AI Bubble Monitor
Is AI a bubble? No single number can say, but the argument keeps coming back to five things you can measure: how much the hyperscalers spend on data centres, how they pay for it, whether revenue keeps up, how their chip suppliers are being paid, and what the market charges for all of it. Here are 13 indicators for those five, each ranked against its own history since about 2016 and averaged into one pressure score, with every definition printed below.
Statements through 2026 Q2 (hyperscalers) and 2026 Q2 (chip suppliers); prices to the Oct 16, 2026 close.
The pressure score by quarter
1 · Spending
pillar score 99How much of their cash are the hyperscalers putting into data centres?
Microsoft, Alphabet, Amazon, Meta and Oracle buy most of the world's AI chips and build the halls they sit in. Capital expenditure is that bill. When it grows faster than the cash their businesses throw off, the build-out stops being paid for out of pocket, which is where most AI bubble arguments start.
How it is measured
The five hyperscalers' trailing-twelve-month capital expenditure divided by their operating cash flow. 100% would mean every dollar of cash the businesses generate goes back into buildings and servers.
How it is measured
Trailing-twelve-month capex as a share of trailing-twelve-month revenue, the five hyperscalers summed.
How it is measured
Trailing-twelve-month capex against the same window a year earlier.
2 · Funding
pillar score 95Is the build-out still paid for in cash, or increasingly with debt?
Free cash flow is what is left after capex. As it shrinks, the gap gets filled by borrowing: bond issues, leases, and joint-venture vehicles that a balance sheet only partly shows. More debt does not mean a bust. It means less room for a disappointment.
How it is measured
(Operating cash flow − capex) ÷ revenue, trailing twelve months, the five hyperscalers summed. Ranked upside down: a thinner margin is more pressure.
How it is measured
The five hyperscalers' total debt (FMP's figure, which includes the lease liabilities it classes as debt) against a year earlier.
How it is measured
Debt issued minus debt repaid over twelve months, divided by operating cash flow over the same months.
3 · Payback
pillar score 99Is revenue keeping up with the spending?
A data centre pays off only if it earns more than it costs to run and to replace. Servers are depreciated over five to six years, so today's capex is tomorrow's depreciation charge. When capex runs several times ahead of depreciation and grows faster than revenue, an expense wave is building that future revenue has to cover.
How it is measured
Trailing-twelve-month capex divided by depreciation and amortisation. Well above 1× means the depreciation charge has yet to catch up with what was bought.
How it is measured
Year-on-year growth of trailing-twelve-month capex minus year-on-year growth of trailing-twelve-month revenue. Positive means spending is outrunning sales.
4 · Suppliers
pillar score 87Are the chip makers being paid as fast as they sell?
Nvidia, Broadcom and AMD are paid by the spenders, and increasingly invest in their own customers, who then buy their chips. That loop is what critics call circular financing. One trace of it is working capital: when receivables or inventory grow faster than sales, customers are taking longer to pay or product is piling up.
How it is measured
Nvidia, Broadcom and AMD's receivables divided by the quarter's revenue, times 91.25: how many days of sales customers still owe them.
How it is measured
The three suppliers' inventory divided by the quarter's cost of revenue, times 91.25: how many days of production sit unsold.
5 · Market
pillar score 47How much are investors already paying for it?
The other half of any bubble is price. We track what the eight companies are worth against their sales, how far they have outrun the S&P 500 over a year, and how far semiconductor stocks sit above their 200-day average.
How it is measured
The eight companies' combined market value (close × diluted shares) over their combined trailing-twelve-month revenue.
How it is measured
The equal-weighted price return of the eight companies over 252 trading days minus SPY's over the same days.
How it is measured
SMH (a semiconductor ETF) against its 200-day moving average: close ÷ MA200 − 1.
Following the money: the deals
not scored · ledger updated 2026-10-07Who invests in whom, and who buys from whom?
Circular financing is the charge that AI demand is partly paid for by its own suppliers: a chip maker invests in an AI lab, and the lab spends the money on that chip maker's chips, so the same dollars show up once as an investment and again as revenue. The deals are public. Each one below links to the announcement or the filing it comes from, and the arrows point the way the money goes.
A loop is not wrongdoing. Vendor financing is old, the amounts are often ceilings ("up to") or letters of intent, and several labs have real revenue behind their contracts. What a loop does tell you is that a slowdown at one end comes back to the other.
Who pays whom
8 loops of up to 3 parties
16 of the 23 deals sit on at least one loop; together they disclose $532.7B, counting each deal once. Amounts are what a party disclosed, often a ceiling.
- OpenAI → Microsoft → OpenAI $250B disclosed + 1 undisclosedOpenAI → Microsoft: compute contract $250B · Microsoft → OpenAI: equity investment
- Amazon → Anthropic → Amazon $109B disclosedAmazon → Anthropic: equity investment $4B · Amazon → Anthropic: equity investment $5B · Anthropic → Amazon: compute contract $100B
- Nvidia → OpenAI → Nvidia $100B disclosed + 1 undisclosedNvidia → OpenAI: equity investment up to $100B · OpenAI → Nvidia: chip order
- Microsoft → Anthropic → Microsoft $35B disclosedMicrosoft → Anthropic: equity investment up to $5B · Anthropic → Microsoft: compute contract $30B
- Nvidia → Anthropic → Nvidia $10B disclosed + 1 undisclosedNvidia → Anthropic: equity investment up to $10B · Anthropic → Nvidia: chip order
- Nvidia → CoreWeave → Nvidia $6.3B disclosed + 1 undisclosedNvidia → CoreWeave: capacity guarantee $6.3B · CoreWeave → Nvidia: chip order
- OpenAI → AMD → OpenAI no amount disclosedOpenAI → AMD: chip order · AMD → OpenAI: share warrant
- Nvidia → OpenAI → CoreWeave → Nvidia $122.4B disclosed + 1 undisclosedNvidia → OpenAI: equity investment up to $100B · OpenAI → CoreWeave: compute contract up to $22.4B · CoreWeave → Nvidia: chip order
The ledger
| Date | From → to | Kind | Amount | What was agreed | Source |
|---|---|---|---|---|---|
| 2026-07-28 | BlackRock funds → Meta | Joint-venture financing | — | A joint venture, 80% BlackRock funds and 20% Meta, for the ~$14 billion El Paso campus. | Meta press release |
| 2026-04-20 | Amazon → Anthropic | Equity investment | $5B | $5 billion now, with up to $20 billion more over time. | Amazon announcement |
| 2026-04-20 | Anthropic → Amazon | Compute contract | $100B | Anthropic expects to spend more than $100 billion on AWS over ten years, securing up to 5 gigawatts of Trainium capacity. | Amazon announcement |
| 2025-11-18 | Anthropic → Microsoft | Compute contract | $30B | Anthropic committed to buy $30 billion of Azure compute and to contract up to one gigawatt more. | Microsoft blog |
| 2025-11-18 | Nvidia → Anthropic | Equity investment | ≤ $10B | Nvidia will invest up to $10 billion in Anthropic. | Microsoft blog |
| 2025-11-18 | Microsoft → Anthropic | Equity investment | ≤ $5B | Microsoft will invest up to $5 billion in Anthropic. | Microsoft blog |
| 2025-11-18 | Anthropic → Nvidia | Chip order | — | Anthropic's compute commitment starts with up to one gigawatt of Nvidia Grace Blackwell and Vera Rubin systems. | Microsoft blog |
| 2025-11-03 | OpenAI → Amazon | Compute contract | $38B | A $38 billion, seven-year agreement for AWS compute on Nvidia GPUs. | OpenAI announcement |
| 2025-10-28 | Microsoft → OpenAI | Equity investment | — | After OpenAI's recapitalisation Microsoft holds about 27% of OpenAI Group PBC, a stake Microsoft values at about $135 billion. | Microsoft blog |
| 2025-10-28 | OpenAI → Microsoft | Compute contract | $250B | OpenAI contracted to buy an incremental $250 billion of Azure services. | Microsoft blog |
| 2025-10-23 | Anthropic → Alphabet | Compute contract | — | Up to one million Google Cloud TPUs, well over a gigawatt in 2026; Anthropic put it at tens of billions of dollars. | Anthropic announcement |
| 2025-10-21 | Blue Owl funds → Meta | Joint-venture financing | — | A joint venture, 80% Blue Owl funds and 20% Meta, funds the ~$27 billion Hyperion campus pro rata; Meta builds and runs it. | Meta press release |
| 2025-10-13 | OpenAI → Broadcom | Chip order | — | 10 gigawatts of OpenAI-designed accelerators built with Broadcom, deployed from the second half of 2026 to the end of 2029. | OpenAI announcement |
| 2025-10-06 | AMD → OpenAI | Share warrant | — | A warrant for up to 160 million AMD shares, vesting as OpenAI's deployments scale and as AMD's share price reaches set targets. | AMD 8-K exhibit 99.1 |
| 2025-10-06 | OpenAI → AMD | Chip order | — | 6 gigawatts of AMD Instinct GPUs from the second half of 2026; AMD expects tens of billions of dollars of revenue from it. | AMD press release |
| 2025-09-25 | OpenAI → CoreWeave | Compute contract | ≤ $22.4B | Three contracts (March, May and September 2025) for GPU capacity, together worth up to about $22.4 billion. | CoreWeave press release |
| 2025-09-22 | Nvidia → OpenAI | Equity investment | ≤ $100B | Letter of intent: Nvidia invests in OpenAI progressively as each gigawatt of Nvidia systems is deployed. | Nvidia press release |
| 2025-09-22 | OpenAI → Nvidia | Chip order | — | At least 10 gigawatts of Nvidia systems for OpenAI, the first gigawatt in the second half of 2026. | Nvidia press release |
| 2025-09-09 | Nvidia → CoreWeave | Capacity guarantee | $6.3B | An order with an initial value of $6.3 billion: Nvidia buys any CoreWeave capacity its customers leave unsold, through April 13, 2032. | CoreWeave 8-K |
| 2025-09-08 | Microsoft → Nebius | Compute contract | $17.4B | Dedicated GPU capacity through 2031 for about $17.4 billion, rising to about $19.4 billion if Microsoft takes more. | Nebius 6-K |
| 2025-07-22 | OpenAI → Oracle | Compute contract | — | 4.5 gigawatts of Stargate data-centre capacity developed by Oracle; neither company disclosed the value. | OpenAI announcement |
| 2025-03-03 | CoreWeave → Nvidia | Chip order | — | CoreWeave's fleet runs on Nvidia GPUs, which its customers' contracts specify; it buys them order by order, with no long-term supply contract. | CoreWeave S-1 |
| 2024-11-22 | Amazon → Anthropic | Equity investment | $4B | A further $4 billion, bringing Amazon's investment to $8 billion; AWS becomes Anthropic's primary training partner on Trainium. | Anthropic announcement |
What insiders and big investors did
not scored · from SEC filingsTwo more traces from filings. Officers and directors report their trades on Form 4 within two business days; most large sales run on 10b5-1 plans set up months ahead, which is why the share under a plan is shown. Fund managers report their holdings on Form 13F, 45 days after each quarter, so that half of the picture is always a quarter or more old.
Insiders: open-market trades
| Company | Sold, 90 days | Sold, 1 year | Bought, 1 year | Insiders who sold | Under a 10b5-1 plan | Last sale |
|---|---|---|---|---|---|---|
| MSFT Microsoft | $24.7M | $35.3M | — | 4 | 70% | 2026-10-13 |
| GOOGL Alphabet | $232.9K | $1.1M | — | 3 | 100% | 2026-09-30 |
| AMZN Amazon | $365M | $429.9M | — | 8 | 100% | 2026-10-01 |
| META Meta Platforms | $102.7M | $124.4M | — | 10 | 99% | 2026-10-12 |
| ORCL Oracle | $5.3M | $99.3M | $3.5M | 9 | 79% | 2026-09-22 |
| NVDA Nvidia | $967.6M | $1.6B | — | 8 | 15% | 2026-09-21 |
| AVGO Broadcom | $250M | $979M | $373.6K | 8 | 77% | 2026-09-23 |
| AMD AMD | $105.5M | $296.5M | — | 7 | 91% | 2026-09-15 |
| CRWV CoreWeave | $335.5M | $335.5M | — | 10 | 82% | 2026-10-13 |
| NBIS Nebius | $155.9M | $406.3M | $50.3K | 10 | 64% | 2026-10-12 |
Tracked managers
| Company | Managers holding | A quarter before | Bought or added | Reduced or sold | Net bought | Held, all managers |
|---|---|---|---|---|---|---|
| MSFT Microsoft | 38 | 39 | 18 | 20 | −$1.8B | $15B |
| GOOGL Alphabet | 52 | 54 | 11 | 35 | +$10.9B | $75.3B |
| AMZN Amazon | 36 | 36 | 11 | 22 | −$793.1M | $21.6B |
| META Meta Platforms | 33 | 31 | 14 | 16 | +$1.6B | $14.9B |
| ORCL Oracle | 15 | 13 | 3 | 10 | −$347.5M | $2.2B |
| NVDA Nvidia | 20 | 20 | 5 | 14 | −$1.9B | $8.2B |
| AVGO Broadcom | 15 | 15 | 7 | 9 | −$2.7B | $2.3B |
| AMD AMD | 9 | 5 | 6 | 3 | +$166.4M | $1.5B |
| CRWV CoreWeave | 2 | 3 | 1 | 3 | −$43M | $137.2M |
| NBIS Nebius | 2 | 1 | 2 | 1 | +$1.2B | $1.3B |
Who is spending, and who is being paid
| Company | Quarter | Revenue TTM | Capex TTM | Capex growth | Capex ÷ cash flow | Free cash flow TTM | Net borrowing TTM | Total debt | Debt growth |
|---|---|---|---|---|---|---|---|---|---|
| MSFT Microsoft | 2026 Q2 | $331.8B | $115.9B | +80% | 63% | $67B | -$3B | $128.8B | +15% |
| GOOGL Alphabet | 2026 Q2 | $445.9B | $132.4B | +98% | 71% | $53.3B | $70.1B | $112.8B | +171% |
| AMZN Amazon | 2026 Q2 | $775.7B | $173B | +61% | 107% | -$11.6B | $75.7B | $223.2B | +67% |
| META Meta Platforms | 2026 Q2 | $228.2B | $89.3B | +71% | 69% | $41B | $23.7B | $112.3B | +127% |
| ORCL Oracle | 2026 Q3 | $71.8B | $75.7B | +176% | 161% | -$28.7B | $37.2B | $155.9B | +48% |
| NVDA Nvidia | 2026 Q2 | $303B | $7.4B | +47% | 5% | $127B | $24.8B | $38.9B | +267% |
| AVGO Broadcom | 2026 Q2 | $89.1B | $1.3B | +146% | 3% | $39.4B | -$5.2B | $59.4B | −7% |
| AMD AMD | 2026 Q2 | $41.3B | $1.7B | +101% | 17% | $8.4B | -$1.5B | $4.3B | +10% |
Capex is purchases of property and equipment; free cash flow is operating cash flow minus capex; net borrowing is debt issued minus debt repaid. Amazon's finance leases and any joint-venture vehicles sit outside capex. Fiscal quarters are mapped to the calendar quarter holding most of each one.
Watch list: the GPU clouds
| Company | Quarter | Revenue TTM | Capex TTM | Total debt | Debt ÷ equity | Debt growth |
|---|---|---|---|---|---|---|
| CRWV CoreWeave | 2026 Q2 | $7.6B | $20.6B | $51.6B | 10.3× | +254% |
| NBIS Nebius | 2026 Q2 | $1.4B | $11.1B | $10.1B | 1.0× | +721% |
CoreWeave and Nebius rent out GPUs and borrow heavily to buy them; their customers include the hyperscalers above. Total debt here includes lease liabilities.
Watch list: credit
Investment-grade corporate bonds (LQD) against Treasuries of similar duration (IEF) over the last 63 trading days. Below zero means corporate credit is lagging, the way it does when spreads widen. Not scored: an ETF price gap stands in for the spread, which no source here publishes.
How the score is made
What this page cannot see
- AI revenue on its own. Cloud and data-centre segment revenue is not in the data source used here, so payback is measured against each company's total revenue.
- Debt outside the balance sheet. Data-centre joint ventures and leases that have been signed but have not started do not appear in total debt.
- The deals themselves. Chip makers investing in their customers, capacity guarantees and long-term purchase commitments are disclosed in filings, not in one figure. The suppliers' receivables and inventory are the trace this page can measure.
- Timing. Statements arrive four to eight weeks after each quarter, so a history point uses figures the market saw a month or so later. The three market indicators move daily.
The same nightly data feeds the market sentiment gauge (the whole market's mood) and the software board.
Questions
Is AI a bubble?
This page does not say, and nobody can until afterwards. It shows how stretched the AI build-out is against its own history: what the hyperscalers spend, how they fund it, whether revenue keeps up, how their chip suppliers are paid and what the market charges, each as a percentile of the quarters since about 2016. A high score says the numbers are unusual, not what happens next.
What does the score mean?
Each indicator is ranked against its own quarter-end history, so 90 on capex ÷ cash flow means that ratio is higher than in 90% of the quarters since the series begins. The indicators are averaged within each of the five pillars and the pillars into one score: 0–24 low, 25–49 moderate, 50–74 elevated, 75–100 extreme.
Which companies are in it?
The spenders are Microsoft, Alphabet, Amazon, Meta and Oracle. The suppliers are Nvidia, Broadcom and AMD. The market pillar prices all eight. CoreWeave and Nebius, the GPU clouds, are shown for their debt but not scored, because their filing history is too short to rank.
What is circular financing, and does this page measure it?
The term describes chip makers investing in AI companies that then spend the money on the same chips, or vendors guaranteeing their customers' capacity. This page keeps a ledger of those deals, each linked to the announcement or filing it comes from, draws who pays whom and lists the loops where money comes back to where it started. It also measures a trace in the numbers: whether the suppliers' receivables and inventory grow faster than their sales.
Why does the score not move with the news?
Ten of the thirteen indicators come from quarterly filings, so they move once a quarter, when the last company in the basket has reported. The three market indicators move with every close.
Where do the numbers come from?
Quarterly income statements, balance sheets, cash-flow statements and daily closes from Financial Modeling Prep, rebuilt every night after the US close. Total debt is FMP's figure, which includes the lease liabilities it classes as debt.
Is a high reading a reason to act?
It is a description of where thirteen series sit against their own history, and nothing more. It is not investment advice and it does not say what to do about it.
The weekly
The week's 13F filings, insider trades and earnings, plus new research on the site. One email a week, no account.
A reading describes where 13 series sit against their own history and is not investment advice. Updated once a day after the US close; the statement indicators move when the last company in a basket reports. Statements and prices from Financial Modeling Prep.