Startups Are Treating Nvidia Compute as the First AI Bottleneck

For the AI startups Sarah Guo backs through Conviction, compute is not an abstract input. It is often the immediate constraint on what they can test and build. Ed Ludlow framed the practical question from repeated conversations with founders: when startups say compute is where they are most focused, are they accessing Nvidia hardware through cloud providers, or can they obtain direct control of clusters themselves?

Sarah Guo said the answer depends on where the company sits in the stack. Conviction’s portfolio spans infrastructure, models, and applications, so the companies do not all need the same kind of access at the same moment. But one of the first things Conviction did after starting the fund, she said, was buy compute for its companies because a venture fund could absorb timing risk that individual startups could not.

Bloomberg’s on-screen portfolio graphic listed Harvey, HeyGen, Cognition, baseten, and Sierra as Conviction companies. Guo said Conviction bought H100s at the time — “a bunch of nodes, cloud” — because the firm expected its startups would all need access and worried about shortages.

The pattern she described is consistent across company stages: experimentation begins on the current frontier. Today, Guo said, that means Nvidia chips. Frontier performance “allows you to do new things.” As companies mature, the work changes. They post-train smaller models, pay more attention to cost, and redesign experiences around the ability to use more tokens on a task. But the starting point, in her telling, remains current-generation chips.

That sequencing matters. Guo was not saying startups ignore cost. She was saying that cost optimization tends to come after proof of capability. First, founders want access to the highest-performing hardware so they can find what is newly possible. Then they compress, post-train, tune, and rework the economics and user experience.

Nvidia’s demand story is not only visible in its earnings commentary. Ludlow put Jensen Huang’s message to investors to Guo from the buyer’s side: Nvidia describes demand as vastly outpacing not only its own supply, but the industry’s aggregate ability to supply AI compute. Sarah Guo said that matches what she is seeing among founders and compute buyers.

She described roughly “two quarters of increasing stress” in the ecosystem around access to supply, at different scales. The shortage is not limited to large model labs. Guo said it is “very hard” to get on-demand, small-scale compute right now, which is exactly where many startups begin. At the other end, she said she has spent time with leaders of companies that serve Nvidia chips in the cloud, trying to buy $100 million of compute at a time.

What stood out to her was not just the amount, but the inversion of normal enterprise sales dynamics: she was trying hard to pay a large sum with a multi-year commitment, and still facing constrained access. She said she had “never been in that scenario before.”

Guo also argued that investors and observers may still be underestimating the demand signal Huang is giving them. Nvidia’s earnings outperformance is visible, and Huang repeatedly says demand is parabolic. Guo said she believes him on the demand side.

Her reasoning starts with long-horizon agents applied to code. She pointed to rapid revenue growth in products such as Claude Code as evidence that people are already using models more productively on long-horizon agent tasks. But code is only one function. “The world is not just code,” she said. Her phrasing on the scale was compressed: if there is “not zero, but small-scale, a few billion to 40 billion” of revenue in a small number of months because models are being used more productively on long-horizon agent tasks, she expects similar patterns across other functions in the knowledge economy.

The compute implication comes from how those workflows are used. Guo said companies can change the user experience simply by being able to use more tokens on the same task. Long-horizon agents, richer task execution, and more experimentation all increase the value of access to frontier hardware before startups begin optimizing for smaller models and lower cost.

The $26.5 trillion total addressable market number for AI, framed around enterprise opportunity in Ludlow’s question, put a much larger figure around the same demand thesis. Sarah Guo called the number and its visualization “funny,” noting the contrast of Starlink alongside Enterprise AI. But she did not dismiss the underlying claim.

Her view of enterprise AI follows a simple automation thesis she attributed to Andrej Karpathy: the clearest way to think about AI use is automation of tasks people already do. If models are given the right tools and harnesses, they can take over work. Because much of Conviction’s investing is in enterprise AI, Guo said she thinks the opportunity is “as big as Elon thinks it is.”

Ludlow pressed on whether Elon Musk truly believes that opportunity or is using it to justify what Ludlow described as “why SpaceX took xAI.” Guo answered by focusing on disclosure choice. What goes into an S-1 is a choice, she said. Musk could have framed the opportunity more broadly or emphasized other parts of SpaceX’s overall TAM. Instead, in her reading, the inclusion of enterprise AI signaled a commitment to make the company’s offerings relevant there.

Guo did not turn that into a settled forecast about every part of the strategy. Her narrower point was that the S-1 language was deliberate, and that she reads it as evidence Musk is committed to enterprise AI.

Asked whether Musk will pull it off, Sarah Guo separated the infrastructure question from the model and application questions. She said the central issue is whether value sits in infrastructure, models, or applications — a question she also sees in deals involving Anthropic and Cursor.

Her answer on infrastructure was the clearest. Guo said “they have the infrastructure,” and that the infrastructure and capability to build more are “extraordinarily valuable.” She said she thinks Musk will make money on that “no matter what.”

The open question, in Guo’s framing, is whether infrastructure ownership is enough. She put the remaining strategic decision this way: do they also need to own the model layer and the application layer? Scarcity can make infrastructure valuable, but Guo did not present that as the whole answer. She left the value split across infrastructure, models, and applications as the unresolved issue.