Newcomer reported a non-exclusive $6 billion Poolside licensing deal with Nvidia and a separate $1 billion investment at a $12 billion pre-money valuation. The amounts describe different things and should not be combined into a $19 billion purchase price.
What is public about the reported terms
The public introduction to Newcomer's August 20 report attributes those terms to a letter obtained from investors. The full report is subscription-only. PeopleAreGeek has not reviewed the underlying agreement, so this article distinguishes the published terms from assumptions about undisclosed rights or completed employee transfers.
A non-exclusive license is not, by itself, acquisition of the whole company. Nor does the amount of the fee establish which source code, data or future improvements another customer can obtain. The contract would be needed to answer those questions.
The three amounts use different denominators
The $6 billion figure concerns licensing consideration. The $1 billion figure concerns a separate investment. The $12 billion figure is a reported valuation before that investment; it is not another cash payment. Adding all three counts unlike quantities as though they were identical.
For an original simplified financing calculation, $12 billion pre-money plus $1 billion of new equity gives $13 billion post-money. Dividing the new investment by that total gives roughly 7.7%. This is illustrative arithmetic, not a verified ownership stake: the actual capitalization, share classes and transaction terms are not established by the public introduction. The license fee does not belong in the equity denominator merely because it appears in the same story.
Why the Model Factory is more than a checkpoint
Poolside's own Model Factory explanation describes a system for training, evaluating and iterating on models. It connects data preparation, evaluation, architecture experiments and reinforcement learning from code execution. That technical description helps explain the kind of capability discussed; it does not reveal the licensing contract's precise scope.
A checkpoint records model parameters at one stage. Reproducing an experiment also requires knowing the data version, settings, evaluation conditions and relevant software. Possessing one resulting file does not automatically provide the process that produced it.
A useful lesson for internal model work
For a small internal training project, record the lineage from input data through a run to its evaluation result. Suppose run A and run B use different data filters and B scores better. Without a versioned record of those filters, a later team cannot reliably investigate the difference or reproduce the comparison.
That is an original documentation example, not evidence of Poolside's measured productivity. The reported deal is a reason to examine how model-building systems are valued, not proof that future models will be faster, freely licensed or portable across every accelerator.
Attribute the public report, keep license fee separate from equity and valuation, avoid asserting accepted job offers or undisclosed contractual rights.