DeepSeek just raised a few hundred billion.

Liang Wenfeng said that if they can spend $20 billion on buying cards this year, that would mean the procurement team’s performance is super strong.
The biggest anxiety for a CEO is not being able to spend the money.
He worked out the math: converting the funds into NVIDIA cards would pay back in ten months. With the bank deposit annualized yield at 2%, his strategy is to buy as many as possible, with no upper limit on spending speed.
Then he said that actually spending all that money is very difficult—cards are unavailable, hard to get, and the prices are also very high. But I’m willing to pay a premium for it because it’s certain.
This part carries more weight than any research report.
He currently has about 20,000 NVIDIA H100 equivalent compute cards, most of which were purchased only in the last one or two months.
Huawei is also quoting him: 16,000 cards, which he calculated is equivalent to 240B-series cards.
But that still isn’t enough. The biggest model parameters in the US are on the order of 49.6B, and China is short by one order of magnitude.
He said training a model at the same level would require 50,000 GB300 cards, or 20,000 Huawei cards—this is just for training, not counting research usage.
The most technically recognized strongest team just got a huge amount of money, and when they ran the numbers, they found the chips are still not enough to spend.
And this is only DeepSeek.
At the same time, Anthropic is negotiating compute capacity with six companies. Kimi K3 went live and was squeezed out in 48 hours, so it directly paused its C-end service and allocated all compute to paying customers. Google just increased its full-year capital expenditures to $205 billion; with $240 billion on the books, it still wasn’t enough to spend, so it raised another $49.6 billion.
One supplier said the lack of inventory is a fact. Two others said the industry is doing well. Five people with completely different stances said the same thing one after another: can’t buy, and prices are still being marked up.
So why have storage and chips been falling like this recently—what exactly is falling? There’s no evidence of any collapse on the demand side.
Not a single word has changed in demand.
Liang Wenfeng’s line, “If we can spend $20 billion, that would be great,” might be the most real AI demand signal this quarter.
This will be more useful than any analyst’s report.