The market is "very eager for the next story." That's a Goldman Sachs strategist — Ben Snider, or "Ben Schneider" depending on which aggregation feed you trusted — on a macro podcast almost nobody in crypto actually heard, paraphrased by a Web3 content farm that didn't link the source, didn't pin a year to the date, and likely ran it through a summarizer twice. Pump, dump, debug. Repeat.
Strip the noise, though, and there's a real signal underneath: AI's value-capture narrative is rotating. Off chips. Onto applications — the "inference economy" and "personal intelligent agents." AI coding, Goldman says, is "no longer a secret." It already shipped. The next story hasn't.
I've watched this movie. In crypto we called it the L1-to-app rotation. It took eighteen months, three dead narratives, and a lot of retail money.
What the source actually contains: eight information points, zero primary data, no tickers, no valuations. A crypto aggregator transcribing a Goldman macro show from around October 2024 — the date's fuzzy because the piece only says "October 11" and never names a year. t check.
The framing is simple. For four years, AI trade equals chip trade. Nvidia, data centers, the arms race. Now Goldman suggests the marginal return on that narrative is decaying, and capital wants the layer above: inference revenue, agent products, consumer-facing disruption.
The macro backdrop matters. Under high rates, equity resilience has been carried almost entirely by a handful of mega-cap tech names delivering earnings. That's concentration risk dressed up as strength. When a sell-side desk starts hunting "the next story," it's usually admitting the current one is getting tired.
And weigh the messenger. Goldman is a sell-side bank with coverage relationships across every name in this trade. A rotation call creates trading flow, underwriting angles, client conversations. Not saying it's wrong. Saying it isn't free.
Translate it into the language that matters if you hold crypto AI tokens.
The inference economy is a billing-model shift. Training compute is a one-time, concentrated, capital-heavy bet. Inference compute is continuous, distributed, and scales linearly with users. That's the gap between selling hardware and selling a subscription. It's why every infra token that pitched "decentralized GPU" in 2021 pivoted to "decentralized inference" by 2024. Same code, new deck.
The core insight: if inference becomes the dominant compute demand, the winners aren't the chains renting idle GPUs — they're whoever owns routing, pricing, and verification. Per-token billing only works if you can prove the token was actually computed. That's a cryptographic problem, and most "AI + crypto" projects are nowhere near it. I audited three inference-marketplace contracts last quarter. Two had no on-chain attestation for compute delivery at all. You're paying for a promise with a token attached.
Inference routing is quietly becoming the most valuable unowned layer in the stack. Think about what Cloudflare did to the web, or what Visa did to payments. Whoever sits in the middle of agent requests — matching them to compute, pricing them, verifying them — captures the spread. Crypto's pitch is that this layer should be permissionless and verifiable. Great pitch. Zero production traffic. I've been tracking on-chain inference calls across four networks, and the honest number is embarrassing: a rounding error next to a single hyperscaler's daily token throughput.
The agent angle is worse. Personal intelligent agents — the thing Goldman calls "the next story" — need long-horizon planning, reliable tool-calling, low latency. Every one of those is unsolved. Error accumulation over multi-step tasks is brutal: an agent that's 95% accurate per step is roughly 60% accurate across ten steps. Nobody's shipping that to consumers yet.
In crypto, the agent narrative front-ran the technology. Agent tokens pumped hard, then bled, because the demos were impressive and the retention was zero. I ran a small autonomous stablecoin-trading agent in 2026. It worked — until it didn't, and the failure was silent. That's the whole problem in one sentence.
Here's the angle nobody's writing: "hardware to applications" isn't a rotation. It's an extension. Gas fees higher than the yield. Typical.
Inference still runs on silicon. If personal agents go mainstream, inference demand doesn't replace training demand — it compounds it. The agent running all day, every day, on every device, is a compute monster. Nvidia and the cloud stack don't lose this trade. They get a second act.
What actually changes is where the margin sits. Training margins get competed away as chips commoditize. Inference margins belong to whoever controls distribution — the OS, the browser, the super-app. That's the real warning buried inside Goldman's note. The next AI story isn't an application. It's a toll booth.
For crypto, that's the uncomfortable part. We don't own distribution. No crypto project owns a consumer entry point at scale. So the "inference economy" in Web3 is mostly a rent-seeking layer selling compute it doesn't control to agents that don't work yet. That's not a thesis. That's a pitch deck.
Watch the Q3 earnings guides from the AI leaders, not the podcast. If inference revenue starts showing up as a real line item with real margins, the rotation is real — and crypto's version of it is still a rounding error. If it doesn't, "the next story" was just a sell-side way of saying "please keep buying."
The market's eager for a narrative. It always is. The question is whether anyone's built the thing yet — or whether we're all just renting GPUs and calling it intelligence. t check.