Reading the AI Labs' S-1s

The two announcements, in full

These are short enough to quote almost completely, which is itself the point. Anthropic, 1 June 2026:

"Today, Anthropic, PBC confidentially submitted a draft registration statement on Form S-1 to the U.S. Securities and Exchange Commission for a proposed initial public offering of our common stock. This gives us the option to go public after the SEC completes its review. The proposed initial public offering will depend on market conditions and other factors. The number of shares to be offered and the price have not yet been set." anthropic.com/news/confidential-draft-s1-sec — Company

OpenAI, 8 June 2026, in a notice with a markedly different register:

"We recently submitted a confidential S-1. We expect it to leak so we're just announcing it. We have not decided on timing yet; it may be a while because there are things we want to do that are likely easier as a private company. But it's a complicated set of tradeoffs and this gives us the option to go public sooner if that ends up being best." openai.com/index/openai-submits-confidential-s-1 — Company

What a confidential submission actually is

The mechanism comes from the JOBS Act of 2012, extended by the SEC to all issuers in 2017. A company hands the SEC a complete draft registration statement — financials, risk factors, the lot — and the staff reviews it privately over several rounds of comment letters. The company fixes what the staff objects to before anyone outside sees a word.

The public filing happens later, and the rule is that the issuer must publicly file the registration statement at least 15 days before the roadshow. When that happens, the previously confidential drafts and all SEC comment correspondence get published too. So the document everyone wants to read does exist. It is simply not available, and the date on which it becomes available is not yet set.

Why both notices are so short: Rule 135

Both announcements carry an identical legal footer, and it is not boilerplate you should skip. Rule 135 under the Securities Act creates a narrow safe harbour: an issuer may announce a proposed offering without that announcement counting as an illegal offer to sell, provided it sticks to a short permitted list — the fact of the offering, its purpose, the amount and type of security, and the anticipated timing. Anything more is "gun jumping."

This is why Anthropic's notice reads like it was written by counsel: it was, and the sentence "the number of shares to be offered and the price have not yet been set" is there because Rule 135 requires that kind of restraint. OpenAI's is the more interesting document precisely because it strains against the form — "we expect it to leak so we're just announcing it" is a candid sentence in a genre that almost never contains one.

There is a large and confident body of writing about these two filings. Much of it contains precise figures — gross margins to the percentage point, loss projections by year, cash-flow break-even dates. None of it can be checked against a filing, because there is no filing. The numbers are some combination of investor-deck leaks, analyst modelling, and repetition.

This is not an argument for ignoring them. Reported numbers from good outlets are often roughly right. It is an argument for never letting a reported number and a filed number sit in the same sentence at the same weight, which is exactly what most commentary does.

The five tiers

Sort these yourself

Below are six claims currently in circulation about these two companies. Assign each a tier before you check. There is no partial credit and two of them are genuinely arguable — the reveal says which.

What each company has published itself

Start with the highest-quality material available: what the companies said in their own names. This is Company tier — curated, but not fabricated.

The reported financials, clearly marked

Everything in this table is Reported. I am including it because you need a working model, not because it is verified. Where sources conflict I have given the range rather than picking.

Where the two businesses genuinely diverge

The lazy framing is "OpenAI is consumer, Anthropic is enterprise." That is roughly true and not very useful. Three sharper distinctions:

1. Channel control versus channel reach

OpenAI owns its primary demand channel. ChatGPT is a direct relationship with hundreds of millions of users, which means pricing power, first-party data, and no intermediary — but also the full cost of serving a consumer free tier. Anthropic has taken the opposite bet: Company Claude is "the first frontier model available on all three of the world's largest cloud platforms." That is reach without ownership. The margin structures those imply are different, and neither is obviously better.

2. Revenue durability

Enterprise API and platform revenue is stickier than consumer subscription revenue but concentrates risk. Reported Anthropic counts 8 of the Fortune 10 as customers. That is an impressive sentence and also a customer-concentration disclosure waiting to happen: when a public S-1 arrives, the revenue-concentration paragraph is one of the first things to read. Consumer subscriptions churn but no single cancellation moves the quarter.

3. How each frames the moat

Read the two companies' own language. Anthropic's Series H release talks about Claude becoming "increasingly indispensable," about enterprises deploying it "in their core operations," and about Claude "learning how businesses actually operate: the context, the processes, the judgment." That is a workflow-embedding moat argument. OpenAI's public positioning leans on consumer ubiquity and model capability. Inference Workflow embedding is the more defensible claim if it is true, because it survives a competitor shipping a better model; capability leads do not.

The most common planning error in AI product economics right now is a reasonable-sounding syllogism: token prices have collapsed roughly 99% in three years; my product consumes tokens; therefore my AI cost of goods sold will fall.

Both premises are true. The conclusion does not follow, and the reason it does not follow is arithmetic rather than opinion. This simulator makes you commit to a number first, then builds the model with your own assumptions, then shows you the gap between the two. The gap is the lesson — not whether you were right.

(1 − d) × (1 + g) = 1

// which rearranges to the only line worth memorising:

gbreakeven = 1/(1 − d) − 1

Here is the part where a real filing is available — just not the one everyone is waiting for. Coursera's Q2 2026 Form 10-Q was filed on 5 August 2026 and you can read every word of it. Note 10, Commitments and Contingencies:

"Our purchase obligations primarily relate to third-party cloud infrastructure agreements, subscription arrangements, service agreements used to facilitate our operations, and paid advertising and sponsorship vendors. As of June 30, 2026, we had approximately $27.5 million in future minimum payments due under our non-cancelable purchase obligations with a remaining term in excess of one year. These are expected to be paid through 2030." Coursera, Inc. Form 10-Q, Q2 2026, Note 10 — Filed

Now put that beside OpenAI's reported figure. As of March 2026, roughly $665 billion in purchase commitments Reported, with longer-horizon estimates running to $1.4 trillion through 2033.

What pre-committed capacity does to pricing

Inference This is the load-bearing strategic consequence and it is worth being precise about, because the intuition most people have is backwards.

The common assumption is that huge commitments mean prices must rise to cover them. The opposite is closer to true in the short run. Contracted capacity is a sunk, take-or-pay obligation: the datacentre is paid for whether or not tokens flow through it. Marginal revenue on idle capacity is nearly pure contribution. So a lab with vast pre-committed capacity has a powerful incentive to fill it — which means aggressive pricing, generous free tiers, and volume deals.

That is a real subsidy, and it is why current API pricing is probably below long-run sustainable levels for the frontier tier. But it is a subsidy with a specific shape and a specific expiry:

• It persists while capacity exceeds demand. Cheap tokens are how you fill a contracted datacentre.
• It inverts when demand exceeds capacity. At that point the scarce resource is serving capacity, and rationing happens by price and by tier — which shows up first as rate limits and priority tiers, not as a headline price rise.
• Public markets change the clock. A private company can run negative gross margin on a strategic tier indefinitely. A public one explains it quarterly.

Where to look when the filings arrive

When either company does publicly file, four sections will tell you more than the rest of the document combined. In reading order:

1. Contractual Obligations table (in MD&A) and the Commitments note. This is where the take-or-pay compute deals become a legally characterised liability rather than a press release. Check whether they sit on the balance sheet or only in the note — that distinction is the entire question of how binding they are.
2. Cost of revenue detail. Whether inference compute is broken out separately, or buried in a single line, is itself a disclosure choice worth noting.
3. Revenue concentration in the risk factors and the segment note. "8 of the Fortune 10" becomes a percentage.
4. Related-party transactions. Microsoft is simultaneously OpenAI's largest shareholder, a major compute supplier, and a distribution channel. Amazon and Google occupy overlapping roles for Anthropic. Circular arrangements between an investor, a supplier and a customer are exactly what this section exists to expose.

OpenAI: the Foundation and the PBC

Company Reported OpenAI completed its recapitalisation in October 2025. The nonprofit, renamed the OpenAI Foundation, holds roughly a 26% equity stake in OpenAI Group PBC, the Delaware public benefit corporation that is the operating company. Microsoft holds roughly 27%; employees, former employees and investors hold roughly 47%.

The critical provision is not the equity split. It is that the Foundation's board holds special voting and governance rights that let it appoint every member of the OpenAI Group board. Microsoft, despite the largest single stake, reportedly holds no board seat.

Anthropic: the Long-Term Benefit Trust

Company Anthropic is a Delaware PBC governed in part by the Long-Term Benefit Trust — an independent body of financially disinterested trustees holding a special class of stock with the right to elect a growing portion of the board, ultimately a majority.

Reported The mechanism is a Class T share class with negligible economic value and substantial governance power, and the anticipated IPO structure would give public buyers ordinary common stock while the Trust retains its Class T rights. I want to be careful here: the existence and general design of the LTBT is well documented by Anthropic itself, but the precise share-class mechanics as they would apply post-IPO are reported, not filed.

With no public filings, we cannot analyse risk-factor language. What we can do is inventory what neither company has addressed voluntarily, in a period when both were actively communicating with the market. That inventory is a genuine analytical object.

1. Neither has published a gross margin

Anthropic published a valuation, a raise size, an investor list and a run-rate. It did not publish a margin, a cost figure, or a burn rate. OpenAI has published none of these directly. Inference For a company whose reported gross margin is in the 33–39% range against software comparables at 70–80%, this is the number with the most explaining attached, and both are letting others estimate it rather than setting the anchor themselves.

2. Neither has quantified copyright exposure

Unknown Both face active litigation over training data. Neither has published a reserve, a range of reasonably possible loss, or a contingency estimate. When a filing appears, the legal-proceedings note and the loss-contingency language will be the most carefully lawyered text in the document — and the phrase to search for is whether they can "reasonably estimate" the loss. A statement that they cannot is itself informative.

3. Neither has addressed key-person dependence in public

Both companies are unusually identified with a small number of individuals. Standard S-1s carry a key-person risk factor. Neither company's public communications acknowledge it. This is entirely normal pre-filing behaviour — nobody volunteers a risk factor — but worth flagging as an absence rather than an absence of risk.

4. Anthropic's compute disclosure is notably specific; its cost disclosure is not

The Series H release names gigawatts, partners, and chip suppliers with real precision: up to 5 GW from Amazon, 5 GW of TPU capacity with Google and Broadcom, GPU access via SpaceX, and memory relationships with Micron, Samsung and SK hynix. Inference That is a company demonstrating supply security to an audience worried about compute scarcity — and simultaneously not saying what any of it costs or over what term. The specificity of the first half makes the silence in the second half more legible, not less.

5. Neither has said anything about pricing trajectory

This is the omission that matters most to anyone building on these APIs. No published commitment on price stability, deprecation windows, or notice periods for enterprise terms. Inference That silence is rational — committing to price floors or ceilings before an IPO would be strategically foolish — but it means every downstream business is planning on an input whose supplier has made no forward commitment whatsoever.

What this means for education specifically

Education has a structural problem that most AI-consuming sectors do not: the willingness to pay per interaction is low and the number of interactions per outcome is high. A learner working through a course may generate hundreds of AI interactions across weeks against a subscription price that is fixed. That is the worst possible shape for exposure to a rising per-interaction cost, and it is precisely the shape that Module 4's arithmetic threatens.

Two consequences follow, and they point in different directions:

Five things to do differently

1. Instrument tokens per completed task, monthly, and put it in the operating review. Not tokens, not spend — tokens per completed task, alongside task success rate. This is the single measurement that distinguishes "we shipped more capability" from "we lost control of the loop," and almost nobody has it. Everything else in this list depends on it existing.
2. Re-underwrite any product where AI usage is unbounded and price is fixed. Identify every SKU where a single user can drive arbitrary inference against a flat fee. Model each at the Module 4 breakeven — flat unit cost, not falling. If a SKU only works on falling unit cost, it does not work.
3. Build the cost levers now, while they are optional. Prompt caching, model routing to smaller SKUs, distillation of high-volume paths, batch processing for anything not interactive. These are the difference between your effective d and the vendor's headline d. They take two to three quarters to build and you want them before you need them, not after.
4. Negotiate for throughput terms, not unit price. When these vendors are public, the negotiable surface shifts from price per token toward committed-use discounts, priority tiers, and rate-limit guarantees. Ask now for notice periods on deprecation and pricing changes, and get them in writing. A 90-day deprecation notice is worth more than a 10% discount.
5. Keep at least two frontier vendors genuinely wired in. Not as a slide — as a routing layer that is exercised in production on real traffic. Company Claude is now available on AWS, Google Cloud and Azure, which makes multi-vendor materially easier than it was. The value of this is not price arbitrage; it is that it converts a supplier relationship into a market.

The watchlist