Transcript 0:00 On Sunday, the world's largest open model sold out. Moonshot AI paused new subscriptions to Qimi K3 three days after launch. Demand over 48 hours had pushed close to the limits of its computing capacity. 0:14 Their exact words are, "Our GPUs are feeling it." Last week I asked, who can run those enormous open models? The first answer came faster than I expected actually. For a moment there, not even the company that built one. 0:29 And the sellout is only the smallest event of the weekend. On Friday, Xi Jinping personally opened China's largest AI conference and told the world to embrace open source AI. 0:40 On Saturday, Alibaba teased a 2.4 trillion parameter model and promised to release it openly. And on Monday morning, Axios reported that the Trump administration is showing signs it could ban Chinese models entirely. 0:54 In only four days, a lot of things happened and changed about what's going on with open model race in the whole world. [upbeat music] Attention Span. Welcome to Attention Span by Turing Post. 1:09 My name is Ksenia, and today we're going to look at what sold out, why the shortage serves both Moonshot and Beijing, and why Washington's answer may be a ban. 1:20 An AI model sold out is a strange sentence, I agree, but it's a good one because it catches your attention. Let's see what exactly does it mean. The weights of Qimi K3 are not out yet. They are promised for July 27th. 1:33 Until then, K3 exists only as a hosted service on Moonshot's own clusters. When usage exceeded forecasts within 48 hours of launch, Moonshot suspended new consumer subscriptions. Existing subscribers are well-protected. 1:49 New spots will reopen in batches as capacity is added. Moonshot also split its membership in two tiers, one for the web and app products, and a separate QimiCode membership for programming workflows. 2:02 That split is an important signal because it signals about economics. K3 was built for exactly the workloads that consume the most inference. 2:11 An agentic coding session can run for hours, making repeated tool calls and holding enormous context. And the regular chat user maybe asks a couple of questions and leaves. 2:23 So every new coding subscriber is not just the user for the model. Each one is a small continuous compute obligation. And remember the 2.8 trillion parameters from last week? 2:34 The router activates only 16 out of 896 experts per token, but all of them still have to be stored and served. When the flagship use case is a model working on its own for hours, capacity runs out fast. 2:50 Now the part that I find most interesting, this is an economics story. Moonshot, the company behind Qimi models, is preparing a Hong Kong IPO. The potential valuation is above $30 billion. 3:03 Annual recurring revenue reportedly reached $300 million in June, up from 200 million in April. Competitors, including DeepSeek, are also raising capital specifically to buy compute. 3:16 Now, when you read those numbers together, the pause does not look like a failure anymore. Quite the opposite. A model too popular to sell is an unusual pitch to investors. It is also a persuasive one. 3:29 The shortage becomes proof of demand right before a listing. The July 27 weights release fits this logic too. Once the weights are public, larger customers can run K3 on their own hardware and skip the queue entirely. 3:44 Open weights become the overflow valve for Moonshot's own capacity problem. Customers with infrastructure serve themselves. Moonshot's clusters serve everyone else. 3:54 And last week we asked, what are open weights for if you cannot run the model yourself? Here is a concrete answer. 4:01 They're for people who can, and this weekend showed that even the model's creator needs those people to exist, but even them are feeling GPU constraint. 4:10 This is the latest message from Databricks CEO, "We host open source models such as Qimi and offer them to our customers. Demand has been so strong that we are running out of GPUs across multiple regions." 4:25 And while Moonshot was rationing GPUs, Alibaba announced Qwen 3.8 with potentially 2.4 trillion parameters, multimodal, claimed to be second only to Claude Opable 5, and this is an unusual part for Qwen, they're going open weight. 4:43 This is just a promise as of now, a preview and a promise, because every capability claim attached to Qwen 3.8 right now is unaudited. But the strategic signal is very important even without the data. 4:57 Qwen's biggest max your models have stayed closed in recent generations. Open model researcher Nathan Lambert pointed out that something is changing in how Chinese labs decide what to release. 5:09 And one more detail, Alibaba holds a stake of roughly 36% in Moonshot. So these are not two unrelated companies racing. This is one ecosystem escalating in public, days apart, in the same direction, towards open. 5:25 Why the same direction? Because three days before the sellout, the direction was announced from the biggest stage China has. 5:33 On Friday, Xi Jinping attended the World AI Conference in Shanghai for the first time and delivered the opening keynote. He urged countries to seize the historic opportunity of open source AI. 5:45 He framed China's open models as a public good. 5:48 And then another thing was announced alongside the speech, because 29 countries had agreed to establish the Shanghai-based World Artificial Intelligence Cooperation Organization, or I call it WAICO, while Xi promised 5,000 AI training opportunities and application centers with groups including ASEAN, the African Union, and BRICS. 6:11 The USA and Europe Would not include Waco does not require its members to adopt Chinese models, but it gives China an institutional forum for the training, standards, and governance discussions that shape which ecosystems gain adoption. 6:27 American export controls still constrain China at the most advanced chip layer, which gives Beijing an incentive to make the model layer portable, inexpensive to adopt, and difficult to contain. 6:38 A downloadable model can run on another country's infrastructure, support local services, and reach developers without a Chinese cloud. 6:47 Waco training programs and standards discussions can provide institutional channels around that technology, so when countries train engineers and build products around Chinese models families, China gains influence even if they host the models themselves. 7:03 Open weights therefore become a distribution strategy for technical standards, developer habits, and international relationships. Jensen Huang is pissed off because of that. 7:14 This is Xi Jinping clearest articulation yet of China's ambition to shape global AI governance. 7:20 The same conference for showcased Huawei's newest AI computing systems, and the Kimi crunch immediately strengthened the argument for domestic inference chips. 7:30 So assemble the week from Beijing side, the head of state declares open models a pillar of foreign policy on Friday. The country's top open model sells out over the weekend. 7:41 The largest tech company announces its biggest model ever will be open, which raises the real question, what is the plan here? And here's my read, and it works on three levels. 7:52 The first level is commercial, and it's the simplest level. A closed Chinese model has no path to global customers. That's clear. 8:00 No Western enterprise would send its data to a closed Chinese API that also trails the frontier. Open ways dissolve that objection. You can inspect the model, host it yourself, and keep your data in your own building. 8:13 For Chinese labs, openness is not generosity. It's the only distribution channel that works, and it is working now. Chinese models have reportedly been overtaking American ones in monthly open model token use. 8:27 The second level is industrial. Almost every company builds their own base models in China because controlling the stack is the default instinct. Open first is the practical extension of that. 8:39 Releases bring feedback, ecosystem, and talent, and Chinese labs still depend on NVIDIA for training. However, everyone wants more of those chips than they can get. 8:50 But inference is different, and inference is where Huawei and a dozen domestic chip lines can already compete. And a free popular open model generates exactly one thing at scale, inference demand. 9:02 So give away the model and you build the market for your own hardware underneath it. The model is free. The compute is where you can get locked in. And the third level is the state. 9:12 Xi Jinping's speech turned a commercial equilibrium into foreign policy. Models as a public good, training programs for the Global South, a new international organization headquartered in Shanghai. 9:24 If the American offer to the world is intelligence as a subscription, the Chinese offer is now intelligence as infrastructure, something you are given to run yourself and build on. 9:35 One is a product, the other is a sphere of influence. So the plan, as far as I can see, is not a directive from the top. 9:43 It's three logics, distribution, hardware, and diplomacy, that all point in the same direction, and the state has now noticed they align. That alignment is what makes this different from the deep seek moment last year. 9:57 This time, openness has capital, chips, and a foreign ministry behind it, which brings us to Monday morning, today morning. 10:05 According to Axios, the Trump administration is showing signs it could ban cutting-edge Chinese AI models. Parts of the administration have attempted de facto bans on foreign open source models before. 10:17 Kimi's rise last week is reportedly reigniting those efforts. I will hold my judgment until there is a policy to read, but it's worth mentioning what a ban of an open model can and cannot reach. 10:30 It can reach hosted APIs, app stores, government procurement, and enterprise compliance departments. It cannot reach the weights themselves. 10:40 After July 27, when Kimi K3 promised to publish their weights, those weights will become files, downloadable, copyable, and runnable by anyone with the right hardware. 10:52 A ban would not remove Chinese models from the world. It would decide which developers build on them openly and which build on them without saying so. So in one weekend, the constraint on open AI moved twice. 11:06 First, from capability to compute. The models are strong enough now that the bottleneck is GPUs. 11:12 Then from compute to permission, because the next question about the model you use may not be whether it is good enough or whether you can afford to run it, but whether you will be allowed to choose it at all. 11:23 The weights drop on July 27. What governance do after that may shape this race more than any benchmark. That's it for today. 11:31 Please tell me in the comments which constraint do you think decides this race, compute or policy? And thank you for watching. I'll see you at the end of this week. [outro music]