Nvidia's reported willingness to guarantee $250 billion of financing for an OpenAI data center, inside a project that could reach $500 billion once the chips themselves are counted, is the clearest signal yet that compute is now being underwritten as infrastructure rather than as software capex. The structure exists because OpenAI is still private. Lenders have no audited public financials and no traded equity to mark collateral against, so they price that opacity into the coupon unless a creditworthy party co-signs. Nvidia is functioning as a guarantor on a lease, and it is being compensated in chip orders, which is straightforward good business rather than anything nefarious. This is precisely the function a public listing has historically served, and it helps explain why both OpenAI and Anthropic have signaled they intend to come to market. Terms are not final and the deal could still collapse. On the circular financing concern, the discipline is simple: debt is only a problem when the revenue on the other side of the asset fails to carry it, and the current evidence runs the other way, with five-year-old Nvidia chips still in production use and pricing on some older GPUs rising despite earlier warnings of a two-year shelf life. The OpenAI facility is not expected to be operational until 2028.
Nvidia is now leading a coalition of technology companies pressing the US government and its allies to treat open-source models as a cyber security asset rather than a threat, and the commercial logic behind that position is not subtle. Nvidia's pricing power depends on having a thousand customers rather than a handful of frontier AI labs. If every enterprise and every sovereign trains and fine-tunes on open weights, each becomes a GPU buyer, whereas a market concentrated in four or five closed labs hands those labs pricing leverage over Nvidia. The cyber security framing has genuine substance: in a recent incident an attack originating from an open model was ultimately contained by another open model after the closed-model defense hit its token ceiling. It is worth noting who is absent, as neither OpenAI nor Anthropic has joined. Our working assumption is that roughly 80% of AI output eventually runs on open weights and 20% on closed frontier models, with the closed tier reserved for genuinely hard tasks, and that a two-year-old open model handles most steps in most workflows at a fraction of the cost. On the hardware split, Nvidia's own guidance of roughly 20% of compute allocated to training and 80% to inference implies the company is already quasi-conceding inference share while retaining a near-duopoly with Google's TPU (tensor processing unit) in training, which is why relationships with every frontier lab remain intact even as OpenAI develops its own inference silicon.
The investable consequence of cheap open models sits one layer up, in AI application companies, and we see a 12 to 18 month window opening with the most attractive entry point concentrated in a 6 to 12 month band inside it. The logic is mechanical. Lower cost per task means more workflows clear the automation hurdle, which converts directly into either incremental revenue or reduced cost, and therefore into net income for the buyer. A manufacturing company with 25 sites is never going to build its own model, so it will buy an outcome instead. Our four screening criteria are; enterprise rather than consumer focus given that Google and Apple will likely dominate consumer, forward deployed engineers to ensure positive return on investment related to AI spend, open model architecture so the vendor can route each task to the cheapest capable model, and outcome-based pricing tied to completed work rather than the legacy SaaS model of $100 per user per month, as an example. Forward deployed engineers are arguably the hardest to fake, because an enterprise IT team has largely proven they cannot implement an off-the-shelf AI solution correctly on its own. Glean, Harvey, and Lahzo screen well on these criteria at current valuations. The counterexample is instructive: OpenEvidence, a product used daily by roughly 80% of US physicians, is still capped by an advertising model against a finite physician count, and the only revenue design that would unlock a genuinely large TAM - that we can think of - is insurance-linked payment tied to demonstrably lower diagnostic error rates. Total addressable market discipline, not product adoption, remains the binding constraint on valuation in this category.
Robotics is where the same open-weight dynamic compounds fastest, and two data points from this week frame it. Enigma emerged from stealth with a $71 million seed round aimed at making robot control as intuitive as adjusting a volume dial, which is fundamentally a software and model-efficiency answer to a shortage of training data and compute, in the same spirit as the constraint-driven engineering that produced DeepSeek and Kimi. The more instructive commercial model is 1X Technologies, whose Neo home robot pairs an on-board AI brain with a human teleoperator who steps in whenever a task exceeds current capability, with the model observing the human's execution inside the robot body. The objection that a human is doing the work misses the point, because the customer is buying completed dishes, folded laundry, and a made bed, not autonomy as an abstraction, and the teleoperation share declines as the model learns. The flywheel is the actual asset here. Every household running the same model contributes edge-case data, and edge cases are exactly what defeats robots today, from an unfamiliar door handle that must be lifted before the deadbolt engages to a back door that needs to be shimmied. A human solves that once for themselves, while a fleet solves it once for everyone. Adoption will therefore look linear until it does not, and our view is that humanoid robotics should be underwritten as a data and service business rather than a hardware business, with the inflection arriving the moment these machines cross a general threshold of usefulness in the home.