Morning Brief, October 9th 2026

Editorial

A revenue comparison without its accounting basis is a valuation argument with a missing premise. Gross sales describe money paid into a transaction. Net revenue describes the seller’s recognized share. Both can rise while the cash left after providing the service remains inadequate. Investors need the bridge between those quantities before deciding what multiple to pay.

Firmus’s withdrawal makes the cost of that discipline visible. An infrastructure plan can remain technically plausible while failing to attract equity on the offered terms. Returning to private investors changes who negotiates the price; it does not remove the need to finance the assets.

My conviction is that AI financing deserves more scrutiny at the contract level. Show which party owns the customer, who bears the computing bill, and which obligations survive if demand disappoints. My speculative extension is that businesses with clear retained margins will obtain better financing terms than those offering faster but less comparable growth. Comparable audited accounts showing the opposite would defeat that argument. Model capability alone cannot settle it. SourcesCB

The parallel run's editorial, on the same byline:

The industry spent this week learning what its revenue numbers are made of.

Start with the $20 billion. On September 29th, OpenAI's annualized revenue was approaching $70 billion; by Thursday it was approaching $50 billion. Nothing happened to the business in between. What changed, per the reporting, is which definition reached which journalist: one figure apparently absorbs cloud-partner sales the way Anthropic's does, one does not. In 1999 the equivalent game was gross versus net bookings, travel sites reporting the full airfare as revenue while keeping the commission. It took an accounting bulletin to stop it, and until then every comparison between companies was a comparison between definitions.

Private AI finance now runs on exactly such numbers. Run-rates annualize the best recent month. Backlogs count option-laden commitments as demand. And the comparison that matters most to investors, OpenAI against Anthropic, turns out to be computed on incompatible bases while both companies head toward public markets where the definitions are no longer theirs to choose. Anthropic's draft prospectus reportedly asks investors to price catastrophic model behavior; the harder near-term exercise may be pricing the revenue line.

Meanwhile the demand side of the buildout produced its first clean casualty figures: PC shipments down 20.1%, Apple trimming flagship orders because memory costs pushed its prices past what buyers would pay. The data centers are not stealing abstract capacity; they are outbidding laptops and phones for the same silicon, and the consumer is the residual claimant. That is a real transfer, visible in units, and it will show up in the memory makers' margins from the selling side next quarter.

My position: the first full audit of a frontier lab's revenue, Anthropic's S-1 if it lands this month, will show a materially less flattering number than the private figures that preceded it, and the market will absorb it anyway, because the alternative to believing the run-rates is having no numbers at all. I would be wrong if the S-1's audited figure matches the leaked $4.6 billion cleanly and the IPO prices above the $2 trillion talk without a discount. Either way, the week established the rule: in this market, ask who counted before you ask how much.