$NVDA | Researcher reports signs of Blackwell use in six Chinese AI models.
Will DePue, a former OpenAI researcher who helped build Sora and Sora 2, says six models from DeepSeek, Moonshot AI, MiniMax and Ant Group show signs they were set up on Nvidia Blackwell chips. Labs can set a model’s starting values on one system and move them to another for training.
•Which models matched:
– B200 or B300: DeepSeek-V4-Flash and MiniMax-M3.
– GB200 or GB300: DeepSeek-V4-Pro, Moonshot’s Kimi K3 and Kimi K3.1-Pro, and Ant’s Ling-3.1-flash.
•Reported hardware tests:
A B300 exactly reproduced some preserved DeepSeek-V4-Flash starting weights. A GB200 closely matched Kimi K3’s pattern—0.999995 correlation, according to his X post, where 1 means perfect correlation. An H100 couldn’t match V4-Flash, and changing the seed by one broke the matches.
•How he checked:
A model’s weights are the numbers training adjusts. In the PyTorch setup studied, a starting number called a seed helps generate those values; the chip affects their arrangement. Traces can survive training. DePue searched roughly 6 trillion seed-and-chip-layout combinations for matches in published weights.
•Alternative explanation:
Specially configured AMD MI300X chips can mimic the layouts. DePue says his tests ruled out that explanation for V4-Flash. He considers it very unlikely for the other five models but cannot conclusively exclude it.
•Estimated scale:
Assuming Blackwell also did the training, his best estimates are 2,000–2,500 GPUs for DeepSeek-V4-Pro and 2,000–3,000 for Kimi K3. Buying comparable clusters could cost roughly $150M–$200M each. These are estimates, not verified purchases or spending.
•Review status:
On X, DePue says senior OpenAI and Anthropic researchers checked the findings and that a full paper and code will follow.
@amitisinvesting Jensen had a flat corporate organization, no successor plan, single point of failure. Then misbehaves allowing himself to leave the company. He not only built one of the most important companies of this century, he also built a house of cards.
$SOFI just launched something most people still haven’t noticed:
SoFi Earn (SOFIE) — a bank-issued tokenized deposit.
4.4% yield.
FDIC insured.
Converts 1:1 into SoFiUSD for on-chain use.
First national bank to offer this to retail consumers.
Quiet product innovation while the stock stays quiet.
$CRWV I'm seeing a lot of confusion over this headline despite this being exactly what the management said they were going to do on the earnings call.
Let me clear up some misconceptions.
First, this is not parent-level debt. Yes, it's still debt and yes they have to pay it back but this allows them to tap into the debt market for much more attractive terms as they talked about in the earnings call. Then if they try to do something like unsecured senior nodes which would be about 16%.
Let me folks that they gave us the payoff which for shorter-term contracts is as short as 6 months and for long-term contracts is for 1 to 2 years. The point is this is a very attractive raise that is going to give them a lot of money.
This is investment grade and is a DDTL. They've done a ton of these and while we will not know the exact terms until this is finalized, there are SEC filings that detail the information on how these work and I encourage investors that are shaken by this to pull it up.
If there's something you don't understand, run it through AI and it actually does a pretty good job here.
But in essence they are going to get 7 to 7.5% on $4 billion. The way these are structured is it's tied to the sofr plus a spread. The spread is fixed at the time of the filing, but usually the total percentage is not fixed which means if the SOFR is increased or decreased then CoreWeave pays that amount more or less depending on whether or not the rate is raised or lowered. Note that rather than floating they do have at least one fixed rate DDTL, so that could be what we have here though we don't know at this time.
This is investment grade. How is that possible? It's because it's tied to an existing project with a customer that is investment grade. Occasionally these are not take or pay contracts but for $CRWV they always are. Usually for these issuances, the money can be used for any corporate purpose, although in this case 90% of the capital expenditures to this particular project are being raised from this issuance.
So in essence, what you have is a structured contract with guaranteed revenue tied to a very attractive investment grade debt issuance.
But what about depreciation? Again, the payoff. Is incredibly short right now. Second, with this type of loan structure, most of the waiting is based on who the customer is and the take-or-pay contract itself.
The company laid all of this out very clearly on the earnings call so I'm not really sure why people are upset unless they just really weren't paying attention, especially because they noted that this was coming in the following weeks.
@SelfMadeMastery Hard to say since the Jensen we thought we knew is not that person. I would hope doing what he can to keep his family out of the spotlight.