On October 26, Anthropic filed its S-1 with the SEC.
At the time, it was investors’ most detailed look at the company’s overall business conditions, financials, risk factors, etc.
It showed a company with rapidly growing revenues, a positive Adjusted Operating Margin, and large net losses due to changes in the fair value of convertible notes (the company’s private valuation has grown as Anthropic’s IPO approaches, which creates a non-cash accounting charge).
In the original S-1, which reflected 2026 figures through only the first two quarters of the year, first-half revenue rose from $1.2 billion to $16.3 billion year over year, Adjusted Operating Margin rose from -159% to 10%, and the accounting-driven net loss rose from $13.0 billion to $80.7 billion.
Absent from the S-1 were the long-term obligations Anthropic has incurred to secure compute resources and infrastructure, which were widely understood to be in the hundreds of billions of dollars.
Then, on November 2, Anthropic filed an amended S-1 with the SEC that cracked the door open a bit further.
The amended filing, which also included Q3 figures, showed total year-to-date revenue of $33.2 billion—with $16.9 billion coming in Q3 alone—a last-nine-month Adjusted Operating Margin of 12%, and a nine-month net loss of $104.0 billion.
More importantly, it also revealed that Anthropic has $895 billion in commitments through 2045, with nearly $450 billion coming due by the end of 2030.
The cost of those commitments may well be enough to keep Anthropic from achieving profitability. As noted in the S-1, Risk Factors, they “have a history of net losses and may not generate sufficient revenue to achieve or sustain profitability in the future.”
[text_ad]
That is, in Anthropic’s defense, a bit of boilerplate for the filing, but they’ve “incurred net losses in each year since [their] founding,” and the scale of their ongoing spending obligations reflects what is increasingly an all-or-nothing bet on artificial intelligence.
The Verge put it best in a November 2 headline: “Anthropic’s business plan: Build God or die trying.”
The expenditures make it abundantly clear that Anthropic is not concerned about short- or intermediate-term fluctuations in profitability; the company is going all-in on building some form of general-purpose AI superintelligence, come what may.
They allude to it in the S-1 filing when highlighting what they see as their “Total Addressable Market” (i.e., where they see an opportunity to generate revenue).
Specifically, Anthropic highlights the opportunity in “knowledge work” (on page 189 of the amended S-1):
“We believe the work that knowledge workers perform represented more than $40 trillion in economic value globally in 2025, and with frontier AI’s potential to make them meaningfully more productive, our near-term opportunity is to expand this value and accelerate economic growth.”
That $40 trillion figure dwarfs the only other economic opportunities Anthropic cites in the S-1: cybersecurity (Grand View Research estimates that the cybersecurity market is $302 billion this year) and life sciences (Anthropic’s S-1 calls out $2 trillion in spending by pharmaceutical companies this year).
If increasing the productivity of knowledge workers (or replacing them) is the profit opportunity—and if it’s the justification for planning nearly half a trillion dollars in spending in the next four years—it stands to reason that we should assess Anthropic’s prospects for success or failure based on progress on that front.
So, how’s that going?
Based on the case studies included in the S-1, not well.
Beginning on page 208 of the original filing (and page 211 of the amended filing), Anthropic outlines the successes that their own clients have had as a result of implementing Anthropic’s AI tools.
While the results from efforts in the software/cybersecurity and life sciences fronts are impressive, results from other segments of the economy are anything but.
To wit, the case study with Bristol-Myers Squibb found that “(t)ime to answer a real-world evidence question” fell from two months to just five minutes, while Salesforce saw its “(e)ffective output per developer year over year” rise by 151% when using Claude Code.
At the same time, PayPal’s “(a)verage time saved per sales employee per week” clocked in at a paltry 45 minutes, while Legora found a time savings of 4.3 “(n)on-billable hours recaptured per lawyer weekly, in an independent study of 31 firms.”
Similarly, the International Rescue Committee, in a 300-user pilot, netted 464 “staff hours returned per week,” or 1.55 hours per user.
And while Anthropic highlighted a case study conducted with Accenture that netted a 7x return on investment (ROI) for “a broadband provider that is a joint Accenture and Anthropic customer” (presumably this case study with Cox Communications), Accenture itself has subsequently capped its own Claude spending at $29 per month per employee. Although it’s worth pointing out that employees can spend more than that limit if they attribute it to a specific internal cost center to justify the use.
In the aggregate, these case studies paint what I consider to be a fairly clear picture of Anthropic’s offerings: Claude and Claude Code are impressive tools for a narrow slice of industry-specific applications while offering little more than limited productivity gains for knowledge workers at large.
Unless and until that changes, and to paraphrase The Verge, it seems that Anthropic is far more likely to “die trying” than they are to “build God,” which leaves the AI bubble on borrowed time.
[author_ad]
