The only layer you can't download

On 27 September 2026 The Information reported that China's Ministry of Industry and Information Technology had asked ByteDance, Alibaba and other companies to report how many of Nvidia's RTX PRO 5500 cards they plan to buy and for what, and that it told some of them the government intends to approve those purchases ^[1]. Reuters, which picked up the story, cautioned that it could not verify it independently ^[1]. ByteDance is weighing an order of around one million units, according to sources cited by the Korean press ^{[2] [3]}, and Nvidia plans to ship some 500,000 a quarter to China from late December ^[2]. A single order would absorb half a year of supply. The RTX PRO 5500 is not a data-centre accelerator but a workstation card, with 84 GB of GDDR7 memory, and its spec sheet makes no mention of NVLink, the interconnect that turns many chips into a single training system ^[4]. Neither the ministry nor the companies have confirmed or denied the report, and some industry executives expect the card to sidestep US export restrictions ^[1]. Nvidia answered with a double complaint, against «outdated US export controls» and against «China's own limits on US imports» ^[1].

The news is best read as an instance of something larger, and that larger thing should be named before its protagonists. Artificial intelligence is centralising in the only layer that cannot be downloaded, the layer of atoms: the silicon, the memory and the watts it takes to switch a model on. China has spent three years giving the world the product of its investment back in the form of open weights, which anyone can copy, and this week it showed that not even China has chips to spare. What can be downloaded gets shared and what cannot gets concentrated, and that asymmetry decides which countries will think with an intelligence of their own and which will rent one.

The country of the home-grown chip goes shopping for someone else's

Ten days earlier, on 17 September, Huawei had shown in Shanghai the roadmap by which China means to depend on no one. The company announced, in the keynote of its rotating chairman David Wang, that the Ascend 960DT will arrive in the first quarter of 2027, three quarters ahead of schedule, and Wang committed the 970 for 2028 and the 980 for 2029 ^[5]. Eric Xu, the other rotating chairman, said that same day what the keynote left out: «Since we don't have enough capacity to even satisfy the demand in China, we don't have a plan to expand into the international market in a fully-fledged way» ^[6]. His other line explains the whole strategy: China «cannot accept a destiny» in which its fate hangs on whether others are willing to sell it chips ^[6]. The shortage already had symptoms. Moonshot AI, which in July released the largest open model in the world, suspended new subscriptions to its service that same month because, according to Omdia analyst Lian Jye Su, it did not have enough chips to serve the surge in demand ^[7].

Until now, the brake was applied by Beijing. In September 2025 the cyberspace regulator ordered Alibaba and other companies to halt the purchase and testing of the RTX Pro 6000D, the card Nvidia had designed for the Chinese market ^[8]. Washington authorised sales of the H200 in February 2026, and in August it emerged that ByteDance and Tencent had each received around 10,000, out of an allowance of up to 100,000, and that Beijing asked them to send the rest to Hong Kong rather than the mainland ^[9]. Nvidia's quarterly report puts it bluntly: H200 sales were restricted by the Chinese government, and the company projects its next quarter without a single dollar of data-centre compute revenue from China ^[10]. Customers based in China still bought US$7.88 billion from it last quarter, twice as much as a year earlier, with H200 shipments below 1% of its data-centre sales ^[10]. What is new this week is that the ministry no longer blocks those purchases but asks for them. The RTX PRO 5500 costs between 85,000 and 90,000 yuan in China, a price similar to that of Huawei's Ascend 950PR ^[2], so the choice is not about price but about availability. No agency in Washington has yet said whether the card falls outside its controls, and for now the executives' expectation is only that.

A summit on superintelligence without a single chip

Donald Trump received Xi Jinping in Washington on 24 September, halfway between Huawei's announcement and the ministry's. The fact sheet the White House published the next day says the two leaders agreed to use the term «super intelligence» rather than «artificial intelligence», to create a Superintelligence Dialogue whose next exchange will take place before November and to open a bilateral channel for reporting incidents ^[11]. The words chip, semiconductor and Nvidia do not appear in the document. The Chinese government's readout does not mention them either, and describes the two countries as holders of «strong AI capabilities» for whom the right thing is to draw on each other's strengths, «not guard against each other» ^[12]. Trump had set the tone that same day on Truth Social, where he wrote that superintelligence would be a big topic of discussion but that he wanted to leave it «exactly where it is», with the Department of Justice as the only barrier ^[13]. Three days after the summit, Beijing asked its companies for their shopping lists.

The two countries that sat down at that table hold almost all of the planet's compute. The UN's independent scientific panel on artificial intelligence calculated in July that the United States accounts for 75% of the computing power of the world's 500 largest AI supercomputers and China for 15%, which leaves most member states «dependent on systems they cannot build, inspect, audit or fully adapt to local context» ^[14]. Superintelligence was discussed between the owners of 90% of the compute. Compute was not discussed.

What China gave back

China has managed its share of that compute in a way the United States has not imitated. At the World Artificial Intelligence Conference in Shanghai in July, Xi Jinping called on the world to «encourage open source, openness, collaboration and sharing», and offered to help the countries of the Global South build capacity ^[15]. Hugging Face found in August that in almost every month of 2026 the largest open model from a Chinese lab was bigger than anything released by an American one ^[16]. DeepSeek released its 1.6-trillion-parameter flagship under the MIT licence, the most permissive there is, and on 10 September added a new version under the same licence ^[17]. Eddie Wu, Alibaba's chief executive, explained in August why a company gives away what cost it billions: «a thriving open source model ecosystem drives greater demand for our cloud computing services» ^[18]. The gift, in the words of the one giving it, is the doorway to a rental. Those are the facts.

An open model is no use switched off. Kimi K3, the largest open model there is, takes up some 1.56 terabytes of memory even compressed, the deployment recipe its maker links to calls for at least eight GB300 accelerators and several nodes for real traffic, and renting what it takes to switch it on costs between 25,000 and 50,000 dollars a month ^[19]. Edition Nº 003 of this publication argued that building on open weights is the region's only road to the frontier, and it still is, though that road ends at machines that are rented. ECLAC measured it from the ledger: Latin America generates 14.4% of the world's traffic in artificial-intelligence solutions and receives 1.12% of the investment, while 86% of the value is concentrated in the first two links of the chain, equipment and compute ^[20]. The region downloads the weights and rents the machines.

The question of whether China will keep publishing has a partial answer, and it is not the one the pessimist fears. The weights keep arriving, with conditions that grow with size. The Kimi K3 licence requires a separate agreement from anyone selling the model as a service with revenue above 20 million dollars a year ^[21]; Qwen3.8-Max's demands a commercial licence above 50 million ^[22], and GLM-5.3's reserves a security review for giants above 10 billion ^[23]. The first two thresholds are drawn right where a Latin American success would begin, so the company in the region that serves a Chinese model at continental scale will stop being a user and become a tenant. The economics push the same way. Zhipu lost 2 billion yuan in the first half on revenue of 954 million ^[24], and ChinaTalk summed up the dilemma: small open models help with going abroad, but they are not going to pay for gigawatt clusters ^[25]. Beijing is discussing limits on foreign access to its most advanced models, «both closed-source and more open versions», though according to Reuters it is unclear when or whether they will take effect ^[26]. Washington threatens that openness from the other side, with sanctions over distillation and, according to the US press, a possible ban on Chinese open models ^[27].

The counterpoint has to be conceded in full. Chinese openness is not closing but narrowing: DeepSeek keeps the MIT licence on its flagship, Alibaba reopened in August its Max model, which it had held back, and Xi put his signature on open source two months ago ^{[15] [17] [22]}. Mid-sized open models today run on a computer with a single consumer card and cover a company's everyday uses, from local agents to programming ^[28]. The million-card report rests, besides, on anonymous sources. The risk is not that China stops publishing tomorrow but that its conditions get written in two capitals that have just shown they can talk about superintelligence without asking anyone.

Brazil has already chosen how to live with that. The government is tendering a supercomputer of its own for just over 1 billion reais, which the science minister expects Nvidia to supply, and in parallel closed a cooperation with the Chinese firms Huawei and iFlytek for another 1.276 billion ^[29]. Having a supplier on each shore is not sovereignty, but it is the realistic form of sovereignty when the chips are handed out by two countries.

THE PLAY

1. Budget for the hardware before the model. The weights download for free and the compute does not, so a sovereign plan that funds a model and rents the machine buys a dependency under another name. Brazil's billion-real tender for a machine of its own is the right unit of measure ^[29].
2. Read the licence as a lease. The 20- and 50-million-dollar thresholds mark the day your success starts paying rent. Choose the base with an eye on where you will be in three years, not on today's leaderboard.
3. Own the stretch that fits on one card. Fine-tuning, the corpus and evaluation on mid-sized models are done on machines of your own. The frontier can be rented, as long as there are two interchangeable landlords and never just one.