That is the only AI question worth real money.
Everything else the industry argues about is a question of timing: which model tops which benchmark, whether this is a bubble, when the buildout corrects. Timing changes the price you pay. It does not determine where the value ultimately goes.
In June 2024, Leopold Aschenbrenner published Situational Awareness: The Decade Ahead. It mapped the industrial mobilization required to build machine intelligence: chips, power, data centers, transformers and capital. He subsequently launched an investment fund and translated the argument into a concentrated strategy with substantial infrastructure exposure.
Two years later, its central insight is no longer contrarian. The buildout is visible in every hyperscaler budget, every clogged interconnection queue, every multi-year transformer quote. Consensus is priced. Investors understand that intelligence needs electrons and steel.
Which means the live question has moved one step further:
When intelligence itself becomes free at the margin, which companies, assets and entirely new categories capture the value?
That is what this publication exists to answer. This founding essay draws the map. The essays that follow test it, route by route.
Where the value goes
When cognition collapses in price as a factor of production, the value does not stay in the intelligence layer. It moves to the complements intelligence consumes, to the assets it cannot reproduce, and to the markets that only become viable once thinking stops being the expensive part.
In his account of the productivity paradox, economic historian Paul David showed that electrification delivered its largest gains only after factories were reorganized around distributed electric motors. Swapping a central steam engine for a central electric motor changed far less. The real gains required redesigning the factory floor, the production flow, and everything around it. His paper, “The Dynamo and the Computer”, was published in 1990. Its lesson is current.
A cheap general-purpose technology does not create most of its value inside the technology itself. It creates value through everything reorganized around it. Cognition follows the pattern. The prize is not in producing intelligence. It is in what intelligence runs on and cannot do without, and in every business that was uneconomic while it needed an expensive human mind on each unit, and suddenly isn’t.
Why the price keeps falling
This is not a prediction. It is an extrapolation of curves that have been measured in public for years.
Epoch AI’s trend data estimates that training compute for frontier language models has grown roughly four to five times per year. Its research on algorithmic progress found that the compute required to reach a fixed level of performance has been halving roughly every eight months. The two effects compound: more physical compute, used more efficiently.
The result shows up in prices. When OpenAI launched GPT-4 in March 2023, its API cost $30 per million input tokens and $60 per million output. Across fixed levels of capability, Epoch now estimates that inference prices have been falling between ninefold and nine-hundredfold per year, depending on the task and the performance threshold being measured.
Capability is improving at the same time. METR’s research on the length of tasks AI agents can complete found a historical doubling time of roughly seven months. The exact estimate will move as evaluations improve, but the direction matters: these systems are not only getting cheaper per unit of thinking. They are becoming useful across longer and more complicated units of work.
You don’t need a date for “true machine intelligence” to underwrite any of this. You only need the observation that curves this consistent, with this much capital behind them, don’t stop on a dime. Take the direction as given. The honest uncertainty is real, and it lives at the end of this essay.
Everyone is looking in the wrong place
The first default move is to own the intelligence itself: invest in the model layer, or build an “AI app for X.”
But the model is the exact part racing toward zero price, and this summer produced the sharpest receipt yet. On 16 July 2026, Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter model, and published the full weights eleven days later. It is priced at $3 per million input tokens and $15 per million output, with cached input at $0.30. On the Artificial Analysis composite intelligence index it scores within a few points of the strongest closed systems, and in blind developer testing it took first place on the Frontend Code Arena, ahead of the closed frontier.
Read that back slowly. A capability that cost hundreds of millions of dollars and several years to reach was matched, priced at parity with a mid-tier Western model, and then given away as a download.
Nor is the mechanism mysterious. In February 2026, Anthropic said DeepSeek, Moonshot and MiniMax had generated more than 16 million Claude exchanges to improve their own models. Epoch estimates that the best open-weight models now trail the closed frontier by only about four months. A frontier release quickly becomes training data for its cheaper successors:
This year’s frontier is next year’s free tier.
Owning the asset whose entire trajectory points at zero is a strange plan for getting rich. The app wrappers are worse: a thin skin over a capability their own supplier can absorb in a release note.
The second default move is to fight about the bubble. Is it one? When does it pop? That is the mainstream conversation, and it is the wrong fight. A bubble is a timing error. It changes what you pay to get in. It does not change where the value goes. You can be dead right that the buildout is overheated and still miss the ten-year migration running underneath it.
The migration is already visible
The last week of July 2026 was the receipts week. All four large US hyperscalers reported within nine days, and three of them raised their spending plans for the year. Alphabet went to between $195 billion and $205 billion, from $180 to $190 billion. Meta lifted the floor of its range. Amazon raised again on higher memory costs, having already published a $200 billion plan in February. Microsoft’s headline calendar-2026 number fell to roughly $175 billion, but on a change to the assumed useful life of its data centers rather than a change in what it intends to build. In aggregate, the four now plan to spend on the order of $760 billion this year, against $413 billion in 2025.
The demand behind the spending accelerated in the same prints. Azure grew 43% in the June quarter and crossed $100 billion in annual revenue for the first time, with commercial remaining performance obligations of $678 billion. Google Cloud grew 82%, to $24.8 billion in the quarter. AWS grew 36.7%, its fastest in eighteen quarters, and Amazon said its AI business and its chip business each now run at a $25 billion annualized rate.
Then the thesis turns physical. Recent reporting found generator step-up transformer lead times exceeding 160 weeks; large power transformers have averaged 128 weeks. Sightline Climate counted roughly 16 gigawatts of data-center capacity intended to come online during 2026, with only about 5 gigawatts under construction at its February assessment. The rest sat in the slower world of interconnection queues, permits, substations, equipment and construction.
And consumption compounds straight through the price decline. OpenRouter, which routes traffic across hundreds of models, reported weekly volume rising from 5 trillion to 25 trillion tokens in six months, more than a quadrillion tokens a year at that pace.
Lay those facts beside one another:
The intelligence is getting cheaper.
The amount consumed is growing.
The physical and institutional systems underneath it are becoming more valuable.
The market is already paying for electrons and steel while everyone argues about which model wins.
Value migrates in layers
It does not migrate once. Everyone now agrees that cheap intelligence needs power, so power attracts capital. But supplying power requires its own constrained inputs: transformers, switchgear, cooling systems, transmission equipment, turbines, and land with a credible route to interconnection.
The same pattern repeats elsewhere. Robots need intelligence, but they also need precision gears, motors, sensors and reliable actuation. Autonomous transactions need agents, but they also need identity, payment rails, permissions and proof. Continuous AI services need models, but they also need trusted data and systems of record.
The obvious complement fills first. The deeper mispricing usually sits one layer behind it. “Buy the picks and shovels” is now consensus. The edge is the pick that makes the shovels.
The map: four routes
Four routes value takes when thinking gets cheap. One claim each. The essays that follow go deep, route by route, wherever the evidence is freshest.
1 · Substitute: what AI can reproduce
Cognition itself, and the layers built on selling it: routine analysis, content production, outsourced back-office operations, per-seat software without proprietary data or workflow control, and any business whose scale requires adding people in proportion to revenue. Value leaves. Prices fall toward the cost of compute.
This route is the donor set, and the trap inside it is valuation. These businesses often look cheap. Their historical earnings look stable. Their customers are slow to switch. But when a company’s economic function is the cost being removed, a low multiple is not protection. Cheap because it is melting is not a bargain.
2 · Amplify: what AI consumes more of
Energy and the grid. Cooling and data centers. Physical actuation. Agent rails and verification. Demand rises. The bottleneck earns the rent. The queue numbers above are this route’s receipts, and they are why the map leads here.
Two things inside it are underpriced. The first is bodies: the market is busy pricing the robot’s brain, which depreciates like the silicon it is, while precision gears and actuators do not rot on the same schedule. The durable money is in the body, not the mind. The second is proof: cheap intelligence makes fabrication and attack cheap too, so identity, provenance and the ability to demonstrate that an autonomous action was real stop being discretionary spend. Turn a billion agents loose and they multiply transactions through rails they cannot rebuild, and verification demands they cannot skip.
3 · Reprice: what AI can’t create
Land and prime space. Authentic experience. Scarce rights and access. Human attention, status, health and longevity. Scarcity re-rates. Surplus flows to the fixed.
Intelligence cannot print more of any of it. When the machine does the thinking, the freed surplus flows toward what a machine cannot be and cannot fake. This is the quietest route and the one with the longest duration.
4 · Unlock: what AI makes viable
Markets that were uneconomic until cognition got cheap. Categories appear. New supply, new demand. Five mechanisms, each with a one-question test:
Cost-collapse services. Anything that needed a costly expert on every unit goes continuous and cheap: the personal CFO, legal triage, health navigation, tutoring. Was this a €200-an-hour human?
Long-tail markets. Segments too small, too custom or too slow to serve profitably become serviceable: micro-vertical software, the one-person enterprise, hyper-local services. Was the TAM “too small”?
Continuous systems. Work done annually because labor was expensive goes always-on: audit becomes continuous assurance, market research becomes standing intelligence. Was the frequency set by labor cost?
Coordination-unlocked businesses. Value blocked by coordination overhead rather than cognition: multi-party logistics, complex procurement, small-scale M&A integration. Did this die in email threads?
Abundance-native primitives. No pre-AI analog at all: agent-to-agent markets, machine-speed negotiation layers, simulation-first design. Highest uncertainty, highest ceiling.
One corollary stitches this route back into the second: every new autonomous category drags a verification demand behind it. The more agents transact, the less optional it becomes to prove who did what. The two routes are one thesis, not two.
Routes 2 and 3 track where existing value re-pools, and most of it is ownable today, largely in public markets. Route 4 tracks what gets created, and most of it is not ownable yet. I can’t hand you the next decade’s winners by name on this route. Most are private, too early, or don’t exist, which is precisely why it is the frontier and not the trade. But a map that only shows what to defend, never what is being built, is half a map. The next companies at today’s giants’ scale will come at least as much from here.
One further theme sits deliberately off the map: hard-money and real-asset expressions of the fiscal path funding the buildout. That is a read on how the expansion gets paid for, not a destination value pools into, so it is not drawn as a fifth route.
How value moves: buildout, digestion, diffusion
None of this lands at once, and the order matters more than any date. Three phases, spanning roughly 2025 to 2035. Markers govern, not the calendar. Each phase announces itself in observable facts, and each phase pays a different owner.
Buildout, roughly 2025 to 2027, and where we are as I write. The physical substrate is the bottleneck, and the money is with whoever supplies the scarce infrastructure. The marker: lead times quoted in years, and interconnection queues that don’t clear.
What is new is that the financing strain has arrived on the balance sheets. Alphabet turned cash-flow negative in the second quarter for the first time, and its long-term debt rose 111% to $98 billion over the first half of the year; Amazon’s rose 81% to $119 billion in a single quarter. Further down the credit ladder it bites harder: on 9 July, S&P cut Oracle to BBB−, one notch above junk, after the company burned $23.7 billion of free cash flow in a single fiscal year building AI capacity. The mobilization is real, and it is increasingly borrowed. That matters for phase two.
Historical rhyme: Canal Mania in Britain, 1790–93. The US railroad boom, 1866–73. The long-haul fibre buildout, 1996–2000.
Digestion, roughly 2027 to 2030, and the market is already rehearsing it. The question turns from “can it be built” to “did it pay,” asked line by line in audited prints. That same July week produced the first split verdict of the cycle: Alphabet raised its spending plan and fell 7% the next day, while Microsoft showed the revenue bridge, Azure past $100 billion with growth accelerating, and was rewarded. Same capital cycle, opposite receptions, decided entirely on whether the revenue showed up.
The gap that bridge has to span is still wide. OpenAI and Anthropic together were reportedly running near $55 billion in annualized revenue by the spring, against three quarters of a trillion dollars of planned annual spend. The figures are not directly comparable, and that is exactly the point: digestion is the phase where the market stops accepting that they don’t need to be.
It will not be comfortable. Most of the durable positions of the next decade get bought in it.
Historical rhyme: the canal share collapse, 1793–97. The railroad panic of 1873 and the six lean years after it. The telecom bust, 2000–02.
Diffusion, roughly 2030 to 2035. The capability stops being the story and becomes the substrate. It shows up in ordinary companies’ margins rather than in vendors’ revenue. Rails, trust, land and attention compound, and route 4 becomes ownable at scale. The marker: agents transacting at scale in reported financials.
Historical rhyme: cheap bulk transport feeding the factory economy. National distribution, then mail-order retail. Dark fibre becoming broadband, cloud and streaming.
Notice what none of this claims: a date when machines get smart. Take the direction as given; that is what the trendlines are for. Argue instead about what is actually contested: the speed of the transition, the bottlenecks along the route, the durability of commercial moats, and who captures the surplus while it moves.
And a clock runs through all four routes. The complements are ownable today. The unlocked categories mostly are not yet; they come within reach as the phases turn. When each category flips from merely buildable to actually ownable deserves its own essay, and it will get one. Without the sequence, the map is a menu. With it, it’s a route.
What would make this wrong
This thesis only works if a chain of assumptions holds, and the load-bearing one comes first:
Defensibility has to survive somewhere above cheap cognition.
The whole map assumes value has somewhere to pool. If intelligence becomes cheap and dissolves every moat it touches, value has no destination. It leaks to customers as consumer surplus. Margins flatten. The map comes back blank.
Call that world the Great Leak. It could look like utopia from the consumer side and rupture from the capital side. I do not think it is the most likely world; physics, land, licenses, networks and trust do not commoditize on the same curve as software. But it cannot be dismissed, because the entire structure depends on it being wrong.
The other failure points follow.
The capability can’t plateau. If agents stay impressive demos that can’t run unsupervised past narrow, checkable tasks, the demand paying for the buildout breaks, and the crash takes the near-term names with it. The most bearish branch, and a real one.
The intelligence providers have to stay in their lane. The map assumes cognition stays cheap and comes from many suppliers. If a handful of labs integrate downstream and capture the complements themselves (the rails, the power, the trust layer), the map redraws around them.
Cheap has to mean cheap, not rationed. Prices must keep falling, and power and chips must scale with demand. If cognition ends up abundant in theory but expensive and rationed in practice, “abundance” quietly becomes “premium.”
Adoption can’t stall. Diffusion assumes firms actually reorganize. Paul David’s factories are the warning: the gains waited on reorganization, not on invention. A long adoption lag leaves the surplus real on paper and unclaimed in practice.
Purchasing power has to hold. Route 4 assumes there are customers able to buy the new abundance. A technology can create enormous productive surplus while disrupting the incomes needed to purchase its output. If the gains concentrate narrowly enough, demand itself becomes the constraint.
And beneath all of that sit two conditions that aren’t investment risks at all. They are the ground the map stands on. The intelligence has to stay controlled and aligned. Everything above assumes AI remains a tool that does what it is pointed at. Lose that and the question stops being where capital migrates, because there won’t be a stable enough world to ask it in. The political order has to hold. This transition displaces a great deal of white-collar work, fast, and it reaches diffusion only if societies absorb the shock without the rupture that smashes data centers, freezes capital markets, or tears up the property rights the map depends on.
Note the uncomfortable symmetry: a safe, aligned, socially stable world with abundant intelligence can still falsify this thesis, if the surplus leaks, if adoption drags, if the providers eat the complements. The catastrophe-avoidance conditions are necessary, not sufficient. The map is worth drawing only in the worlds where we get to keep drawing.
How the map gets tested
The map changes when the evidence changes, and every change comes with its reason. That is the method, not a disclaimer.
The essays that follow develop it in three directions: deep dives into the routes where value concentrates, essays on the mechanisms that create new categories, and scenario pieces that take seriously how the thesis could fail. The sequence follows the strongest evidence, not a fixed curriculum.
Once a quarter, a watchlist records which scenarios have become more or less plausible, and why. At every monthly close, a short note records what moved and what was done about it. When the map is wrong, those are the two places it will say so first.
One disclosure matters: I have skin in the game and invest behind this framework. That is not proof the map is right. It is why the framework exists, and why I want it challenged.
So return to the opening question. Where will the trillion-dollar companies of the next decade come from? They will not win by owning the smartest model. They will own the scarce systems the models depend on, control the assets the models cannot reproduce, or build the markets that become viable only when thinking is no longer the constraint.
Intelligence will become abundant. The largest fortunes will be built around what it makes scarce — and what it makes possible.
The map in full: whenintelligenceisfree.com. The same map, held as positions and marked to market every month: the book. The same read, applied to the decision on your desk: advisory.