Nvidia rented Poolside's team for $12B and left the company standing

Nvidia has licensed Poolside's Model Factory and hired 109 of the company's employees in a deal reported at $12 billion, and it did not buy the company ^[1]. It went from investor in Poolside to licensor of the factory and employer of the people who built it ^[2], which leaves the corporate entity intact and the engineering bench empty.

Less than a month before the deal, Poolside's Eiso Kant told the Latent Space podcast that fewer than 70 people had built the model and fewer than 115 across engineering and research had done the whole effort, a count he said generously included himself ^[15][7]. Take 109 out of a ceiling of 115 and at most six technical staff remain ^[8]. The founders describe the transaction as "not an acquisition and not an acquihire" ^[3].

Latent Space, which reported the deal, also supplies the taxonomy. It has been calling Windsurf-Google, Character-Google, Scale-Meta and Instacart-OpenAI execuhires: deals where the executives leave with the buyer while the employees take a payout and are left holding what remains of the company ^[4]. Poolside inverts the direction, with the founders staying and the staff going ^[5]. According to the report's headline, founders stay for $1 billion and employees go for $6 billion out of the $12 billion total ^[18], which works out to roughly $55 million per hired employee ^[19].

The stated reason is compute, and it reads literally. Poolside says it had a six-week window at the end of last year to raise $2 billion for a 40,000-GB300 cluster due to come online in January, did not close in time, and lost the cluster ^[10]. That implies about $50,000 per GPU of committed cost ^[16]. The company also says 10,000 to 20,000 GB300s would have produced a model rivalling the current frontier, but next year's frontier requires more than an order of magnitude more, and the binding constraint is no longer only capital but physical data center space and contracted compute ^[11]. At the same implied rate, the frontier-rivalling cluster was a $500 million to $1 billion line item ^[17], and the one after it is unreachable without a power contract.

For anyone whose plan depends on an independent code-model vendor, that is the arithmetic that matters. A team of 115 can build a competitive coding model ^[7]; a team of 115 cannot finance the cluster the next one needs ^[11]. When continuity rests on roughly a hundred named people rather than on a balance sheet, the vendor's roadmap is really a hiring decision at a larger company. Diligence should follow: who employs the researchers, what happens to your licence and your weights if they are hired en masse, and whether the counterparty keeps shipping when its shell outlives its staff ^[5].

Poolside's own parting position is the tell. It wrote that human-level capability will be fully commoditised by open source models while superintelligence likely will not ^[12], and that today's coding revenue is the intelligence-bound, low-margin kind, with durable value sitting in experiment-bound scientific discovery instead ^[13].

Watch PIC, the infrastructure company spun out of Poolside in January 2026, whose ambitions the report frames as scaling to 7GW of neocloud capacity ^[20]; watch the founders' new direction, which they say they are not ready to share ^[6]; and watch whether anything recognisably from the licensed Model Factory appears in Nvidia's own products. The dispatch covers 19-20 August 2026 ^[14].