The company that sells the shovels just showed up at the dig site with its own crew. Nvidia — the chip maker that profits from every AI gold rush — is reportedly paying $6 billion to license Poolside's "Model Factory" software, making job offers to 109 of its engineers, and investing another $1 billion at a $12 billion pre-money valuation. The target: a trillion-parameter open-weight model that can challenge DeepSeek and Kimi on one side and OpenAI and Anthropic on the other.
The deal was first reported by Newcomer based on a letter Poolside sent to its investors, then corroborated by The Information, Bloomberg and the Wall Street Journal. It is not yet official, but every credible outlet treating it as done.
What Nvidia Actually Bought
Strip away the headline and the deal has three parts. First, Nvidia gets a non-exclusive license to Model Factory — the software pipeline Poolside built to train its open-source Laguna coding models. Second, Nvidia offers jobs to 109 Poolside employees, the engineers and researchers who built Laguna, who are expected to join the Nemotron project, Nvidia's own open-weight model line. Third, Nvidia invests $1 billion in what remains of Poolside at a $12 billion pre-money valuation.
The three co-founders — Eiso Kant, Jason Warner (ex-GitHub CTO), and COO Margarida Garcia — stay. Poolside continues to operate independently, and the license is non-exclusive, so the startup can keep building.
Why Poolside? Its Laguna models were positioned as the West's answer to DeepSeek and Qwen in open coding models. The latest Laguna S is a 118-billion-parameter mixture-of-experts model with roughly 8 billion active parameters per token — genuinely competitive, trained entirely on Nvidia hardware. It is one of the few Western open model programs still shipping on a regular cadence, which matters for the bigger story below.
The Soft Acquisition Playbook
The structure is the story. Nvidia did not buy Poolside. It licensed the technology, hired the team, and invested in what remained — a "soft acquisition" that secures the IP and the talent while keeping the company legally independent.
This is the third time Nvidia has run this exact play, after Groq and Enfabrica. The pattern is now a recognizable playbook: pay a large license fee for the crown-jewel software, absorb the core engineers, take a minority stake in the shell, and skip a formal merger review. For a company of Nvidia's size, avoiding antitrust scrutiny is not a neutral side effect of the structure — it is often the whole point.
Think of it as renting the recipe and hiring the chef instead of buying the restaurant. You get the food, the kitchen, and the staff, without the regulatory bill and without the messy ownership consolidation.
Why Nvidia Needs to Own a Model
For years Nvidia's Nemotron line was a demo: open models built to show off GPU performance rather than to win. The Poolside move marks a deliberate escalation to a strategic position.
The stated rationale, per the WSJ, is geopolitical. Nvidia believes the United States has fallen behind China in open AI because American labs over-invested in proprietary models. Chinese labs — DeepSeek, Kimi (Moonshot), Alibaba's Qwen — have shipped frontier-class open-weight models on a steady cadence. Western labs went roughly 11 months without shipping an open-weight model in that parameter class. The Poolside deal is Nvidia's answer: take ownership of the most credible Western open-model pipeline and scale it to trillion-parameter scale.
The deeper signal is structural. The infrastructure monopolist is entering the model layer, which means it is now competing directly with its own biggest customers. Nvidia already halved a $250 billion commitment to OpenAI this month; it has invested in dozens of model companies and is now building its own frontier model to undercut them. The "shovel seller" logic — stay neutral, sell to everyone — is quietly being replaced by a vertically integrated one.
And Nvidia holds one card no rival model lab can match: compute. It owns the most advanced, most abundant training infrastructure on earth. If the hardware margin subsidizes the model business, no independent lab can compete on training cost. Poolside's own near-death experience proves the point — it lost a 40,000-GB300 cluster because it could not raise $2 billion in a six-week window. Nvidia has no such constraint.
What Changes in the Competitive Landscape
If Nemotron 4 ships at a trillion-plus parameters, three things follow.
First, the open-weight race becomes a two-front war. The US now has a state-of-the-art-backed open champion to match DeepSeek and Kimi, while also pressuring OpenAI and Anthropic from below. Open-weight models are far cheaper to operate and trivially customizable — the economics that made DeepSeek a pricing shock now arrive with Nvidia's distribution and capital behind them.
Second, frontier pricing pressure returns. The frontier pricing war that started with OpenAI's GPT-5.6 price cut assumed a closed-model oligopoly. An open model at the same capability tier, subsidized by hardware margins, resets the ceiling on what any proprietary API can charge.
Third, the neutral-supplier era ends. Nvidia now has skin in the model game. Enterprises that bought GPUs while licensing OpenAI or Anthropic models must watch where the conflict of interest lands: the company selling the chips now also sells the open model that runs on them.
What Developers and Enterprises Should Do
Start treating open-weight models as a first-class deployment option again. Benchmark the current Laguna S and the Nemotron line against your coding and agent workloads — the gap that justified premium API pricing is narrowing, and the TCO math for self-hosting improves every release cycle.
Re-audit your supply-chain assumptions. A vertically integrated Nvidia means the neutral middle ground you relied on — chips from one vendor, models from another — is getting thinner. Model a scenario where your inference runs on open weights, on hardware you control, and keep that exit open even if you never take it.
And watch the Nemotron 4 release cadence. If a trillion-parameter open model ships within the next two quarters, the economics of every API-first application you run change. The gold rush is not just about who digs — it is about who owns the pickaxe, the recipe, and the crew. Nvidia just acquired all three.
Frequently Asked Questions
Is Nvidia acquiring Poolside?
No. Nvidia is licensing Poolside's Model Factory technology (non-exclusive), offering jobs to 109 employees, and investing $1 billion at a $12 billion pre-money valuation. The three co-founders stay and Poolside remains independent.
How much is Nvidia paying?
$6 billion for the license plus $1 billion in new investment at a $12 billion pre-money valuation — about $7 billion in total commitments, per reports.
What model will Nvidia build?
Larger, more capable versions of its open-weight Nemotron line, with reports pointing to a trillion-plus parameter flagship (Nemotron 4).
Why is Nvidia doing this?
Per the WSJ, Nvidia believes the US has fallen behind China in open AI because American labs focused on proprietary models. The deal aims to close that gap and compete with DeepSeek, Kimi and Qwen — while also challenging OpenAI and Anthropic, since open-weight models are cheaper to run and easier to customize.
How does this affect Nvidia's customers?
Nvidia now competes with some of its biggest customers (OpenAI, Anthropic, Microsoft, Oracle). It already halved a $250 billion commitment to OpenAI this month. Enterprises should expect a thinner neutral middle ground between chip vendor and model provider.
Who is most threatened by Nvidia's move?
On the open-weight side, DeepSeek, Kimi and Qwen now face a US-backed challenger backed by the world's largest compute fleet; on the closed side, OpenAI and Anthropic face fresh pricing pressure, because open models are cheaper to run and easier to customize.
