Nvidia paid $7B for the parts of Poolside that a $2B venture round could not finance

Take the structure apart and the two halves are pricing different things. Six billion buys a non-exclusive license to the internal system Poolside used to build its models ^[1], and one billion buys equity at a twelve billion pre-money, which is 7.7 percent of the thirteen billion post ^[2][2], so Nvidia is paying six times as much for a right that excludes nobody as for a stake in the company that holds it ^[4], and the license on its own is about 46 percent of what Poolside's newest round says the entire business is worth ^[3]. The investor letter's insistence that this is neither an acquisition nor an acquihire is doing real work ^[4], because 109 engineering offers into the Nemotron effort ^[3] set against six billion of license money is 55 million a head if you insist on reading it the other way ^[7].

The raise that missed is the part worth sitting with. Two billion against 40,000 GB300s is fifty thousand dollars a unit ^[5], owed on a cluster scheduled to come online in January, inside a six-week window that closed before the money did ^[5]. What Poolside was asking a syndicate to underwrite was not equity risk in the usual sense but a prepayment on a delivery slot, repayable only if the next training run went well. The capital that eventually showed up was three and a half times the capital that did not ^[1], which locates the constraint in who can sign rather than in how much exists.

Harry Stebbings, a Poolside shareholder, offers the competing read: neo labs are out of favour, universally, across every investor he speaks to, and nobody was surprised Poolside struggled ^[8]. If that is the whole story, the market had a view on the asset and acted on it. This is probably wrong, but the more interesting version is that both things are true at once, and the discriminating variable is balance sheet: Rory O'Driscoll's count is that only four or five entities in the United States can finance a state-of-the-art frontier model, that no VC owns more than one or two percent of OpenAI or Anthropic, and that Nvidia has become the fifth member of that club ^[13].

The seed arithmetic is where the letter stops being a company story. Stebbings puts the outcome at roughly 15x for a seed position and calls it pretty great ^[9]; O'Driscoll points out that 15x on a nine billion outcome implies an effective entry of six hundred million rather than sixty, with the dilution of capital-intensive rounds accounting for the gap ^[11], and that 50x would have required an exit near sixty-three billion ^[12], seven times what actually happened ^[8]. Note also that the source uses a twelve billion pre-money and a nine billion outcome without reconciling them ^[9].

For Nvidia, seven billion is about ten percent of the roughly seventy billion of free cash flow the SaaStr write-up attributes to this year ^[17][6], and O'Driscoll's own test does not clear on all of it: funding a neocloud or an open-weight lab sells chips, funding Mercor at a twenty billion valuation buys training data, which does not ^[15][18]. A tenth of cash flow spent on ecosystem options is a tenth not spent on capacity or buybacks.

The thesis breaks if another open-weight lab closes two billion or more without a strategic anchor in the next two quarters. Then it was the asset, not the ceiling.