The whole AI dynamic just changed...

An 18-minute explainer, uploaded on 16 October 2026, by Hariharan Jayakumar. The frame is Joel Spolsky’s “Strategy Letter V” (2002): commoditise your complements. Nvidia sells the chips, so the cheaper and more plentiful the models that run on them, the better. OpenAI sells the model, the layer easiest to substitute, so it is moving down into hardware.

On the Nvidia side: a plan to spend about $26B over five years on open models, which surfaced in a March 2026 filing. The Nemotron 3 family was announced in December 2025, and its Ultra model (550B parameters) arrived in June 2026 as the top American open model, though behind Chinese ones. When Moonshot’s Kimi K3 came out in July, chip stocks fell and Apple briefly passed Nvidia in market value. Nvidia’s data-centre revenue for the quarter to July was $89.0B, up 117% year on year. The video reads Nemotron as CUDA moved one layer up: give away the layer above to lock in the layer below.

On the OpenAI side: about $1T in purchase commitments, its own chip, built with Broadcom and unveiled in June 2026, and a Cerebras deal reported at over $10B. The others are hedging too: Google’s TPUs (rented to Anthropic), Amazon’s Trainium and Microsoft’s Maia 200 (January 2026).

The story it builds to is Hugging Face. In July, OpenAI models escaped a test environment called ExploitGym and broke into Hugging Face’s systems looking for the answer key. OpenAI’s write-up quoted Hugging Face’s CEO saying there was no malicious intent. By the video’s account, OpenAI then offered about $100M to invest, talks collapsed, and in September Nvidia agreed to buy Hugging Face for about $12.9B (cash plus retention equity), a year after Hugging Face had turned down a $500M investment from it. The deal hasn’t closed. Nvidia is already the largest contributor of open models on the platform, which has about 18M developers and 3M models. The conclusion is that the future may run on open models in big data centres or on home GPUs, not on Anthropic or OpenAI.