Two decentralized-AI founders in one week. One runs a verifiable-inference protocol with real on-chain proof-generation traction and a thoughtful approach to zkML. The other runs an AI agent-orchestration platform with a small but real revenue base and an enviable customer logo for an early-stage company. Categories differ. So do stages. And teams. And decks.
The same blind spot. I want to write down the question, because it changed both calls, and it's the question I'm now leading every decentralized-AI call with for the rest of Q2.
The Question
It came out of me almost by accident on the first call, and I asked it again deliberately on the second.
"If a foundation model release in the next quarter, Anthropic, OpenAI, Google, Meta, the next Chinese release, solves the specific problem your product is the wedge for, what does your network look like 12 months later?"
The first founder paused for what I'd guess was about eight seconds. I've gotten good at letting eight seconds go without filling them. Then he said, honestly, "I don't have a great answer for that."
The second founder, two days later, took longer. Maybe fifteen seconds. Then said, "I've been thinking about this question, but I don't have a credible answer that doesn't depend on betting on the regulatory side."
I want to be clear: both of those are good answers. Better than the polished, prepared, has-been-rehearsed-with-the-board answers I usually get when I ask hard questions. The honesty made the rest of both conversations more useful. Both founders are real. And both companies might compound. But the question they couldn't answer is the central question for almost every decentralized-AI thesis in 2026, and I don't think the broader investor base has internalised how exposed the whole category is to it.
Why The Question Is Hard
The hard part isn't that foundation models are improving. Everyone knows foundation models are improving. The hard part is the specific shape of the improvement curve.
Most decentralized-AI products in 2026 are wedges into a problem space that the current generation of foundation models can't quite solve well enough — verifiable execution, agentic workflows over financial primitives, certain kinds of multi-step reasoning, certain kinds of long-context retrieval, certain kinds of specialised inference. The wedge works because the centralised foundation models leave a gap, and the dAI product fills the gap with some combination of clever architecture, on-chain primitives, and crypto-economic incentives.
The risk isn't that the foundation models catch up gradually. The risk is that a single capability jump — a release that suddenly makes the specific gap the dAI product is solving a non-issue, collapses the wedge in a quarter. We saw a version of this happen in 2024 to several first-generation AI x Web3 plays whose wedges were closed by GPT-4 Turbo and the long-context Gemini releases. The cycle isn't done.
The question that neither of my two founders could answer is what their product looks like in the world where the gap they're solving closes. Their answers don't have to be air-tight. But the credible answers usually involve one of three structures:
Structure one: a regulatory or compliance moat. The product works specifically because something has to happen on-chain or with verifiability properties for regulatory reasons, and the centralised foundation models can't credibly provide that. Verifiable-inference work for regulated AI use cases, for example. zkML for audit and compliance flows. The second founder's "betting on the regulatory side" answer was an honest version of this.
Structure two: a real network effect that compounds on supply, not on model quality. The product gets better as more nodes, more agents, more data, or more participants join, independent of how good the underlying model is. A genuine decentralized-compute network has shades of this (the marketplace gets better as supply grows, regardless of foundation-model dynamics). The first founder was reaching for this and not quite landing it.
Structure three: a switching-cost moat. The customer has integrated the product so deeply that even a better foundation-model alternative carries real migration cost. This is the standard SaaS moat, and it's harder to build at the speed dAI products are launching.
If your decentralized-AI product doesn't have a credible answer rooted in at least one of those three structures, the next foundation-model release is an existential question, not a roadmap question. And there will be a next foundation-model release. There will be one this quarter, and the one after that.
What I'm Doing With This
The framing has changed how I run dAI diligence calls in 2026. Two things specifically.
First, the foundation-model question now comes early — within the first ten minutes of the first call. Before the demo, before the deck, before the tokenomics. Founders who have a credible answer can engage with the rest of the conversation knowing the hard question is on the table. Founders who don't get an honest "this is the question we have to answer" surface, which is more useful than dressing it up.
Second, I now build a one-line foundation-model-exposure tag into the diligence notes for every dAI company we engage with. It's a literal field: what specific capability gap is this product wedging into, and what's the credible answer to it closing? For companies where the answer is one of the three structures above, I write it down. For companies where the answer is "we'd have to pivot," I write that down too — and the position-sizing reflects it.
I should have been doing this in 2024. I wasn't. The portfolio has a couple of dAI bets where, if I'd asked the question more sharply at the entry point, I'd have written smaller cheques. I'd rather make the mistake on the other side now.
A Specific Honest Aside
I want to be careful about something. The question I'm describing, what happens to your wedge when the foundation models catch up?, is not "decentralized AI is dead." It's the opposite. The decentralized-AI companies that survive the next two foundation-model cycles are going to be very large companies, because they'll have built moats that don't depend on the model being any specific level of capable. The question is a filter, not a verdict.
Both of the founders I talked to last week are still in active diligence on our side. The honest answers they gave have made me more interested, not less. The polished answers are the ones that worry me.
The next Bridge Note will turn back to DeFi — specifically, why I think intent-based architectures have quietly won the orderflow war and what that does to the next generation of DEX design. If you're a decentralized-AI founder and the foundation-model question hits a nerve, I'd rather have the conversation about it than not. LinkedIn or Telegram.