Underwriting Foundation Model Platforms: An Interview with AI Infrastructure Investor Shaurya Mehta

Money is pouring into systems that promise better reasoning, faster work, and new products that did not exist a few years ago. In 2024, organizations were projected to spend $235 billion on AI, and that level of commitment has raised the standard for what investors need to prove before they underwrite a platform.

Shaurya Mehta invests in AI infrastructure using a process designed for noisy markets: map the ecosystem early, conduct disciplined diligence when an opportunity becomes real, and keep updating the thesis after the check clears. He also served as a judge for the Coding with AI Hackathon, which keeps him close to how builders actually experience tools when they have to ship.

When you start tracking a foundation model company before an investment is actionable, what are you actually trying to learn?
I am trying to learn whether the advantage is structural or temporary. Early on, I map the ecosystem to understand who is pushing the research frontier, what bottlenecks are emerging, and which signals keep getting stronger over time. In a market that moves fast, it is easy to confuse attention with traction. I look for indicators that compound: talent density, feedback loops between usage and improvement, and distribution that keeps widening even as competitors copy features.

That early work is also where you earn the right to move quickly later. If you have already done the hard thinking on why a platform should keep pulling ahead, you are not building conviction from scratch when a round opens. You are pressure testing a thesis you have been refining for months.

When the opportunity to invest became real, what did your diligence process focus on?
It has to integrate technical truth with financial truth. I lead diligence that combines management and customer conversations with an integrated operating and financial model. I evaluate core use cases across consumer and enterprise, then translate that into what the business needs to sustain leadership: compute requirements, cost structure, and the capital intensity required to stay at the frontier.

A major part of that work is getting specific about training and inference economics at scale. I build valuation scenarios, benchmark outcomes against public and private comparables, and stress-test what happens when assumptions move against you. The goal is not to produce a perfect forecast. It’s to understand what must be true for the upside case to hold and where the real downside risks actually sit.

AI is everywhere now. How do you separate real adoption from noise when you are underwriting a platform?
I look for repeat behavior and expanding use cases, not one viral moment. Broad usage matters because it changes what users expect and what competitors are forced to match. When 88% of respondents say their organizations are using AI regularly in at least one business function, the baseline has shifted. You are not underwriting a toy. You are underwriting infrastructure for work.

That is why I care about what the product becomes inside companies: how it shows up in workflows, whether it keeps earning trust, and whether reliability improves as usage grows. Adoption becomes meaningful when people stop talking about the tool and start depending on it.

After you invest, what does staying close actually look like in practice?
It means treating the thesis like a living document. After the investment, I track progress through product releases, research updates, competitive developments, and shifts in enterprise adoption. I refresh financial assumptions and update valuation and risk frameworks as new data comes in. The work does not end at signing, especially in a sector where the product and the market can change quickly.

Discipline matters here. If you only look at information that confirms what you already believe, you will miss the moment the story changes. I try to stay grounded by defining which signals would force an update, then checking those signals consistently.

You spend time with founders and other investors as well. How does that shape the way you evaluate opportunities?
It pushes you toward clarity. Rooms like the GGW Sharks founders and investors panel have a way of stripping a narrative down to what matters: what is the wedge, why now, and what stays defensible when a better funded competitor shows up. If a team cannot explain the system constraints and the path to scale in plain language, that usually shows up later as execution risk.
It also reinforces that the best diligence questions are practical. What breaks under load. What gets cheaper with scale. What customers would actually miss if the product disappeared tomorrow.

When you look ahead, what do you think matters most for investing in these platforms in 2026?
The winners will be the ones who can keep compounding without losing trust. That means better models, yes, but also the unglamorous parts: cost curves, reliability, and the ability to serve real workloads consistently as usage expands. In 2024, private investment in generative AI reached $33.9 billion, and that level of capital will keep raising expectations for proof, not promises.
I wrote about this in my HackerNoon piece, The Coming Shift: Why Multi Agent Systems Will Redefine Work in 2026. The thread is the same one I come back to as an investor: markets get excited by capability, but they stay loyal to systems that behave predictably, scale cleanly, and keep getting better after the headline cycle moves on.