Jason Warner spent Thursday writing to his shareholders about a company that would keep his name on the letterhead and lose nearly everyone who worked for him.
A License, Not a Buyout, the Letter Insists
Poolside's letter to investors is specific about what this deal is not: "not an acquisition and it is not an acquihire" (The Next Web, 2026). What Nvidia gets instead is a $6 billion non-exclusive license to Model Factory, the system Poolside built to train and evaluate its models, plus a separate $1 billion equity stake at a $12 billion pre-money valuation, according to a copy of the letter reviewed by The Wall Street Journal (Whelan, 2026).
Nvidia extends job offers to 109 of Poolside's engineers and researchers, close to the entire team chief executive Eiso Kant has said built the company's Laguna model (The Next Web, 2026). Poolside's three co-founders, Kant, Jason Warner, and chief operating officer Margarida Garcia, do not join Nvidia, and the company keeps operating independently (Whelan, 2026).
Poolside is roughly three years old, founded in 2023 by Kant and Warner, the former chief technology officer of GitHub. Its leadership pointed to a missed window to secure a 40,000-chip cluster and a failed 2025 fundraising effort as the reckoning that led, in their telling, to this arrangement (Whelan, 2026).
A Shape Nvidia Has Used Before
This is not the first Nvidia deal built to look like a license and a hire rather than a purchase. Reporting on Nvidia's earlier deal with chip startup Groq describes a similar shape: Nvidia took the company's top engineers while Groq brought in new leadership of its own (The Next Web, 2026). A license, a minority stake, and targeted hiring do not trigger the review a full acquisition would, and reporting on this deal has already framed it as a template other compute vendors may reach for next (Yahoo Finance, 2026).
This Model Skips the Laptop Entirely
Poolside trained Laguna on Nvidia server chips, and the team now works inside Nvidia on models built for the same scale (The Next Web, 2026). Nvidia's next full Nemotron release is rumored to carry more than a trillion training parameters, a scale that puts it among the largest models in the world, though still behind the biggest Chinese open-weight releases (Whelan, 2026). Nvidia wants an open-weight model built to compete with DeepSeek and Kimi K3, not a product sized for a browser tab on someone's laptop (Whelan, 2026).
Jensen Huang laid out the reasoning in July, in his first post on X, titled "Open Weights and American AI Leadership." His argument: the United States does not win the artificial intelligence race with one closed frontier model. It wins by building an open ecosystem that reaches every layer of enterprise and sovereign computing before Chinese open-weight models get there first (Whelan, 2026).
Nvidia Already Built the Hardware This Strategy Needs
Nvidia built the machine this deal needs before the deal existed. In May, at Computex, Jensen Huang introduced RTX Spark, a superchip pairing a 20-core Grace CPU with a Blackwell GPU. Nvidia's top-tier configuration carries up to 128 gigabytes of unified memory and a claimed 1 petaflop of FP4 performance, though the company has not said whether that figure holds at sustained power draw or only in a brief boost mode, and lower-priced SKUs cap memory at 64 gigabytes (Tech Insider, 2026). Six laptop makers, Asus, Dell, HP, Lenovo, Microsoft, and MSI, have eight confirmed RTX Spark laptop models launching this fall, with Acer and Gigabyte expected to follow (Kingy AI, 2026). Nvidia's own developer documentation states the chip runs 120 billion parameter language models locally with up to a million tokens of context (NVIDIA, 2026).
This is a consumer chip built to run the kind of open-weight model Poolside's engineers now work on, at a smaller scale than the trillion-parameter flagship.
A Ladder Nvidia Built in Pieces
Nvidia assembled this ladder over several months rather than announcing it as one plan. Nemotron 3 Nano, a 4 billion parameter model released in April, targets Jetson edge chips, DGX Spark, and RTX GPUs (NVIDIA, 2026). Nemotron 3.5 Lightning, a 30 billion parameter model with only 3 billion parameters active at inference, shipped in August and runs on RTX PCs, DGX Spark, and Jetson, scaling up to RTX Pro workstations and full data center GPUs (NVIDIA, 2026). Poolside's engineers now build the top rung, a flagship model too large for any of that hardware today.
Three Questions This Deal Leaves Open
Nvidia has not said whether the flagship built from Laguna and Poolside's engineering team ships as open weight like the rest of the Nemotron family, or moves behind a more restricted license. Nvidia has not published a distillation pipeline connecting that flagship to the Nano and Lightning models already shipping, so the ladder may reflect shared architecture rather than a planned handoff. And neither Nvidia nor Poolside has addressed whether a license-plus-hire structure built to avoid acquisition review invites a different kind of scrutiny once regulators notice the pattern repeating.
Whelan, Robbie. "Nvidia Strikes $6 Billion Deal to Take On China Heavyweights." The Wall Street Journal, 24 Aug. 2026, www.wsj.com.
"Nvidia Pays Poolside $6bn to License Its Model Factory and Hire 109 Staff." The Next Web, 2026, thenextweb.com.
"Nvidia's $7 Billion Poolside Deal Reveals a Licensing Playbook That Sidesteps Acquisition Scrutiny." Yahoo Finance, 2026, finance.yahoo.com.
"Nvidia RTX Spark: 1-Petaflop Chip Hits Intel, AMD." Tech Insider, 2026, tech-insider.org.
"NVIDIA RTX Spark: Specs, OEMs, Availability & Evidence." Kingy AI, 2026, kingy.ai.
NVIDIA. "NVIDIA and Local AI Community Fuel Open Source Models and Intelligent Agents." NVIDIA Blog, 2026, blogs.nvidia.com.
NVIDIA. "Nemotron 3 Nano 4B: A Compact Hybrid Model for Efficient Local AI." Hugging Face, 10 Apr. 2026, huggingface.co.
NVIDIA Developer. "Run AI Locally on NVIDIA." NVIDIA Developer, 2026, developer.nvidia.com.
