The Trillion-Dollar Question

In 2026, the four largest cloud providers will spend about $700 billion on infrastructure. Amazon has guided to two hundred billion, more than double its 2025 outlay. Alphabet, a hundred and seventy-five to a hundred and eighty-five, against ninety-one the year before. Meta raised its guidance for the year to as much as $145 billion, citing component costs and extra data center buildout. Microsoft is tracking above 120 billion.

Against 2025, that is not the kind of increase budgets normally describe. It is closer to a doubling — Alphabet's line roughly doubled, Amazon's more than doubled — on a base that was already a record.

These are the largest numbers in commercial computing history. On their own, they are also almost useless.

I want to be direct about why, because it determines what this book is for. Capital expenditure tells you how much is being spent. It tells you nothing about who will own the thing that was worth owning when the spending stops. Those are different questions, and the gap between them is where riches are made and lost. The railways were built. Most of the railway companies went bankrupt. Fiber was laid across America in the late nineties, and the fiber is still there, still carrying traffic, still valuable—and the firms that laid it were wiped out so completely that buyers who spent nothing on construction bought their assets for pennies on the dollar. The buildout was real. The demand was real. The returns went somewhere else.

So the question worth two hundred billion dollars of Amazon's money is not how much. It is: when the money is spent, who ends up owning what is worth owning?

That question is the spine of this book. Every chapter is an attempt to answer it for one layer of the stack.

The reflexive comparison is to 1999, and I want to dispose of it early because it is both the most common frame and the most misleading one.

In 1999, the spending ran ahead of the demand. Companies built capacity for a projected internet that didn't yet exist, financed by equity raised on the strength of that projection. The circularity was the whole mechanism: the capex justified the valuation, and the valuation funded the capex. When the projection failed to arrive on schedule, both halves collapsed at once.

That is not what is happening now, and pretending otherwise will cost you.

The current buildout is demand-backed in a way the telecom buildout never was. The hyperscalers aren't building for a market they hope will appear; they are building against order books they cannot currently fill, from enterprise customers already running production workloads and already paying. GPU allocation is rationed. SK Hynix has sold out its memory capacity for the year. TSMC's advanced packaging capacity — the CoWoS process that chiplet-based accelerators need — is fully booked through 2027. You cannot manufacture that kind of scarcity out of enthusiasm. Enthusiasm produces gluts, not queues.

This is why "it's a bubble" is the wrong warning and, worse, an expensive one. It gives an executive permission to wait. And waiting, in an environment where the binding constraints are physical and the lead times are measured in years, is not caution. It is a decision to buy at the back of the queue.

But — and this is the harder thing to hold in mind simultaneously — demand being real does not mean the returns will land where the money is going. Those are independent propositions. Broadband demand was real too. The 1999 traffic forecasts were, in the end, roughly correct; they were simply five years early, which is the same as being wrong if you financed on a three-year horizon. Real demand arriving late and distributed to owners different from those who paid for the infrastructure is the ordinary outcome of a buildout. It is not the exception.

So the useful posture is neither bull nor bear. It is: the buildout is real, it will mostly get used, and almost none of the people funding it will capture the value it creates. Your job is to be specific about which parts are which.

Here is the method this book uses, and it is worth stating plainly because it is the only thing separating analysis from commentary.

In a capital boom, value accrues to whatever is genuinely scarce. Not to whatever is expensive, or fashionable, or receiving the most investment — to whatever cannot be easily produced more of. Everything else is a pass-through. Money flows through the abundant layers on its way to the scarce ones, leaving the abundant layers holding revenue without margin.

Apply that to the current stack and the picture reorganizes itself immediately.

There is an enormous amount of capital going into AI applications — wrappers, copilots, vertical assistants. Almost none of it is going into anything scarce. Models are increasingly commoditized, and open-weight releases are accelerating that. The infrastructure is rented. The engineering is replicable in a quarter. The customer relationships are the only defensible asset, and they were defensible before AI existed. This layer will produce revenue, some of it very large, and it will produce very little durable value, and the two facts will be confused with one another for at least another eighteen months.

Meanwhile, the genuinely scarce things sit further down and are mostly physical.

High bandwidth memory is scarce. Memory now runs to sixty-three percent of an AI chip's component cost, up from fifty-two, and on some accelerators, memory and packaging together approach two-thirds of unit cost. SK Hynix has sold out its 2026 capacity outright. New fabs do not arrive until 2027 or 2028. You cannot software your way around this.

Advanced packaging is scarce. Expanding CoWoS capacity takes eighteen to twenty-four months and billions of dollars, and TSMC is the only supplier at the required volume.

And power — power is the constraint that will define the second half of this decade. US data centers consumed 183 terawatt-hours in 2024, by the IEA's estimate, more than four percent of national electricity, and Goldman Sachs Research projects their share of peak summer demand rising from four percent in 2025 to more than eight percent by 2027. In Ireland, data centers consumed 23% of national electricity in 2025—nearly as much as every household in the country combined—a figure that, in any other context, would be described as an emergency.

Now hold two timelines against each other, because this comparison is the single most important fact in the chapter. A data center takes eighteen to thirty-six months to build. The median wait for a new power project to reach the grid is five years, and in the worst markets, seven or eight. Those numbers do not reconcile. There is no version of the next four years in which silicon supply is the binding constraint for very long, because power will overtake it. Sometime in 2027 or 2028, the question stops being whether you can buy accelerators and becomes whether you can energize them.

That means the most prime asset in AI infrastructure may turn out to be an interconnection agreement—a piece of paper negotiated with a grid operator that nobody in the industry was competing for in 2024.

That is what following the constraint gets you. It moves attention from the layer attracting capital to the layer that will bottleneck it.

Running that method across the stack, four candidates survive. They organize the rest of this book.

Silicon and power. The layers that are physically scarce and cannot be conjured on demand. Custom silicon is the live question here — the hyperscalers designing their own accelerators report thirty to fifty percent lower total cost of ownership on workloads they control. That is not a procurement decision. It is a bet that the generality of a merchant GPU is a tax you stop paying once you know exactly what you are running. Part I takes this apart.

Agents that own workflows. Not copilots. The distinction is the subject of Part II, and I think it's the most consequential and most misread transition in enterprise technology right now. Two years of assistive AI produced enormous adoption and remarkably little P&L impact for a structural reason: a system that recommends leaves the accountability, the headcount, and the process exactly where they are. A system that acts does not. Value accrues to whoever owns the workflow end to end, and most enterprises are currently staging toward a destination they haven't decided to reach.

The data position. Who is legally and technically permitted to compute on what. Open-weight models moving into the enterprise change the leverage in every vendor negotiation. Synthetic data changes what you can train on when the regulator says no. Federated learning, homomorphic encryption, and zero-knowledge machine learning let parties compute jointly without showing each other their data—turning a compliance cost into a structural advantage—part III.

Governance capacity. The least glamorous and, I will argue, the most reliably underpriced. Between 2023 and 2025, the share of S&P 500 companies disclosing AI risk went from twelve percent to eighty-three percent. The share disclosing board-level oversight of AI reached twenty-two percent. The share of directors with disclosed AI expertise is 2.7 percent.

Read those three numbers together. Nearly every large company now says AI is a material risk. About a fifth have put a board process around it. Almost none have anyone on the board who understands it. That gap is not a compliance statistic — it is a leading indicator of which organizations will be able to deploy into regulated environments and which will spend 2027 in review. Governance has stopped being the function that slows deployment and has become the one that permits it at all — part IV.

Parts V and VI extend the argument outward and then back. Part V follows the convergences where AI stops being a software story — post-quantum cryptography, commercial quantum, AI-driven materials discovery, programmable biology, and the energy systems that underwrite all it. Part VI is about allocation: the metrics that actually predict durability and a set of dated, falsifiable predictions I am willing to be judged on.

Two camps are divided on the buildout, and it is worth being precise about where they diverge because the disagreement is narrower than it appears.

The bear case is a bullwhip argument. Hyperscalers, facing scarcity and unable to forecast when supply will catch up, are over-ordering. When supply arrives — 2027, 2028 — orders reverse, capacity gluts, prices collapse, and everyone who financed against 2026 pricing spends 2029 explaining themselves.

The bull case is that every analyst except Nvidia has structurally underestimated demand. Inference, in particular, is accelerating faster than models predicted as applications reach production — it overtook training in data center revenue for the first time in late 2025, and that shift is early. Enterprise spending has not yet moved from pilot to production at scale. When it does, the current buildout looks conservative.

Notice that these camps do not disagree about demand. They disagree about timing and discipline. And once you see that, the question stops being macro and becomes operational.

The thing that resolves it is use — the share of the capacity you paid for that is actually doing work. A disciplined hyperscale deployment runs most of its fleet most of the time. An undisciplined regional one can sit closer to half. On a facility costing hundreds of millions, that gap is tens of millions a year in power and depreciation burned against unused capacity, and it compounds annually.

The bear case is correct about undisciplined deployments. The bull case is correct about disciplined ones. There is no aggregate answer, which is precisely why "is it a bubble?" is the wrong question and has been the whole time. It invites you to take a view on the market when what will determine your outcome is how well you run your own deployment.

One more thing before we get into the stack, because it is the piece most likely to be absent from the strategy deck you are being shown.

AI compute capacity is overwhelmingly concentrated in the United States and China. Europe, a multi-trillion-dollar economy with genuinely excellent AI research, has built very little independent compute, and its public commitments are more than an order of magnitude smaller than a single American hyperscaler's quarterly capital expenditure. To reach actual autonomy, those commitments would need to grow by a factor for which there is no mechanism to deliver within the window that matters.

For an executive outside the United States, the implication is uncomfortable and usually left unstated: you can operate inside the US ecosystem and accept that your AI strategy carries US regulatory and foreign policy exposure, or you can spend years and billions building indigenous capability. Nearly everyone will choose the first. Very few will say so out loud or price the exposure.

I raise it here because it is the clearest available example of what this book is trying to do. The trillion-dollar number is not the story. The story is what the number is buying, who ends up holding it, and which dependencies you get without noticing.

The next three chapters stay in Part I and stay physical. Chapter 2 is about silicon ceasing to be a commodity. Why hyperscalers concluded that designing their own accelerators was strategically necessary; what the cost advantage actually consists of and how large it honestly is; and why CUDA remains the moat that makes displacement slower than economics alone would predict. Chapter 3 is about the constraints nobody priced: memory, packaging, and above all power, and what it means that grid timelines run twice as long as construction timelines. Chapter 4 is about the inference inflection — the moment the economics of AI stop looking like a capital event and start looking like an operating cost that never ends, and what that does to who wins.

Then we leave the physical layer and go up the stack, and the question stays the same the whole way.

Who ends up owning what is worth owning?

Barbara Bickham is Founder & CTO of Trailyn Ventures and the author of Where the Value Accrues, serialized first in Unbound, weekly deep-tech intelligence for executives, investors, family offices and LPs.

Sources: