The numbers
| Company | 2026 capex plan |
|---|---|
| Amazon | about $220B (raised from $200B on 30 Jul 2026) |
| Alphabet (Google) | $195-205B (raised from $180-190B on 22 Jul 2026) |
| Microsoft | about $175B for calendar 2026 (29 Jul 2026) |
| Meta | $135-145B (Q3 2026 results) |
Company definitions differ: Meta and Microsoft include finance leases, and Amazon's total covers more than AWS (robots, chips, satellites). Microsoft's figure fell from about $190B only because of an accounting change, not lower spending.
Why it is unprecedented
Fact Measured against the size of the US economy, the peak years of past build-outs were: railroads about 2.2% of GDP, highways and the 1990s telecom boom about 1.1%, canals 0.66% and electrification 0.5%. One academic estimate puts AI at about 3.6% a year on average from 2025 to 2032.
- Speed. The four biggest spenders roughly tripled their yearly capex in two years. Fact
- Concentration. A handful of companies and one main chip supplier account for most of the spending, and they are also among the largest companies in the stock market. Fact
- Power. The IEA expects data centres' electricity use to more than double by 2030, to about 945 TWh, roughly what Japan uses today. Fact
- Financing is changing. The giants used to pay from their own cash. Now they also sell record amounts of bonds and use joint ventures and private credit, and Goldman Sachs expects about a third of 2026 hyperscaler capex to be funded with debt. Fact
Why are they spending so much?
The companies give four main reasons. Each is their thesis, not a proven outcome.
- Demand is bigger than supply. Amazon said in July 2026 that it will fall short of AI demand in 2026 and 2027. Google Cloud grew 82% in a year and reported a $514B backlog of signed contracts; AWS reported $496B. Fact
- Under-investing is the bigger risk. Mark Zuckerberg said in 2025 that if Meta ends up "misspending a couple of hundred billion dollars" that would be unfortunate, but the risk is higher on the other side: being late to superintelligence. Fact
- AI will do work, not just answer questions. The bet is that AI agents will write software, run customer service, do research and later drive cars and robots, so the market is a share of wages and services, not of software budgets. Our view
- Whoever has the most compute wins. AI labs believe bigger models trained on more compute keep getting better, so they sign years of compute in advance: OpenAI announced the $500B Stargate project in January 2025 and Sam Altman said in November 2025 that it had about $1.4 trillion of compute commitments over roughly eight years. Fact
How is it supposed to pay off?
| Who | How the money comes back |
|---|---|
| Cloud providers (Amazon, Microsoft, Google, Oracle) | Rent out AI computing by the hour and sell AI services to businesses, like cloud storage and servers before it. |
| Meta | Better ad targeting and more time spent on its apps; later AI assistants, smart glasses and business agents. |
| AI labs (OpenAI, Anthropic, xAI) | Subscriptions, API sales to developers and companies, ads and shopping inside chat, and agents that do whole jobs. |
| Chip makers (Nvidia, AMD, Broadcom) | Paid up front: they earn the money now, as long as their customers keep buying. |
| Further out | Robots, self-driving cars, AI-discovered drugs and computing in space. See the moonshot explorer on each company page. |
Value the AI-heavy companies yourself: SpaceX · Tesla · Anthropic · OpenAI
Payback calculator: how much AI revenue does it take?
The sliders are assumptions: the defaults are ours, the choice is yours. Assumption
To pay for $730B a year of data centres that last 5 years, earning 10% on the money tied up in them, with running costs at 40% of sales, AI customers need to spend about $1.52T a year. That is 3.7 times Apple's yearly revenue (about $416B in 2025), or about $185 for every person on Earth.
Formula: (yearly capex + return x capex x life / 2) / (1 - running costs). At a steady spending pace, depreciation equals one year of capex, and on average half the equipment's cost is still tied up. It is a rough check in the spirit of Sequoia's "AI's $600B question", not a forecast. Our view
Why are investors worried?
Spending runs far ahead of revenue
AI revenue is growing fast but is still far below the hundreds of billions a year going into data centres. If revenue doesn't catch up, the assets earn too little to justify their cost.
Chips age fast
AI chips improve every year, so today's GPUs may be worth little in 4-6 years. Several companies have lengthened the useful lives they assume for servers, which lowers yearly depreciation and raises reported profit. Critics such as Michael Burry argue this flatters earnings.
Fact Microsoft is extending the assumed life of its data centre buildings from 15 to 25 years from fiscal 2027.
More of it is borrowed
Bond sales, joint ventures and private credit now fund a growing share. Debt must be repaid even if AI disappoints, and some of it sits off the balance sheet where it is harder to see.
Fact Meta's Hyperion data centre joint venture raised about $27B of debt against a roughly $30B asset. S&P cut Oracle to BBB-, one notch above junk, citing its AI spending, and its 2055 bonds traded near 77 cents on the dollar in September 2026.
Circular deals
Suppliers invest in their own customers, who then buy their chips. That can make demand look stronger than it is.
Fact Nvidia announced in September 2025 that it intends to invest up to $100B in OpenAI as OpenAI deploys Nvidia systems; AMD gave OpenAI warrants for up to 160 million AMD shares alongside a 6 GW chip deal.
Prices could collapse
If cheaper models or chips arrive, AI prices fall and expensive capacity earns less.
Fact On 27 January 2025, after China's DeepSeek released a low-cost model, Nvidia lost about $589B of market value in one day, a record.
Power and permits
New campuses need gigawatts of power and grid connections that can take years; some communities and states are pushing back.
Is it paying off for customers?
Many companies are still testing AI rather than earning money from it. If their returns don't show up, they stop paying.
Fact An MIT study in August 2025 found that about 95% of company generative-AI pilots it looked at had no measurable profit impact yet.
Bull case and bear case
| The optimists say | The skeptics say |
|---|---|
| Demand still exceeds supply, and backlogs are in the hundreds of billions. | Backlogs include long contracts with a few AI labs that are themselves losing money. |
| The spenders are the most profitable companies in history. | Their free cash flow is shrinking and borrowing is rising. |
| Every big technology build-out (railroads, electricity, the internet) eventually paid off for society. | Many of the investors who paid for those build-outs lost money; society gained, shareholders often didn't. |
| AI will do work worth trillions in wages. | Competition and falling prices may pass most of that value to customers, not to the companies building data centres. |
Our view Our view: both columns are partly right. The technology is real; the open question is whether today's spenders earn enough on it, and how fast.
Lexicon: the words in the AI boom debate
- Capex (capital expenditure)
- Money spent on long-lived things such as buildings, chips, servers and power lines. It is paid now but counted as a cost slowly, through depreciation.
- Opex (operating expenses)
- Day-to-day running costs: salaries, electricity, rent. Counted as a cost in the year they are paid.
- Depreciation
- Spreading the cost of an asset over the years it is useful. A $5B cluster with a 5-year life costs $1B a year in profits.
- Useful life
- How many years a company assumes a server or building will earn money. A longer life means lower yearly depreciation and higher reported profit today.
- Hyperscaler
- The giant cloud companies that build data centres at huge scale: Amazon (AWS), Microsoft (Azure), Google and, for its own use, Meta.
- Data centre
- A building full of servers, cooling and power equipment. AI data centres are measured in megawatts or gigawatts of power, not square metres.
- GPU / AI chip
- The processors that train and run AI models, mostly from Nvidia, plus AMD and companies' own chips such as Google's TPU and Amazon's Trainium.
- Gigawatt (GW)
- A billion watts: roughly the output of a large power plant, enough for several hundred thousand homes. New AI campuses are planned at 1-5 GW each.
- Training vs inference
- Training builds a model once, at huge cost. Inference is every time someone uses it. Inference is where AI revenue is earned.
- Compute
- Computing power, usually rented by the hour or bought as chips. AI companies sign contracts for years of compute in advance.
- Raising debt
- Borrowing money, by taking a loan or selling bonds, that must be paid back with interest whatever happens. It does not dilute shareholders but adds risk if profits fall.
- Issuing equity
- Selling new shares to raise money. Nothing has to be paid back, but each existing share owns a smaller slice of the company (dilution).
- Bond
- A loan sold to investors that pays interest and returns the money at a set date. Its price falls when lenders worry about being repaid.
- Credit spread
- The extra interest a company pays over a government bond. A widening spread means lenders see more risk.
- Credit default swap (CDS)
- Insurance against a borrower failing to repay. Its price is a live gauge of how worried lenders are.
- Investment grade vs junk
- Credit ratings. BBB- is the lowest investment-grade rating; below it, bonds are called high yield or junk and many funds cannot hold them.
- Off-balance-sheet financing (SPV)
- Debt raised by a separate company (a special purpose vehicle or joint venture), so it does not appear as the main company's debt, even though the main company relies on the asset.
- Private credit
- Loans from investment funds rather than banks or the bond market. Faster and more flexible, but less transparent.
- Vendor or circular financing
- When a supplier invests in or lends to its own customer, who then uses the money to buy the supplier's products. It can inflate demand that would not exist otherwise.
- Backlog (remaining performance obligations)
- Revenue customers have signed contracts for but not yet been billed. A sign of future demand, but contracts can be delayed or cancelled.
- Free cash flow
- Cash from operations minus capex: what is left over for shareholders, debt and savings. Heavy AI spending has pushed it down, or below zero, at several giants.
- Return on invested capital (ROIC)
- Yearly profit divided by the money invested to earn it. The AI build-out pays off only if this ends up above the cost of the money.
- Run-rate
- A recent period's revenue scaled to a year. Useful when revenue grows fast, but it can overstate a full year.
- Market capitalisation
- Share price times the number of shares: what the stock market says the whole company is worth.
- Bubble
- When prices rise far above what the future cash flows can justify, fed by the expectation that prices will keep rising. Usually only obvious afterwards.
Questions people ask
Is AI a bubble?
Nobody can know in advance. The spending is real and mostly paid for by very profitable companies, and cloud providers say demand still exceeds supply. But the amounts are larger than any past build-out relative to the economy, more of it is now borrowed, and revenue from AI is still far smaller than the money going in. The calculator above shows how much AI revenue it takes to make the spending pay.
When will the AI bubble burst?
Nobody knows, and it may deflate slowly rather than burst. Signals to watch: cloud growth slowing while capex keeps rising, companies cutting capex plans, falling prices for renting AI chips, wider credit spreads on AI-linked debt, and AI labs struggling to raise their next round.
What is AI capex?
Capital expenditure on AI: the money companies spend on data centres, chips, servers, networking and power to train and run AI models. The four biggest spenders plan over $700 billion of total capex in 2026, most of it for AI.
How is the AI boom different from the dot-com bubble?
In the late 1990s much of the fibre build-out was paid for with debt by young telecom companies, many of which went bankrupt. Today most AI spending comes from the most profitable companies in the world, paid from their own cash. The similarities: very high expectations, spending far ahead of revenue, and a growing share of borrowing and circular deals.
Sources
- Alphabet Q2 2026 results (22 Jul 2026)
- Amazon Q2 2026 results (30 Jul 2026)
- Microsoft FY2026 Q4 results (29 Jul 2026)
- Meta Q3 2026 results, SEC 8-K
- Nvidia Q2 fiscal 2027 results (26 Aug 2026)
- Brookings: The $10 trillion question, financing the AI build-out (Van Nieuwerburgh, Sep-Oct 2026)
- Goldman Sachs: global AI investment forecast to exceed $1 trillion in 2026
- IEA: Energy and AI (April 2025)
- Sequoia Capital: AI's $600B question (2024)
- OpenAI: the Stargate Project (21 Jan 2025)
- OpenAI and Nvidia partnership (22 Sep 2025)
- OpenAI and AMD partnership (6 Oct 2025)
- Axios: Oracle rekindles the data center debt risks (25 Sep 2026)
- Sherwood News: Zuckerberg on "misspending a couple of hundred billion" (Sep 2025)
- Sam Altman on X, 6 Nov 2025 ($1.4T compute commitments); MIT NANDA, The GenAI Divide (Aug 2025); Michael Burry on depreciation (Nov 2025); Bloomberg and Reuters on Nvidia's market value, 27 Jan 2025.
Not investment advice. Stonkometer is an educational calculator, not investment advice. Valuations depend entirely on the assumptions you choose. We are not a broker or adviser, and nothing here is a recommendation to buy or sell any security. Do your own research and consider a licensed adviser. Full disclaimer