Hyperscaler Balance Sheets: What a Balance Check Does and Doesn't Show

The IMF's October 2026 Global Financial Stability Report puts the borrowing of the biggest AI spenders in front of a wide audience. Its accompanying blog says it in one line: "Hyperscalers are borrowing more; investors have been leveraging to invest in AI stocks." Third-quarter reporting has started: Goldman Sachs held its call on 13 October. Before the technology companies' new filings arrive, we ran our balance check on the last balance sheet that Microsoft, Alphabet, Meta and Nvidia filed, and wrote down what that check can tell you about borrowing and what it cannot.

What a balance check is

A balance sheet lists what a company owns (its assets), what it owes (liabilities) and what belongs to its shareholders (equity). Because of the way books are kept, what the company owns has to equal what it owes plus what its shareholders own. A balance check runs that test on the figures taken from a filing: do the three agree?

If they do, the figures are consistent with each other and were very likely read correctly. If they do not, either we misread the filing or, rarely, there is a problem in it. A passing check is not a verdict on the company. The longer explanation is in what the accounting identity is.

Four filings, checked

These are the latest balance sheets we hold for the four companies, taken from the filings linked in the table. Microsoft's is an annual report, because its fiscal year ends in June; the other three are quarterly reports. Microsoft, Alphabet and Meta are commonly counted among the hyperscalers. Nvidia makes chips used in the build-out. Each of the four passed our check on the data we hold today.

"Equity per $1 owned" is equity divided by total assets, said the way you would say it to a friend: of every dollar of assets, how many cents belong to the shareholders. Total liabilities are the figures the filers state themselves. Equity includes any noncontrolling interest. Not financial advice.

What this does not show

Liabilities are not debt. Total liabilities count everything a company owes: supplier bills, taxes, leases and borrowings, among other things. Borrowing is a subset. Where a filer reports a long-term debt line, our API returns it as liabilities.long_term_debt, and the filing's notes describe the terms. We have left those figures out of this page on purpose.

The check cannot see what is not on the balance sheet. The equation holds for any correctly kept set of books, however much the company has committed to elsewhere. An obligation a company has not recorded as a liability does not move either side of the equation, so a passing check says nothing about it. Such arrangements are usually described in the notes to the financial statements, not on the face of the balance sheet. Press coverage of the IMF's report, such as this piece from Asia Times, quotes an analyst on commitments that sit off the hyperscalers' balance sheets. That is the part of the story that a balance check is blind to.

Our check can be wrong too. As of 13 Oct 2026 we have tested 38,568 filings, and 995 of them fail. A failure can be a misread on our side rather than a real accounting error, and telling the two apart is the point of the exercise. The write-up of those failures shows what they look like.

Run the same check on the next filings

When each company files its next quarterly report, its balance sheet will change. This script reads the latest balance sheet we hold for each of the four from our no-key demo endpoint and tests whether assets equal liabilities plus equity. For these four filings they agree exactly, to the precision each filer reports, so the script tests for exact equality. Other filings can be off by a rounding unit, and a wider run should allow for that. It prints the filing date and the result:

import requests

for ticker in ("MSFT", "GOOGL", "META", "NVDA"):
    r = requests.get(f"https://balanceproof.dev/api/demo/{ticker}", timeout=30)
    r.raise_for_status()
    d = r.json()
    assets = d["assets"]["total_assets"]["value"]
    liabilities = d["liabilities"]["total_liabilities"]["value"]
    equity = (d["equity"]["total_equity_incl_nci"]["value"]
              or d["equity"]["shareholders_equity"]["value"])
    balances = assets == liabilities + equity
    print(ticker, d["filing_date"], balances)

With a free key, swap the request for the keyed endpoint and send the key in an X-API-Key header. The rest of the script is unchanged:

r = requests.get(
    f"https://balanceproof.dev/api/company/{ticker}",
    headers={"X-API-Key": "YOUR_KEY"},
    timeout=30,
)

Run on 16 Oct 2026, both versions printed the filing dates in the table and True for all four. The demo is limited per address, so it suits a quick look rather than a loop. If you build AI agents, the same check is the check_balance_sheet tool in our MCP server. It returns a verdict and says on what basis it was reached. There is also a Python package, pip install balanceproof.

Where this fits

We cover about 6,246 companies (as of 13 Oct 2026). We check every filing we can test against the same equation, and we say so when one does not add up. For these four, the check passed. What each company owes, and to whom, is the next question, and the answer is in the filings, which are linked above and below.

Next step: get a free API key (1,000 calls per month, no card) and run the script on any ticker you are reading this earnings season. This is not investment advice.

Sources

• IMF, Global Financial Stability Report, October 2026.
• IMF blog, Global Financial Markets Face Rising Risks From AI, Bond Yields, and Cyberattacks, source of the quotation above.
• Asia Times, IMF warns of debt-fueled AI spending as bubble fears grow, 14 Oct 2026.
• Goldman Sachs third-quarter 2026 earnings call, transcript, 13 Oct 2026.
• SEC EDGAR filings: Microsoft 10-K, Alphabet 10-Q, Meta 10-Q, Nvidia 10-Q.