A $35 billion order, set against production capacity of 3 million cards. Chinese customers reported demand for 6 million.
ByteDance, Alibaba, and Tencent placed the order together: RTX PRO 5500, at $11,400 per card. Reuters reported it on October 20; all three companies declined to respond to requests for comment. NVIDIA told its partners that global production of this card would top out at 3 million units in 2027.
The shortfall is obvious. If every card produced worldwide went to China—and nowhere else—it would still cover only half the demand.
What partners were told in private was even more direct: only 5 to 10 percent of the requested volume might be approved.
This isn’t even a top-tier card. It’s a workstation card, with 84 GB of memory and 600 watts of power consumption. Chips with the Blackwell architecture, like the B300, can’t be sold to China, while H200 sales are reviewed case by case. After going around in circles, the RTX PRO 5500 has become the most powerful card available within the rules.
A lot of people think export controls are about computing power. What they really restrict is memory.
Inference and training are fundamentally different. Training can wait in a queue. Inference can’t. When a user clicks, the system has to calculate. The model weights have to stay in memory; if there isn’t enough, you have to split the workload across machines, reduce concurrency, or batch requests together. Batches pile up behind one another, latency rises, and users leave. At every step in that chain, everything is waiting on the capacity of the memory on the same card.
The numbers make sense. Alibaba Cloud’s external revenue grew 45 percent that quarter, and AI-related revenue has grown by triple digits for twelve consecutive quarters. The group’s capital expenditure approached $10 billion, up 75 percent year over year. Tencent’s capital expenditure in the second quarter was 52.8 billion yuan, up 176 percent. Whatever else is happening, the money is genuinely being spent—and increasingly, it’s going toward inference.
The spokesperson’s comment was interesting. They said that since restrictions were introduced in 2022, Chinese workstation chip and graphics card manufacturers have achieved “unprecedented growth.” That’s officialese, but it says a lot: the flip side of export controls is that they’ve handed this entire market to domestic manufacturers.
So what does that $35 billion buy?
I think it buys time. The restrictions are drawn up by model, so demand shifts to models that aren’t on the list. This card isn’t on the list today. What about next year? And this order was reported just when the restrictions were at their tightest. They put several years’ worth of demand on the table at once, because nobody knows what will still be available under next year’s rules.
That also explains the strange figure. Chinese customers reported demand for 6 million cards. Do they really need 6 million? Maybe not. Nobody wants to fall behind. Report too few and you might not get a place in line. At this point, an order is a queue number, not necessarily a measure of actual demand.
No one is willing to say exactly when deliveries will arrive. The official timeline calls for early samples in December and volume shipments in the first quarter of next year. Partners have only been given a verbal range for how many they might get; there’s never been a written commitment.
Tokenized NVIDIA shares on Binance trade around the clock. In the hours after the news broke, the market priced in both figures: $35 billion and a little over $3 billion. Next quarter’s data center revenue will show which one is real.
The information that really matters isn’t in the order. It’s who gets the first batch of sample cards in December. That’s the next number we’ll be able to verify.
$NVDAB $TAO $BTC #股票财报季 #英伟达 #比特币 #Binance Square
ByteDance, Alibaba, and Tencent placed the order together: RTX PRO 5500, at $11,400 per card. Reuters reported it on October 20; all three companies declined to respond to requests for comment. NVIDIA told its partners that global production of this card would top out at 3 million units in 2027.
The shortfall is obvious. If every card produced worldwide went to China—and nowhere else—it would still cover only half the demand.
What partners were told in private was even more direct: only 5 to 10 percent of the requested volume might be approved.
This isn’t even a top-tier card. It’s a workstation card, with 84 GB of memory and 600 watts of power consumption. Chips with the Blackwell architecture, like the B300, can’t be sold to China, while H200 sales are reviewed case by case. After going around in circles, the RTX PRO 5500 has become the most powerful card available within the rules.
A lot of people think export controls are about computing power. What they really restrict is memory.
Inference and training are fundamentally different. Training can wait in a queue. Inference can’t. When a user clicks, the system has to calculate. The model weights have to stay in memory; if there isn’t enough, you have to split the workload across machines, reduce concurrency, or batch requests together. Batches pile up behind one another, latency rises, and users leave. At every step in that chain, everything is waiting on the capacity of the memory on the same card.
The numbers make sense. Alibaba Cloud’s external revenue grew 45 percent that quarter, and AI-related revenue has grown by triple digits for twelve consecutive quarters. The group’s capital expenditure approached $10 billion, up 75 percent year over year. Tencent’s capital expenditure in the second quarter was 52.8 billion yuan, up 176 percent. Whatever else is happening, the money is genuinely being spent—and increasingly, it’s going toward inference.
The spokesperson’s comment was interesting. They said that since restrictions were introduced in 2022, Chinese workstation chip and graphics card manufacturers have achieved “unprecedented growth.” That’s officialese, but it says a lot: the flip side of export controls is that they’ve handed this entire market to domestic manufacturers.
So what does that $35 billion buy?
I think it buys time. The restrictions are drawn up by model, so demand shifts to models that aren’t on the list. This card isn’t on the list today. What about next year? And this order was reported just when the restrictions were at their tightest. They put several years’ worth of demand on the table at once, because nobody knows what will still be available under next year’s rules.
That also explains the strange figure. Chinese customers reported demand for 6 million cards. Do they really need 6 million? Maybe not. Nobody wants to fall behind. Report too few and you might not get a place in line. At this point, an order is a queue number, not necessarily a measure of actual demand.
No one is willing to say exactly when deliveries will arrive. The official timeline calls for early samples in December and volume shipments in the first quarter of next year. Partners have only been given a verbal range for how many they might get; there’s never been a written commitment.
Tokenized NVIDIA shares on Binance trade around the clock. In the hours after the news broke, the market priced in both figures: $35 billion and a little over $3 billion. Next quarter’s data center revenue will show which one is real.
The information that really matters isn’t in the order. It’s who gets the first batch of sample cards in December. That’s the next number we’ll be able to verify.
$NVDAB $TAO $BTC #股票财报季 #英伟达 #比特币 #Binance Square