I've spent a career working with businesses who assess risk for a living, and with the founders and leaders that think about it constantly. And when they go public, which is the dream and ambition of so many entrepreneurs, they actually disclose their material risks in the risk factors section of an S-1. So I read Anthropic's risk factors first, and what I found is the clearest evidence I've seen that the cost of work is falling fast. When work gets cheaper, the economy doesn't do less of it; it does far more. I think we're maybe in the third inning of what that means, and most of the productivity is still ahead of us.
That's my premise. People have limits. We need breaks, and we only have so many hours. Now one person can orchestrate the work of agents, and agents don't have those limits. Their work could add the equivalent of millions of new workers to the global economy. That's my view, not a forecast. And it could bring a productivity boom on an order of magnitude we've never experienced.
Anthropic filed its S-1 with the SEC on October 26, 2026. Here's how I read it: what could break, how it's covered, then the economics.
What could break
Risk factors are the section where a company lists, in its own words, what could hurt the business.
Anthropic's are frank. It says it has "incurred net losses in each year since our founding", and its operating loss was $693 million for the six months ended June 30, 2026. It expects "significant net cash used in operating and investing activities for the foreseeable future". Its risk factors also spend heavily on compute: the chips, electrical power and data center capacity a frontier model depends on.
- Source: S-1, pp.33–38, 64, 137, 162.
That is the honest starting point.
How it's covered
As of June 30, 2026, Anthropic says it had raised an aggregate of $141.1 billion of capital, and held $112.4 billion in cash and cash equivalents plus $19.3 billion in short-term investments. For a business whose central risk is the cost and supply of compute, that balance sheet is the coverage. It buys time, capacity and staying power.
- Source: S-1, p.11.
What the S-1 shows about the economics
Founded in 2021, and selling Claude publicly only since early 2023, Anthropic has turned AI into revenue at remarkable speed:
- Revenue: $16.3 billion for the six months ended June 30, 2026, up from $1.2 billion a year earlier.Source: S-1, MD&A, p.137; founding and launch, p.62.
- Profitability: positive on an adjusted operating basis, with a 10% Adjusted Operating Margin (non-GAAP). Still loss-making under GAAP, with a (4)% operating margin.Source: S-1, Summary Financial Data, p.25.
- It sold to enterprise customers early. In 2023, customers with $100,000 or more of revenue on a trailing-twelve-month basis made up 90% of Anthropic's revenue. That group numbered about 1,500 at the end of 2025. As of September 30, 2026, there were more than 11,900 of them, up from more than 950 a year earlier.Source: S-1, pp.135, 143.
- Expansion inside accounts: the S-1 reports Net Revenue Retention of approximately 900% as of September 30, 2026.Source: S-1, p.11.
Cheaper work means more work
This is the line I keep rereading. Anthropic says Claude Opus 5.5 "reduced customer costs per task by approximately 40% compared to Claude Opus 5". That figure is for one model pair, not a rate for every generation. Anthropic also says lower costs per task drive greater usage. It adds: "Our compute efficiency allows us to deliver these reductions while preserving our margins." That last point is the company's claim, and it's the one I'd watch.
- Source: S-1, p.11.
Pricing elasticity is the ability to cut price and gain more in volume than you give up. It fits the pattern known as the Jevons paradox: when using something gets cheaper, total use can rise enough to outweigh the savings (Jevons, The Coal Question, 1865). Turning users into revenue, with margins it says it is keeping, gives Anthropic room to go downstream. It could take cheaper intelligence to smaller businesses and everyday people, not only the enterprise. That's my read, not a plan the filing states.
Agents join the workforce, tasks first
People and agents using Anthropic's tools can do more. Agents take tasks before they take jobs, and they don't clock out. A person with good agents does the work of a bigger team. That's augmentation before substitution.
You can see the same pattern forming in the physical world:
- Freight. Driverless semi-trucks are already hauling commercial freight in Texas, and Aurora says they add the potential for around-the-clock capacity.Source: Aurora Q2 2026 shareholder letter, pp.4, 7 (sec.gov).
- Warehouses. Amazon has run distribution centers since at least 1997. It has built warehouse robotics since 2012, and by mid-2025 had deployed more than a million robots across more than 300 facilities. Amazon says these robots "work alongside our employees", and that its DeepFleet AI model is set to improve robot travel time by 10%.Sources: Amazon FY1997 Form 10-K405, MD&A (sec.gov); Amazon, Jun 30, 2025 (aboutamazon.com).My view: for most of those years, the limit on distribution throughput was how fast people could move. AI-coordinated robotics, working alongside people, starts to lift that limit.
- Making things. Jeff Bezos's new company, Prometheus, is separate from Amazon. It is building AI to help engineers design and manufacture physical products faster.Source: CNBC, unofficial transcript, Jun 11, 2026 (cnbc.com).
- Factories. Manufacturing is next. That's my view, not a forecast with a number attached.
Why it doesn't show up in the numbers yet
Capability shows up first. Productivity shows up later, once companies rebuild how the work gets done around it. Economists call that lag the productivity J-curve (Brynjolfsson, Rock and Syverson, NBER w25148, nber.org). That's why I say "maybe the third inning": the S-1 shows the adoption, and the pull-through comes after.
The way we measure an economy grew out of the Depression. In January 1934, the Commerce Department sent the Senate a report called National Income, 1929–32, planned and supervised by the economist Simon Kuznets. Its first chapter starts with people: "Year in, year out the people of this country, assisted by the stock of goods in their possession, render a vast volume of work toward the satisfaction of their wants." Kuznets was candid about the limits of his own measure. He warned that a nation's welfare can scarcely be inferred from a measurement of national income (S. Doc. No. 124, 1934, pp. 1, 7). Wartime planning added gross national product, and today the standard is gross domestic product, GDP (BEA, 2000; BEA, 1991). Economists measure labor productivity as output per hour worked (BLS). What an agent produces doesn't show up as a new job on a payroll report.
There's a measurement point too, and I'll be clear it's my opinion. Our expectations for output were set by human working hours. An agent doesn't keep those hours. And when the price of a task falls sharply, more of the gain can show up as lower prices and more work done than as a bigger number on the scoreboard, at least at first. I'm making a point about how we read output, not citing an estimate.
Others see the same early stage. Box CEO Aaron Levie, reacting to the filing, pointed to "how early we are in the build out of AI." (x.com) On All-In, Jason Calacanis said of AI use at work: "we're still talking about 10 and 20% here. It's nascent." (x.com)
I expect the next one to two years to start showing it, as agents join the workforce.
The bear case is real
The strongest arguments against this view deserve a fair hearing:
- The losses are real. Revenue at this scale with GAAP losses means the full-cost economics are still being proven.
- The constraint may be supply, not demand. Chamath Palihapitiya wrote on November 1: "Power is THE binding constraint in AI/SI." (x.com)Sarah Guo flagged a "mismatch b/w consumption based revenue and long term compute obligations". (x.com)
- Commoditization. Open-source and rival models can catch up fast, and that competition could squeeze any one lab's margins.
- Conversion. The S-1 itself warns that "free users may be satisfied with the limited functionality available in free tiers." (S-1, p.80)
My answer:
- No permanent lead. No lab is guaranteed a permanent lead, and company-level margins are the part most at risk. But the economic point doesn't depend on any one company winning. If competition pushes the price of work down faster, cheaper work still means more work.
- Supply. The supply constraint is the real one. It's why the 1–2 year view is an expectation, not a promise.
What would change my mind
- Cheaper tasks stop pulling in more usage.
- Enterprise customers stop expanding.
- Compute and power can't keep up, and the timeline slips.
- A couple of years from now, we still can't see it in how much work gets done.
Where this leaves me
Frontier AI takes compute, capital and people who can build it. The S-1 shows how much capital is already behind one American lab. I'm more convinced the AI productivity wave is real, and this S-1 is the clearest evidence I've seen. The reasons are customer expansion, a falling cost per task, margins the company says it is preserving, and a deep balance sheet. Together they give room on price, and price is how intelligence reaches more businesses. By the S-1's own account, that is still early.
Keeping companies like Anthropic building in America matters for the country's productivity. I think it helps keep the US at the frontier of AI.
Know what could break, then how it's covered, then the economics. That order works for a public company. It works for a family and a business too.
These are my personal views, not investment advice or a recommendation to buy or sell any security. As of this writing, neither I nor Sovra Capital holds a position in Anthropic, and that may change.
Sources: Anthropic, PBC Form S-1, filed with the SEC Oct 26, 2026 (sec.gov); page numbers are the S-1's printed pages. Other sources are named inline. Full citations for the measurement history: U.S. Department of Commerce, National Income, 1929–32 (study planned and supervised by Simon Kuznets), Senate Document No. 124, 73rd Congress, 2nd Session (1934), pp. 1, 7 (fraser.stlouisfed.org); BEA, Survey of Current Business, Jan 2000 and Aug 1991 (bea.gov); BLS Handbook of Methods (bls.gov).