Inside the 2026 AI buildout, five companies are on pace to spend more than Belgium’s annual GDP, with power capacity, not chip supply, now deciding how much compute they can deploy.
AI capex 2026 has become shorthand on earnings calls for a buildout that analysts now peg between $775 billion and just over $1 trillion, depending on where the counting starts and stops.
The binding constraint on that spending is whether local utilities can deliver the electrons, because power availability now determines deployed capacity, inference economics and the practical return on each GPU purchased.
TL;DR
The Big Five hyperscalers will spend an estimated $775-800 billion on AI infrastructure in 2026, confirmed across Q1 2026 earnings calls, per CFA-affiliated research from Al Capital Advisory.
Dell’Oro Group puts total global datacenter capex above $1 trillion for the year, with memory pricing inflation compounding the gap between planned and actual spend.
Power, not GPUs, is now the gating factor, the IEA projects datacenter electricity demand will double between 2022 and 2026, and permitting and grid interconnection queues now run longer than chip lead times.
J.P. Morgan estimates hyperscaler capex specifically will hit $697 billion in 2026, a figure that excludes neocloud and sovereign buildouts layered on top.
The counterargument, that this is a bubble inflating on circular financing and overbuilt capacity, has real evidence behind it and deserves to be taken seriously rather than dismissed.
Inside The Headline Number And Why Three Analysts Disagree On It
Ask three research shops what AI capex 2026 actually totals and you get three different numbers, which is itself the story. Al Capital Advisory‘s CFA-affiliated research puts the Big Five hyperscalers at $775-800 billion for the year, a figure the firm says was “confirmed in Q1 2026 earnings” calls from Microsoft (MSFT), Alphabet (GOOGL), Amazon, Meta (META) and Oracle.
J.P. Morgan‘s banking division, writing in its own capital markets research, lands lower at $697 billion for hyperscaler spend specifically. The bank calls it “one of the defining capital deployment themes” in corporate finance this year, per the bank’s infrastructure financing note.
Dell’Oro Group‘s estimate is the outlier on the high side. The firm’s October tracking shows global datacenter capital expenditure “on track to exceed $1 trillion in 2026,” a figure that folds in non-hyperscaler buildouts, sovereign AI programs and the neocloud segment that J.P. Morgan and Al Capital largely exclude from their hyperscaler-specific totals, according to Dell’Oro’s report as summarized by Converge Digest.
Inside those definitions is the real difference between $697 billion and $1 trillion-plus. It is not a rounding error, but a genuine fight over what counts as “AI capex” versus ordinary cloud capacity expansion. Until hyperscalers break out AI-specific line items with more granularity, every number in this space carries an asterisk.
What is not in dispute is direction. Futurum Group‘s framing of the cycle as a “$690B Infrastructure Sprint” and the CapexIndex live dashboard tracking Alphabet, Amazon, Microsoft, Meta, Oracle and Apple all show the same trajectory, quarter-over-quarter acceleration with no sign of a plateau through the back half of 2026, per Futurum’s analysis.
Inside The Capex Absorption Ratio, And What It Reveals About Who Is Actually Exposed
One of the more useful metrics to emerge this year is what Axis Intelligence calls the Capex Absorption Ratio. It measures how much of a company’s operating cash flow is being consumed by AI infrastructure spend rather than returned to shareholders or banked for flexibility.
The firm’s tracker found that the Big Four hyperscalers spent 99% of Q2 2026 operating cash flow on capex, according to Axis Intelligence’s capex tracker. That is an extraordinary number. These companies are, in effect, plowing almost every dollar of cash generated by their existing businesses straight back into GPUs, power contracts and datacenter shells.
That leaves essentially no cushion if AI revenue growth disappoints or financing costs rise. A 99% absorption ratio is not sustainable financing behavior for a mature business. It is the behavior of companies that believe, or are betting, that the payoff window is short and the penalty for under-investing now is permanent competitive disadvantage.
This is where the AI capex 2026 story starts to bleed into capital markets risk. When nearly all free cash flow is absorbed by one spending category, companies turn to debt, off-balance-sheet vehicles and structured financing to keep building without visibly gating shareholder returns.
J.P. Morgan’s own research flags this explicitly, noting that “capital needs are large, build timelines are long and cash-flow profiles differ from traditional investment-grade” borrowers. That is banker language for a simple reality, this does not look like financing a toll road, it looks like financing a bet.
Why Power, Not Nvidia, Is Now The Choke Point
For the first two years of the frontier AI buildout, the scarce resource was silicon. Nvidia (NVDA) GPU allocation determined who could train what and when, and lead times on H100 and then Blackwell-class chips set the pace for the entire industry.
That constraint has not disappeared, but it has been overtaken by a slower-moving, harder-to-fix problem, there is not enough electricity where the datacenters need to be. Building new generation and transmission capacity takes years, not quarters, and that makes available power capacity the limiting input for compute deployment and the cost per unit of inference.
The International Energy Agency‘s projection that global datacenter electricity consumption will double between 2022 and 2026 is the baseline figure cited across the industry research cited by Futurum Group. Inside constrained regional grids, that figure understates the problem because datacenter demand is not spreading evenly across grids.
It is clustering in a handful of regions, Virginia, Texas, Ireland, parts of the Nordics, where fiber, land and existing substations already exist. That means the marginal megawatt in those specific places is vastly more expensive and slower to secure than the national average would suggest.
Intel’s own messaging at OCP 2026 leaned into this reality rather than around it. The company framed its October announcement around “open, ecosystem-driven” infrastructure specifically to help organizations “scale AI infrastructure for emerging workloads.”
That language is corporate-speak for acknowledging that infrastructure, not model architecture, is now the gating variable, per Intel’s announcement. Nvidia’s own GTC 2026 keynote reinforced the same shift in emphasis. SemiAnalysis’s writeup of the event describes Nvidia introducing three entirely new systems aimed squarely at inference efficiency and power draw per token, rather than raw training FLOPs.
That is a signal that the company sees the market’s attention shifting from “can you train the model” to “can you run it profitably at the power budget you actually have,” of the conference.
A Market Table: Who Is Spending What, And On What Basis
The following figures come from the primary and tier-1 sources gathered for this piece. Inside the table, readers should treat the ranges as directional rather than precise, given the definitional disputes outlined above.
Metric Figure Source Big Five hyperscaler AI capex, 2026 $775-800 billion Al Capital Advisory, CFA-affiliated research Hyperscaler capex (narrower definition) $697 billion J.P. Morgan banking research Global datacenter capex, all segments Exceeds $1 trillion Dell’Oro Group, via Converge Digest Capex Absorption Ratio, Big Four, Q2 2026 99% of operating cash flow Axis Intelligence capex tracker Projected datacenter electricity demand growth Doubling, 2022-2026 International Energy Agency, cited by Futurum Group Etched AI chip startup valuation offers $40 billion-plus TechCrunch, Oct. 5 2026 OpenAI pre-IPO round under discussion $30 billion at ~$1.4 trillion valuation TechCrunch, citing Bloomberg, Sept. 29 2026
Also Read: OpenAI’s $30 Billion Raise Would Value It at $1.4 Trillion
Who Is Financing This, And The Quiet Shift Toward Structured Debt
Inside the financing stack, the hyperscalers have enough balance sheet strength to self-fund a large share of this buildout, but even they are reaching for external capital at a pace that would have been unusual three years ago. J.P.
Morgan’s note on financing AI infrastructure describes a market where “capital needs are large” and “build timelines are long.” The bank uses that language to justify why it expects a wave of project-finance-style structures, with special purpose vehicles isolating datacenter debt from the parent company’s balance sheet, something closer to how power plants and toll roads are financed than how software companies traditionally raise capital.
Below the hyperscaler tier, the financing story gets more interesting and more fragile. Crunchbase’s Q3 2026 data shows global venture funding totaled $159 billion for the quarter with close to 6,000 startups funded.
The quarter set a record for billion-dollar rounds as the AI funding race intensified. Inside that total, infrastructure-adjacent names are commanding valuations that look detached from revenue.
AI chip startup Etched is reportedly fielding funding offers at a valuation above $40 billion, more than double its prior mark from just months earlier, per TechCrunch’s Oct. 5 report. European neocloud operator Verda closed a $189 million Series B, per Data Center Dynamics, one of several signs that capital is still chasing the physical layer of AI even as some analysts warn the segment is overbuilt.
OpenAI’s own fundraising sits at the top of this pyramid. The company is reportedly in talks to raise at least $30 billion in a pre-IPO round at a valuation near $1.4 trillion. 29 report citing Bloomberg.
That figure, which Fathom has previously covered, makes sense only if a substantial share of that capital is earmarked for compute commitments rather than operating costs.
Inside The Counterargument
The strongest case against treating this as a durable, justified investment cycle is that a meaningful share of the capex is funded by financing structures that create the appearance of demand rather than reflecting it.
SemiAnalysis itself has pushed back on the more alarmist framing, publishing a piece specifically titled to rebut the claim that “half of 2026 US datacenter capacity” will be “delayed or canceled.” It argues that figure has circulated widely without solid grounding, per the newsletter’s own rebuttal.
That piece cuts both ways, it suggests the bear case is overstated, but its very existence confirms that serious doubts about overbuilding are widespread enough in financial and social media that SemiAnalysis felt compelled to respond directly.
Inside the more durable skeptical case, no single circulating claim is necessary. A 99% capex absorption ratio, as Axis Intelligence documents, is not how businesses behave when they have high confidence in near-term payback.
It is how businesses behave when competitive dynamics force them to spend regardless of near-term return, for fear that a rival’s superior model capability becomes unanswerable if they under-invest even briefly. That dynamic, sometimes called a capacity arms race, can produce enormous overbuilding relative to actual end-demand because no single company can afford to be the one that blinked first.
If enterprise AI revenue growth fails to keep pace with the capex curve, the hyperscalers have locked in multi-year power and construction commitments against revenue that may not materialize on schedule. Computerworld’s reporting on executive sentiment suggests enterprise AI revenue growth is already under scrutiny, with “tech execs getting wise about ROI from AI” and tying spend to outcomes rather than just adoption, per Computerworld.
Bank of England Governor Andrew Bailey has made a version of this argument publicly, warning that the AI investment boom carries a real risk of market shock, a warning Fathom covered in October. Axis Intelligence’s 99% Capex Absorption Ratio does not prove the buildout is wrong.
It is entirely possible that AI capability and enterprise adoption continue compounding fast enough to justify the spend, but the honest version of this story has to hold both facts at once, the infrastructure is real and the financial structure underneath it has features consistent with genuine excess.
Sovereign Capital Enters The Race
The AI capex 2026 story was, until recently, almost entirely a private-sector phenomenon dominated by five or six US hyperscalers. Inside sovereign capital allocation, that is changing.
South Korea’s science ministry announced plans for a 4.7 trillion won, roughly $3.5 billion, program to develop a frontier AI model starting in March 2027. The effort is explicitly framed as a national competitiveness play rather than a commercial venture, according to Reuters’ Oct. 6 report.
Samsung has separately committed $1 billion to KKR’s Helix AI datacenter buildout, a deal Fathom covered when it was announced. The pattern across East Asia increasingly resembles the national champion strategy South Korea has used in semiconductors for decades, now redirected toward frontier model capability.
China’s approach is structurally different but arguably more coordinated. The Diplomat’s reporting on Beijing’s AI governance push describes “an impressive ability to enact and coordinate regulations, policies, technical standards” at a pace that contrasts with what the piece calls Washington’s “fragmented approach,” per The Diplomat.
This matters for the capex numbers because sovereign and quasi-sovereign capital does not behave like venture or hyperscaler capital. It is far less sensitive to quarterly ROI pressure and far more willing to treat infrastructure spend as a strategic cost center.
That means the floor under global AI capex may be higher and stickier than a purely commercial analysis would suggest, even if US hyperscaler spending growth eventually decelerates.
Regulation Is Arriving At The Exact Moment Spending Peaks
Inside the EU, the AI Act’s enforcement timeline converges awkwardly with the capex cycle. From Aug. 2, 2026, the EU AI Office and member state authorities became responsible for implementing, supervising and enforcing the Act’s remaining provisions, according to the European Commission’s own digital strategy page.
That same date also triggered a requirement under Article 57 for each member state to establish at least one national AI regulatory sandbox, per artificialintelligenceact.eu’s tracking of the implementation calendar.
The practical effect for companies building datacenter capacity inside the EU is a compliance overhead that did not exist eighteen months ago. It is layered on top of permitting and grid interconnection delays that already made European AI infrastructure slower and costlier to build than its US equivalent.
Aleph Alpha’s push into sovereign, open-weight European models, which Fathom covered in a recent piece on its Kolibri release, is partly a bet that this regulatory and infrastructure gap creates room for a European alternative to US hyperscaler dependency. That remains a minority position.
Meanwhile in the US, the regulatory conversation is fragmenting downward rather than consolidating upward. New York City Council heard testimony from AI whistleblowers and policy leaders from OpenAI, Anthropic, Google and Meta on Oct. 5.
Local lawmakers were explicitly trying to position the city “to lead the nation” on AI safety rules, according to PYMNTS’ coverage of the hearing. City and State New York’s reporting on the same hearing was blunter, headlining that the “leading AI companies fail to impress” councilmembers, per City and State.
That hearing sits alongside Reuters’ reporting that OpenAI and Anthropic separately told Australian parliament they would welcome mandatory data breach reporting rules for AI agents. It is a notably cooperative posture from two companies that have otherwise resisted binding external oversight.
Methodology
Inside this methodology, the piece draws on primary and tier-1 sources gathered between Oct. 5 and Oct. 6, 2026, official statements and research from Al Capital Advisory, J.P. Morgan, Dell’Oro Group (via Converge Digest), Axis Intelligence, Futurum Group, SemiAnalysis, the European Commission’s digital strategy office, artificialintelligenceact.eu, and wire reporting from Reuters and TechCrunch.
Capex figures for 2026 are drawn from Q1 and Q2 2026 earnings-derived analysis published by third-party research firms rather than from direct line-item disclosure by the hyperscalers themselves. None of the hyperscalers break out “AI capex” as a standalone reported figure, which is the central reason estimates range from $697 billion to over $1 trillion depending on scope.
Fathom could not independently verify the Capex Absorption Ratio methodology beyond what Axis Intelligence has published publicly. It treats that figure as directionally credible but not independently audited.
Similarly, the sovereign AI spending figures from South Korea and China are based on government announcements and may not reflect final appropriated or disbursed amounts. Where this piece distinguishes hyperscaler-specific spend from total global datacenter capex, that distinction follows the framing used by the cited source. It has not been independently reconciled across sources, given each firm uses different inclusion criteria for what counts as “AI” versus general cloud infrastructure.
Conclusion
Inside the next earnings cycle, watch three things from here. First, whether Q3 and Q4 2026 earnings calls start breaking out AI-specific capex with more granularity, which would resolve the $697 billion to $1 trillion-plus dispute.
Second, whether the Capex Absorption Ratio falls below 99% for any of the Big Four, a signal that spending discipline is returning. Third, whether grid interconnection queues, not GPU allocation, start appearing explicitly in hyperscaler risk disclosures.
That would confirm power has formally replaced silicon as the sector’s binding constraint.
Read Next: Samsung Commits $1 Billion to KKR’s Helix AI Buildout