--- title: "One Rack, Two Price Tags" date: 2026-11-05 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-nvidia-51-09-1105 language: en --- # One Rack, Two Price Tags > On November 3, the Financial Times published a short report. Five unnamed people familiar with the matter said Nvidia had in recent months made a new request of cloud computing customers around the On November 3, the Financial Times published a short report. Five unnamed people familiar with the matter said Nvidia had in recent months made a new request of cloud computing customers around the world, one it had never made before. The request is a single sentence: in contracts signed from now on, at least 51 percent of services must be sold to non-Chinese companies. Turned around, Chinese customers can account for no more than 49 percent. The line is short, but it cuts many businesses in two. It was never put on paper. According to the same sources, the instruction was passed on verbally by Nvidia sales staff and regional teams, with no formal document and no clause number, handed down one layer at a time. It applies worldwide, not only to data centers in Asia. But it covers only future contracts; leases already signed and running are not affected retroactively. That boundary matters, because it means most data centers' schedules are, for now, untouched. ## Why Fifty-One The number fifty-one sits exactly on the line called "a majority." A majority means the business of that data center is, on the books, decided by non-Chinese customers, and Chinese customers, however many, are the minority side. It is softer than an outright ban and much harder than a note saying "please mind the risk." The first would cut the business off; the second nobody would take seriously. Stopping at a majority happens to leave room to keep trading. ## A Line Passed on by Word of Mouth To understand what the line governs, first see what the cloud actually sells. The chips never leave the country. The machines stay in data centers abroad, and Chinese companies send training jobs over the network, then collect the results. The equipment does not move; the computing power crosses the border. For two years, many companies treated this route as a way around controls. Now Nvidia wants data centers to read the customer list before signing a new contract. Nvidia's public response neither confirmed the number nor denied it. It said only that data centers and cloud platforms built on American chips are the foundation American model and application developers need to compete globally. That is a dignified sentence, and a vague one. It dresses a commercial clause as industrial infrastructure, while leaving the specific question unanswered: if a data center is full of Chinese customers, does it still count as that foundation? ## Two Price Tags for One Rack Price is more concrete than any clause. According to the Financial Times, citing people who know the deals, a rack of Nvidia B300 chips sells for about 600,000 dollars on the international market, a price the trade can see. The same rack changed hands in mainland China in October for 18.5 million yuan. This machine cannot come in, so the price grew a second curve of its own, steep enough to be hard to recognize. What can legally come in is another tier of chip. The B300 is one of Nvidia's most advanced products and is on the US export control list, so it cannot be sold to China. It is what American companies such as OpenAI and Anthropic use: multiple chips in one rack, a small factory once the power is on. Two prices for one machine, and what lies between them is not just freight and tariffs. One is a public quote; the other is the price of scarcity itself. Whoever pays it is not really buying a machine. They are buying time. That also explains why data centers abroad have become important. If the machines cannot be brought in, the work is sent out. When the road for sending work out also gets a gate, the pressure keeps moving in the same direction. ## The Whitelist Keeps Narrowing The new request did not come from nowhere. Earlier this year, Nvidia set up a whitelist: only companies that passed stricter compliance checks could buy its AI chips. Those outside the list could only wait, or look elsewhere. According to the Financial Times, after the whitelist was created, the number of Asian customers approved to buy fell by more than half. Most of those removed were cloud providers that rent computing power to others; they hold the most machines and send them the widest range of places. Screen the buyers first, then the renters: both doors sit on the contract side. The flow of chips is checked once, and the flow of services produced by those chips is checked again. The gaps are filled one square at a time. Tighten the buyer at one end and the renter at the other, and the two lines do the same thing: chips may stay abroad, but who is served inside those foreign data centers is now a question someone asks. The person asking wears no uniform and sends no notice; the question comes up only when a contract is discussed. ## The Trouble for a Singapore Company The pressure lands first on data center operators. A computing company headquartered in Singapore told the Financial Times that outside the United States it is hard to find enough non-Chinese customers to fill that 51 percent. It added one sentence: if the ratio is strictly enforced, business will certainly suffer. That is a plain assessment. A data center is built for a load; its customer mix is not something a knob can change. Land, power, then machines: a data center takes years. The customer list changes by the month. Asking a heavy-asset industry to predict the nationality of its customers years ahead, at the moment of signing, is hard in itself. Harder still are the providers in between. They set no rules and are not named by the rules, yet at every renewal they must judge for someone else and carry the consequences for someone else. ## The Line Is Drawn on Contracts It is worth noting that the request does not amount to halting chip sales to China. It governs whom overseas cloud providers rent computing power to. That is the services trade, not the goods trade. The two are often confused. One is whether a machine can clear customs; the other is whether a machine that cannot clear customs can still be used across the sea. Now there is one more sign on the second road, small lettering, very visible. For Chinese companies, the options are shrinking but have not disappeared. They can keep talking to data centers abroad, break jobs into smaller pieces, or wait for domestic chip capacity to climb. Each road has its own cost and its own clock. For Nvidia, the problem is not easy either. It has to find a way to keep doing business between a large market and a thickening set of compliance demands. The fact that the instruction was verbal is itself a trace of that difficulty. A rule that never reaches paper is harder to deal with. You do not know when it will be cited, or how far enforcement must go to count as compliant. The answer only becomes clear the moment the contract is signed. One rack, 600,000 dollars; another price, 18.5 million yuan. Between the two numbers lies the truest distance in the computing trade today. The line is short on paper. Drawn into business, it is very long. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-nvidia-51-09-1105 Free to quote with attribution and a link to the original.