Deus

🚨 Another major scandal is coming out of NVIDIA. And this one has nothing to do with Jensen Huang's affairs. According to a Wall Street Journal report, Nvidia was offering billions in financial guarantees to AI companies that could not afford its chips, so they could keep buying them. Some of these deals were being pushed by Jensen's own son-in-law, Nico Caprez, who became a vice president in January. His team was telling small cloud companies and data center builders that Nvidia could arrange credit support for them within a couple of weeks. The goal was simple: turn them into long term buyers of Nvidia GPUs. And the size of Nvidia's commitments has exploded. Its own filing shows $530 BILLION in commitments and guarantees, up from $184 billion just one quarter earlier. Inside the company, CFO Colette Kress was raising concerns. She kept asking how these guarantees were priced and what Nvidia was really getting back. In an August meeting, Jensen cut her off and called her "a deer in the headlights." She was also the one who cut Nvidia's guarantee on OpenAI's Ohio data center from $250 billion to $105 billion. Now Jensen is gone, and the CFO who questioned these deals is running Nvidia. To understand why this is a big deal, you need to understand the AI funding loop. Here is how it works: 1. Nvidia puts money into an AI company, or guarantees its loans and leases. 2. That company uses the money and credit to build a data center. 3. The data center is filled with Nvidia GPUs. 4. Nvidia reports those GPU sales as revenue. 5. If the company cannot rent out all its computing power, Nvidia agrees to buy the unused part. SO NVIDIA IS THE INVESTOR, THE LENDER, THE SUPPLIER AND THE BACKUP CUSTOMER, ALL IN THE SAME DEAL. The money goes out of Nvidia and comes back to Nvidia as sales. CoreWeave is the clearest example. Nvidia owns a stake in it, sells it chips, and agreed to buy up to $6.3 billion of its unused computing capacity through 2032. OpenAI is the biggest. Nvidia is guaranteeing up to $105 billion for its Ohio data center, supplying all the chips inside it, and invested $1.5 billion in the company building it. Nvidia's own filing for the quarter ending July 26 shows where the money is tied up: • Supply commitments for memory and chip wafers: $279 BILLION, up from $119 billion. • Guarantee on the OpenAI Ohio data center: up to $105 BILLION. • Deals to buy unused cloud capacity from neoclouds: $36 BILLION. Nvidia's actual liabilities on its balance sheet: $91 BILLION. That means the commitments outside the balance sheet are almost 6 times bigger than the debts on it. Nvidia does not lose this money unless something breaks. A guarantee only costs Nvidia if a customer cannot pay. But this is where it gets uncomfortable. The WSJ says Jensen himself had started to panic. He warned executives that the biggest cloud companies were moving to their own chips to depend less on Nvidia. At the same time, the wafers and memory Nvidia had committed to buy were coming at him like an "avalanche." So the biggest customers were trying to leave, Nvidia was locked into buying huge amounts of supply, and smaller customers were being kept alive with Nvidia's credit. And things were already starting to crack: • Nvidia paused some of its financing deals in August. • Nearly half of the $80 billion in data center bonds issued this year were rated junk. • Neocloud stocks crashed this week, with Nebius down over 20% and CoreWeave down 17% at the lows. • OpenAI raised less than it wanted, at a lower valuation. And now Kress is the one deciding whether this continues. If she pulls back on the financing, the companies that depended on Nvidia's money will have to borrow elsewhere, in a debt market that is already nervous. If they cannot, they stop buying chips. And if they stop buying chips, Nvidia's revenue drops. That is the risk of a loop. It works perfectly when money keeps flowing, and it hurts on the way down.

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.