ASI Brief

AI brief, Monday 26 October 2026

Anthropic files to go public and reports the most misleading loss in AI this year, Meta reveals what its new share class was actually for, Verizon buys $2bn of AI compute, and a coal producer explains why the interconnection queue is good for its business.

Anthropic files to go public, and reports an $80.7bn loss that mostly is not one

Anthropic filed a Form S-1 with the SEC on 26 October, applying to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. The filing carries no price range, no share count and no offering size — all are left blank, as they normally are in a first filing — so anyone quoting a valuation today is not quoting this document.

The headline number will be the net loss: $80.7bn for the six months to 30 June 2026. Read the line above it before repeating it. Over the same six months:

So the great majority of the loss is a non-cash accounting consequence of the company becoming more valuable. The same effect, smaller, runs through earlier years: net losses of $1.5bn, $8.3bn and $42.0bn for 2023, 2024 and 2025, of which $1.0bn, $5.4bn and $34.3bn were that fair-value line. Real costs are real — compute and infrastructure alone was $8.9bn in the half, and 2025 included a $1.5bn charge for the Bartz settlement — but they are not what the headline measures.

Two other things worth noting. Customers spending $100,000 or more on a trailing-twelve-month basis went from "more than 950" to "more than 11,900" in the year to 30 September 2026. And Anthropic is incorporated as a Delaware public benefit corporation, which it describes as "a novel capital and corporate governance structure" requiring it to balance stockholders' pecuniary interests against the interests of those materially affected by its conduct and a stated public benefit purpose.

Why it matters. Expect the $80.7bn to be reported as a cash bonfire; it is closer to the opposite, and the operating trend is the thing to watch. The structural question is the more interesting one: a public benefit corporation with a mandate to weigh non-shareholder interests is about to be priced quarterly by public markets, and nobody has tested whether that structure survives the experience.

Source: Anthropic, PBC, Form S-1, filed 26 October 2026 (SEC EDGAR).

Meta says what its new share class was for: about $95bn of equity

Yesterday's brief noted that Meta's board had approved a non-voting Class C share class, effective January 2027, and that the stated reasons began with "the scale of investment required to compete, sustain growth, expand infrastructure". Five days later, the mechanism is on file.

Meta announced two concurrent offerings on 26 October: $22.5bn of Class A common stock sold in connection with forward sale agreements, and 250,000,000 depositary shares each representing a twentieth of a share of new Series A mandatory convertible preferred stock, raising about $12.5bn — roughly $35bn together, with over-allotment options on both. The document then describes what comes next: a planned at-the-market programme of about $60bn, to launch "upon the commencement of trading of the Company's previously announced Class C common stock". Meta is careful to say no agreement for that programme exists yet and there is no assurance it will be established.

The subtitle on the filing is "Supports Meta's continued investment in AI infrastructure and personal superintelligence". It restates 2026 capital expenditure of $135–145bn, and sets out the balance sheet behind it: $97bn of operating cash flow in the first three quarters, about $82bn of cash and marketable securities at the end of the third quarter, and Aa3/AA− credit ratings that it says let it raise long-term debt efficiently.

Why it matters. The Class C share class was not housekeeping. It is the vehicle the $60bn at-the-market programme runs on, which is why the board's rationale led with the scale of investment rather than with governance. Taken together this is up to about $95bn of equity to part-fund a capex programme already guided at $135–145bn for one year — from the company with $97bn of nine-month operating cash flow. That is the clearest statement yet that the AI build-out has outgrown what even the most cash-generative companies in the world fund out of cash flow.

Source: Meta Platforms, Issuer Free Writing Prospectus, 26 October 2026 (SEC EDGAR).

A telephone company just bought $2bn of AI compute

Verizon's third-quarter results, out on 26 October, put an AI deal in the headline alongside the subscriber numbers. The company says it "signed an agreement valued at $2.0+ billion with AWS that establishes a commercial framework to scale distributed compute and meet the real-time demands of agentic and generative AI workloads".

For scale: Verizon's own capital expenditure was $11.9bn across the first nine months of 2026, which it describes as disciplined investment in mobility and broadband. A single compute framework worth over $2bn sits against that.

Why it matters. Most AI infrastructure numbers come from the five or six companies building the infrastructure. This one comes from a buyer — and not a technology buyer, but a network operator whose business is selling connectivity. The interesting word in the quote is "distributed": the workloads being provisioned for are real-time and agentic, which is an argument about where inference has to physically sit, not just how much of it there will be. Watch for the same line item appearing in other non-technology incumbents' results over the next two quarters; that is what demand broadening looks like before it shows up in anybody's revenue.

Source: Verizon Communications, third-quarter 2026 results, Exhibit 99 to Form 8-K, 26 October 2026 (SEC EDGAR).

The interconnection queue is a coal producer's best argument

Alliance Resource Partners, one of the larger US coal producers, reported on 26 October. Its chief executive Joseph Craft made the connection between AI demand and his own order book explicitly:

"Electricity load in the markets we serve is growing while permitting, interconnection and supply chain constraints are slowing the pace at which new generation can come online. We believe this dynamic will extend the runway for the existing coal fleet, with the impact becoming increasingly visible in our customers' coal burn as data center-driven demand materializes."

The commercial follow-through is in the same release: 4.2 million tons of committed and priced volume added during the quarter through 2028, and 31.7 million tons committed and priced for 2027.

Why it matters. Yesterday's brief measured the gap between the data-centre pipeline and the megawatts actually energised, using two utilities' filings. This is the same gap from the other end, and it shows the gap is not only a delay — it is an incumbency. If load arrives faster than new generation can be permitted and interconnected, the plants that are already connected run more, and the oldest ones retire later. That is a carbon consequence of the AI build-out that shows up in a coal producer's contracted tonnage years before it shows up in anybody's emissions disclosure.

Two caveats, because this is a seller describing its own market: an executive with 31.7 million tons to place has every reason to narrate a long runway, and the claim is testable — coal burn is reported, so the next few quarters will either show it or not.

That interconnection constraint is also the whole distance between a power proposal and a power contract for a single site. Ark7's Project Soperton in Georgia, for instance, publishes its power status as a written supply proposal still awaiting a signed service agreement — the queue Craft is describing is what makes those two different things. Ark7 is an advertiser on this site and we are paid if you open that link; see About.

Source: Alliance Resource Partners, third-quarter 2026 results, Exhibit 99.1 to Form 8-K, 26 October 2026 (SEC EDGAR).