The $7.6 trillion wager: what the 2026 AI infrastructure boom actually costs, and who is really paying for it

Every AI capex story published this year leads with the same reflex: marvel at the size of the number, then move on to whether Nvidia's stock is overvalued. That framing misses what is actually happening. The 2026 AI buildout is not primarily a technology story. It is the largest coordinated capital allocation decision in the history of private enterprise, and it is being financed, priced and stress-tested in ways that will look familiar to anyone who has studied a monetary cycle rather than a product cycle.

Four companies are underwriting a number larger than the GDP of every country on earth except two or three. They are doing it simultaneously, for reasons that have less to do with any one product roadmap and more to do with a shared fear: that being under-built is the only mistake in this cycle that cannot be corrected later. That fear is rational, expensive and increasingly visible on ordinary household electricity bills. Here is what the money actually buys, who is exposed if the bet does not pay off, and a number nobody in this conversation is publishing: what a single AI query actually costs to run, versus what you are charged for it.

The scale nobody has fully priced in

Amazon, Microsoft, Alphabet and Meta are guiding to roughly $725 billion in combined capital expenditure for calendar 2026, up 77 percent from approximately $410 billion in 2025. Individually, Amazon leads at around $200 billion, Alphabet at $175 to 185 billion, Meta at $115 to 135 billion and Microsoft near $190 billion once memory and component cost inflation is factored in. Add Oracle's roughly $50 billion and the five-company total for 2026 alone approaches $700 to 690 billion by several independent analyst counts.

Zoom out further and the number gets harder to hold in your head. Goldman Sachs has revised its combined capex baseline for the four largest hyperscalers to $5.3 trillion between fiscal 2025 and fiscal 2030, and separately models a broader $7.6 trillion aggregate spend across compute, data centers and power between 2026 and 2031. Every hyperscaler has now overshot its own prior-year guidance for two straight years running, which is the detail that should concern anyone modelling this as a bubble that simply deflates on schedule. Consensus estimates have been consistently too low, not too high.

Where the capital actually lands

The instinct is to assume this money buys chips. Some of it does, and Nvidia's numbers confirm it: fiscal 2026 revenue reached $215.9 billion, up 65 percent year over year, with data center revenue alone close to $194 billion and non-GAAP gross margins holding in the 71 to 75 percent range. That margin profile, on hardware, is the reason Nvidia briefly became the first company in history to cross a $5 trillion market capitalisation in October 2025, then pushed past $5.5 trillion by May 2026.

But chips are only the most visible line item. A growing share of every capex dollar is now going somewhere less glamorous: the power system required to run those chips at all. Analysts estimate that memory alone, driven by scarce high bandwidth memory needed to feed each GPU, now consumes roughly 30 percent of hyperscaler data centre spending, a fourfold jump from three years earlier. Microsoft's chief financial officer has directly attributed billions of dollars of the company's 2026 capex increase to rising memory and component costs, not additional chip volume. The three companies that make that memory, SK Hynix, Samsung and Micron, all crossed a combined $1 trillion-plus valuation milestone in 2026, a psychological marker for an industry that spent the previous two decades as a byword for brutal commodity cycles.

The bottleneck is not silicon, it is electricity

This is the fact most coverage of the AI boom still underweights. The hard constraint on how fast this buildout can proceed is no longer chip supply. It is grid capacity. US interconnection queues, the process by which a new facility applies to draw power from the public grid, now hold more than 2,000 gigawatts of pending generation and load, roughly double the entire installed capacity of the US bulk power system. The typical wait from application to commercial operation now runs four to five years nationally, and reported waits in the highest-demand corridors, Northern Virginia, Phoenix and Dallas, run four to seven years, with some regional utilities quoting waits approaching a decade for the largest projects.

That bottleneck is why the largest operators increasingly build power generation on site rather than wait in line, a workaround the industry calls behind-the-meter generation. It is also why the cost of that scarcity is now showing up somewhere unexpected: household electricity bills across 13 US states. PJM Interconnection, the grid operator covering the mid-Atlantic and parts of the Midwest, saw its capacity auction price rise from roughly $29 per megawatt-day in 2024-25 to $270 in the following auction, then to a regulator-capped $329 in the most recent round, an increase PJM's own independent market monitor attributes primarily to data centre demand. That single auction result will add an estimated $16.1 billion in capacity costs for the year beginning June 2026, and residential customers in the Washington DC area, western Maryland and Ohio are already seeing monthly bill increases in the range of $16 to $21. Multiple states, including Virginia and Pennsylvania, have moved to create separate electricity rate classes specifically for data centres in response.

This is the part of the story that converts an abstract capex number into a live political and regulatory risk. A capital cycle that quietly redistributes cost onto residential ratepayers invites exactly the kind of state-level intervention that can slow permitting, raise costs further, or both.

The revenue reality check

None of this spending is speculative in the sense that nobody is paying for compute. It is speculative in the sense that the paying customers are not yet generating anywhere near enough revenue to justify the infrastructure built on their behalf. OpenAI's annualised revenue run rate sat near $25 billion by mid-2026, while the company carried an $852 billion valuation following a $122 billion round in March, backed by Amazon, Nvidia and SoftBank among others, implying a revenue multiple above 30 times. Anthropic overtook OpenAI on both revenue and valuation within the same year, reaching a reported $47 billion annualised run rate and a $965 billion valuation after its Series H closed in late May 2026, a run rate that had stood at roughly $1 billion in December 2024. The two companies also differ structurally: Anthropic's revenue mix is roughly 80 percent enterprise and developer spend, which investors typically value at a premium to consumer subscription revenue, the segment that dominates OpenAI's mix.

Both numbers are genuinely extraordinary growth curves by any historical technology company standard. Both are also a rounding error against the $725 billion the infrastructure layer is spending this year alone to serve them. The bet embedded in every hyperscaler earnings call is that inference demand, not training demand, will close that gap over the next several years as agentic workloads scale far beyond today's chat interfaces. Whether that demand arrives on schedule, before the debt and depreciation obligations underwriting today's data centres come due, is the single largest open variable in this entire capital cycle.

The number nobody is publishing: what a query actually costs

Here is the gap in the public conversation that Decentralised News set out to close. Every AI company disclosed enough for outside researchers to estimate the electricity a single query consumes. Epoch AI's widely cited 2025 analysis put a typical mainstream chatbot query, on GPT-4o-class hardware, at roughly 0.3 watt-hours, a figure OpenAI's own leadership has separately corroborated at 0.34 watt-hours. Heavier reasoning models tell a different story. University of Rhode Island researchers measured GPT-5-class extended reasoning queries averaging close to 19 watt-hours, with complex prompts running as high as 40 watt-hours, roughly sixty times the lightest case.

What none of the major labs publish is the other half of that equation: how that marginal electricity cost compares to what a user or enterprise customer is actually charged per query, once subscription fees and API pricing are blended in. That comparison is the entire point of the Bottleneck Doctrine that underpins DN's instrument family. When a number is consensus-priced, meaning everyone assumes they already understand it, that is usually the number worth actually computing.

Run the model below with your own usage pattern. The proprietary metric it produces, the Energy Share of Price, tells you what fraction of your AI spend is genuinely the cost of the electrons, and what fraction is the capex tax: the chip depreciation, cooling, networking and margin layered on top of every token a data centre manufactures. In DN's modelling, that share typically lands in the low single digits to low double digits, depending on model tier, which is the clearest evidence available that today's AI pricing is still substantially a subsidy on capital, not a reflection of marginal cost.

What this means if you invest through the crypto and DePIN lens

DN's readers are watching this cycle from an unusual angle: not as cloud investors, but as participants in the decentralised compute and DePIN thesis that treats idle GPU capacity as a tokenisable, tradeable resource. The logic is straightforward. If hyperscaler compute is capacity-constrained and priced at a substantial capex premium over its marginal energy cost, as the Energy Share of Price above demonstrates, then any credible decentralised alternative that can undercut that premium has a genuine economic argument, not just a narrative one. That is the same structural gap DN's Compute-Backing Ratio instrument was built to interrogate: whether DePIN and AI-token valuations are actually backed by verifiable compute revenue, or whether they are pricing in a capacity crunch that has not yet translated into paying demand.

For readers looking to position around this theme directly, spot and derivatives exposure to AI-infrastructure and DePIN-linked tokens is available through most major exchanges, including Bybit, OKX and MEXC, all of which list the broadest current selection of decentralised compute and AI-agent tokens. As with every instrument in this series, this is not financial advice. The capital cycle described above is real and measurable; whether any individual token captures value from it is a separate, much harder question that deserves its own due diligence.

The uncomfortable summary

The AI infrastructure boom is not a story about whether the technology works. It plainly does. It is a story about whether $7.6 trillion of committed capital, financed substantially through vendor arrangements between the same handful of companies that both supply and consume this compute, can be repaid by a revenue base that today represents a low single-digit percentage of the annual spend. Electricity, not chips, is now the binding constraint on how fast that bet can even be tested, and the cost of that constraint is already landing on electricity bills far from any data centre boardroom. The honest answer to whether this is the best investment in history or the most expensive one is the same answer a careful analyst gives about any capital cycle this size: it depends entirely on what happens to demand over the next three years, and nobody, including the companies spending the money, actually knows yet.

Frequently asked questions