Read Anthropic's S-1 Like a Controller

By The AI Lab and Nathan Mapp

November 3, 2026

A printed financial filing on a desk with a calculator and reading glasses

Last week our community posted a thread on Anthropic's recently released S-1 filing. The perspective that anchored the conversation came from Nathan Mapp, a Controller in our community who is one of our trusted voices on where AI and finance meet. In the AI Lab Slack, a bunch of us along with Nathan picked the filing apart for the better part of a week. This is a summary of the insights from that conversation, including the final Q3 numbers from the amended S-1 Anthropic filed on November 2.

Usual disclaimer: this is not investment advice, and neither the AI Lab or our community members are your investment or financial advisors.

Ignore the headline loss (yes, really)

Ignore the headline GAAP losses. As of June 30, Anthropic's accumulated deficit was $132.9B. About $113B of that is accounting marks on convertible notes and related instruments, which get revalued every time the company's valuation goes up. Only about $19-20B is actual operating losses since 2021.

You can see this from Amazon's notes also: they are $6.2B of principal but sit on the balance sheet at $108.6B. So every time Anthropic gets more valuable, GAAP books a bigger loss on instruments like that. The $23.3B Q3 net loss is mostly the same thing again. That's why it is best to analyze this company from the cash flow statement.

The compute bill

The amended S-1 released November 2nd puts total compute and infrastructure commitments at about $895B, up from about $587B in the first filing. Q3 revenue was about $17.0B, so that's roughly 13x annualized revenue. About $447B comes due before the end of 2030 and about $70B in the next 12 months. Around $214B carries termination rights or other conditions, which still leaves about $680B owed, more or less, no matter what.

Against that, Anthropic had about $147B of liquidity at September 30 (cash, short-term investments and an undrawn $15B credit line) and roughly $7B of operating cash flow for the first 9 months of the year. If you annualize Q3's operating cash flow, the $895B they have committed works out to something like 70 to 90 years' worth. No private round or funding source on earth could fund that, which is why they have to go public now.

The financing gets worse when the business does

A lot of Anthropic's backup financing only shows up if they keep hitting targets. Google's $20B unlocks when revenue targets are met, and Amazon's $15B when compute milestones are. Broadcom's $42B facility, which exists to help pay rent on Broadcom's own chips, charges 9% interest that is paid in more debt instead of cash (called "PIK," or payment in kind), and it can convert into stock at a 10% discount. If they miss their revenue targets, the interest jumps to 13% and the discount to 15%. By estimate, this one facility alone could mean roughly 4-5% dilution for shareholders at today's roughly $1T valuation, and more dilution if their interest compounds.

This means funding can get more expensive and more dilutive at the exact moment the business is weakest, which is a combination reserved only for the world's most constrained and valuable resources (in this case, compute power). This dilution also happens on top of dilution that's already coming: about 2.9B employee stock options and RSUs, roughly $26B of stock compensation expense that starts at the IPO, Amazon's notes converting at $2-5 a share, and a 2% share issuance to charity every time the stock rises 50%.

The product goes stale fast

More than 80% of Q3 revenue came from models released in the prior 6 months, and each new model can accomplish the same work as prior models with fewer tokens. This efficiency and access to cheap intelligence is what customers pay for, but by way of example the move from Opus 5 to Opus 5.5 cut cost per task by about 40%. This is great for customers but means in the case of Opus models that usage has to grow 1.7x just for revenue to stay flat. Essentially, growth has to keep compounding extremely rapidly just to outrun efficiency gains.

It also means most of the revenue has to be won again every quarter, a far cry from recurring revenue models propping up software business models. Approximately 80% of revenue is pay-as-you-go usage, and only ~35% of that is locked in under contract minimums. The number to watch going forward is a familiar one: revenue growth rate. Already, Q3 added about $5.4B of new quarterly revenue, which is down from $6.8B in Q2. Not a crisis, revenue growth is a number that cannot slow down for Anthropic.

On the positive side

Customer concentration fears look overblown. No single customer was more than 10% of revenue so far in 2026, and the top five fell to 21% of revenue within two quarters. The dependency that's real is the channel: 48% of revenue flows through the Amazon, Google and Microsoft cloud marketplaces.

And the revenue engine is remarkable. Gross margin (that is, what's left after the cost of serving compute to customers), went from 12% in the first half of 2025 to 85% a year later. That happened during the same stretch that revenue grew 14x and net revenue retention (NRR) ran over 900%. This roughly means existing customers were spending nine times what they spent a year earlier. Also at the same time, revenue per dollar of compute went from $0.40 to $2.08 in five quarters, and Anthropic disclosed that more than 90% of the code written internally in September was written by Claude. So the virtuous feedback loop being seen is better models make serving cheaper, then cheaper serving frees up compute for the next model, then that next model comes out more efficient again and further compresses costs and improves margins.

Serving now costs Anthropic about 15 cents per dollar of revenue. If total compute costs continue to trend in this way and hold their efficiency, the original $587B of commitments could conceptually support up to $3.9 trillion (with a T!) of revenue over the life of the contracts. Or, said another way, the large spending commitments which look scary are also what make the upside possible.

Where the next leg of revenue comes from

The natural follow-up is: what's the best case for Anthropic actually succeeding? The answer from our discussion appears to be that they have barely tapped revenue beyond large companies (enterprise).

74% of first-half revenue in 2026 came from customers spending $100K+ a year, which is how Anthropic defines "enterprise." The other 26% came from 40 million smaller customers, which roughly calculates to $18 each a month. Much of those 40 million are probably small businesses or growing startups (like the AI Lab!), so true consumer revenue is likely smaller still, which also means it is as-yet untapped.
Subscriptions, including the Pro and Max plans that individuals buy, were only $3.2B, or 20% of first-half revenue.
Even big companies are early. Fortune 500 customers spend less than 0.1% of their total expenses on Claude.

From there, the "bull case" from the perspective of Anthropic and their investors would likely follow a path like this:

The efficiency loop keeps compounding.
Compute already committed is "unlocked" to also serve consumers and professional services, which would be mostly new revenue.
The best models and efficiency methods are kept proprietary and in-house, and Anthropic is able to keep them secure from external appropriation and distillation attacks.
Efficiency methods kept in house allow Anthropic to more cheaply maintain compute serving capacity while training new models, giving them an edge on peers who are pressured into buying more compute at inflated spot prices to keep up.
Internal model breakthroughs compound long enough that open-source models fall behind their current 4-5 month capability lag. This lets Anthropic's model-commoditization cycle stretch out and relieves competitive pressure.

What could break things

Anthropic is limited by how much trust they can create. Using healthcare as an example, this is where disruptors like tech companies and frontier labs have struggled most. Many of the Magnificent 7 companies have tried healthcare with poor results, because they show up selling shovels, or bring activity without defined outcomes. What wins there is trust, because trust is what gets you into the actual workflows. Using a kitchen analogy, consultants sell recipes, AI companies sell raw ingredients, but the value is in what the chef (the doctor, the lawyer, the expert) does with them. And frontier AI labs have shown little to no ability to create durable trust which would allow them similarly durable access to the workflows where their technology would be applied.

Beyond the need to create trust, Anthropic's biggest risks are mostly outside the company:

Open-weight models trail frontier model capabilities by 4-5 months. Once a model is "good enough" for a task, there's little reason to keep paying frontier prices, and the savings are hard to ignore. Anthropic doesn't own what customers build with Claude, so customers can take that work to a cheaper model later. This same dynamic feeds model distillation, where competitors train cheaper models on a frontier model's outputs and then compete directly.
Power. The US could be 15+ GW short on data center power by 2030, even if every planned build is delivered on time.
Data sovereignty and IP concerns. More companies want control of their own data and are building and training their own models. There are also ongoing concerns of IP ownership over model outputs, and Anthropic disclosed they are even offering uncapped IP indemnities for certain customers.
Insurance and legal exposure. Some catastrophic AI risks (even some disclosed in the S-1 itself) may be too big to insure, which raises the question of how companies using these models themselves will be covered.

Putting all of this together, it appears Anthropic can never pace the frontier, regardless of what they signal publicly. The compute is locked in, the financing is tied to growth, and the product goes stale in 6 months, with open source competitors 4 months behind. Staying ahead means enormous upside, while falling behind means paying for $895B of compute at inflated commodity prices while financing gets more expensive and more dilutive.

What this means for you

AI keeps getting cheaper. One model generation cut cost per task by about 40%, and open models trail the frontier by only a few months. Capable AI is becoming a cheap, everyday tool, and the cost of getting started is close to zero.
The skill gap is an opportunity. When the tool is cheap and everywhere, the advantage goes to whoever knows how to use it well. That's a skill anyone can build right now, without waiting for a corporate rollout or permission from leadership.
Trust is worth more as the tech gets cheaper. As the technology commoditizes, organizations built on trust, fairness and human judgment will be the ones that get invited into the workflows that matter and get access to the economic value.
Stay skeptical. More than 90 pages of the S-1 are risk factors, but these are only the risk factors that were fit to be explicitly written by the company also building the tools. The best advice is: learn how these tools work, use them to create value for yourself, and apply your own critical thinking to everything they produce.

Sources: Anthropic's S-1