolitical shifts. It has direct impacts on basically everything about how AI plays out. The conversation in the halls of power about the AI race shapes what types of regulations we're going to see. Practical competition around models impacts the products that we get and how fast we get them. Putting it simply, whether you are interested in the global AI race or not, the global AI race is interested in you, and has now for a year been one of the biggest topics in AI.
2025's first big story was, of course, the DeepSeek freakout when the company dropped their R1 reasoning model in a new app and wiped out nearly $600 billion from Nvidia's market cap — the single biggest loss in pure dollar terms in a single day in history. Now, reinforcing my point that this actually has impacts for you, part of why DeepSeek was such a profound moment is that users who downloaded that app, for most of them, it was the first time they were using a reasoning model. And even though OpenAI had released o1 at that point, and o1 was certainly better than R1, o1 was at the time buried behind the paywall. DeepSeek not only shaped markets, it impacted what OpenAI had to do next.
Throughout the year, the realization that China was not nearly as far behind on AI came to the fore. And heading into this year, while we knew that to be the case, the model releases from Chinese labs have been demonstrating that they are not just interested in being cheap discount versions of the real thing. When Moonshot released Kimi K2.5 a couple of weeks ago, it was absolutely nipping on the heels of the top models from the major Western labs. And I can tell you for a fact, as OpenClaw and agentic use cases have taken hold, there are lots of folks who are looking to the Kimi model as a very high-performance and much cheaper alternative to Opus or GPT. The gap has closed sufficiently that you're starting to see articles like this one from the BBC: 'Is China quietly winning the AI race?' And it's not just Moonshot that shows that Chinese model quality is increasing dramatically. ByteDance's new video model, Seedance 2.0, is currently going viral after its release earlier this week. I covered it in some detail on Tuesday's show, but the model is impressing not just for its high-quality visuals, but its naturalistic sound effects and background music, which are generated at the same time the images are rather than as a post-process. Indeed, that technical breakthrough has not come to Western models yet. And that's what has people really taking notice. The ability to innovate disrupts the mental model that most have held for the Chinese labs.
At the beginning of this year, Google DeepMind CEO Demis Hassabis discussed the state of the Chinese industry. And while he acknowledged that Chinese labs were mere months behind, he added, 'The question is, can they innovate something new beyond the frontier? I don't think that's been shown yet.' In the world of Seedance and Kimi K2.5, it's not clear how true that is, or at least how much longer that will be true for. It also feels like just the beginning of a very fast-moving year. On Tuesday, rival Chinese lab Zhipu, which rebranded itself to Z.AI, was rumored to be behind a new stealth model on OpenRouter under the codename Pony Alpha. First impressions were positive, with OpenRouter saying it delivers strong performance across coding, reasoning, and role-play, and that it was optimized for agentic workflows with high tool-calling accuracy. Now, this is only the beginning of a big wave of model releases and promotions heading into the Lunar New Year holiday. Alibaba, Baidu, and ByteDance are all said to be releasing models. And there's also a marketing war to capture a growing AI consumer segment. Alibaba is, for example, spending $432 million on discounts to promote their agentic shopping platform. And Tencent is following suit with massive giveaways. And beyond the big four Chinese giants, there's also the lurking shadow of DeepSeek. We haven't seen a new generation model from DeepSeek since they shocked the world a year ago, with each model release being an iteration on top of their V3. It's widely rumored that V4 is expected to be released in time for the New Year celebration.
Now, so far, as close as China is getting, we haven't actually seen a frontier Chinese LLM that goes beyond the state-of-the-art in the West. Even researchers close to the Chinese industry are skeptical that they can actually overtake the US labs. Justin Lin, the technical lead for Alibaba's Qwen models, recently told CNN the odds of China overtaking the US was, quote, 'below 20%.' And I think 20% is already very optimistic. Still, there's something happening with the Chinese labs that is unignorable. Even Wall Street is paying attention, with JPMorgan issuing a buy signal for Zhipu and MiniMax earlier this week. Both stocks surged after the note was published, and they were already in the middle of a huge rally. Indeed, the China market dimension of this is another really interesting dimension. Basically, the release of these new models in China right now is doing for Chinese markets what the big AI models were doing in 2024 and early 2025 before US markets got all nervous about Tim Alman's deal spree. Robert of Bagwan wrote, 'A stark contrast has emerged between the two AI industries in US and China. While Anthropic or Gemini wiped off hundreds of billions of software market value at their new products, the launch of cutting-edge Chinese models lifts everyone else. The software sector in China didn't collapse like their American counterparts. Seedance 2.0 has already sent Chinese film and online literature stocks to a new high. Why is that? One reason is that the software industry in China is far from being mature in the first place, and so the industry can leapfrog to a new era with no baggage. The same logic may also apply to film. We don't have Hollywood just yet. And perhaps the more fundamental reason is more cultural and political. In a society like China's, there is an implicit trust that the best companies cannot let a whole industry wither and die overnight without paying some price. Common prosperity, bro.'
Now, models are of course only one part of the AI race story. And in another area, chips, the status is far more in flux. In January, Zhipu announced that they'd train their first model exclusively on Huawei chips. The actual model was kind of unimpressive and nowhere near the state-of-the-art, but the training run was a proof of concept that Huawei does have now a complete hardware and software stack capable of training an AI model — a zero-to-one moment and still a meaningful milestone in that way. Meanwhile, Nvidia's H200 chips have been fully approved for export into China. We're yet to see reports of the first H200-powered training cluster firing up, but that could be just months away. The big change isn't so much about capabilities, but rather scale. The H20s that China previously had access to have slower memory and aren't as useful in building large training clusters. Chinese labs have reportedly ordered H200s by the hundreds of thousands, so we'll likely see the first Chinese models trained on Western-style mega-clusters later this year.
The relaxation of export controls is still a huge debate in Washington. Senators Elizabeth Warren and Jim Banks are planning to introduce bipartisan legislation to ban the sale of certain chips in the coming days. Count that as another 2026 prediction coming true very, very quickly. The legislation would include an outright two-year ban on Blackwell exports, as well as giving Congress a veto over export approvals in general that could feasibly revoke export licenses for H100s as well. Anthropic CEO Dario Amodei was reportedly crucial to getting the bill moving. He met with Senator Warren on Tuesday prior to the latest announcement. Now, ultimately, it will come down to how much support the bill has across party lines, with Senator Warren commenting, 'A strong bipartisan showing on the need to protect our defense is key. Selling these chips to the Chinese is bad in the short run and worse in the long run.' Still, for now, the H200 exports are going ahead, and administration officials are seeking to calm nerves in Washington. During a hearing on Tuesday, Commerce Secretary Howard Lutnick said the license terms are very detailed. They've been worked out together with the State Department, and those terms Nvidia must live with. When asked whether he trusts China to abide by those terms, which include a ban on military use, Lutnick deferred to the president's judgment.
Now, as tempting as it is to see the global AI race story as only one of China and the US, that ultimately is incomplete. Sam Winter-Levy and Anton Le wrote a recent piece in Foreign Affairs called 'The AI Divide' about how the US-Chinese competition could leave most countries behind. But there is at least one part of the world that has the ambition to serve as a third AI power. I'm speaking, of course, about the UAE. Last week at the World Government Summit in Dubai, G42 CEO Peng Xiao unveiled what he called the world's largest AI chip. The moniker was down to the chip using Cerebras' wafer-scale manufacturing process. So, the chip is physically larger than Nvidia chips. The more interesting part was the discussion around how the UAE sees their role in the AI future. Xiao described the ambition to serve AI compute to the 4 billion people who live between Milan and Singapore, representing 40% of the world's population. He also discussed G42 in terms of being a trustworthy third party neutral to both China and the US. In particular, he touted UAE law that protects corporate data in the same way that most countries treat embassies.
Since then, the headlines keep coming. Bloomberg reports that G42 plans to commence a billion-dollar data center project in Vietnam. Meaning if data center diplomacy is the new geopolitical paradigm, then the UAE has just entered the ring. In addition, Semaphore reports that OpenAI is in talks with G42 to create a specialized version of ChatGPT for the UAE. Sources said the model would be tuned for local language, political views, and speech restrictions. Developing a model that speaks in fluent Arabic has been one of the long-running discussions in Middle East AI, so that part isn't so surprising. However, the idea that OpenAI is considering baking in approved political views for a foreign nation is something that certainly has people's attention. And going back to the idea that this is all interconnected, one of the reasons that they might be willing to do it is as a compromise to block G42 from proliferating Chinese models on their infrastructure.
TL;DR: the AI race is not a simple contest between superpowers, although that is a huge dimension of it. There are plenty of neutral actors and third parties looking to stake a claim as AI geopolitics hits high gear. Maybe the most interesting new wrinkle of 2026 is the beginning of a second space race. When Elon Musk declared his ambition to build a gigantic data center in space last week, the idea was largely met with skepticism or derision in the West. People wrote it off as just an excuse for why SpaceX was actually acquiring xAI, which was assumed to be financial. Not so in China, where state media reported on Thursday that officials plan to launch their own orbital data center project over the next five years. The nation's major space contractor, China Aerospace Science and Technology Corporation, said that they would, quote, 'construct gigawatt-class space digital intelligence infrastructure.' The report said the project would, quote, 'integrate cloud, edge, and terminal device capabilities and achieve the deep integration of computing power, storage capacity, and transmission bandwidth.' Kind of words, but the important point is that they are enthusiastic about this, whereas we're all yucking it up over here. The Chinese contractor has vowed to make China the dominant space power by 2045. Interestingly, both Musk and the CCP are tying their ambitions to China's industrial base. The South China Morning Post reported that Musk has been sounding out multiple Chinese solar providers to supply his satellites. They noted that the discussions are yet to translate into orders, but it could be an interesting twist. Musk had previously said that Tesla will dramatically ramp up solar production, resulting in 100 gigawatts of supply across multiple new and existing facilities. The US currently produces less than 5 gigawatts annually, so this would be a huge step up in supply. Now, of course, both things can be true at once, but what seems clear is that Musk is looking for a solution that can get satellites in orbit in short order. And so, what started as an incredulous idea has quickly transformed into another front in the global AI power struggle. The Wall Street Journal is taking the issue seriously, with the editorial board publishing an opinion piece on the topic on Monday. After discussing various barriers to Musk achieving his goal, they wrote, 'The real danger is political interference that hamstrings America's space development and gives China an upper hand.' The piece was a very overt call for the FCC to get out of the way and approve Musk's massive space project.
And yet, even before we get there, there are plenty of political fights in the US that are firmly stuck on the ground. Late last week, New York state lawmakers threw their hat in the ring of what promises to be the single least effective AI policy for accomplishing its stated goals. I'm talking, of course, about a data center moratorium that would, in the case of New York, pause data center development for three years. This policy was floated by Bernie Sanders in Washington recently, but has gathered approximately zero popular support among the Democratic caucus. Now, it's a little bit beyond the scope of the show. When Bernie Sanders announced his call for a data center moratorium, his stated reason was to ensure that the benefits of AI would flow to everyone. My skepticism, of course, comes from the simple laws of supply and demand that suggest that the single fastest way to ensure that access is limited to only the people who can pay is to limit the amount of compute available. Which is not to say, by the way, that states, municipalities, and governments of all levels shouldn't have an extremely significant role in shaping the nature of how the data center buildout happens. The way that local consumers have traditionally footed the bill for power grid expansion is completely inappropriate in this context. And so, I do think there is a lot of room for governments to get involved here. It's just not with big bullheaded policies like this. Clearly the White House is interested in some version of this, as we're getting reports around some pact to be announced where the hyperscalers who are building these data centers announced their commitment along these lines.
Now, the data center issue had been generally viewed as a big winning platform for the Democrats heading into the midterms. But new polling is suggesting that that might not be the case. Politico writes that a recent survey found that, quote, 'most voters are blasé, even mildly positive, about the possibility of having a data center in their area, associating them with new jobs and other economic benefits. Only 32% of Democrats oppose data centers in their community, which is only mildly above Republicans at 25%.' Which is not to say that there aren't politics here. Another dimension of this, which I'm tracking, but we'll leave for another day, is the post-Trump GOP positioning, where AI is becoming one of the issues by which different parts of the Republican Party are trying to differentiate themselves.
Anyways, friends, like I said at the beginning, whether you are interested in the global AI race or not, the global AI race is interested in you. These things that happen in DC or Beijing impact what decisions are made in San Francisco, and all of it shows up in the models we have access to and the policies that surround them. Hopefully this brings you up to speed around where we stand in early 2026 on these issues. For now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.
