The AI capex bubble in 2026: $725B going in, and where the revenue actually is

AI capex spending in 2026 has reached numbers that don't have historical comparables. The five biggest spenders — Amazon, Alphabet, Meta, Microsoft, and Oracle — will commit roughly $725 billion to AI infrastructure this year alone, up 77% from 2025's record of around $410 billion. Stargate, the OpenAI / SoftBank / Oracle joint venture, is on track to deploy $500 billion across 10 gigawatts of data center capacity by the end of 2025. Goldman Sachs estimates total cumulative AI capex over the coming years will exceed $1 trillion. The question every founder, VC, and CFO is now asking out loud is whether AI capex of this scale will produce the revenue to justify it — and the honest answer in May 2026 is "not yet, but with caveats." This is our read on where the AI capex bubble actually stands, who's making money, who isn't, and what that means for buyers. For our editorial view of the products that actually work, our Top 100 AI Tools is the next stop.

We are a content business that catalogs the products consuming this capex. We are not neutral observers; we have a view from inside. We will be specific about what the data says and what it doesn't, and we will tell you which framings of "AI bubble" are real and which are vibes.

TL;DR

• Big Tech AI capex: roughly $410B in 2025, ~$725B planned for 2026, $1T+ projected for 2027 (Goldman Sachs, CreditSights, Morgan Stanley).
• Hyperscaler cloud AI revenue is real and growing fast: AWS at ~$150B annualized (+28% YoY), Google Cloud at ~$80B (+63%), Azure AI at ~$37B run-rate (+123%) per Q1 2026 earnings.
• Sequoia's David Cahn calculated a $600 billion revenue gap that needs to be filled annually to justify the spend; that gap is widening, not closing, in 2026.
• The MIT Project NANDA study (July 2025) found 95% of enterprise GenAI pilots produce zero measurable P&L impact, on $30-40B of corporate spending.
• The DeepSeek shock of January 27, 2025 wiped $1 trillion from US AI-related market cap in a single day — including a $588.8B single-day loss at Nvidia, the largest single-day loss in stock market history.
• Pat Gelsinger (former Intel CEO): "Of course" we are in an AI bubble. Sam Altman has publicly agreed.
• The bear case is real. The bull case is also real. The honest version is that timing the pop is harder than calling that one exists.

How we read the numbers in 2026

Most of the "AI bubble" coverage in 2026 oscillates between two unhelpful poles — a "this is dot-com 2.0 and the world is ending" frame and a "Big Tech is profitable, nothing to see here" frame. Both leave out specifics. We're going to walk through the specifics.

Our methodology is to cite named sources with dates, distinguish AI capex from AI revenue from AI productivity gains (which are three different conversations), separate hyperscaler economics from enterprise economics from model-company economics, and stake a position on which of the bear cases are credible and which aren't.

We've worked with the public Q1 2026 hyperscaler earnings, the most recent Sequoia / David Cahn analysis, the July 2025 MIT NANDA study, Goldman Sachs' Top of Mind reports (both the original June 2024 and the 2025 update), and named industry commentary from Pat Gelsinger, Satya Nadella, Sam Altman, and the Wall Street analyst community. Source links are in line.

The capex numbers, specifically

The 2026 number that matters: the Big-5 hyperscalers — Amazon, Alphabet, Meta, Microsoft, and Oracle — will collectively spend approximately $725 billion on capex this year, up 77% from 2025's record of around $410 billion (per Q1 2026 earnings, aggregated by Tom's Hardware, CNBC, and Statista). Breaking it down by company:

The capex-to-sales ratios are the most informative line in the table. Oracle at 86% means the company is spending more on infrastructure than the entire size of its revenue base. Meta at 54% means more than half of every dollar Meta earns this year is being committed to AI infrastructure rather than returned to shareholders or invested elsewhere.

For historical context: Morgan Stanley's Todd Castagno calculates that the dot-com peak in 2000 saw aggregate tech capex-to-sales of roughly 32%. The AI boom in 2026 is at 34% and projected to reach 37% by 2028. The capex intensity is genuinely worse than the dot-com era, which is the strongest single fact in the bear case.

Beyond the hyperscalers, Stargate — the OpenAI / SoftBank / Oracle joint venture — is on track to deploy roughly 7 GW of data center capacity (toward a 10 GW commitment) with over $400 billion committed of a $500B total, per OpenAI's announcements through 2025 and 2026. The Abilene, Texas flagship campus is operational; Michigan, Ohio, and additional Texas sites are in construction. CoreWeave, the largest of the GPU-focused "neocloud" providers, plans to spend $30-35 billion in capex in 2026 alone, on top of its rapidly growing $5B+ annual revenue base.

The neocloud category itself tripled to ~$23B in revenue in 2025 — Lambda, Nebius, Crusoe, Civo, Nscale, and CoreWeave have collectively become a meaningful infrastructure tier between the hyperscalers and end users.

The revenue numbers, specifically

Here is where the post needs to be careful, because there are several different revenue questions and they have different answers.

Hyperscaler AI cloud revenue is genuinely large and growing fast. Per Q1 2026 earnings: AWS at roughly $150 billion annualized (+28% YoY), Google Cloud at $80 billion (+63%), Azure AI at $37 billion run-rate (+123%). That's approximately $267 billion in annualized AI-attributable cloud revenue across the three Western hyperscalers — and it's growing at a pace that no comparable industry segment is matching.

Model-company revenue is exploding from a small base. OpenAI went from $2B ARR in 2023 to $6B in 2024 to $20B by end of 2025, and was at a $33B run-rate by May 2026. Anthropic grew faster still: $9B end of 2025, $14B February 2026, $30B by April 2026 (Anthropic's reported number; OpenAI's CRO has publicly disputed it by roughly $8B due to a gross-vs-net accounting question on AWS and Google Cloud revenue, so the conservative read is ~$22B). Claude Code alone reached $1B in annualized revenue within six months of its 2025 public launch.

Enterprise AI deployment revenue is the embarrassing number. The MIT Project NANDA study from July 2025 examined 300+ enterprise AI initiatives and found 95% delivered zero measurable P&L impact. On $30-40 billion in enterprise GenAI spending in 2024-25, the median company saw no return at all. Five percent of pilots ARE extracting real value — millions in measured impact — but the median deployment is a write-off.

Productivity gains exist but aren't yet at scale. McKinsey's 2026 analysis identifies 2.8-4.7% revenue-equivalent productivity gains from AI deployment in banking and financial services, and 2.6-4.5% in pharma and advanced industries. BCG's "future-ready companies" (the top 5% by AI integration) expect 2x revenue growth and 40% greater cost reductions than laggards by 2028. These numbers are real, but they describe outcomes years out, not 2026 hyperscaler returns.

The $600 billion question, mathematically

The most-cited analytical frame for the AI capex bubble debate is Sequoia partner David Cahn's "AI's $600B Question," published June 2024, which updated his September 2023 "$200B Question."

The math: take Nvidia's projected run-rate data center revenue, multiply by 2 (Nvidia is roughly half the total cost of ownership for an AI data center; the rest is power, buildings, networking, cooling, backup generators), then multiply by 2 again (data center operators need approximately 50% gross margin to make the spend pencil). That's the annual revenue the AI industry needs to generate to justify the capex it's deploying.

In late 2023 the answer was $200B. By mid-2024 it was $600B. In 2026, with Nvidia data center revenue running at a substantially higher rate and Stargate-scale buildouts in flight, the implied number is meaningfully larger — and the gap between that number and what the AI industry actually generates in end-user revenue is wider.

Cahn himself has stayed bullish on the long-term trajectory ("2025 is the Year of the Data Center," per his Substack), arguing that the buildout is a time-lag problem rather than a bubble — but he's been careful not to claim the gap is closing on its current trajectory.

The hardware-depreciation tell

One specific signal in the 2025-26 data is worth flagging because it's underdiscussed: hyperscalers are quietly stretching their AI-chip depreciation schedules from the historical 3-year norm to 5 years.

The H100 and Blackwell GPU generations have economic lives of roughly 3-5 years before they're superseded by a more efficient architecture. Stretching the depreciation schedule to 5 years lets companies amortize the capital cost over a longer accounting period, which makes near-term earnings look better — but it's also a tell that the companies expect the hardware to remain economically useful longer than historical patterns suggest, which is the opposite of what a healthy compute market would imply.

If 2027-28 demand doesn't materialize to fill the data centers being built in 2025-26, the H100 and Blackwell installed base becomes increasingly hard to monetize before it's superseded. That's the cleanest version of the depreciation-risk argument; it's a real risk, and it's why the bear case can't be dismissed as vibes.

The DeepSeek wild card

The single most disruptive event in AI economics since the original ChatGPT launch was the release of DeepSeek R1 on January 27, 2025. A one-year-old Chinese startup released a frontier-grade reasoning model trained, by their reported figures, on $5.6 million in compute — versus the hundreds of millions to billions American labs typically spend.

The market response was immediate and brutal. The Nasdaq lost approximately $1 trillion in market cap in a single day. Nvidia alone lost $588.8 billion, the largest single-day market cap loss in stock market history, more than doubling the previous record (Meta's 2022 drop).

The bull rebuttal — articulated most loudly by Microsoft CEO Satya Nadella in an X post the same day — was Jevons Paradox: when a resource becomes more efficient and accessible, total demand for it goes up, not down. Cheaper inference means more applications, more users, more queries, more total compute consumption.

Nadella's framing has held up reasonably well over the eighteen months since. Inference demand has continued to grow even as model efficiency has improved. But the DeepSeek shock revealed something the bull case can't fully explain: the market priced in a tail-risk scenario where the capex assumptions break, and the magnitude of the reaction (~$1T in one day) reveals that institutional investors are not nearly as confident in the AI capex thesis as their behavior in normal trading suggests.

The energy constraint

The capex-and-revenue debate is somewhat downstream of a more concrete constraint: power. The AI data center buildout has run into the limits of the U.S. electrical grid faster than most forecasts anticipated.

Microsoft's 2024 deal with Constellation Energy to restart the Unit 1 reactor at Three Mile Island — the site of the 1979 partial meltdown — is the most-cited example. The 20-year power purchase agreement gives Microsoft 835 megawatts of dedicated carbon-free electricity for AI workloads. The reactor (now renamed the Christopher M. Crane Clean Energy Center) is ~80% staffed as of early 2026 with a 2027 restart target. Meta announced a 6.6-gigawatt nuclear procurement strategy in early 2026 to support its "Prometheus" data center. Amazon, Google, and Oracle have made similar moves.

Power constraints function as a different kind of pressure on the capex thesis: even if you have the capital, even if you have the GPUs, you may not have the megawatts. That's a hard ceiling that doesn't show up in the spreadsheets and is shaping where data centers can actually be built. The "AI bubble" debate underweights this; the data centers that get built will produce revenue, but the rate of build is genuinely constrained by grid capacity in a way the dot-com fiber buildout wasn't.

Why this isn't (just) dot-com 2.0

The dot-com comparison is the most common bear framing and we want to be careful about it. The similarities are real: capex-to-sales ratios above the 2000 peak, hyper-concentrated market structure (S&P CAPE ratio at 38 with AI-related names driving most of the gain), and breathless infrastructure spending against revenue that isn't yet there.

The differences are also real. The companies making the capex bets in 2026 are profitable, cash-generative incumbents — not pre-revenue startups — and they're funding the buildout partly with retained earnings, not pure debt. The technology has demonstrable productivity use cases at the high end (developer tools, customer support, content workflows) even if 95% of pilots fail. And the demand-side curve for AI inference has been robust and growing, not the speculative "eyeballs" demand that dot-com priced in.

The honest version: this isn't a clean dot-com analog. Some pieces look identical (capex intensity, concentration risk, hardware-obsolescence exposure). Some pieces are genuinely different (incumbent profitability, real revenue at the model layer, productivity proof points at the high end of the enterprise distribution). The buyer who confidently calls this "exactly the same as 2000" and the one who confidently calls it "completely different from 2000" are both selling something. We don't think either is right.

What this means for AI tool buyers in 2026

The buyer's question is not "is there a bubble" — Pat Gelsinger, Sam Altman, and most of the Wall Street analyst community have already said yes. The buyer's question is "what does this mean for the products I'm evaluating and the contracts I'm signing this quarter?" Three practical implications:

Pricing will not get cheaper. It may get more expensive. The hyperscalers and the model labs are bleeding cash on infrastructure they need to recoup. Wells Fargo's May 2026 analyst note recommending investors "buy the bubble" included a corollary thesis that prices for AI capacity will rise, not fall, as the capex burden flows through to end-user pricing. The 2024-25 era of free-tier expansion and aggressive token-price cuts is over. Plan for it.

Vendor stability matters more than it has in any prior software cycle. The AI tool category has more cash-burning startups than any comparable software category in history. Some will become enduring businesses; many will be in our AI Graveyard by 2028. When evaluating a multi-year contract, the question of whether the vendor will exist in three years is now a real input.

The ROI bar is rising. If 95% of enterprise GenAI pilots produce no measurable P&L impact, your AI buying decision needs to be built around the 5% pattern that does work — narrow, well-defined use cases with named owners, integrated into existing workflows, with KPIs tracked. The "deploy it and see what happens" approach that worked for the SaaS era will not work for the AI tool era, because the products are expensive enough and unreliable enough that "see what happens" produces "see no return."

For our editorial view of the products that actually clear that bar, our editorial review methodology and Top 100 AI Tools are the place to look. For the broader AI agents conversation (which is the buyer category most directly affected by the capex-revenue gap), our AI agents 2026 piece is the companion read.

Frequently asked questions

Is the AI capex spending in 2026 a bubble? Several credible voices have said so directly. Pat Gelsinger (former Intel CEO) has stated "of course" we are in an AI bubble. Sam Altman has publicly acknowledged a bubble is underway. Goldman Sachs' Jim Covello has argued that the spending is producing inadequate returns. The honest answer is that the capex side has clearly bubble-like characteristics; the revenue side is growing fast enough that timing the pop is harder than calling that one exists.

How much are Big Tech companies spending on AI in 2026? The five largest Western hyperscalers — Amazon, Alphabet, Meta, Microsoft, and Oracle — are committing approximately $725 billion to capex in 2026, up 77% from 2025's record ~$410 billion. Most of that is AI infrastructure. Goldman Sachs and CreditSights project the trajectory to top $1 trillion annually in 2027.

What is the David Cahn $600B AI question? David Cahn at Sequoia Capital calculated in June 2024 that approximately $600 billion in annual AI revenue is needed to justify the capex being deployed. The math takes Nvidia's projected run-rate data center revenue, multiplies by 2 for total cost of ownership, and multiplies by 2 again for required 50% gross margin. The gap between $600B and what the AI industry actually generates in end-user revenue has been the bear case's central frame since.

What was the DeepSeek shock? On January 27, 2025, the Chinese AI lab DeepSeek released its R1 reasoning model, reportedly trained on $5.6 million in compute versus the hundreds of millions to billions American labs spend. The market response wiped approximately $1 trillion in US AI-related stock market cap in a single day. Nvidia alone lost $588.8 billion, the largest single-day market cap loss in stock market history. It was a stress test of the AI capex thesis that the market visibly failed.

What did the MIT NANDA study find? Published in July 2025, MIT Project NANDA's "The GenAI Divide: State of AI in Business 2025" examined 300+ enterprise AI initiatives and found 95% delivered zero measurable P&L impact, on roughly $30-40 billion in enterprise GenAI spending. The 5% that did succeed were concentrated in back-office automation: BPO replacement, agency-cost reduction, and structured operational workflows.

Is the AI capex bubble the same as the dot-com bubble? Some specifics overlap (capex-to-sales ratios above the 2000 peak, concentrated market structure). Some don't (the companies making the bets in 2026 are profitable cash-generative incumbents, and AI revenue is growing rapidly at the hyperscaler and model-company layers in a way dot-com infrastructure spending wasn't). Confident "exactly the same" or "completely different" claims should both be treated with skepticism.

What is Jevons Paradox in the AI context? The argument — most prominently from Microsoft CEO Satya Nadella — that as AI becomes cheaper and more efficient, total demand for it will grow faster than per-unit costs fall, so aggregate compute consumption will increase. The framing has held up reasonably well through 2025-26 even as efficiency has improved. It's a real counter to the DeepSeek bear case, but it doesn't resolve the question of whether end-user revenue will grow fast enough to justify the current capex level.

What does the AI capex bubble mean for tool pricing in 2026? Probably not lower prices. The hyperscalers and model labs need to recoup the buildout, and analyst commentary in 2026 increasingly assumes end-user pricing will rise, not fall. The 2024-25 era of aggressive free-tier expansion and token-price cuts is largely over for the leading models. Plan multi-year procurement accordingly.

Should I sign a multi-year AI tool contract right now? For incumbent vendors with diversified revenue (hyperscalers, established SaaS giants who've added AI), yes — the procurement math hasn't changed. For pure-play AI startups with concentrated revenue and high cash burn, multi-year contracts now carry vendor-stability risk they didn't in 2024. The 2026 buyer needs to evaluate not just the product but the company's ability to exist in three years.

Where to go next

The AI capex picture in 2026 is the macro environment our editorial work lives inside. For our view of the products and categories worth buying inside this environment, our Top 100 AI Tools is the place to start, and our methodology explains how we make the calls.

For the related industry-trend reads in our blog, our analysis of Google AI Overviews killing the open web covers the downstream impact on the content business that catalogs these products, and our AI agents 2026 piece covers the AI category most directly affected by whether the capex thesis pays off.

The honest version: nobody knows how this resolves. The capex is real, the revenue is real, the gap between them is real, the productivity gains are real, the 95% failure rate is real. Pat Gelsinger thinks the cycle has 2-4 more years before efficiency inflection points hit. Sam Altman has agreed there is a bubble while continuing to lead the company that's most aggressively expanding it. The job of the buyer in 2026 is to make decisions inside this uncertainty, not to wait it out.

— The ToolDirectory.AI editorial team