Read a country the way you read a company: GDP is the economy’s revenue, taxes are the government’s revenue, and bonds are how it covers the loss. Then follow the newest borrower in the room. Big Tech is spending about $750 billion on AI data centers this year, roughly every dollar its businesses throw off, and borrowing the rest from the same savers who fund governments.
World GDP, 2026
~$127T
the planet’s “revenue” (IMF forecasts, summed)
World debt
$365T
about 3.1× world GDP (IIF, mid-2026)
U.S. federal debt
$40.4T
+$2.5T in the last fiscal year, ~$6.9B a day
Big-4 AI capex, 2026
~$750B
Amazon, Alphabet, Microsoft, Meta guidance
Numbers move fast. Everything here is dated: market prices are from October 23, 2026; budget figures are the U.S. fiscal year that ended September 30, 2026; company figures are the latest filings (Meta has reported Q3; Alphabet, Microsoft, and Amazon report October 28–30). How to read the labels.
There are really two companies. USA Inc. is the whole economy: GDP is its revenue and household net worth is its equity. The federal government is a smaller company inside it whose revenue is taxes, which run about 17% of GDP. Countries rarely talk about “market cap” because nobody can buy the whole thing; the closest stand-in is the interest rate lenders demand.
| Company term | U.S. economy (USA Inc.) | U.S. federal government |
|---|---|---|
| Revenue | GDP $30.9T (2025) | Tax receipts $5.40T (FY2026) |
| Expenses | — | Outlays $7.40T |
| Net income | — | −$1.99T deficit (−37% “margin”) |
| Interest expense | — | $1.05T, 19% of revenue |
| Assets | Household & nonprofit assets $217.8T | $6.06T (cash, student loans, buildings) |
| Liabilities | Nonfinancial debt, all sectors, $84.1T | $47.8T (bonds $30.3T + pensions & veterans $15.5T) |
| Equity | Household net worth $195.9T | −$41.7T “net position” |
| Off balance sheet | Owes foreigners $22.4T net (NIIP) | Social Security + Medicare 75-yr gap $88.4T |
| “Stock price” | The dollar (DXY 101.75) | 10-year Treasury yield 5.00% |
| Leverage | Debt ÷ GDP ≈ 2.7× | Public debt ÷ tax revenue ≈ 6.0× |
Federal balance sheet from the FY2025 Financial Report (Sept 30, 2025); household and sector debt from the Fed’s Z.1, Q2 2026; budget from the September 2026 Monthly Treasury Statement. Leverage ratios derived.
Allowed: $32.6T public debt × (≈4.5% growth − 3.5% avg rate)assumed nominal growth; average rate on marketable debt is 3.52%
≈ $0.33TActual primary deficit, FY2026$1.99T deficit − $1.05T interest
$0.94TResult: the debt ratio rises about 2 points a yearCBO: 100.6% of GDP in 2026 → 120% in 2036
≈ +2 pts/yrIf the average rate climbs toward the 5% the market charges today, the allowed gap shrinks to zero. That is why rates matter as much as spending.
The planet produces about $127T a year and owes $365T. The U.S. and China are half of world output. Governments are now the largest borrower class, and almost every big government runs a deficit.
IMF Oct 2026 forecasts at IMF exchange-rate assumptions. Euro area uses the IMF April 2026 vintage. “Everyone else” is the remainder (derived).
Total ≈ 310% of world GDP. Up about $10T in the first half of 2026. The U.S. owes $111.8T of it across all sectors; China $72.5T.
IMF Oct 2026, all levels of government (so the U.S. figure includes states and counts the deficit on a calendar-year basis, 8.5% vs. 6.2% for the federal fiscal year). China shown both on the IMF standard measure and the IMF “augmented” measure that adds local financing vehicles; see §5.
Revenue engine: consumers & tech
Government
Market
Biggest absolute borrower. Pays market rates; foreigners and money funds are the marginal buyers.
Revenue engine: factories & investment
Government
Market
Debt is domestic and mostly owed to state banks. The risk is a slow Japan-style squeeze, not a run.
Revenue engine: exports & services
Government
Market
Low total, uneven inside: Germany 63%, France 119%, Italy 140%. France now borrows above Italy.
Revenue engine: savings & overseas assets
Government
Market
Highest debt ratio, cheapest money, but rates are rising for the first time in 30 years.
Revenue engine: finance & services
Government
Market
Pays the highest long rates in the G7. London custody is also why “U.K.” is #2 on the U.S. holder list.
IMF Oct 2026 general-government ratios. Dollar amounts are derived from ratio × GDP, except U.S., China, Japan, U.K. debt, which use the IMF-based dollar figures reported by Al Jazeera, and euro-area debt (Eurostat €14.44T at $1.1205). China’s augmented dollar debt is 135.3% × $21.5T (derived).
For every $1.00 the government took in during fiscal 2026, it spent $1.37. The gap was borrowed, and more than half of what was borrowed went to paying interest on earlier borrowing.
Monthly Treasury Statement, final September 2026. Health is mostly Medicaid. Net interest is budget function 900 ($1.05T); CBO’s “net interest on the public debt” is $1.14T. Customs duties are net of ~$145B refunded after the Supreme Court struck down the emergency (IEEPA) tariffs; gross collections were $318B.
Assets $6.1T
Liabilities $47.8T
Not on the sheet: Social Security and Medicare’s 75-year shortfall, $88.4T in present value. Also not on it: the power to tax a $31T economy.
THE REFINANCING TREADMILL
3.52% → 5%
Average rate on today’s marketable debt vs. the 10-year yield the market charges for new debt. $9.7T matured in FY2026 and had to be rolled.
RATE SENSITIVITY (DERIVED)
+$326B / yr
Extra interest, once fully rolled, for every 1 point higher on $32.6T of public debt. The $7.1T of bills reprice within a year.
CBO, NEXT 10 YEARS
$56T · 120%
Debt held by the public by 2036 and its share of GDP. Interest reaches $2.1T a year. If rates run 1.5 points higher: 133%.
Foreigners own about 29% of the debt held by the public. The rest is mostly Americans’ own savings, parked in money-market funds, banks, pensions, and the Fed. China has quietly cut its holdings by more than half since 2013.
Fed Z.1 Financial Accounts at holder valuation, so the Fed shows $4.08T here vs. $4.57T at face value. Excludes the $7.85T the government owes its own trust funds.
Treasury TIC data. Belgium, Luxembourg, Cayman, and the U.K. are partly custody centers for other countries’ money. China peaked near $1.32T in Nov 2013; it holds $590.7B now.
| Holder | Amount | Where that money originally came from |
|---|---|---|
| Foreign investors | $9.33T | Trade surpluses (Japan, Taiwan, Gulf oil exporters), recycled through central banks, pension funds, and custody hubs. Foreign official holders are shrinking ($3.92T → $3.76T in a year); private foreigners are growing. |
| Federal Reserve | $4.57T | Created as bank reserves during QE. QT ended Dec 1, 2025; the Fed bought ~$40B a month of bills for reserve management, then paused in August 2026. |
| Money-market funds | $3.28T | Household and corporate cash earning ~4% in bills instead of bank deposits. |
| Households & nonprofits | $2.83T | Direct holdings: savings bonds, TreasuryDirect, brokerage accounts. |
| Banks & credit unions | $2.04T | Deposits. Regulations count Treasuries as the safest liquid asset. |
| Mutual funds, ETFs | $2.55T | 401(k) and IRA contributions into bond funds. |
| State & local, pensions, insurers | $3.43T | Tax receipts held in reserve, pension contributions, insurance premiums. |
On paper China’s government owes 68% of GDP. Count the LGFVs that cities created to borrow off the books, and the IMF puts it near 127%. The engine that serviced that debt was selling land to property developers, and that engine has lost half its output.
Official (central + local)Ministry of Finance
¥96T · 68%IMF augmented totalat ¥6.67 per $ ≈ $26.6T (derived)
~¥177T · 127%LGFV ¥71.4T and the 51%/7% split are IMF estimates (Feb 2026 Article IV, via BOFIT); the Finance Ministry’s narrower “hidden debt” count was only ¥14.3T at end-2023. Local bonds have since grown to ¥59.9T (Aug 2026).
All urban land is state-owned. Cities sell 70-year use rights to developers.
2021. Cities used it to fund projects and to service LGFV debt.
Banks lend to city-owned firms on land collateral and an implied guarantee.
Useful, but most earn too little cash to pay the interest.
Evergrande ordered liquidated (2024); Country Garden restructured; land sales −52%.
¥10T plan (Nov 2024) moves hidden debt into cheaper official bonds, cutting rates >2.5 pts.
2024 back-calculated from the 2025 −14.7% change. Jan–Aug 2026: ¥1.38T, down 28.6% from a year earlier.
China’s “corporate” line is inflated by LGFVs, which are counted as companies.
Europe’s problem is not the total, it is the spread between members, and France has become the weak link. Japan carries the most debt of any big economy but owed almost nothing in interest until rates started rising.
Eurostat. Euro area total €14.44T (88.8%). The ECB’s bond portfolios (≈€2.9T of public debt) are running off, so markets must absorb more.
Debt service is 37% of tax revenue (derived). The 10-year JGB yield hit a 30-year high (3.1% at an Oct 6 auction) and the BOJ hiked to 1.25% in September. Rising yields at home give Japanese investors a reason to sell U.S. bonds: Japan’s Treasury holdings fell $116B in a year.
FRANCE
5.04%
10-year yield, highest since 2002; ~1.7 pts over Germany. Debt 119%, deficit 5.4%, minority government. Moody’s (Aa3, negative) was due to rule Oct 23.
GERMANY
€204B
Planned 2027 borrowing after loosening its debt brake: a €500B infrastructure fund plus defense spending above 1% of GDP exempt. Interest rises from €42B to €81B by 2030.
UNITED KINGDOM
£3.0T
Net debt (94% of GDP), passed £3T in September. Debt interest £60.8B in the first half of the fiscal year. Budget due Oct 28.
INDIA · BRAZIL
83% · 97%
Government debt to GDP. Both run deficits near 7.5–8%, but India grows ~7% a year, which shrinks the ratio; Brazil’s keeps rising.
Until 2024 Big Tech’s cash machines threw off far more than they could reinvest, so the surplus went to buybacks and cash piles. In 2026 nearly all of it goes into data centers, buybacks have stopped at Alphabet and Meta, and the gap is filled with bonds.
2024–25 actuals include finance leases (Platformonomics); 2026 is company guidance midpoints (Microsoft’s $175B is ~$190B before an accounting change); 2027 is analyst consensus. 2023 total: ~$155B. Add Oracle: $90–95B guided for its fiscal 2027.
2025 operating cash flow, big 4derived from company totals
$580B2025 capex72% of cash flow
$416B2026 hyperscaler capex vs. cash flowGoldman estimate: $750B vs. $778B
~96%What fills the gap and keeps dividends paidhyperscaler bonds, 2026 forecast (Goldman)
~$250BVanguard series (data to Jul 31, 2026). Consensus has spending flattening around $1.1–1.2T a year after 2027. That plateau, not a cliff, is what the market is priced for.
Morgan Stanley’s map of the ~$2.9T of data centers to be built from 2025 to 2028: about half comes from Big Tech’s own cash flow; the other ~$1.5T is borrowed or raised from outside. Hover the ribbons.
Channel amounts and the hardware/buildings split: Morgan Stanley (Mar 2026). The left column names who mainly stands behind each channel; it carries the same amounts and is a label, not a separate measurement. Did those funds have to sell something to lend? See below.
| Lender | Example | Whose money it really is |
|---|---|---|
| Bond funds | PIMCO anchored Meta’s $27.3B Hyperion bond (6.58%, due 2049, rated A+) | 401(k)s, pensions, and foreign buyers; $448B of foreign money flowed into U.S. bonds in the year to June 2026 |
| Life insurers | Athene (Apollo), Global Atlantic (KKR) buy private AI and data-center loans | Retirees buying annuities: sales topped $100B a quarter for 11 straight quarters |
| Private-credit funds | Blue Owl owns 80% of Hyperion; Blackstone led CoreWeave’s GPU-backed loans | Pensions, endowments, and increasingly retail investors through non-traded BDCs |
| Banks | JPMorgan and MUFG led a $38B loan for Oracle/OpenAI data centers | Deposits, then sold on to loan funds and private credit |
| Sovereign funds | Abu Dhabi’s MGX in OpenAI; Singapore’s GIC co-led Anthropic’s $65B round | Oil revenue and national savings |
| Suppliers | Nvidia put $30B into OpenAI; owns ~13% of CoreWeave; set up a $500B+ financing platform with Apollo, BlackRock, Blackstone, KKR and others | Nvidia’s own profits, which come from the same customers it is financing |
Mostly no. A fund’s cash comes from four taps, and only one of them means selling something else. Spending doesn’t destroy money either: it lands in someone else’s account and looks for a home again.
NEW SAVINGS · NO SALE
$990B
U.S. household saving per year (Aug 2026 rate): paychecks flowing into 401(k)s, pensions, and bank accounts. Plus $464B of annuities bought in 2025, which insurers lend out.
REPAYMENTS · NO SALE
≈ $1.7T
Corporate bonds repaid or called in a year: $2.2T were issued in 2025, but the stock grew only ~$0.5T. Add coupons, dividends, and $1.14T of Treasury interest. That cash arrives on its own and has to be reinvested.
ROTATION · YES, A SALE
$5.8B + $9.2B
SoftBank sold all its Nvidia shares and part of T-Mobile in 2025 to fund OpenAI, then took a $40B bank bridge loan for its next $30B check. Funds do the same at the margin: sell Treasuries or other bonds, or draw on ~$7.9T parked in money-market funds.
BANK LOANS · NO SALE
$38B
The Oracle/OpenAI data-center loan led by JPMorgan and MUFG. A bank loan creates a new deposit instead of handing over someone’s savings. Capital rules set the limit, not a pile of cash.
The pools overlap (some Treasury interest ends up as household saving), so read them for scale, not as a sum. Repaid bonds are derived: SIFMA’s 2025 issuance minus the growth in bonds outstanding. AI bar: Morgan Stanley’s ~$1.5T of outside funding for 2025–28, divided by four. Hyperscaler bonds: Capital Group.
That cash needs a home again, and it buys the next bond.
What gets used up is real stuff: chips, power, concrete, and engineers’ time. So the question is never “is there enough cash?” but “what yield do savers want for holding another $1T of AI IOUs instead of Treasuries?” That price is rising. Meta’s $30B bond drew a record $125B of orders in October 2025; by July 2026 hyperscaler order books were under 2× covered. Hyperscalers sold $241B of bonds in Jan–Aug 2026, against $321B of new 10-year Treasuries.
Buying shares from another investor sends the company nothing: for every buyer there is a seller, and only the price moves. Cash reaches AI firms only through their customers, new bonds and loans, or new shares (IPOs and private rounds).
Some of the money goes around in a loop: a supplier invests in a customer, which then buys the supplier’s product. The IMF counts 14 circular deals since 2023 carrying about $300B of equity and $650B of debt, against $1.4T of revenue commitments.
into OpenAI’s 2026 round
via Oracle, Microsoft, CoreWeave clouds
booked as Nvidia data-center revenue
Same shape: Amazon’s $50B into OpenAI with OpenAI committing $38B to AWS; Microsoft and Nvidia’s $15B into Anthropic with $30B of Azure commitments; AMD’s warrants for up to ~10% of AMD as OpenAI deploys its chips.
Roughly 60 cents of each AI data-center dollar buys chips and servers, and most of that lands at one company. Nvidia’s profit last year equaled almost a third of all U.S. federal corporate income tax.
McKinsey’s split of $5.2T of AI data-center capex through 2030.
Compare: U.S. corporate income tax receipts were $381B in FY2026. Fiscal year ended Jan 2026; latest quarter ended Jul 26, 2026.
An investment “works” when its users pay enough each year to cover wear and tear (depreciation), running costs, and a normal return to whoever lent the money. Here is roughly what that bill looks like around 2030, and four ways it could get paid.
Chips & servers: 60% of 2026–30 capex (~$5.4T), worn out over 5 yearsconsensus capex path from §7
$0.64TBuildings & power: 40% of 2025–30 capex (~$5.8T), over 20 years
$0.12TReturn to lenders and shareholders: ~9% on ~$3T still on the booksassumption
$0.27TElectricity, staff, and other running costsassumption
$0.15TRevenue AI needs to bring in each year
≈ $1.2TOthers’ versions: Sequoia’s “$600B question” (2024); JPMorgan $650B a year for a 10% return; Bain $2T by 2030. Today’s run-rate from the two largest labs is about $115–125B (OpenAI ~$40–50B, Anthropic ~$74B, counted differently).
$200 a month × 12the premium plan price
$2,400 / yr× 300 million Americans
$720BShare of the ~$1.2T bill covered
60%Equivalent to
4.6% of all U.S. payIt gets closer than you’d think, but it means every man, woman, and child paying a car-payment-sized bill. Today OpenAI has ~50M paying subscribers worldwide, mostly at $20. U.S. employee compensation was $15.7T in 2025.
AI does work worth 7.6% of U.S. pay; vendors keep half$15.7T × 7.6% × 50%
$600BJobs-equivalent at ~$100k average pay$15.7T × 7.6% ÷ $100k (assumption)
≈ 12M peopleShare of the U.S. workforce12M ÷ 170M in the labor force (Sep 2026)
≈ 7%That 7% matches Goldman’s estimate of 6–7% of U.S. workers displaced if AI is widely adopted. What it would do to unemployment is in §11. Cutting jobs at Big Tech itself barely matters: Meta’s ~8,000 layoffs in May 2026 save a few billion a year against $135–145B of capex. The savings have to come from customers.
12M jobs is about 7% of the 170M-person U.S. labor force. Whether that becomes a boom or a recession depends on one race: do the displaced workers find other paid work before their spending dries up?
TODAY · SEP 2026
4.2%
Unemployment: 7.1M people out of work in a 170.3M labor force. Payrolls grew just 29K in September.
IF NOBODY IS REHIRED
≈ 11.2%
(7.1M + 12M) ÷ 170.3M. Above the 2009 (10.0%) and 1982 (10.8%) peaks; below April 2020 (14.7%).
SPREAD OVER 5 YEARS
200K / month
Small next to the 1.6M layoffs and 5.2M hires that already happen every month. The churn absorbs it only if hiring keeps up.
PAY AT RISK
$1.2T / yr
7.6% of U.S. pay, about 4% of GDP. For scale, the 2007–09 recession cut payrolls by 9M, about 5%.
Scenarios hold the labor force at 170.3M and add 12M (or 6M) to today’s 7.1M unemployed. Goldman’s base case is gentler: 6–7% of workers displaced over about a decade, with unemployment only ~0.5 point above trend during the transition, because most people find new jobs.
Displaced workers cut back first: rent, cars, restaurants.
The top 10% own 88% of U.S. stocks and spend a smaller share of each extra dollar.
Restaurants, stores, and builders lose customers, so the job loss spreads beyond the first 12M.
About $0.3T a year of income and payroll tax rides on $1.2T of pay (my estimate), on top of today’s ~$2T deficit.
Fewer sales means fewer seats and less usage: the bust in §14.
In a normal recession the economy makes less. Here AI keeps doing the work, so GDP could hold up for a while: a “jobless boom.” The problem is buyers. If the income from that output flows to a smaller group who spend less of it, the output goes unsold and the boom turns into a slump. In past technology shifts, new kinds of work eventually absorbed the displaced, but over decades, not years.
Early signs: employment of 22–25-year-olds in the most AI-exposed jobs fell 13% relative to others after generative AI spread, while older workers in the same jobs held steady (Stanford, Aug 2025).
The happy path isn’t layoffs. It is the same workers producing more, selling more, and getting paid more, so there are buyers for the extra output and enough revenue left over to pay the AI bill.
Illustration: 2% a year vs. 2% plus 1.4 points, which compounds to Goldman’s ~15% productivity lift over ten years. The gap is ≈ $1.9T a year by 2030 and ≈ $5.7T by 2036. The ~$1.2T AI bill comes due around 2030, while the gain is still small: that timing is the squeeze. For reference, productivity grew 2.7% a year in 2000–07, 1.5% in 2007–19, and 2.2% in 2019–25 (BLS).
Acme Support (made up): 1,000 agents answering customer tickets for software companies, 250 days a year.
| Before | With AI | |
|---|---|---|
| Agents | 1,000 | 1,000 |
| Tickets per agent per day | 40 | 80 |
| Revenue at $15 a ticket | $150M | $300M |
| Pay (10% raises) | $100M | $110M |
| AI seats and usage | — | $15M |
| Office, sales, other software | $35M | $55M |
| Profit before tax | $15M | $120M |
| Corporate tax at 21% | $3M | $25M |
The extra tickets only sell if clients need them: new small clients who couldn’t afford support before, and existing clients who are growing (the loop below). Doubling is the optimistic end: a study of 5,179 support agents measured +14% tickets per hour on average and +34% for novices.
The lab pays its cloud; the cloud pays Nvidia, utilities, and builders; their lenders collect interest. This is the AI debt being paid.
Faster support cuts churn. It uses AI too, ships features faster, and sells more subscriptions.
Buys more BetaSoft seats, so BetaSoft sends Acme more tickets.
Acme’s +10% pay, BetaSoft’s new engineers, Gamma’s new warehouse staff.
Travel, home repairs, health care, more apps, and more of Gamma’s products.
Gamma sells more because households buy more, and the loop runs again.
Software can grow without much material. Tables can’t. If AI helps a furniture maker sell twice as many tables, the extra GDP comes with lumber, trucks, power, and eventually landfill. Whether that adds up economy-wide depends on whether the tables are new demand or won from a rival.
Made-up firm; per-table figures are assumptions.
| Before | With AI | |
|---|---|---|
| Tables sold | 10,000 | 20,000 |
| Price | $500 | $450 |
| Revenue | $5.0M | $9.0M |
| Lumber & materials ($150 a table) | $1.5M | $3.0M |
| Workers | 40 | 55 |
| Wood used (~40 kg a table) | 400 t | 800 t |
| AI tools (design, ads, scheduling) | — | $60K |
Unlike software, the extra sales need more inputs: twice the lumber, glue, finish, trucking, and factory power, plus more workers. AI helps find buyers and cut waste, but it doesn’t make wood. And it is a tiny cost line here, so a goods boom adds little to AI’s own revenue.
Only if total table demand grows. If Oak & Co. just wins customers from rivals, the industry uses the same wood. Demand grows when tables get cheaper or go out of style faster. That is the Jevons paradox: efficiency makes a thing cheaper, so people use more of it (Jevons on coal, 1865; Microsoft’s Satya Nadella on AI, 2025).
12.1M tons in 2018, 5.6× the 1960 level. 80% went to landfill, 19.5% was burned for energy, and almost none (40K tons) was recycled. Faster replacement makes this line steeper.
| Boom type | GDP | Jobs | Materials & waste | Emissions | Pays the AI bill? |
|---|---|---|---|---|---|
| Services and software Acme, BetaSoft (§12) | up | kept, with raises | little extra | some, from data-center power | yes, if pay rises with output |
| Physical goods Oak & Co. | up | more in mills and trucking | more lumber and landfill | up, unless power and freight are clean | a little: AI is a small cost line |
| Job cutting §11 | flat, then down | −12M | less buying | down, as in any recession | only briefly |
The giants survive a bust; their suppliers and lenders take the losses. If Alphabet, Amazon, Microsoft, and Meta stop building, their free cash flow jumps back into the hundreds of billions overnight. The pain lands on the companies that borrowed to sell to them.
Bond demand fades, rates stay at 5%, or a lab misses revenue targets. OpenAI plans to burn $25B (2026) → $57B (2027).
Labs scale back cloud contracts. Oracle’s backlog and CoreWeave’s debt depend on them.
Neocloud bonds, GPU-backed loans, private credit, data-center SPVs. Insurers and BDC investors take losses.
Nvidia’s $89B-a-quarter data-center sales fall; chip, power, and construction orders follow.
Mag 7 = 35% of the S&P 500; households own $52T of stocks, so the wealth effect hits spending.
Tax receipts fall, safety bids pull yields down; how far depends on inflation.
S&P 500 HIT (DERIVED)
≈ −30%
If the Mag 7 fall 50% and the other 65% of the index falls 20%. Dot-com: S&P −49%, Nasdaq −78%.
ECONOMY (IMF)
−1.8%
U.S. output lost by 2028 in the IMF’s AI-pullback scenario; −1.1% for the world. AI investment is ~25% of 2026 U.S. growth (Morgan Stanley).
DOT-COM PRECEDENT
~20¢
What telecom bondholders recovered per dollar after $500B+ of mostly debt-funded fiber; under 9% of the fiber was in use in 2001. The fiber still got used, a decade later.
| Borrower | Exposure | In a bust |
|---|---|---|
| Alphabet, Microsoft, Amazon, Meta | ~$500B of bonds outstanding; Alphabet alone holds $242.5B cash | Stocks fall, bonds fine. Cutting capex frees cash; buybacks resume. |
| Oracle | ~$125B debt, BBB− | Most exposed big name: spending ahead of contracts from one customer. |
| Neoclouds (CoreWeave and peers) | CoreWeave ~$52B of debt and leases | Restructuring likely; GPUs as collateral fall in value fast. |
| AI labs | OpenAI $852B valuation, Anthropic $965B; tens of billions in yearly burn | Down-rounds or mergers into a hyperscaler; commitments renegotiated. |
| Off-balance-sheet | $1.2T (Moody’s) to $1.65T (Nikkei) of leases and purchase deals | Lease contracts are where “hidden” leverage surfaces. |
| U.S. Treasury | $40.4T | Deficit widens as tax receipts fall. Each 1-point Fed cut saves ~$71B a year on $7.1T of bills. |
2026 broke the textbook. An Iran war (since February 28) sent oil above $110, inflation back to 3.7%, and the Fed raised rates in September for the first time since 2023. The 10-year yield touched 5.36%. Gold, which you might expect to soar, is 28% below its January record.
FED FUNDS
3.75–4.00%
+0.25 on Sep 16 under Chair Kevin Warsh. Markets price ~80% odds of another hike in December.
10-YEAR · 30-YEAR
5.00% · 5.33%
Peaked at 5.31% (close) Oct 5. Term premium ~+0.9 pt, vs. −0.45 average in 2014–21. Mortgage 7.32%.
GOLD
~$3,990
Record ~$5,590 on Jan 28. +64% in 2025. Central banks bought 863 tonnes in 2025; gold is now 27% of global reserves, above Treasuries (22%).
SILVER
~$54
Record $121.62 on Jan 29; −55% since. Sixth straight supply deficit expected (46 Moz), but solar demand is forecast −19%.
Treasury par yield curve, selected dates. Oct 22: 10-year fell 17bp in one day as AI stocks sold off (Nasdaq −2.6%, chip index −6.7%) after Meta raised its capex outlook.
Why gold fell anyway: 10-year real yields hit ~2.95%, the highest in about 18 years, which raises the cost of holding a metal that pays nothing; the dollar firmed; and investors sold gold to meet margin calls when tech fell.
| Scenario | Fed | 10-yr | Dollar | Gold | Silver | Stocks | Credit spreads | U.S. deficit |
|---|---|---|---|---|---|---|---|---|
| AI pays off productivity lifts growth and profits | holds; neutral rate higher | stays ~4.5–5% | firm | lags | industrial bid | up | tight | narrows |
| AI bust, inflation cools a 2001 replay | cuts fast | falls | weaker | rises as real yields fall | falls first | −30% or worse | widen | jumps |
| AI bust + oil shock stagflation, today’s risk | stuck | sticky | mixed | eventually | industrial drag | down | widen a lot | jumps |
| Debt scare investors demand more to hold Treasuries | pressured to ease | 5.5%+ | weaker | strongest case | follows gold | down | widen | interest spiral |
CBO: if rates run 1.5 points above its baseline, deficits are $6.0T larger over 2026–36 and debt reaches 133% of GDP. The BIS warns of “fiscal dominance”: heavy debt pushing central banks to go easy. The IMF says yields “may be normalizing back toward pre-GFC levels.”
$40T, +$2.5T/yr
U.S. federal debt, growing about $7B a day.
$1.05T
U.S. net interest in FY2026, more than defense ($965B).
3× GDP
World debt ($365T) vs. world output ($127T).
68% → 127%
China’s government debt, official vs. counting the city-owned LGFVs.
~$750B
Big-4 AI capex in 2026, ≈ all of their operating cash flow.
$120B → $1.2T
AI revenue today vs. roughly what it needs to earn by ~2030.
12M jobs
Rough labor-equivalent if half that revenue comes from replacing work.
5% · −28%
The 10-year yield, and gold from its January peak: 2026 rewrote the playbook.
Units and dates. $ = U.S. dollars; T = trillion, B = billion. U.S. budget figures are fiscal years (Oct–Sep). International ratios are IMF general government (all levels of government, calendar year), which is why the U.S. deficit is 8.5% there and 6.2% on the federal fiscal-year basis. Currency conversions use IMF rates of Oct 22, 2026 (¥6.674/$, $1.1205/€, ¥158.2/$, $1.338/£).
Derived figures. Leverage ratios, the debt-stability arithmetic, rate sensitivity, China’s augmented total in yuan and dollars, Big Tech’s 2025 cash flow, the S&P drawdown, and every line in §10–13 are my arithmetic on the cited inputs. The companies in §12–13 are made up to show the mechanics. Assumptions are labeled where they appear.
Scenarios are not forecasts. The ~$1.2T annual bill and the three revenue worlds use round, stated assumptions (5-year chip life, 20-year buildings, ~9% return, ~$150B running costs). Once spending plateaus, chip depreciation converges on chip spending (~$0.7T a year) whatever life is assumed; a shorter life mainly pulls costs forward during the growth years, which is Burry’s point. The rates matrix shows typical directions, and 2026 has already shown they can break (gold fell during an AI selloff).
Conflicts. Hyperscaler bond totals range from $93–121B for 2025 depending on which companies and currencies are counted; OpenAI and Anthropic report revenue differently; China’s debt ratio ranges 68–135% by definition. Where sources disagree, the page names the measure used.
Educational illustration only, not investment advice. Sources accessed October 23–24, 2026. Alphabet, Microsoft, and Amazon Q3 2026 results (due Oct 28–30) and the Fed’s Oct 28 decision are not reflected.