_{ANTHROPIC S-1 DROPPED}

The Anthropic S-1 dropped this week, and it was the most anticipated IPO filing since… SpaceX’s in May.

In the first half of 2026, Anthropic posted a net loss of $80.7 billion, more than the entire GDP of Iceland. In those same six months, it generated $16.3 billion in revenue, committed to spending $637 billion on compute, and disclosed that its own technology might destroy civilization.

Buried in the risk factors, the S-1 warns that sufficiently advanced AI systems could, if not developed, deployed, and operated with appropriate safeguards, "pose catastrophic or existential risks to humanity, particularly if such systems were misused, replicated, weaponized." 

I do a deep dive below, but if you want the TL/DR:

But first, a word from our sponsors…

_{PRESENTED BY TOLTIQ}

The Busiest Deal Teams This Year Aren't at PE Firms

According to PitchBook's Q2 2026 Global M&A Report, corporate-led M&A held above $890 billion last quarter, most of the $1.3 trillion global total, while buyout value fell 35.7% from the first quarter. More strategic deals means more corporate development teams running diligence, usually with far fewer people than a PE firm would put on the same process.

That leaves a small team reading thousands of documents, while the legal, finance and business teams all need answers from the same files. ToltIQ turns the data room into one record the whole company can work from. Upload the full VDR, ask questions across every document, and each answer comes back with the page it came from. Your colleagues work in the same workspace instead of passing around versions, and after signing, the record stays in place for the integration team that inherits the deal.

See it on your own data room. Schedule a demo with their team today.

WTF is Anthropic

The first thing Anthropic tells you in its S-1 is that it is "first and foremost a research lab." Not a software company or a platform, but a research lab that happened to build the fastest-growing revenue base in the history of technology. Not bad for a bunch of nerds.

The stated mission is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity." This is not the boilerplate language some legal associate threw in v1 of the S-1. The corporate structure of the entire company including five classes of stock, an independent trust that controls the board, and a Founder LLC that controls half the votes is downstream of that sentence.

Anthropic was incorporated in January 2021. It is a Delaware public benefit corporation, or PBC. Here I was thinking every public company is a public benefit corporation, who else am I delivering shareholder value to?

The product is Claude, a family of frontier AI models. The company sells access to Claude through three products: Claude Platform (an API and developer tooling), Claude Code (a coding agent), and Claude Cowork (a knowledge-work agent that was released as a research preview in January 2026 and began merging with Claude.ai in September 2026).

Distribution runs through AWS, Google Cloud, and Microsoft Azure, plus systems integrators (Accenture, Deloitte, KPMG, PwC) and data platforms including Salesforce and Snowflake. The company recently launched something called Ode with Anthropic, described as an AI services company built with private equity partners and aimed at mid-sized companies.

The S-1 describes two long-term concepts that tell you how management thinks about the endgame. The first is the "Virtual Collaborator", an AI that functions as a domain expert across any field, available to any company or individual.

The second is a "country of geniuses in a data center", the notion that the aggregate cognitive output of sufficiently advanced AI systems would be comparable to the output of the most capable human workforce imaginable. These are not product features. They are the intellectual frame through which the company believes investors should evaluate the size of the market it is going after.

The revenue model is called "pay-for-intelligence." Customers pay per token, priced by model. It is consumption-based meaning you pay for what you use, metered at the level of individual inputs and outputs. The filing defines Run-Rate Revenue as trailing 28-day consumption fees multiplied by 13, plus monthly subscription fees multiplied by 12. While not the most outrageous thing seen in an S-1, that still goes to community adjusted EBITDA.

The Numbers

I am almost positive that Michael Burry was salivating when he opened this S-1. I am not saying whether there is an AI bubble or not, I have no idea (though the Chinese stock market may imply there is), but some of these numbers are truly ridiculous.

Revenue grew from $41 million in 2023 to $386 million in 2024, a casual 10x. Then to $4.6 billion in 2025, another casual 12x. Then to $16.3 billion in the first half of 2026 alone. H1 2026 revenue is 14x H1 2025.

For the quarter ended June 30, 2026, revenue grew 1,368% year over year. During that same quarter, compute and infrastructure costs grew 300%. Revenue is outrunning the primary cost of producing it by more than four to one.

The operating loss is collapsing. In H1 2025, Anthropic lost $2.1 billion on operations. In H1 2026, $693 million, on revenues that were 14 times larger. That trajectory implies operating profitability in the relatively near term on a GAAP basis, assuming growth continues on anything close to the current curve.

The net loss of $80.7 billion in the first six months of 2026, while absurd is not a business loss but rather is an accounting entry. Don’t you hate when accountants get involved? 

The S-1 shows the driver: $72.4 billion in fair value changes on convertible notes, derivative liabilities, and SAFEs. As Anthropic's valuation has risen dramatically, instruments issued to early investors at lower valuations are marked to their current value and the difference runs through the income statement as a loss. The company has not lost $80 billion. It has become significantly more valuable, and the accounting rules require it to record the difference.

That isn’t to say they are not burning money though. Personnel costs in H1 2026 were $3.2 billion for six months. That figure includes stock-based compensation, which was $1.2 billion for all of 2025 and $2.2 billion in H1 2026 alone, suggesting that the cost of retaining the people building the models is itself enormous and accelerating. Sales, marketing, and partnerships expense was $3.2 billion in H1 2026, up from $120 million in H1 2025. That is a 26x increase for those of you following along at home. The company is spending aggressively to acquire and distribute. The question the filing does not fully answer is how much of the 14x revenue growth is attributable to that spending and how much is organic from existing customers expanding.

I know a newsletter that would happily be a part of the partnership budget…just saying.

Cutting Through the BS

I mean the balance sheet, obviously. This is a finance newsletter, be respectful.

As of June 30, 2026:

• Cash, cash equivalents, and short-term investments: $131.7 billion

• Total assets: $147.4 billion

• Total liabilities: $117.5 billion

• Redeemable convertible preferred stock: $157.5 billion

• Total stockholders' deficit: ($127.6 billion)

• Accumulated deficit: ($132.9 billion)

• Total capital raised to date: $141.1 billion

Starting with cash, because it is king after all, $131.7 billion on the balance sheet means Anthropic is not raising this money because it needs it immediately. The IPO is a liquidity event for early investors and employees, and a public market listing that establishes a price and a currency.

The company has more cash than the GDP of most countries, and it is burning significant amounts of it on compute, but the operating loss trajectory suggests the burn will become manageable well before the cash runs out.

The compute commitments are disclosed as a total that exceeds a stated amount, but the filing leaves the number blank in the draft. The duration is clear: a portion due before December 31, 2030, and most of the remainder falling between 2031 and 2045. Multi-decade compute contracts. Anthropic has locked in infrastructure through mid-century. Glad Skynet will have someone to go to for compute when the time comes.

Per company filings and reporting across the counterparties, the fixed commitments total approximately $637 billion across seven counterparties: Broadcom ($161.2B, TPU equipment leases, 5 years), SpaceX AI ($120.5B, NVIDIA GPUs, cancellable on 90 days' notice after early 2027), Google ($114.5B+, April 2026 through July 2033), Amazon ($110B+, May 2026 through April 2036), Microsoft ($44.9B+, November 2026 through May 2033), Nscale ($44.6B), and Lambda ($41.3B).

Not counted in the fixed total is an additional estimated $125-150B in Google TPU compute for 2028-31, and a potential $20B+ AMD commitment if a full initial tranche is purchased.

The largest single line is Broadcom, a chip supplier, not a cloud provider. Anthropic is not simply renting compute from hyperscalers. It is leasing TPU equipment directly and building toward owning the infrastructure layer.

The SpaceX AI commitment, NVIDIA GPUs, cancellable after early 2027, is a hedge with an exit option built in. The Google and Amazon commitments run through the mid-2030s with no disclosed cancellation terms.

Now for those checking the leaderboard, a practical question may jump out at you. How on earth can a loss making company with annualized $33 billion of revenue spend roughly 20 times its current annual revenue on compute over the life of these contracts? Obviously the answer is growth, but it’s hard to fathom a company of this size being at an inflection point like it’s the HVAC company your MD pitched for a role up to the whole Street. Whether this commitment proves prescient or catastrophic depends entirely on whether the revenue trajectory holds.

The Customers

As of September 30, 2026:

• Customers with $100K+ TTM revenue: more than 11,900

• Customers with $1M+ TTM revenue: more than 1,900

• Customers with $10M+ TTM revenue: ~250

The filing does not use the word "enterprise" as a segmentation category, which appears to be a deliberate choice. The $100,000 annual spend threshold is how the company defines what it considers enterprise-scale adoption.

Net Revenue Retention: approximately 900%.

This number requires some unpacking because it is so far outside any established benchmark that it reads as a misprint.

NRR measures what existing customers spent this period as a percentage of what they spent the prior period. A figure of 100% means customers are spending the same amount. 130% means existing customers are spending 30% more.

The best enterprise software companies in history at their peaks, think Snowflake, Datadog, CrowdStrike, have reached 150-180%. 900% means that, in aggregate, Anthropic's existing customer base is spending nine times what it spent a year ago. The company added no new customers; it didn’t have to because existing ones spent nine times more.

The explanation is not just loyalty. Companies are discovering Anthropic's models for specific use cases, then they expand usage across functions, teams, and workflows as they build on the API.

The 28-day consumption model means every new internal deployment is immediate incremental revenue. A company that started using Claude for customer support, then added it to document drafting, then engineering, then analysis, does not sign a new contract for each. It just uses more tokens, and the meter runs.

TTM revenue per employee as of September 30, 2026: $6.6 million, up from $1.4 million a year earlier. The workforce is not growing as fast as the revenue. The company is becoming more productive per head by a factor of nearly five in twelve months. The S-1 describes this as a product of AI systems doing more of the work that human employees would otherwise do, so the models are helping build the models.

Closing the Loop

The S-1 describes a dynamic the company calls "closing the loop." Each Claude model contributes to building the next Claude model. The current Claude assists in research, code writing, documentation, evaluation, and the generation of training data for its successor.

The effect, as the company describes it, is that model improvement compounds: the better Claude gets, the faster the next Claude can be built, which gets better faster, and so on.

The filing also notes this creates a specific risk: if a model has systematic flaws, those flaws could propagate forward into successor models that were partially built using the flawed model's outputs.

This is also, by implication, an argument for the safety-first architecture. You want the models doing this recursive self-improvement to be as carefully aligned as possible, because errors scale.

The compute strategy is a direct expression of this compounding thesis. Anthropic runs a single compute pool across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. The company directs workloads across chip types to optimize for price and performance. The filing notes the company has in the past requested more compute from cloud providers than they were willing to supply. At the scale Anthropic is operating, there is a limit to how much any single provider can allocate. The multi-chip, multi-cloud strategy is both a hedge and a necessity.

The S-1 also cites Claude Opus 5.5 cutting customer costs per task by approximately 40% versus Claude Opus 5. This is both good and bad.

Good because falling costs per task drive adoption and usage. Customers buy more when each unit of output is cheaper. Bad because Anthropic is effectively giving away 40% of its value per token. The NRR of 900% suggests the volume effect vastly outweighs the price compression, but if anyone at this company cares about making money, the long-term margins are a real question.

Lawyers Always Win: The Governance

This section alone must have made all of the lawyers’ year. I have never seen this many letters in terms of share classes in my life. I assume all partners and even associates are adding pools to their Hampton houses after writing this monstrosity.

TL/DR: Anthropic has five classes of common stock. Three of them carry no economic rights.

Class A — one vote per share. Listed on Nasdaq as ticker ANT. This is what the public buys and will be going into your index fund. It carries economic rights to dividends and liquidation proceeds. There is one vote per share on matters other than the election and removal of Class T directors.

Class N — no votes. Converts to Class A upon transfer to an unaffiliated third party. Held primarily by strategic partners. Also no economic rights separate from any dividends or liquidation rights the board might declare. The S-1 notes that an Amazon convertible note will convert into Class N shares at the IPO.

Class E — issued to employees (excluding co-founders) after the offering. Limited voting rights. No economic rights. None outstanding at the time of the offering.

Class F — one share, held by the Founder LLC. The members of the Founder LLC are the seven co-founders: Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah. Class F carries 50.1% of voting power on all matters that are not the election or removal of Class T directors, but no economic rights.

Class T — held by the Long-Term Benefit Trust. The LTBT is an independent body of five financially disinterested trustees, expected to have expertise in AI safety, national security, public policy, and social enterprise. The trustees are not named in the draft. Through Class T stock, the LTBT elects and removes the Class T directors, who together form a majority of the board of directors. The LTBT has no authority over day-to-day management or strategy. Its sole governance role is selecting the board majority.

The LTBT and the Founder LLC will vote together pursuant to a Stockholders' Agreement. The company qualifies as a controlled company under Nasdaq rules. It intends to rely on controlled company exemptions for its nominating and corporate governance committee, which will not be fully independent. The audit and compensation committees will be fully independent.

What this means practically is a person who buys Class A shares in the Anthropic IPO will own economic exposure to the company's revenue and profits. That person's vote, however, only applies to a small fraction of total voting power. The co-founders, through the Founder LLC's single Class F share, hold 50.1% of voting power on most matters. The LTBT controls the board majority through Class T. A Class A shareholder cannot remove a Class T director. A Class A shareholder cannot, even in combination with every other Class A shareholder, override the Founder LLC's 50.1% on non-board matters. There is no realistic path to a hostile takeover. There is no realistic path to shareholder activism forcing a strategic change. (Bill Ackman shakes fist at sky)

The co-founders have pledged to direct 80% of their personal equity in Anthropic to philanthropy. This is not a legal constraint on the company, but a personal commitment by the individuals.

This will eventually be a case study in your kids’ corporate finance class because it goes against how incentives are typically structured and actually creates an agency problem. While some will tell you profit is evil, incentivizing the people who run corporations to act in the interest of the people who own them is usually a good thing.

Here, however, the people who hold the most power over the company's direction (the Founder LLC members) have reduced their financial incentive to optimize for personal wealth maximization. The structure is actually explicitly designed to decouple governance control from financial interest.

Five percent of fully diluted shares (846,840,960 Class N shares) were contributed to the Anthropic Public Benefit Fund prior to the offering. The board may issue up to an additional 2% of outstanding shares each time the stock price rises 50% above the prior milestone, starting at the IPO price. Every major price appreciation event is contractually linked to an additional charitable contribution of shares.

Delivering Shareholder Value (ish)

Anthropic's PBC structure is operative, not decorative. Under Delaware PBC law, directors are required to consider, manage, and balance three things: stockholder interests, the interests of those materially affected by the company's conduct, and the Company Purpose as stated in the charter. You know, the crazy part is that to run a successful profit making company, usually doing all of those things is a good starting point.

As mentioned in my preamble, the Company Purpose is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."

The practical effect is that the board has legal cover, and legal obligation, to make decisions that are not purely profit-maximizing, provided those decisions serve the stated purpose.

The S-1 names this as a risk factor: public benefit duties may conflict with short- or long-term financial interests. It also names it as a feature: the structure lets the company gate deployments, limit access to advanced models, absorb costs in service of safety research, and operate under a set of obligations that a standard Delaware corporation could not as easily justify to shareholders.

The Mythos model is the current example. Released through Project Glasswing to a small set of trusted organizations rather than to the general public, it is cited in the S-1 as an illustration of precautionary deployment practices. The filing states these measures "may become more frequent and more costly as capabilities advance."

That sentence is both a risk factor and a strategic commitment: the more powerful the models get, the more carefully the company says it will gate them, and the more that care will cost.

Think of it as if Lockheed made a nuke, realized it did too good of a job, and said actually rather than generate revenue from this crazy thing we made, we are going to incur extra costs to ensure it doesn’t accidentally destroy the world.

No Risks To See here

The risk factor section is extensive, in case my essay above didn’t give you that impression. The most material ones, in the order the filing presents them are shown below.

Model leadership is self-reinforcing in both directions. The company's success depends on developing the most advanced models. If it falls behind, the filing warns of a "self-reinforcing or compounding cycle" of slower growth. Less revenue, less compute, less talent, worse models, less revenue. The inverse is also true, which explains the compute commitments locked through 2045.

Training runs can fail. Large training runs are expensive and can underperform. The company is betting significant capital on each training run producing a better model. Sometimes they do not.

Closing the loop propagates errors. Using current models to build future models creates the risk that systematic flaws scale forward. The filing explicitly acknowledges this.

Compute is constrained. Anthropic has sought more compute than cloud providers were willing to supply. The supply chain for advanced chips is not elastic. Cross-default provisions in the company's debt could accelerate indebtedness if the compute commitments create financial pressure.

Revenue is consumption-based with limited long-term contracts. Most revenue is not locked in under multi-year agreements. Customers can reduce usage. The NRR of 900% suggests they have been doing the opposite, but there is no contractual floor.

Cloud providers are also competitors. AWS, Google Cloud, and Microsoft Azure distribute Claude. AWS and Google are also investors. AWS and Google also build their own AI models. The filing describes the conflicts of interest explicitly and names them as ongoing risks to the business.

Governance concentration. Class A holders may have limited influence over significant company decisions, including a change of control. But hey, at least you won’t get any calls for proxy solicitation.

Public benefit duties may conflict with financial interests. The board may make decisions that harm short- or long-term financial performance in service of the company's stated purpose, and shareholders have limited recourse.