A founder I respect a great deal pitched me his second company last week. It's an AI x Web3 play. I asked him which of the three markets he was actually in. He looked at me like I'd asked the wrong question. Two years ago I might have agreed with him. In 2026 I think it's the only question that matters.
The "AI x Web3" thesis as a single trade, the one a16z, Pantera, Multicoin, and most of the credible Web3 funds (including, partly, ours at PRIM3) underwrote between 2023 and 2025, was useful when nothing was working yet. It bundled three completely different products under one narrative, raised capital against the narrative, and bought time for the underlying categories to figure out which of them was real. We're now far enough in that the bundle isn't doing useful work anymore. The three markets have separated. They have different unit economics, different customer profiles, and different timelines. Only one of them is compounding the way the headline implies, and, for what it's worth, it isn't the one I'd have predicted when I started funding this space.
Market One: Decentralized Compute
The first market is the one that sounded least exciting in 2023 and looks most defensible in 2026. Decentralized GPU and inference infrastructure, Bittensor, io.net, Aethir, Render, Akash, the second-tier wave coming in behind them. Plus, the AI-native infrastructure work I'm closer to as Investment Director at ChainGPT Labs (full disclosure on that role, I'm conflicted on the category).
The unit-economics story is straightforward. Hyperscaler GPU capacity is structurally rationed and structurally priced above marginal cost. There are a lot of GPUs in the world that aren't owned by hyperscalers. Aggregating them, scheduling jobs against them, and delivering enterprise-credible inference at a 30–60% discount to AWS/Azure/GCP is a real product with real demand from a real customer set: applied-AI teams whose burn rate is dominated by inference costs, not training costs.
What's compounded in this market between mid-2025 and early 2026 is the customer-side story. Eighteen months ago, decentralized-compute networks were primarily serving crypto-native users, model training for token-launch projects, inference for on-chain bot traffic. In 2026 the customer mix has shifted. I've now seen unit-economics-checked numbers from three different teams in this category that confirm: a non-trivial portion of their inference volume comes from off-chain AI companies with no crypto exposure, who got there through the cost arbitrage, not the ideology.
That's the bridge. When customers show up for the unit economics rather than the narrative, the category is real.
Market Two: The On-Chain Agent Economy
The second market is the one everyone wants to talk about and the one I'm most cautious on in 2026: autonomous AI agents transacting on-chain. The pitch is that as agentic AI matures, agents will need wallets, agents will need to settle micropayments, agents will need to coordinate without a human in the loop, and Web3 rails are the natural settlement layer for that economy.
I want this to be true. Most of the people I respect in this category want this to be true. The early experiments are real and the engineering work is impressive. The numbers, however, are tiny. The on-chain transaction volume attributable to genuinely autonomous, customer-driving agents in early 2026, separating out human-orchestrated bot traffic, MEV, and protocol-internal agent activity — is, by my read, in the low single-digit millions of dollars a month, not the hundreds of millions the narrative implies.
That's not nothing. It's the same shape as DeFi in 2018. It might compound. But the gap between the narrative and the unit economics is wider here than in either of the other two markets, and the time-to-real-revenue is longer. When founders pitch us on agent-economy plays in 2026, the question I now ask is what their customer-paid revenue line looks like if you strip out crypto-internal transactions. The honest answers are mostly small. The dishonest ones are the ones that quote "agent-attributable on-chain volume" without that subtraction.
I'll be wrong about this if the agentic-AI breakout in the foundation-model layer comes a year earlier than I'm pricing. It might. It also might not. The bridge here is the most uncertain.
Market Three: AI-Native Consumer Crypto
The third market is the one we've been quietly building inside at PRIM3 through ChainGPT and a few other portfolio bets, and it's the one I think most outside-the-space VCs misread.
This is AI as a consumer interface to crypto. Wallets that compose transactions from natural language. Trading copilots that explain on-chain positions to non-technical users. Compliance and tax tools that read your full on-chain history and produce something a human accountant can sign off on. Personalised research and screening over the universe of tokens, protocols, and on-chain businesses. The work I've been involved with at ChainGPT sits squarely here.
The TAM math is straightforward: there are roughly 200–300M people globally who hold some crypto, depending on which source you trust (the Triple-A 2025 report puts global ownership at ~580M, but the "actively transacting" cohort is smaller). Almost all of them are bottlenecked by interface complexity. AI is the obvious wedge for unlocking the next layer of users. The product surface is also one of the few places where a credible product can be built without billions of training-compute dollars, fine-tuning over open models plus protocol-specific context is enough for most of the work.
The bridge here is less technological than it is regulatory and trust-related. An AI agent that composes transactions on your behalf is a fiduciary surface. The teams that win this market will be the ones that handle the fiduciary question credibly — KYC, custody, transaction-approval, transparency about what the AI is doing on your behalf, and not just the ones that ship the slickest natural-language interface.
So Which Of The Three Will Compound
If I had to put a single ranking on it for 2026–2028: compute first, consumer second, agents third. Compute because the unit economics already work and the customer mix is broadening. Consumer because the wedge is real and the TAM is unambiguous, even if the regulatory bridge isn't fully built. Agents because the engineering is impressive but the customer-paid revenue line is still small and the time-to-meaningful-scale looks longer than the funding cycle currently pricing it.
The reason this matters for founders: pitching "AI x Web3" as one thing in 2026 will get you a smaller cheque than pitching one of the three markets clearly. The reason it matters for allocators: an LP looking at an AI x Web3 fund's portfolio should now ask which of the three markets the GP is overweight to — not because there's a single right answer, but because the answers tell you which version of the thesis they're actually underwriting.
When I sit on AI x Web3 panels — and I'll be on a couple in Q2, including one I wrote about for Cointelegraph recently — the conversation that lands is the one that admits the three-market separation has already happened. The conversation that doesn't land is the one that still treats AI x Web3 as a single trade.
The Forbes-Quotable Line
If I had to compress it to one sentence: "AI x Web3" was a useful narrative when nothing was working; in 2026 it's three different markets with three different timelines, and the founders who don't pick one are getting outcompeted by the ones who do.
The next Bridge Notes will go deeper on the DeFi side — specifically the restaking second act, which is the part of the DeFi 2026 story I think the largest funds are most underwater on. If you're building in any of the three AI x Web3 markets I've described, or you think I've miscounted to two or four, push back: LinkedIn or Telegram.