What VCs are getting wrong about AI infrastructure bets

The AI infrastructure funding cycle has produced some genuine category winners and a long tail of companies that seemed essential in 2023 and look increasingly optional today. The pattern of what went wrong is consistent enough to be worth examining, because the next wave of infrastructure bets is already being made, and many of them are following the same logic.

The layer cake problem

AI infrastructure investing tends to get organised around layers: compute, foundation models, orchestration, tooling, applications. The implicit assumption is that each layer is durable and that winning a layer means building a lasting business. This assumption has held at the compute and foundation model levels. Below that, it has been far less reliable.

The problem is that the foundation model providers don't see the layer cake the same way investors do. They see a product they want to make as capable as possible, which means absorbing capabilities from adjacent layers whenever it makes sense. Vector databases, prompt management, evaluation tooling, agent orchestration: all of these were independent categories that are now features, either native to the models or offered directly by the providers.

Investing in a layer that a foundation model provider can absorb in a product update is not infrastructure investing. It's feature investing, and features don't build durable businesses.

What actually creates durable infrastructure value

1. Provider neutrality at scale

The infrastructure companies holding their value are the ones that sit above any individual model provider and get more valuable as the number of providers increases. Tools that help enterprises manage multi-model deployments, switch between providers, or evaluate outputs across models are genuinely useful in a way that doesn't depend on any single provider's roadmap. The switching cost works in their favour.

2. Compliance and governance as a wedge

Enterprise AI adoption is now substantially gated by legal and compliance requirements that model providers are structurally poorly positioned to solve. Data residency, audit trails, content filtering tuned to specific regulatory regimes, access controls that integrate with existing enterprise identity systems: these are genuine infrastructure problems that aren't going away and that hyperscalers aren't focused on solving at the required level of specificity.

3. Domain-specific data pipelines

The most defensible infrastructure plays are the ones that own the data pipeline for a specific industry. Medical records, legal documents, financial filings, industrial sensor data: getting the ingestion, cleaning, and structuring right for these domains is hard, slow, and not something a general-purpose provider will do well. Companies that have built this for a specific vertical own something that compounds.

The question to ask before writing the cheque

For every AI infrastructure bet, one question cuts through most of the noise: would this still matter if OpenAI, Anthropic, and Google each added this capability natively in the next eighteen months? If the answer is no, you're not looking at infrastructure. You're looking at a company that's capitalising on a temporary gap in the model providers' roadmaps.

Temporary gaps are real businesses. Some of them are worth funding as such, with clear eyes about the timeline and the exit path. But they shouldn't be valued or underwritten as if they're building durable infrastructure. The ones that were, and that raised at infrastructure multiples on feature-level defensibility, are where most of the AI infrastructure write-downs of the next two years will come from.