Watch where the venture money goes, not where the hype is loudest. The funding rounds landing this week tell a clear story: investors are cooling on companies that merely use AI and warming to the ones that decide how AI gets trained, governed, deployed, trusted and wired into the physical economy. The froth is moving from the shop window to the plumbing.
The evidence this week
The week’s rounds skew infrastructural. SciFin emerged from stealth with a $44m seed, AIR disclosed $50m across two seed financings, and Empirik launched with $21m, while the big names, Andreessen Horowitz, General Catalyst and Sequoia, keep circling the layer beneath the apps. The common thread is not a clever chatbot; it is the machinery that makes AI work, safe and connected.
Zoom out and the direction is unmistakable. Nvidia’s data-centre revenue alone ran to $89bn in a single quarter, and a single lab, Anthropic, has committed tens of billions to compute (around $30bn to Azure plus a $35bn Nvidia-backed buildout in Texas). That is the demand signal the infrastructure bets are chasing: the spending is concentrated in the picks-and-shovels layer, and investors are following it down the stack.
Why the shift makes sense
A thin wrapper around someone else’s model is easy to build and easy to copy, which makes it a poor bet once the novelty fades. The durable value is in the hard, unglamorous layers: training data, governance, security, and the connective tissue between AI and real-world systems. Investors have watched a hundred “ChatGPT for X” startups get flattened the moment the underlying model added the feature for free, and they have learned the lesson.
The read for builders and investors (not investment advice)
Bull: if you build or back the infrastructure layer, the wind is at your back and the buyers have deep pockets. Governance, security, eval, orchestration and data tooling are all being pulled along by the hyperscalers’ and labs’ enormous capex, and that spend is contracted years ahead.
Bear: “infrastructure” is not automatically safe. Much of the picks-and-shovels layer is itself richly valued and dependent on the same circular AI capex holding up. If the labs’ spending slows, the plumbing suppliers feel it fast, and a lot of today’s infra darlings are priced for a buildout that has to keep compounding.
Neutral: the useful signal is durability, not the label. Ask whether a company owns something a model maker cannot trivially absorb: real proprietary data, a workflow it controls, a distribution advantage, or genuine switching costs. If the answer is no, “infrastructure” is just a nicer word for another wrapper.
What this means
If your plan is an app that wraps a frontier model, the bar just went up: you need something the model maker cannot copy for free. If you are building or backing the boring infrastructure, the money is flowing your way, but check it is not simply another leveraged bet on the same AI capex cycle. Either way, “we use AI” has stopped being a pitch. It is table stakes, and the money knows it. (None of this is investment advice; do your own research.)
Related on Top Tool Stack: Anthropic’s $35bn Compute Bet · Nvidia’s Fine, the Market Sold Off Anyway
Did you know: “ChatGPT for X” was the hottest pitch of 2024. Two years on it is closer to a red flag, because the fastest way to kill your startup is to build a feature the model maker can add for free.