Nvidia's $6 Billion Deal for Poolside's Model Factory Is a Wager on Open-Source Models Driving Chip Demand

Nvidia (NASDAQ: NVDA) recently struck a $6 billion licensing agreement with AI start-up Poolside, extending job offers to more than 100 of the company's employees in a move reported by Bloomberg on Aug. 20. The deal landed just days after a note circulated from Goldman Sachs' One-Delta desk argued that "useful intelligence per dollar is exploding," making it harder for makers of large language model to defend premium pricing.

For a chief supplier of the silicon powering the AI labs, Poolside brings model-making capabilities, while the frontier labs face pressure from open-source competition. Nvidia sells compute capacity, and widely available, capable models drive broader AI adoption and usage, which fuels the heart of its business.

Image source: Getty Images.

At Vercel, a model routing platform popular with software developers, open-weight model share has gone from 33% of tokens to 54% in under two months.

The cost of intelligence is declining at all tiers

The day after the Nvidia-Poolside deal was reported, OpenAI cut the price of its frontier GPT-5.6 Sol model by more than 20%, while still offering two cheaper alternative tiers.

A couple of weeks ago, Space Exploration Technologies released Grok 4.6, which scores just behind the most powerful models on benchmark performance tests, but at a cost that's roughly 60% lower.

At the deep-discount level, Chinese lab DeepSeek recently upgraded its V4 Flash model, which now costs roughly $0.66 per million output tokens during its off-peak hours, a window that aligns well with American users. That's a fraction of the $50 per million output tokens charged by Anthropic's most expensive model. DeepSeek's V4 Flash model topped Vercel's AI gateway in weekly token volume in July, accounting for nearly a fifth of its traffic.

On the same day Nvidia's agreement was announced, a mystery model named Ox Alpha appeared on the OpenRouter and OpenCode platforms with free preview pricing and a 1 million-token context window. It quickly became the second-most-used model on OpenCode, handling over 5 million sessions in its first three days. No lab has claimed it, though speculation points to the model having come out of one of China's leading labs.

Nvidia's not stopping at Nemotron

With the Poolside deal, Nvidia gets access to coding models geared toward government and defense customers; it also gets the researchers who built them. The chipmaker continues to assemble the pieces that it will need in order to become an American open model contender. This follows the internal development of its open-weight Nemotron models, and an open letter that CEO Jensen Huang organized in July urging against "premature restrictions on open models." That letter was signed by the leadership of a host of key AI sector players.

Poolside's investor letter tells the seller's side of the story. The start-up missed a six-week window to raise $2 billion late last year and thereby lost its opportunity to lease a 40,000-chip computing cluster that was set to come online in January. Management said the company needed that capacity to keep developing frontier AI models. Its response was to sell distribution rights to Nvidia.

This news comes just as the largest AI labs are preparing to go public. I think the move is smart positioning, giving Nvidia, at the very least, more leverage with its largest customers.

Nvidia benefits when AI usage spreads, and cheaper intelligence is a powerful way to make that happen. As the model wars heat up, I imagine Huang is content to let the labs duke it out while his team continues to develop internally. We'll learn more about its plans when the company reports second-quarter results on Wednesday, Aug. 26.

Bryan White has no position in any of the stocks mentioned. The Motley Fool has positions in and recommends Goldman Sachs Group and Nvidia. The Motley Fool has a disclosure policy.