Last October — and I am putting a deliberate asterisk on "last," which we will return to — a client forwarded me a two-paragraph item that a Web3 news aggregator had run under a Goldman Sachs byline. The headline was immaculate: AI trading is rotating from hardware to applications, and the market is "very eager for the next story." I read it three times on my phone while standing in line for coffee, and by the third read I was not thinking about AI at all. I was thinking about the date. There was not one. There was a month and a day, "October 11," floating free of a year, like a transaction with no block height attached. In my world an unconfirmed transaction is not money; it is a promise waiting to be validated. A claim without a timestamp is not data — it is a rumor wearing a suit.
That was the discovery that hooked me: not the thesis, but the seam where the thesis had been stitched onto a podcast I could not locate. I have spent twenty-eight years watching how narratives get minted, and the ones that move the most capital are rarely the ones with the most evidence. They are the ones with the cleanest seam. This note had a beautiful seam — a strategist's name, a podcast brand, a rotation thesis — and almost no binding. So I did what I do with any smart contract I am asked to trust: I stopped reading the marketing and started reading the bytecode. What I found was not a lie. It was something more instructive than a lie. It was a signal wearing the costume of a fact.
Let me lay out what the note actually said, stripped of its gloss, because the compression is the point. A Goldman Sachs strategist — the piece calls him Callahan — argued that the AI trade is migrating from the chip arms race toward what he named the "economy of inference" and "personal intelligent agents." He noted, almost as an aside, that AI coding "is no longer a secret," that it is already a broadly recognized application. And he flagged that the ripple effects of personal agents across the entire consumer-facing supply chain "will be very worth watching." The backdrop he sketched was a high-rate environment where equity resilience rests heavily on the earnings of a handful of large-cap technology names, and a Q3 earnings season that everyone was holding their breath for.
That is eight information points. No first-party data. No named instruments. No quantified estimates. It is a second-hand transcription of a macro podcast, run by a crypto outlet, which makes it a cross-domain relay: a sell-side narrative handed to a retail-facing audience with the original context sanded off. And there were tells. The strategist's name, "Ben Schneider," is almost certainly a corruption of Ben Snider, a longtime Goldman strategist. The podcast name, "Macro Dialogue," does not match any brand Goldman actually publishes under. There is no link, no date, no transcript. In my line of work, when the provenance of a record is this thin, we do not treat it as a record. We treat it as a lead.
So here is my first act of discipline. I am not a Goldman client. I am a protocol PM who audits infrastructure for a living, and I read this not as an investment recommendation but as a signal about where narrative capital is about to move — and what that migration means for the people building verifiable systems underneath it. The protocol is cold; the evangelist is warm. And the coldest part of this document is the part that never got typed: the year. Without a year, I cannot anchor the state of the technology it describes. Was personal agency a research curiosity or an early production system when the words were spoken? The note will not tell me, and that refusal is itself a data point.
Start with the phrase that carries the most weight and the least definition: the "economy of inference." In plain engineering terms, inference is the forward pass — the moment a trained model answers a request. Training is a one-shot, centralized, capital-intensive event; inference is continuous, distributed, and scales roughly linearly with the number of users you serve. That structural difference is why the phrase matters. A business that sells training compute sells a cyclical, lumpy, capex-driven product. A business that sells inference sells something that looks, on a good day, like a metered utility — per-token, per-call, recurring revenue. The valuation paradigm shifts from cyclical to SaaS. That is the entire reason a sell-side desk would bother rotating its language. It is not a technical observation about models. It is an accounting observation about how the same silicon gets billed.
But let me be precise about where the money actually sits in an inference stack, because this is where most retail-facing summaries go soft. When you serve a model, your cost is dominated by the KV cache — the stored attention state that lets the model remember context without recomputing it — plus the prefill of your prompt and the decode of each generated token. Every serious optimization in the last two years is an attack on one of those three surfaces: quantization to shrink the weights, speculative decoding to draft-and-verify tokens, continuous batching to keep the GPU saturated, and paged attention to stop the KV cache from fragmenting your memory. The reason "economy of inference" is a real thesis and not a slogan is that these techniques are measurable. You can watch dollars per million tokens fall on a chart. You can audit it. You cannot audit a vibe, but you can audit a cost curve, and the cost curve is the only honest part of the rotation story.
Here is the insight the note never states: the reason AI coding crossed the chasm first is not that coding is the most valuable application — it is that coding is the only application with a cheap, verifiable oracle. A coding agent can generate a patch and then run the test suite. The compiler, the linter, the continuous integration pipeline — they hand back a ground truth that costs almost nothing to obtain and cannot be argued with. That is a reinforcement learning environment that pays you in truth. When I audited early ERC-20 implementations back in 2017, the thing that let me find a gas optimization flaw was exactly this property: the EVM does not care about your intentions. It executes. Coding agents scaled because their environment is as merciless as the EVM, and merciless environments are where learning happens fastest. Every other domain is still waiting for its compiler.
Now contrast that with the "personal intelligent agent," which the note elevates to the next story without ever explaining why it has not shipped. I want to give you the arithmetic, because the arithmetic is the whole story. Suppose an agent executes a task in twenty sequential steps, and suppose each step is 95% reliable — which, for open-ended real-world tasks, is optimistic. The probability that the entire chain succeeds is 0.95 raised to the twentieth power, which is roughly 36%. Two-thirds of your attempts fail, not because any single step is bad, but because errors compound multiplicatively. This is the long-horizon problem, and it is not a marketing problem. It is a mathematics problem. You cannot narrate your way out of a geometric series.
And that is before you price the latency. A coding agent tolerates a ten-second pause because the human is already in a "wait for the build" posture. A personal agent that books your flight, negotiates with a hotel API, and confirms with your calendar has to do all of that inside a conversational cadence a human will actually tolerate, while paying for every token it burns. The unit economics of a resident, always-on agent — one that watches your inbox, your calendar, your subscriptions — are brutal, because unlike a coding session, it never stops. Continuous inference against continuous value is a hard trade to close, and nobody has shown me the closing. In DeFi Summer we learned the same lesson in a different syntax: a strategy that works beautifully in a backtest dies the moment it has to pay gas on every step. Curiosity is the only leverage in a market this narrative-driven, and the most useful curiosity is the one that asks what the agent costs while it sleeps.
Which brings me to the part of the note that is technically the most interesting and rhetorically the most dangerous: the claim that investment is rotating "from hardware to applications." Read carelessly, that sounds like a verdict that compute demand is peaking. I think that reading is wrong, and I think it is wrong in a way that reveals the note's actual thesis. The rotation is happening in the narrative, not in the physics. Training compute is a one-shot injection; inference compute is a tap that stays open and grows with every user you onboard. If personal agents ever work, they convert the compute business from a bonfire into a utility bill. The hardware narrative is not being falsified — it is being extended into a longer, flatter, more durable demand curve. The sell-side is not saying the picks are worthless; it is saying the picks have stopped glittering and the shovels might now be sold by the hour. When I spent six months in 2022 mapping how separated execution and consensus layers could relieve the congestion that killed so many NFT projects, I learned that the loudest claim — the "death of monolithic chains" — was never about a funeral. It was about a reallocation. The old thing did not die. It got a new job.
There is a second, quieter claim in the note that I want to pull on, because it is the one with the most genuine information gain. The strategist says the "consumer-facing supply chain" will feel the agent's ripple. Think about what that actually names. Search, where an agent that answers directly does not need to render a page of ads. Commerce, where an agent that compares and buys collapses the funnel that advertising was built to monetize. App distribution, where an agent that spans applications makes the operating system's storefront a tollbooth with no road under it. Advertising itself, where an agent becomes the first consumer that is immune to persuasion. Every one of those is a trillion-dollar incumbent whose moat was attention, and an agent is, structurally, a machine for not paying attention. The note says the ripple "will be worth watching." I would say it will be worth surviving.
I have watched this exact movie before, in a different theater. In DeFi Summer, we convinced ourselves that the protocol would beat the platform because the protocol was more open. Then we learned that distribution beats ideology every time. The L2 stack wars taught the same lesson with a colder face: the real difference between the OP Stack and the ZK Stack was never the cryptography — it was who could convince more teams to deploy a chain first. Whoever owns the entry point owns the economics, regardless of how elegant the bytes are underneath. If personal agents become the entry point, the winners will not be the labs with the best model. They will be the players who already own an operating system, a browser, or a super-app — the ones with a default surface to plant the agent on. The most useful question is the one nobody is asking: who holds the doorway?
And I want to name the crypto-native version of this risk, because it is the one my readers will actually feel. Post-ETF, Bitcoin stopped being peer-to-peer electronic cash and became a macro asset that trades on the New York open. That transformation did not require anyone to break the protocol; it only required the market to decide that the protocol's meaning was price. The same drift is available to agents. An agent economy that runs on verifiable credentials — proof that this autonomous process is authorized, that this data was consented to, that this action is attributable — is a genuinely decentralized future. An agent economy that runs on walled-garden APIs and opaque model providers is just the old platform economy wearing a new interface. Both are called "personal agents." Only one of them is yours.
This is where my day job touches the thesis. In 2024 I ran a pilot connecting autonomous agents to decentralized identity protocols, and the finding that surprised even me was how quickly the interesting problem stopped being capability and became accountability. When an agent books the wrong flight, who is liable — the model provider, the tool vendor, the user, or the credential issuer? When an agent acts on your behalf across three jurisdictions, whose privacy law applies? These are not philosophical footnotes. They are the load-bearing walls of any agent economy that wants to scale past a demo. Verifiable credentials — cryptographic attestations that a given agent has a given authority over a given scope of data — are the only mechanism I have seen that lets you answer those questions without asking a platform to be honest about it.
And that is the equity question hiding under the whole rotation. A personal agent that only the wealthy can afford to run continuously is not liberation; it is a butler. If the "economy of inference" lowers the marginal cost of an always-on assistant toward zero, it becomes a genuine equalizer — the same way cheap computation once put a supercomputer in every pocket. If it does not, it becomes one more layer where the people with the best entry points extract rent from everyone standing in line behind them. Art is the glitch that proves we are human, and the glitch in this agent economy will be the moment a user realizes the assistant was never theirs. Chasing the frontier where code meets belief means insisting on the first version and refusing to celebrate the second.
Now let me turn the note against itself, because this is the part the sell-side cannot say out loud. The single most revealing sentence in the whole document is not about inference or agents. It is the admission that the market is "very eager for the next story." That is not a bullish signal. That is a confession that the current story has stopped working. When a desk tells you the market wants a new narrative, it is telling you the old narrative's marginal catalyst is exhausted — and that it has a product to sell you at the moment of maximum suggestibility. The rotation thesis is not a discovery about technology. It is a discovery about attention, and attention is the thing being repriced. I learned to read markets this way during the 2020 arbitrage I stumbled into on a small governance token: the tell was never in the price, it was in who needed me to believe the price.
So the contrarian reading is this: the "next AI narrative" is not the application layer. It is the narrative layer itself. The blind spot in every summary of this note — including the one you just read — is that we are all arguing about which horse wins the race while the desk quietly sells tickets. The applications may genuinely be where value accrues. But the claim that they are, arriving from a party with underwriting relationships and a distribution channel, is a claim to be audited, not believed. And the second blind spot is even simpler: the winner of the agent era will almost certainly not be the best agent. It will be whoever controls the surface the agent runs on. That is a distribution fight dressed as a capability fight, and it is the oldest story in technology, wearing its newest costume.
So where does that leave the builder, the auditor, the person who has to ship something real before the story changes again? I think it leaves us with a cleaner question than the one the note asked. The note asked which narrative comes next. The better question is who will be able to prove their claims — who can show, with a timestamped, verifiable, on-chain record, that their inference economy actually earns, that their agent actually acts on your behalf and not on the platform's. In the silence of the chain, we hear the future — not because the chain is loud, but because it is the only place where a claim arrives with its own proof of when it was made.