Wealth Is Context

In May of last year, on a stage at Sequoia's AI Ascent, Sam Altman described his ideal product. Not a smarter model. A "very tiny reasoning model with a trillion tokens of context that you put your whole life into." Every conversation you have ever had, every book you have ever read, every email. All of it in the window, appending forever.

Read it again, slowly. The CEO of the model company put the model in the diminutive. Tiny. A part. In his own telling, what makes the product valuable is not the reasoning engine. It is the trillion tokens. It is your life, made available to the machine.

That sentence is the trade of the decade getting priced in public, and most founders are standing on the wrong side of it.

The claim

Wealth is context.

It is a data play, but not the one we were sold. Two companies with identical data are not equally wealthy. Two people with identical experience are not equally wealthy. Existence is not the asset. The asset is two properties stacked on top of existence: availability, meaning your AI can reach the data at the moment of decision, and reliability, meaning what arrives can be trusted.

Data that exists but cannot be reached is dead stock. Data that can be reached but cannot be trusted is a liability, and a measured one. We will get to the measurement. Context is data that is both reachable and trustworthy. That is the whole definition.

We heard half the sentence

In 2006, Clive Humby said data is the new oil. Everyone remembers the slogan. Almost nobody remembers the clause that came with it: valuable, yes, but if unrefined it cannot really be used. Oil has to become gasoline before it moves anything.

The world heard "hoard barrels." Twenty years of data lakes, data warehouses, data swamps. Storage as strategy. Then the refinery finally showed up. AI is the refinery, and the moment it arrived it exposed which barrels held crude and which held rainwater.

The industry quietly admitted this in June 2025, in one week, in two tweets. Tobi Lütke said he preferred "context engineering" over prompt engineering: "the art of providing all the context for the task to be plausibly solvable by the LLM." Karpathy seconded it within days: the delicate art and science of filling the context window with just the right information for the next step. Prompt engineering died because the prompt was never the scarce thing. The scarce thing is what the prompt is allowed to know.

Watch the money reach the same verdict from the other direction. In 2019, a16z published "The Empty Promise of Data Moats" and declared raw data worthless as a defense, and they were right about the thing they were looking at: static rows, hoarded barrels, "we have lots of data, therefore defensibility." In December 2025, Foundation Capital published a thesis that the next trillion-dollar platforms will be systems of record for decisions, not objects: the exceptions, the overrides, the precedents, the reasoning that today evaporates in Slack threads and conference rooms. Same asset class, opposite verdict, six years apart. Both correct. The asset changed. Data was never the wealth. Context is.

Why now: rented capability, owned context

Here is the mechanism, and it is brutally simple. Capability is rented. Context is owned.

Every model capability gets matched within months. Reasoning, coding, memory, voice: each lab ships, the others clone, the price falls. Memory was supposed to be the lock-in, the feature that finally made your AI subscription sticky. Instead, by early this year, the vendors were shipping import tools for each other's memories. Stop and look at what that means. The companies that own the models are fighting over who gets to hold your context. They are telling you, with their engineering budgets, what they believe the asset is.

The model is the engine, and engines are leased to everyone at the same price. Your context is the only component of the whole system that is yours, that compounds, and that no competitor can rent.

The numbers say organizations have not absorbed this. Gartner projects that through 2026, companies will abandon 60 percent of AI projects unsupported by AI-ready data, and that 63 percent of data leaders either lack the data practices AI requires or do not know whether they have them. S&P Global watched the share of companies scrapping most of their AI initiatives jump from 17 to 42 percent in a single year. McKinsey finds 88 percent of organizations now use AI while only about 39 percent can attribute any bottom-line impact to it, and the strongest predictor of impact is not model choice. It is workflow redesign: the unglamorous work of making the company's own context flow to the point of decision.

Same models. Same vendors. Same prices. Wildly different outcomes. The difference is not intelligence, because intelligence is for sale to everyone. The difference is what each buyer can feed it.

Unreliable context is negative wealth

Availability is only half the asset, and the other half has teeth.

In February 2025, a research team built a benchmark called RAGuard to test what happens when retrieval systems feed models misleading context. Every system tested performed worse than its zero-shot baseline. Read that carefully: bad context made the models dumber than no context at all. Chroma's context-rot research found the same shape from another angle. Even the best frontier models degrade as raw input grows, on trivially simple tasks. You cannot dump the data lake into the window and call it wealth.

And badly governed context does not stay inside the benchmark. Air Canada was held liable when its chatbot invented a bereavement policy, and the tribunal was unimpressed by the argument that the bot was somehow a separate entity. Deloitte handed back part of a 440,000 dollar government fee after fabricated citations surfaced in an AI-assisted report. Unreliable context is not zero. It is negative. It compounds the same way trustworthy context does, just in the other direction.

So the definition holds on both edges. Context is data that is reachable and trustworthy. Lose the first property and you own dead stock. Lose the second and you own a slow leak.

The moat reading

I keep one filter taped above every defensibility question: is this hard to do, or hard to get? Hard to do is anything bottlenecked by intelligence, and AI is repricing all of it toward zero. Hard to get is anything bottlenecked by elapsed time: licenses, networks, capital relationships, compounding operational data. AI compresses the time it takes to do things. It does not compress the time it takes for things to happen.

Context is hard to get. It accumulates at exactly the speed your business runs and no faster. Ten years of decision traces takes ten years. The pricing exception you approved in March, the client you fired and why, the hire that worked when the resume said it should not: these accrue one at a time, in real time, and no amount of compute can parallelize the calendar. A competitor can rent your model tomorrow. They cannot rent your history.

Which is why the head start is the moat, and why the cheapest day to start capturing context was yesterday.

Judgment, stored

I have written that the lever is cheap and the judgment is the moat. I have written that generation got cheap and verification did not, and that the asymmetry is the whole story. Both still hold. But both locate the wealth in a faculty, and a faculty has a flaw as an asset class: it lives in a person. It leaves in the elevator. You cannot put judgment on a balance sheet.

Context is judgment, stored. The pattern library, serialized. The override and the reason for the override, written down where the next decision can find them. Verification, the expensive half of the new economy, runs on context the way an engine runs on fuel: a brilliant verifier with no access to what actually happened is just taste with nothing to taste. When I said the constraint moved from access to specificity, this is where the specificity comes from. Nobody assembles context they do not have.

So the progression across two years of these essays resolves cleanly. Experience was never commoditized; it was repriced. Judgment absorbed the premium. And context is the form judgment takes when you want it to outlive the meeting, the employee, and eventually the founder.

The two ledgers

The organizational ledger looks like this. Every decision your company makes either deposits context or evaporates it. The deposit is not the decision record alone, it is the rationale, the rejected alternative, the "why not" that no model can infer later. Captured at the moment of decision, organized for retrieval, and governed so the right workflow sees the right slice: that is institutional memory converted from anecdote into capital. A battery that charges on capture and discharges into every future decision, leaving the company smarter after each cycle.

Governance is what makes the battery safe to wire in. The rule I use is binary: if you cannot state why a workflow needs a field, the AI does not get it. Reliability is a governance act, not a model feature. The companies in those abandonment statistics did not fail because the models were weak. They failed because they pointed strong models at context they could neither reach nor trust.

The personal ledger runs on the same physics at smaller scale, and it is sharper, because nobody is coming to build it for you. Your notes, your decisions, your half-finished thinking: either machine-reachable or not. Take two operators of equal ability. One opens a blank window every morning. The other works with a system that holds years of their context: what they decided, what they rejected, how they think. The second is wealthier in a way that compounds daily, in a currency the first cannot mint, and the gap widens every time the models improve. Altman's trillion tokens is this exact ledger, productized. Build yours anyway, and keep it portable, because whoever holds your context holds you.

The objection that makes the case

The strongest counterargument is the Bitter Lesson: compute eats hand-built structure, context windows grew from four thousand tokens to ten million in three years, models keep getting better at messy input. Surely capability swallows curation, and the cleanup is a temporary tax.

Three answers. First, the Bitter Lesson eats how, not what. It kills scaffolding, orchestration tricks, clever retrieval pipelines. It cannot conjure facts the model was never given. Infinite IQ does not know your March pricing exception unless someone wrote it somewhere a model is permitted to read. Second, the empirical record runs the other way: bigger windows did not dissolve the problem, they renamed it, and context rot is the new name. Third, even in the limit, the thesis strengthens. The better models get at exploiting context, the higher the return on every unit of context you have made available and trustworthy. Capability does not depreciate context. It raises the interest rate paid on it.

That is the resolution of the metaphor we started with. Oil was wrong in the one way that matters: oil burns once. Context is capital. It compounds, it earns interest, and the rate just went up.

The audit

Two questions, then a move.

Where do your decisions and their reasons live tonight? If the answer is "in my head" and "in Slack," you are wealthy in a currency only you can spend and AI cannot spend at all.

If your best person left tomorrow and your AI kept running, what fraction of their judgment would still be in the building?