Anthropic's S-1: what AI lab money means for your Moodle

Short answer: Anthropic’s S-1 shows a company whose sales grew fourteen-fold in a year, whose operating loss is small next to that growth, and whose headline $80.7 billion loss is almost all an accounting charge on its own convertible notes. It also shows a business tied to compute contracts that run to 2045 and to prices it says it may have to cut, or rethink. For a school or university running AI inside Moodle, the lesson is not “AI will get expensive” or “AI will get cheap”. It is that per-token prices are set by companies under heavy financial pressure, and your setup should be able to switch provider without a rebuild.

This post covers what the filing says, how it sits next to OpenAI and Google, and what to do about it in Moodle. The model lineup and current prices are in our October 2026 model roundup; we do not repeat them here.

Nothing here is investment advice. We build Moodle plugins; we do not pick stocks.

What did Anthropic actually file?

On 26 October 2026 Anthropic announced that it had filed a registration statement on Form S-1 with the SEC for a proposed IPO of Class A common stock.¹1 Anthropic, Anthropic files S-1 with the SEC for proposed initial public offering (26 October 2026). The S-1 itself says the company has applied to list on the Nasdaq Global Select Market under the symbol “ANT”.²2 Anthropic, PBC, Form S-1 registration statement, SEC EDGAR (filed October 2026). The price range, share count and size of the raise are blank and will be filled in later amendments, so check EDGAR for the latest version.

What do the numbers say?

Every figure in this table is from the S-1’s consolidated statements of operations, in US dollars, millions. The press numbers you have seen ($16.3 billion revenue, $80.7 billion loss, $72.4 billion charge) are these, rounded. The only disagreement we found between reports is rounding of the prior-year figures.

H1 means the six months ended 30 June. The S-1 marks the H1 figures as unaudited; the full-year figures are audited.²

Three facts from the filing explain the table:

• Compute is the big cost, and it is falling as a share of revenue. Compute and infrastructure was 201% of revenue in H1 2025 and 55% in H1 2026, according to the S-1.²
• The $72.4 billion is a valuation effect. The S-1 says the change “was primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation”.² The separate $8.8 billion other expense comes from fair value accounting on note purchase agreements.
• The operating loss is $693 million on $16.3 billion of revenue for the half.

Our read: the $80.7 billion is real in GAAP terms, but it says more about how much investors now value the company than about how much cash it burns running models. The number to watch is the operating line, and that is close to break-even.

Who pays Anthropic, and how concentrated is it?

The S-1 says no single customer was more than 10% of revenue for the nine months ended 30 September 2026, and the top five customers were about 25%. It also says 79% of revenue came from outside the top five in the September quarter, up from 64% in the March quarter.² Customers paying $100,000 or more over twelve months went from more than 600 to more than 6,000 in a year, and they made up 74% of H1 2026 revenue.²

Education gets a few lines. The S-1 lists education among the industries using Claude and describes a Claude for Education offer for universities and K-12.² It is not a revenue line on its own. Our read: schools are a small customer for every frontier lab, and they get priced like one.

What do the risk factors say about prices?

This is the part that matters to anyone paying per token. The S-1 says Anthropic may change its pricing because of competition, regulation or how the market matures, and that “we may be required to reduce our prices, offer shorter contract durations or offer alternative pricing models, such as outcome or task-based pricing”.² It warns that open-weight models “could exert pricing pressure on our offerings”.²

It also says the opposite pressure exists. Agentic products are “structurally more compute-intensive”, and if Anthropic cannot set prices that cover that, its compute efficiency “could deteriorate as agentic usage grows”.²

On compute, the filing describes long-term contracts with minimum purchase and take-or-pay terms that “obligate us to pay for resources regardless of actual utilization”, with most of the remaining spend falling between 2031 and 2045.² The total committed amount is left blank in the public S-1. We will not fill it from press reports.

Our read: a listed AI lab has two forces on price. Competition and open weights push per-token prices down. Fixed compute bills and quarterly reporting push the company to earn more per customer, which tends to show up as tiers, surcharges, long-context pricing and “conditional” prices rather than a higher headline rate.

Who actually controls Anthropic?

Anthropic is a Delaware public benefit corporation. Its stated purpose is “the responsible development and maintenance of advanced AI for the long-term benefit of humanity”, and the board must balance that against shareholder returns.²

The Long-Term Benefit Trust is five financially disinterested trustees. Through a single class of stock, Class T, which has no economic rights, the trust can elect and remove directors who make up a majority of the board.² That makes Anthropic a “controlled company” under Nasdaq rules, and the S-1 warns buyers that the trust “may exercise its power in a way with which you disagree”.²

For a customer, this is mildly good news: a board that does not answer only to quarterly results is less likely to make sudden moves to hit a number. It says nothing about what a million tokens will cost.

How does OpenAI compare?

OpenAI finished its restructure on 28 October 2025. The nonprofit, now the OpenAI Foundation, controls a for-profit called OpenAI Group PBC and appoints its board. The Foundation holds about 26% of the equity and Microsoft about 27%, and OpenAI committed to $250 billion of Azure spending, The Register reported.³3 The Register, OpenAI non-profit will run for-profit that has yet to make a profit (28 October 2025).

On listing, Sam Altman told Fortune in September 2026, “I would say not 2026”, according to Proactive Investors, which also reports that OpenAI confidentially filed for a potential IPO in June.⁴4 Proactive Investors, OpenAI won’t go public in 2026, Altman confirms (14 September 2026). OpenAI’s revenue and loss figures in circulation are leaks and forecasts that disagree with each other, so we leave them out.

The difference that matters: Anthropic’s numbers are in an SEC filing, with audited full-year figures and unaudited half-year ones. OpenAI’s are not public. Until OpenAI files publicly, comparisons of the two companies’ finances are guesses.

And Google?

Google is the opposite case: a profitable company spending at a scale nobody else can. Alphabet’s third quarter 2026 results, released 28 October 2026, show:⁵5 Alphabet, Alphabet announces third quarter 2026 results, Form 8-K exhibit 99.1 (28 October 2026); also on Alphabet investor relations.

On the earnings call, CFO Anat Ashkenazi kept full-year 2026 capex guidance at $195 billion to $205 billion and said capex will “increase significantly in 2027”. She put the Google Cloud backlog at $561 billion, per the Motley Fool transcript.⁶6 The Motley Fool, Alphabet (GOOGL) Q3 2026 earnings call transcript. Google Cloud head Thomas Kurian mentioned TPU work with Anthropic on the same call. The labs are each other’s suppliers, customers and rivals at once.

Our read: Google can afford to price Gemini aggressively for years because search and cloud pay for it. Anthropic must show a path to profit on its own. OpenAI is somewhere in between, with Microsoft behind it. Three different money machines mean three different reasons to change prices, at different times.

What does this mean for a school running AI in Moodle?

Your bill is set by someone else’s balance sheet

Most institutions pay per token, through a provider plugin, with no contract that fixes the rate. The S-1 says plainly that pricing models may change. Google’s own prices already step up after promotional periods, as our model roundup showed. Plan for the price per request to move, in either direction, at least once a year.

Moodle is already built to switch

This is the good part. Moodle’s AI subsystem separates placements (where users see AI, such as the editor or course assistance) from providers (who runs the model). The developer docs put it bluntly: “Placements do not know about Providers, and Providers do not know about Placements. Everything should go via the Manager.”⁷7 Moodle HQ, AI subsystem, Moodle developer docs.

The provider list in current Moodle covers Amazon Bedrock, Anthropic Claude, Azure AI, DeepSeek, Gemini, Ollama and OpenAI, and the docs note that many other services “offer an OpenAI API-compatible layer”, which the OpenAI provider can reach through its API endpoint setting.⁸8 Moodle Docs, AI providers. Ollama runs open models on your own server. So swapping vendors is a settings change, as long as you have not built your teaching workflow around one vendor’s special features.

Where lock-in actually comes from

Lock-in with Moodle AI is rarely the API. It is:

• Prompts tuned to one model. A system instruction that works well on one model can get worse on another. Keep your prompts in version control and test them on two providers.
• Contracts and data terms. If your data protection review approved only one vendor, that vendor is your only option in practice. Get a second one approved before you need it.
• Vendor features outside Moodle. Assistants, file stores or agents set up in a vendor’s console do not move with you.

Budget controls are blunt

Moodle’s core AI settings include a site-wide and a per-user rate limit, both counted as requests per hour, according to the core language strings.⁹9 Moodle, core AI language strings, GitHub. That stops runaway use. It does not cap spend, because a request with a long course page costs far more tokens than a short one.

Today, the practical controls are:

1. Use a cheap, fast model for summarise and explain. They rarely need the top tier.
2. Turn on both rate limits and set them low at first.
3. Set a hard monthly spend limit in the vendor’s own billing console, if it offers one.
4. Keep a local model behind Ollama as a fallback for when a price rises or a budget runs out.

Local models are the hedge, not the default

A small open-weight model on your own hardware keeps student text in-house and has a cost you can predict: the server. The S-1 itself names open-weight models as a source of price pressure on Anthropic.² It will usually be weaker than the cloud tiers; as a fallback and a bargaining position, it is still worth having.

Where custom plugins come in

Core Moodle counts requests, not tokens or money. A per-course token budget, a cost report by department, or routing each request to the cheapest provider that can handle it are plugin jobs, either a provider plugin that wraps others or a local plugin that reads the AI action logs. That is the kind of work we build in days, not months. For the wider view on how AI is changing Moodle work, see Moodle developer experience in the AI era, and for where Moodle stands against other platforms, Moodle vs other LMS in 2026.