Poolside's 40,000-Chip Cluster Slipped. NVIDIA Hired 109 Engineers and Licensed the Factory.

Poolside couldn't keep the 40,000-chip cluster it had reserved. By the time the company's six-week fundraising window closed, the reservation was gone. That missing cluster is the through-line for a deal Eric Newcomer first reported via a letter to investors: NVIDIA is licensing Poolside's training factory and hiring most of its engineering staff, in a structure Poolside's founders are calling "not an acquisition and not an acquihire."

The numbers, as the Latent.space write-up walks through them, are roughly $6 billion in a non-exclusive license for the training factory plus $1 billion in investment at a $12 billion pre-money valuation. Crypto Briefing's deal structure read confirms the split, and The Decoder puts the headcount moving over at 109 engineers, out of a technical staff that Poolside co-founder Eiso Kant has pegged at under 115.

The "reverse-execuhire" frame, which Latent.space is pushing and the founders are borrowing, is the deal's structural curiosity. In a normal acquihire, the rank-and-file stay with the buyer and the founders take the payout and leave. Here the rank-and-file leaves for the buyer (NVIDIA) and the founders stay with the original company. The kernel is a talent-plus-factory transfer, with the lab shell kept alive for a different purpose.

That purpose becomes legible only once the compute math is on the page. Poolside's stated production floor, per the same pod write-up, is that 10,000 to 20,000 GB300-class chips are enough to train a model that competes with the current frontier. The 40,000-chip cluster Poolside had reserved was already the floor, not the ceiling: the next training cycle, the one that would put Poolside at next year's frontier, requires more than an order of magnitude more than that. Poolside's pitch is that the binding constraint has shifted from capital to physical data-center space and contracted compute. The cluster it lost was the first concrete instance of the new constraint: the chips were available, the build-out and the power were not.

NVIDIA is the only buyer that can absorb a factory license plus a 109-person training team and wire both into a chip-supplier's roadmap. That is what makes this a template rather than a one-off: a sub-frontier lab at the $1 to $2 billion raise tier, attempting a training run that now requires a hyperscale cluster it cannot host, runs out of options before the run starts. The talent goes where the compute already lives, and the chipmaker is the only address that satisfies both halves of the constraint.

What the founders keep is the part of the company that is not a model factory. Poolside Infrastructure Company (PIC), which the Wikipedia entry for Poolside AI dates to a January 2026 spin-out, is the infraco shell. PIC's stated target is 7 gigawatts of neocloud capacity, a figure that should be read as a forward claim, not delivered build. PIC just hired Justin Lagasse as CFO, per the PR Newswire release, and the build-out is the second-order beat of this deal: the model-factory talent is now an NVIDIA team, and the GPU buyer is being repositioned as a power-and-build buyer. The Forbes read on the West Texas data-center project is the physical-infrastructure context for that build-out.

The 7-gigawatt number is also the falsifier for the pattern. If a sub-frontier lab at this tier can close a cluster on a normal six-week timeline, this template breaks. If, as in Poolside's case, the cluster slips and the next training cycle needs an order of magnitude more than what slipped, the template holds. PIC's role is to make the cluster not slip next time by owning the data-center build directly.

Poolside had under 70 people on the model itself and under 115 across engineering and research combined when the deal closed, and the write-up notes the company is still hiring across applied research, evals, and post-training. The next chipmaker-lab deal at the $1 to $2 billion raise tier will test the template.