Information cutoff: October 27, 2026. Anthropic filed its preliminary Form S-1 on October 26, 2026; offering terms and some disclosed figures may be amended.
Commercial disclosure. I founded and operate Archive Signal, a separate business that sources, structures and licenses proprietary datasets. Its buyers include AI developers, creating a commercial interest in AI spending growing and in the market assigning value to rights-cleared operational data. This note argues that operational context and verified outcomes will become more valuable as model capability expands. See the standing disclosure for the business relationship, trading policy and use of AI tools.
Every junior buy-side analyst I know appeared to disappear into Anthropic’s S-1 last night.
Fair enough. The numbers are absurd.
Revenue increased from $386 million in 2024 to $4.59 billion in 2025, then to $16.29 billion in the first six months of 2026. The second quarter alone produced $11.56 billion of revenue, up 1,368 per cent from a year earlier.
By September, Anthropic had more than 11,900 customers generating at least $100,000 of trailing-twelve-month revenue, more than 1,900 generating $1 million and roughly 250 generating $10 million. Net revenue retention was approximately 900 per cent.
That is not normal software growth. It is what demand looks like when a new productive input goes from interesting to indispensable almost overnight.
The operating performance was stronger than nearly everybody expected, too. Anthropic generated $2.08 of revenue for every dollar of compute and infrastructure expense in the second quarter, up from $0.40 in the first quarter of 2025. Q2 produced $974 million of GAAP operating income and an adjusted operating margin of roughly 19 per cent.
The $80.7 billion first-half net loss now ricocheting around social media is mostly an accounting spectacle. Approximately $72.4 billion came from fair-value changes to convertible notes, derivative liabilities and SAFEs as Anthropic’s valuation rose. The company’s first-half operating loss was $693 million, while operating cash flow was positive $3.93 billion.
These figures come from Anthropic’s preliminary Form S-1. So yes: the financial story is real.
But I think Wall Street is still reading the wrong document.
Anthropic filed a software prospectus. What it revealed is an industrial company.
The model is becoming the easy part
That sentence needs a qualification.
Building a frontier model is obviously not easy. It requires exceptional people, enormous amounts of capital and a research organization operating at the edge of what is technically possible.
But for the customer, access to capable intelligence is becoming easier, cheaper and more substitutable.
Anthropic says Opus 5.5 reduced customer cost per task by approximately 40 per cent compared with Opus 5. Its models are available through Amazon, Google and Microsoft as well as Anthropic’s own products. Competitors are improving. Open-weight models continue to advance. Model routing makes it increasingly possible to use different systems for different jobs.
That does not eliminate the model companies’ advantages. It changes where everybody else should look for theirs.
If capable intelligence can be rented by the token, then merely having access to a model is not a moat. The durable value moves toward the things the model cannot obtain on demand:
- the right operational context;
- permission to use it;
- an interface through which the model can act;
- the physical capacity to carry out the action; and
- evidence that the action produced the intended result.
That is the Sorites thesis in one list.
And Anthropic just filed 335 pages explaining why it is right.
Anthropic’s growth loop runs through its customers
The most important section of the S-1 is not the income statement. It is Anthropic’s description of what it calls the “Improvement Loop.”
Enterprise deployments expose the company to hard problems, edge cases and domain-specific challenges. Feedback from customers using Claude in production improves the models, platform and applications. Those improvements produce wider adoption, which produces more feedback.
Read that again without the corporate language.
Real work is becoming part of the model-development system.
This matters because the first generation of large models learned from a world that had already been recorded: books, websites, code repositories, images and other digital material. The next generation must learn from work that is still poorly captured.
How does an experienced technician know the bearing is failing before the sensor crosses a threshold? Why did a repair that looked correct fail three weeks later? Which construction sequence worked in unstable ground? Why did a machine operator abandon the documented procedure and make a small correction that saved the part?
These are not trivia questions waiting somewhere on the internet. They are episodes embedded in physical systems, human judgment and time.
A language model can read the work order. It cannot infer from the work order alone whether the repair held.
A camera can record the worker. It cannot automatically tell you what force was applied, what the worker noticed, whether the result passed inspection or whether the component failed six months later.
A simulation can generate a plausible environment. It cannot guarantee that its assumptions captured the ugly, irregular conditions that make real deployments difficult.
The improvement loop only becomes valuable when somebody closes it.
The valuable record begins after the agent acts
AI agents are moving from describing operations to intervening in them. They can identify a fault, contact a driver, open a work order, change a schedule, recommend a part and escalate an exception.
That is impressive. It is also only half a system.
The action tells us what the agent did. The outcome tells us whether it was right.
This distinction will become more important as AI pricing moves beyond tokens and seats. Anthropic acknowledges that agentic products may require pricing based on tasks or outcomes. The moment a vendor wants to charge for a completed task, the parties need to agree on what “completed” means.
Was the problem actually solved? Did the machine return to service? Was the customer satisfied? Did the same failure recur? Was production faster, or did the intervention merely move the bottleneck somewhere else? Who is responsible when the agent followed the approved workflow and the result was still wrong?
Outcome pricing requires outcome evidence.
That evidence is not merely a billing mechanism. It is the highest-value training and evaluation material in the loop. It connects conditions, decisions, interventions and results.
We have been arguing that the work order is not the repair, the camera is not the dataset and the archive is not useful merely because it is large.
Anthropic’s filing does not dance around that thesis. It walks directly into it.
The company says production feedback improves the model. It says agentic systems will tackle increasingly complex workflows. It says it plans to enter robotics and physical work. It even defines powerful AI partly through its ability to control laboratory equipment, robotic systems and other machines.
Once the model touches the physical world, verified reality becomes part of the stack.
The cloud has a substation
The other revelation is just as important.
Anthropic presents its compute as a single, fungible pool spread across AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. Software dynamically routes workloads toward the capacity expected to be most productive.
It sounds frictionless.
The filing makes clear that it is anything but.
The pool depends on accelerators, memory, networking, storage, cooling, water, electrical generation, transmission, substations, grid interconnections, permits, construction crews, equipment leases and enormous financing arrangements. Chips may be supplied by one party, financed by another, installed in a third party’s facility and operated by somebody else.
The named obligations include roughly $161.2 billion across Broadcom-related equipment leases, at least $114.5 billion with Google, at least $110 billion with Amazon and at least $44.9 billion with Microsoft. The $161.2 billion Broadcom figure combines arrangements; the 2026 five-year equipment lease alone is approximately $125.2 billion excluding interest. These figures should not be carelessly added to every other announced capacity estimate—the terms, maximums, delivery conditions and time periods differ—but the scale is clear enough. Anthropic’s April description of its expanded Google and Broadcom partnership provides additional context for the capacity buildout.
Anthropic is not simply buying cloud services. It is coordinating one of the largest industrial buildouts on earth.
The company warns about shortages of chips and memory, construction delays, electrical power, water, permitting, grid upgrades, equipment obsolescence and the possibility that contracted capacity arrives too early, too late or in the wrong configuration. A March service disruption was caused partly by rapid demand growth and compute constraints.
This is where the prospectus stops looking like SaaS.
Anthropic may have software margins in a given quarter, but it also has utility exposure, project-finance exposure, construction exposure and the risk of being locked into a generation of equipment while the technology changes underneath it.
The AI economy does not float above the physical economy.
It is becoming one of its largest customers.
Cheaper intelligence creates more physical demand
It is tempting to assume that falling model costs reduce infrastructure demand. So far, the opposite appears to be happening.
Lower cost per task makes more tasks worth attempting. Better models support longer reasoning chains, more tool calls, more retries and more ambitious agentic workflows. Customers do not merely run yesterday’s workload more cheaply. They find new work to hand over.
Anthropic’s 900 per cent net revenue retention is the early evidence. This is not primarily a seat-expansion story. Roughly 80 per cent of first-half revenue came from consumption. When intelligence becomes useful enough, organizations consume much more of it.
The market is therefore not bounded by the existing software budget. It reaches into the cost of the work itself.
That is why the fact that Anthropic represented less than 0.1 per cent of total quarterly expenses among the relevant publicly traded Fortune 500 customers matters. The company is not simply competing for a larger slice of IT spending. It is trying to become an input into finance, engineering, customer service, research, logistics, manufacturing and eventually physical execution.
The prize is enormous.
So is the burden of proof.
The more consequential the task, the less willing an operator will be to accept “the model said so” as evidence that the work was done correctly.
This is where our thesis gets teeth
For the last several months, we have been developing a view that looks unfashionably physical beside the model race.
We believe the next scarce AI assets will be found inside the real economy: industrial histories, expert corrections, failed attempts, inspection evidence, machine behaviour and verified outcomes. We believe these assets must be collected with clear rights, worker participation and commercial terms that do not quietly transfer the learning advantage to a software vendor. We believe major projects should preserve a second asset alongside the infrastructure itself: a governed record of how it was built, operated and improved.
The S-1 does not merely leave room for that thesis.
It nails it.
Anthropic is telling investors that its models improve through deployment, that it intends to push into physical work, that agent economics may move toward completed tasks and outcomes, and that its own growth is constrained by the most physical inputs imaginable.
Intelligence is getting cheaper. Reality is not.
Power is scarce. Interconnection is slow. Skilled physical work is difficult to reproduce. Industrial systems are fragmented. Operational rights are negotiated contract by contract. Verified outcomes arrive late and often live in somebody else’s database.
Those are not peripheral frictions around the AI market.
They are the market.
Canada should notice what it already owns
Canada is unlikely to outspend Anthropic, Amazon, Google or Microsoft in a general-purpose compute race.
That is not the same as having nothing valuable to contribute.
Canada has abundant energy, industrial sites, mines, forests, ports, utilities, transportation systems, advanced manufacturing, large infrastructure programs and generations of operational history. It has experienced people doing work that has never been adequately recorded because nobody previously had a reason to turn it into machine-usable evidence.
The strategic opportunity is not to dump those records into a national data lake and congratulate ourselves for owning terabytes.
It is to build governed improvement loops around the physical economy.
Who can use the records? For what purpose? Under what consent? Who owns the derived material? Can the operator export the history? Can a foreign model provider learn across customers? Do workers and businesses that created the knowledge participate in its value? Does the record include the eventual result, or only the instruction that was issued?
These sound like data-governance questions. They are actually questions about industrial power.
If the same vendor controls the agent, the workflow, the action log and the outcome record, it may eventually control the cleanest account of how the operation learns.
The contract is where that future is decided.
After the model race
The obvious winners from Anthropic’s growth are the companies selling accelerators, memory, networking, power, transformers, cooling, construction and data-centre capacity.
The less obvious winners will help intelligence operate in environments where failure has a cost:
- industrial integration and operational technology;
- permissions, identity and security for agents;
- evaluation and audit systems;
- task and outcome verification;
- rights-cleared data origination;
- specialist implementation firms; and
- industrial companies that retain control of their operating histories.
The vulnerable businesses are equally visible. Thin model wrappers will struggle. Seat-based software whose main purpose is moving information from one screen to another will face pressure. Data vendors selling decontextualized volume will discover that a billion unlabeled records can be worth less than a thousand episodes with clear provenance and a verified result.
The model companies may still become some of the largest businesses in history. Anthropic’s filing makes that possibility much easier to take seriously.
But it also shows what they need from the rest of the world.
They need electricity. They need equipment. They need sites. They need people who understand physical systems. They need access to real work. And if they want agents to do more than generate plausible instructions, they need somebody to preserve what happened next.
Wall Street opened the S-1 looking for software margins.
What Anthropic handed us was a map of the physical economy after intelligence gets cheap.
We have been building on that map all along.
Sources and methods
- Anthropic Form S-1, filed October 26, 2026
- SEC filing index and exhibits
- Anthropic’s filing announcement
- Anthropic on its Google and Broadcom compute expansion
Financial, customer and infrastructure figures are drawn from Anthropic’s preliminary prospectus and may be amended before an offering. Infrastructure arrangements have different durations, conditions, minimums and maximums and should not be treated as directly comparable. The conclusions about physical deployment, operational data rights and verified outcomes are the author’s analysis. This is market analysis, not investment advice.