What happened
36Kr reports that NVIDIA recently invested $1 billion in Poolside and agreed to pay $6 billion for a non-exclusive license to the startup’s model-factory technology. The report says NVIDIA also sought to bring most of Poolside’s engineering team into its orbit while leaving the company formally independent. It links the transaction to NVIDIA’s broader open-weight strategy, including its Nemotron models and partnerships with other AI companies.
36Kr’s English-language report, based on an authorized QbitAI article by Cheng Qian, says NVIDIA recently invested $1 billion in Poolside and committed a further $6 billion for a non-exclusive right to use Poolside’s “model factory” technology. The article says the transaction was finalized during the week of publication. It provides no contract, filing, joint statement, payment schedule, valuation, or direct confirmation from NVIDIA or Poolside, so the figures and structure are not independently confirmed here. Its “$40 billion” headline is unsupported by the body’s deal figures, which total $7 billion.
According to 36Kr, Poolside was founded in 2023 by software developer Eiso Kant and Jason Warner, a former GitHub chief technology officer. The report says Poolside typically trained a model in five to eight weeks, had a research-and-development team of no more than 70 people, and ran 10,000 to 20,000 experiments per month. 36Kr says this internal model-development system attracted NVIDIA. It further says Poolside’s remaining team would mainly comprise its two founders and an operations executive, without explaining the precise employment, licensing, or governance arrangements.
The report places the transaction within NVIDIA’s open-weight strategy. It says NVIDIA has released the Nemotron 3 series since late 2025 and publicly released model weights, training data, training recipes, post-training methods, and training software for GPU clusters. 36Kr says NVIDIA formed the eight-member Nemotron Alliance in March, including Mistral, Perplexity, Thinking Machines Lab, Cursor, LangChain, Reflection AI, Black Forest Labs, and Sarvam. NVIDIA contributes DGX Cloud computing resources, while alliance members contribute technology and data toward a future Nemotron 4 model; the report provides no independent technical evaluation.
36Kr describes NVIDIA’s open-model activity as broader than language models, citing Isaac GR00T for robotics, Cosmos for physical-world simulation, Alpamayo for autonomous driving, and Clara for biomedical applications. It says Poolside’s transaction is intended to strengthen NVIDIA’s ability to support open-source or open-weight model development while maintaining cooperation with closed-model companies such as OpenAI and Anthropic. The source does not establish whether these initiatives share technology, funding, personnel, or commercial terms with the Poolside arrangement.
Read the primary source: eu.36kr.com ↗
Why it matters
The reported transaction would give NVIDIA a direct stake in the systems and talent used to develop AI models, while preserving commercial demand for the company’s GPUs and data-center infrastructure. It also illustrates how AI infrastructure companies may support multiple competing model developers rather than rely on one winning model. The deal’s financial structure, licensing scope, and legal treatment remain unclear from the source.
The central significance of the reported deal is strategic rather than simply financial. 36Kr frames NVIDIA as seeking influence over model development while continuing to sell the infrastructure required to train and run models. If accurate, the arrangement would give NVIDIA access to a model-production system and experienced personnel without requiring it to choose one external model provider as the market winner. That could reinforce NVIDIA across competing open-weight and closed-source ecosystems.
The report connects this strategy to a structural change in AI development. It describes OpenAI, Anthropic, and DeepMind as focusing primarily on closed models, while DeepSeek and Kimi K3 represent a stronger open-model alternative. This is the article’s analysis, not an independently established market taxonomy. Its practical implication is that open-weight models may create additional compute demand if more organizations adapt, train, and deploy them, allowing NVIDIA to benefit even if no single model dominates.
The arrangement also matters for how AI companies acquire capabilities. 36Kr says the reported structure combines a non-exclusive technology license with recruitment of key personnel while leaving Poolside operating under its own name. The article says this is not a direct acquisition and therefore did not require antitrust review. That is a significant legal characterization, but the source gives no jurisdiction, regulatory record, or outside legal assessment. It should be treated as the report’s description, not a settled conclusion about competition law.
For open-source AI, the distinction between access and control is important. 36Kr presents NVIDIA’s Nemotron program as unusually open because it reportedly publishes weights, data, recipes, post-training methods, and training software. However, the source does not identify the applicable licenses, confirm that every component is reusable, or explain whether Poolside’s model-factory technology will have comparable terms. A company can support open-weight models while retaining control over infrastructure, tooling, data, or commercial distribution; those details will determine developers’ practical independence.
The reported transaction could affect competition and concentration. NVIDIA’s infrastructure is already central to AI development, and a large license or talent arrangement with a model builder could extend that influence into model production. The source says the license is non-exclusive, which could preserve Poolside’s ability to work with other parties. But the article offers no evidence about exclusivity in related services, data, personnel, or future models, leaving the competitive effect unresolved.
What to watch next
The key verification points are formal disclosures from NVIDIA or Poolside, the exact license and payment terms, the status of Poolside’s team, and whether the arrangement produces a new open-weight model or training system. Readers should also distinguish between models that publish weights and projects that publish training data, recipes, software, and permissive licensing. The source does not establish when any resulting system will be released or how it performs.
The first priority is primary confirmation. NVIDIA and Poolside disclosures would need to establish whether the reported $1 billion investment and $6 billion license are separate commitments, how the amounts are paid, what rights the license grants, and whether the transaction includes equity, services, warrants, or other obligations. The source provides none of those details and does not identify the relevant corporate or regulatory filings.
The second issue is Poolside’s operating status and personnel. 36Kr reports that the startup’s remaining team would mainly consist of three people after NVIDIA recruited more than 100 engineers, but it does not explain whether those people become NVIDIA employees, contractors, or members of a continuing Poolside entity. Confirmation would clarify whether this is primarily a licensing transaction, a talent acquisition, or a hybrid arrangement with different legal and commercial consequences.
The third issue is whether the arrangement produces a concrete, publicly usable AI system. The report says NVIDIA wants to use Poolside’s technology for future open-model development and links it to the next Nemotron generation, but gives no release date, model size, training dataset, benchmark, license, or availability plan. Any future announcement should be assessed for reproducibility and openness: published weights alone would not establish that the process, data, code, and downstream rights are open.
The source also says Poolside planned a 2-gigawatt data center in Texas, that a project partner withdrew in April, and that a subsequent $2 billion financing fell through. These claims concern the startup’s computing position but are not independently documented in the article. Further reporting should establish whether the data-center plan remains active, canceled, or replaced, and how much of Poolside’s reported model-development capacity depends on external cloud or NVIDIA infrastructure.
Finally, readers should watch NVIDIA’s broader open-weight portfolio and alliances. The article names Nemotron, Isaac GR00T, Cosmos, Alpamayo, Clara, and the Nemotron Alliance, but does not show that the Poolside deal changes their ownership, licensing, performance, or availability. Meaningful follow-up would be a verifiable product or model release, detailed licensing documentation, independent evaluations, or regulatory disclosures, rather than further predictions about which model ecosystem will prevail.
