What an AI valuation actually means, and what it hides

What does a $190 billion valuation actually mean? Less than people assume, and the number is arrived at in a way that would surprise anyone expecting it to reflect the whole company.

Private AI valuations are being quoted like market capitalisations. They aren’t the same thing, and the difference matters when you’re trying to judge whether any of this is sane.

A valuation is a price on a slice, not the whole

When a company raises at a valuation, investors have bought a small percentage at an agreed price. Multiply that price by all shares outstanding and you get the headline figure.

Nobody has tested whether the remaining 95% could be sold at that price. In a public market, every share trades at the quoted price continuously. In a private round, one slice did, once.

A private valuation is the price of the last small transaction, extrapolated across everything that didn’t trade.

Work through a real one

Databricks raised $5 billion at a $190 billion valuation, up from roughly $134 billion six months earlier, with run-rate revenue past $7 billion and second-quarter growth above 80%.

Figure	Value	What it tells you
Valuation	$190B	Price of the latest slice, annualised across all shares
Raised	$5B	About 2.6% of the company changed hands
Run-rate revenue	$7B+	A recent period multiplied out, not audited annual revenue
Growth	80%+ year over year	The number carrying the multiple
Implied multiple	~27x run-rate	Defensible at 80% growth, not at 30%

That’s a real business with real customers, which separates it from most bubble comparisons. The question isn’t whether the revenue exists. It’s whether 27 times it is a sensible price.

CNBC reported the proceeds are earmarked for products that help businesses build and run AI agents, which is the demand story the multiple is underwriting.

Worth noting the valuation rose 42% in six months while revenue grew 80% annually. Those are compatible, and they mean the multiple expanded roughly in line with the business rather than ahead of it.

Run-rate is the word doing the most work

Annualised run-rate takes a recent period and multiplies it out. For a company growing 80% a year, that flatters the figure substantially, because it projects the best month across twelve.

It also says nothing about retention. Revenue that recurs and revenue that arrived once look identical in a run-rate number, and only one of them is worth a high multiple.

None of this is deceptive. It’s simply a metric chosen by the company, reported on its own timing, with no auditor attached.

Chief executives use the moment for wider claims too. Databricks boss Ali Ghodsi paired the round with an assertion that artificial general intelligence has already arrived, per Forbes. Treat statements attached to a fundraise as positioning.

Who bought matters more than the number

The single most useful detail in any round is whether new investors participated or existing ones marked up their own position.

Investor demand ran far ahead of the raise, with chief executive Ali Ghodsi telling TechCrunch the company set out to raise $1 billion and saw $15 billion of interest. That gap is the clearest read on sentiment available.

Coatue led the Databricks round with Blackstone, MGX and T. Rowe Price alongside, plus new names including Sixth Street Growth, BOND, Clearlake, Point72, Premji Invest and TPG, per CNBC.

Fresh outside capital at a higher mark is a genuine repricing. Existing holders writing up their own stake is an accounting entry, and the two get reported identically.

The terms nobody publishes

Here’s the part that makes headline valuations genuinely unreliable, and it rarely appears in coverage.

Late-stage rounds often carry liquidation preferences, guaranteeing investors get their money back before anyone else if the company sells. Add ratchets that issue extra shares if a later round prices lower, and the economics diverge from the headline.

A company can raise at a higher valuation on worse terms and report it as progress. Without the term sheet, which is private, you can’t tell the difference from outside.

Employees feel this most directly. Preferences sit ahead of common stock, so an exit that looks successful at the headline number can return considerably less to the people holding options.

Compute commitments are the hidden liability

AI companies carry an obligation ordinary software companies don’t, and it doesn’t show up in a valuation headline.

Long-term compute contracts commit real money regardless of usage. Nvidia’s arrangement of up to $500 billion of infrastructure financing, with the vendor guaranteeing collateral value, exists because those commitments are hard to fund otherwise.

The Motley Fool flagged the obvious catch, that a guarantee is only worth the guarantor’s willingness to honour it when it’s called.

The case that these prices are reasonable

The sceptical framing above needs a counterweight, because the bullish case isn’t stupid.

Demand is real and growing. Epoch AI tracks frontier compute expanding 5x per year since 2020, and that spending happens because customers are paying for outputs at the other end.

These companies also have revenue, which is the thing dot-com comparisons keep missing. A high multiple on growing revenue is a different animal from a high valuation on none.

And private markets have stayed open, letting companies raise repeatedly without the disclosure burden of listing. Whether that’s healthy is arguable. It’s certainly a choice they’re making deliberately.

Investor appetite for AI-linked companies has stayed strong through the year, Reuters noted, which is why second rounds inside twelve months keep clearing at higher marks.

The awkward part of that reading is what it implies about price discovery. If capital is abundant and the asset is scarce, valuations tell you about capital supply as much as about the company.

How to read the next AI valuation headline

Four questions get you most of the way, and none of them is the valuation itself.

Ask what percentage was actually sold, whether new investors led, what revenue basis the multiple uses, and whether the growth rate justifies it if it halves. Our piece on what a genuine correction would look like covers the signals worth watching.

The same scepticism applies a layer down, at the vendors selling into these companies. Our four questions that check an AI claim is the equivalent test for a product rather than a round.

The number that would tell you most is the one nobody publishes: a down round at a company with genuine revenue growth. That would separate a repricing of expectations from a business failing, and it hasn’t happened yet.