Source the 250 constraint-pushers shaping AI ecosystems

Corporate innovation teams still start their sourcing from sector directories and named lab brands. The result is a short list that looks the same as everyone else's. Sarah Guo, founder of Conviction and one of the sharper observers of the AI frontier, offers a different unit of analysis.

Quick answer: AI frontier ecosystem sourcing works better when you map the operators pushing real constraints—the roughly 250 researchers and founders Guo identifies as the actual engine of progress—rather than scanning sector tags or coastal hub rankings. The physical systems those people depend on (energy, permitting, compute supply chains) reveal which ecosystems are worth sourcing from and which are costume.

Guo told Patrick O'Shaughnessy on Invest Like the Best something that should land differently for a corporate sourcing team than it does for a typical VC audience: the frontier is not a category. It is "the people that are actually like showing up in the morning and pushing this whole thing forward"—roughly 250ish researchers and founders, by her estimate. That number is small enough to map. Most corporate teams are not mapping it.

The useful unit is who is pushing a constraint, not what sector they are in

Sector tags ("AI," "energy storage," "robotics") are how databases organize themselves. They are not how breakthroughs propagate. The people Guo describes are defined by the constraint they are attacking, not the label on the slide deck.

This has a direct equivalent in the sourcing work we cover in this series. The Waterloo co-op pipeline produces operators, not just graduates. The CMU Robotics Institute spins out founders who have spent years running physical hardware against real failure modes. AGH University in Kraków feeds manufacturing and materials talent into a cluster that is not well-indexed by any major startup database. The Network School selection process, which we examined closely, is selecting for a similar property: people who build something rather than people who describe building something.

The practical test Guo's framing implies is this: can you name the operators in a given node, or can you only name the institutions? If you can only name the institutions, you are one step removed from the real map. You are reading the brochure.

Directories are indexes, not diligence. They surface company names and funding rounds, but they do not tell you whether the team is still working the problem or whether the technical founder left eighteen months ago. The companies most worth finding are private, active, and off-market. They are often invisible to a standard database scan precisely because they have not needed to market themselves.

The bottleneck is physical, so the ecosystems that matter are where power, permits, and testbeds exist

Guo is direct on where the frontier actually stalls. It is not a technology or capability problem. "I think it is a regulatory problem and an alignment problem," she argues, meaning political and social alignment, not AI safety. The hyperscaler infrastructure lead she references says nothing moves the needle at sufficient scale before 2030. That constraint is not a software problem.

For corporate sourcing teams, this translates into action. If the real bottleneck is physical—permitting, gas interconnection, nuclear siting, or data-center construction tacit knowledge—then the ecosystems worth watching are the ones that have already negotiated some version of those constraints.

Austin and ERCOT sit at this intersection. We covered the Austin energy ecosystem in detail. The grid structure creates both genuine demand-side pressure and a testbed that companies in other markets cannot replicate cheaply. Laramie and the University of Wyoming present a quieter version of the same dynamic through a mountain-west energy cluster where geological and regulatory conditions are the actual product.

Japan is where this logic concentrates most clearly for a manufacturing or energy mandate right now. The GX program (Green Transformation) has committed ¥20 trillion to the transition, and the GX Acceleration Agency is making equity investments in deep-tech startups, often without disclosing amounts. That opacity makes the cohort nearly invisible to standard databases. Hokkaido's hydrogen infrastructure investment and the Kansai manufacturing cluster, covered as an industrial innovation laboratory, are both shaped by the same physical logic: the location has a grid, a port, a testing right, or an industrial feedstock that cannot be conjured in a WeWork.

The sourcing implication: filter first by "can this company actually operate here?" before filtering by sector. Companies that have solved the physical setup problem are worth more, and they are harder to find.

The company worth finding treats the missing input as the product

Guo highlights Sunday Robotics as a worked example of this pattern. The founders identified that the limiting input for robotics AI was cheap, varied, real-world data, not more simulation or more GPUs. So they built a data collection system rather than waiting for an internet-of-robot data to materialize on its own. Two young founders, she notes, have contributed "most of the interesting ideas in robotics AI over the last four years." The company went from cardboard in a Stanford basement to a manufactured full stack, with a home beta expected by the end of this year.

The pattern generalizes. NEDO-backed component makers in Japan are sometimes doing the same thing by identifying a missing input to a larger system and building the company around supplying it. Flow batteries designed for frequency regulation rather than bulk storage treat grid volatility as the product constraint, not a market condition to wait out. University Technology Transfer Office spinouts that are physically tied to a national testbed, a reactor, a port, or a manufacturing line are building from an input that outside competitors cannot simply buy.

This is also why the Journal of Corporate Finance research from Warwick University (Farida, Fidrmuc, and Zhang, 2026) is relevant. Acquisitions of private targets produce more patents and greater innovation synergies than acquisitions of public targets, and the outcome is specifically associated with acquirers' expertise in identifying innovative private targets. The company treating the missing input as the product is almost never on the first page of a sector search. It is off-market because it does not need to market. Finding it is a sourcing skill, not a browsing skill.

If your team is trying to build that sourcing skill systematically, the startup sourcing mandate template we published is a starting point for translating a vague strategic interest into a specific, checkable criteria set.

Open weights and mid-stack tools are how AI capability enters industrial operations

Frontier model APIs are too expensive, too slow, and too sensitive for many industrial applications. Guo is blunt about the open-source policy question: "if you did restrict use of open-source models in the United States, like you'd basically just restrict law-abiding American businesses." The cat is already out of the bag.

For corporate sourcing teams, this reframes what to look for. The companies worth finding in the industrial AI space are usually not building foundation models. They are doing domain post-training on a base model and selling the result to a buyer who has a plant, a port, a grid, or a lab. The input is proprietary operational data. The moat is knowing what to train on and how to validate it against a real process.

Guo's frame is that intelligence is becoming "too cheap to meter," and it will arrive through an ecosystem of businesses rather than from one lab's imagination. Find the business that has domain data and a real customer with a physical operation, not the business that has a better chatbot.

Compute independence is a multi-node supply chain, not a Bay Area story

Compute independence "looks a great deal like energy independence," according to Guo. Working backward from a live data center, you need power, chips, glass, labor, cooling, and financing. The thin sieves, the points where supply is genuinely constrained, exist outside the US. Comparative advantage in those components is real and durable.

Japan sits at multiple nodes in this supply chain. Kyushu's semiconductor ecosystem has TSMC's Kumamoto fab as its most visible example, but the supplier cluster around it matters more for sourcing purposes. Japan-to-North America materials and equipment flows are another node, a corridor we mapped in detail. The installer and builder layer, the companies that actually site, construct, and commission data centers and grid infrastructure, is a third. It is almost entirely absent from any VC-oriented startup database.

Guo's TSMC mug line makes a sourcing frame out of geopolitics. The point is that a thin sieve in one material or process step can determine who wins a supply chain. The company controlling a thin sieve is often small, private, and operating in a geography that coastal-hub-focused analysts do not cover. It is exactly where a corporate sourcing team with a physical mandate should be looking, and exactly where a standard database scan will miss.

What Guo's map means for how you actually source

The 250-person graph is useful precisely because it is small. A sourcing team that can work from a live map of constraint-pushers and trace which physical ecosystems those people depend on will build a pipeline that looks nothing like one built from a directory.

A directory-first approach gives you everyone who has registered a company, filed a patent, or raised a round that got into Crunchbase. A constraint-first approach gives you the teams actually running against the bottleneck your mandate cares about. The first list is long and stale. The second is short and worth acting on.

The practical steps are not complicated, but they require discipline:

Guo's bet and this blog's bet are the same: the frontier is a small set of operators plus the places that can actually host power, factories, ports, and test rights. The sourcing work is finding where those two things overlap.

If you want a starting point matched to your mandate, Innovation Scout takes a description of your goals and returns a short list of active, relevant companies. You can try it at chibit.io/scout.

FAQ

How do you find the 250 people Guo is talking about, practically?

The 250-person map is not a published list. It is Guo's shorthand for a density of researchers and founders whose work is genuinely pushing the frontier forward. Practically, you trace it through lab affiliations, paper co-authorship, spinout networks (CMU, Stanford, ETH, Tokyo Tech, KAIST), and the programs that select for operators rather than pitch decks, like NEDO's J-Startup cohort or Network School. The institutions are a starting point, not the answer.

Is this approach relevant for corporate M&A teams, or just VCs?

Corporate M&A teams arguably need the operator map more than VCs do, because they are sourcing for a specific mandate rather than portfolio diversification. A head of corporate development with a "find energy storage targets in East Asia" mandate cannot use a 3,000-company database efficiently. Starting from the constraint and the operators attacking it narrows the field to a manageable short list before any diligence starts.

Why do physical ecosystems matter for AI sourcing specifically?

AI infrastructure requires power, cooling, land, and labor at a scale that is not location-agnostic. Training runs and inference clusters are constrained by grid capacity and interconnection timelines, not just by chip availability. Companies building AI-adjacent products for industrial applications also need access to real operational environments, a port, a plant, or a grid node, to generate the training data that makes their product defensible. Physical ecosystem quality is therefore a proxy for how far along a company actually is.

How is sourcing from a constraint-first map different from using a startup database?

A startup database answers "who has registered, raised, or filed?" A constraint-first map answers "who is actively working the problem your mandate requires?" The difference shows up in two ways: the companies on the constraint-first list are more likely to be genuinely active, and they are more likely to be off-market, meaning they have not been pre-approached by every other corporate development team running the same database scan.

What makes a physical ecosystem worth sourcing from versus one that is just marketing itself well?

The 4-signal scorecard for evaluating innovation districts covers this in detail. The short version: measure pilots, not vibes. A real sourcing-grade ecosystem has testbed access, anchor institution output that includes spinouts (not just graduates), visible policy that resolves a physical constraint (a permitting pathway, a grid connection program, a procurement framework), and companies that have shipped to a real customer in that environment. An ecosystem that has a brand but no pilots is not on your short list.

Quotes from Sarah Guo are drawn from her appearance on Invest Like the Best, "She Knows the 250 People Building AI. Here's What They Actually Believe," interviewed by Patrick O'Shaughnessy. Quotes are from the transcript and lightly cleaned for readability.