Anthropic's S-1 Is an Accounting Test for AI in 2026

Anthropic confidentially submitted a draft S-1 to the SEC on June 1, 2026, and as of today the public version still hasn't shown up on EDGAR. When it does, the first number everyone grabs will be revenue. The number I care about is where the training cost sits on the income statement. If the cost of training the frontier model gets booked mostly as research and development instead of cost of goods sold, the gross margin at the top of the page looks great while the business underneath it is still losing money.

That one line is going to set the tone for the whole AI trade this fall.

What the Data Shows

Fed chair Kevin Warsh put numbers on the boom in public at Jackson Hole on August 28, 2026. He said reports put annualized token sales for the two leading labs alone at more than $100 billion, an increase of 500 plus percent from a year ago, and that the four-quarter change in investment in equipment and intangibles has been around 9%, its highest growth rate since 2021. He also said more than half of that capex growth can likely be ascribed to the AI buildout. So the Fed chair named the revenue and the spending in the same speech, and never named the cost of producing that revenue, because nobody outside the labs can.

Barclays took a swing at it in a research note in late August 2026. Coverage of that note puts training spending at about 48% of AI lab revenue in 2026, down from 96% in 2024, and heading to 35% in 2027 and 30% in 2028 on their projections. It also runs two hypothetical frontier labs at $100 of revenue each and gets adjusted gross margins of 55% and 38%. I haven't read the note itself, only the write-ups, so treat those as reported estimates and not filed numbers.

Here's the thing about training spending. A lab has to keep training to keep the leading model, and having the leading model is a big reason it has the revenue in the first place. Money you spend every year just to protect this year's revenue looks a lot like a cost of that revenue. Book it in research and development anyway and you get a better gross margin, a friendlier tax treatment, and a line that investors have been trained to read as an investment in the future.

Why It Matters for Your Portfolio

This isn't a bookkeeping detail, because the AI trade is priced off margins. Nasdaq 100 futures closed at 29,232.50 on September 3, 2026 and are down 0.95% over the last 5 closes, while gold futures closed at 4,484.50 and are up 1.21% over the same stretch. That's not a market leaning into a big filing. If the first real income statement from a frontier lab shows most of the training bill sitting below gross profit, then the multiple on everything that sells into that lab gets re-read, and that means chips, cloud, and power.

What I'm Watching

Two things, both out of the filing itself. First is the accounting policy on training costs and how much of it lands in cost of revenue, because a small slice versus most of it is the difference between a profitable-looking business and a loss. Second is the trend across all of the periods shown, not just the newest one, since a single step up in usage can carry one period and then flatten out.

The public S-1 is the trigger. Until it hits EDGAR I'm watching, not calling it.

Related Reading: AI Data Center Debt Is Sitting in Insurance Money, 2026 and The AI Bubble's Early Tripwires Are Already Flashing

Howard is a full-time trader based in New Jersey with 13 years of experience across Forex, crypto, equities, and futures. He started Position Note to document his trades and analysis in public. All positions are disclosed. Nothing here is personalized investment advice.