What Anthropic's S-1 tells us about where AI is going

Love them or hate them, Anthropic’s S-1 shows where one of the most valuable companies in the world is putting its attention: on the opportunities that drive the push for greater intelligence, and on the hard problems between that intelligence and those opportunities.

What the intelligence is for

The labs have to keep pushing the boundaries of intelligence, so ask what more intelligence is for. There has to be a reason other than reaching AGI first or being better than every other intelligence provider on the planet.

The filing answers that with opportunity. It opens with four areas: biology and medicine, economic growth, spreading that growth around the world, and better government services. It says Claude is already “cutting the time it takes to write clinical trial reports from months to minutes”, and that Anthropic has made discoveries with Claude, including a “molecular machine” with the “potential to be developed into a novel biotechnology tool”. It talks about “compressing decades of scientific progress into years” and “accelerating drug discovery, disease detection”, and its Model Hardware Standard “enables AI agents to operate lab and manufacturing equipment”. Nine of the ten largest pharmaceutical companies in the Fortune 500 are customers, and the founders’ letter talks about bringing forth “an age of miracles”.

Those opportunities are what push for greater intelligence. Biotech and finding cures is one of the main purposes Anthropic sits behind, and that is a purpose I will stand next to any day. Getting the intelligence to those opportunities is where the hard problems are.

From research company to scale

Take a research company, add infrastructure, then try to apply scale, and you need distribution. The filing says it plainly: “Frontier AI is most valuable in real-world settings, as deployment across enterprises is where economic impact is greatest.” Anthropic’s answer on distribution is clear. It works through Accenture, Deloitte, KPMG and PwC (and we all know how much we love them), runs applied AI teams with forward-deployed engineers inside its customers, has committed $100m to train 10,000 engineers to lead deployments, and takes nearly half its revenue through the Amazon, Google and Microsoft cloud marketplaces. Over and above that, it is moving into verticals: the filing names healthcare and life sciences as new markets, and expects frontier AI to reach robotics, advanced manufacturing, logistics and materials science.

Customers paying Anthropic $100,000 or more a year went from 20 at the end of 2023 to 200 a year later, about 1,500 at the end of 2025 and more than 11,900 by September 2026, including more than half the Fortune 500. Among the listed Fortune 500 companies in that group, spending with Anthropic averages less than $1 for every $1,000 of their total expenses. Anthropic says so itself: “the gap between what Claude can do and how broadly it is used represents one of our largest growth opportunities.”

Infrastructure and scale are the real problem

Distribution, infrastructure and scale are a serious problem for Anthropic and for every other lab. Compute, chipsets, network, efficiency and latency are real non-functional requirements, and they are not getting easier. Anthropic spent $8.9bn on compute and infrastructure in the first half of 2026 and has about $498bn of compute commitments it cannot cancel. Its own risk factors say it has “significantly underestimated our compute requirements” in the past, that “the cost of compute could rise significantly”, and that as agentic work grows, “the aggregate compute required to serve our customers could increase disproportionately relative to the associated revenue”. Energy is part of it too: Anthropic has committed to cover the electricity price increases its data centres cause for consumers. The filing names the concentration of its infrastructure in the US, grid constraints and data centre moratoriums as risks, and says it is “expanding our data center capacity internationally as part of our efforts to diversify our infrastructure footprint and reduce concentration risks”. It will get worse before it gets better.

Anthropic already runs compute “as a single, fungible pool” across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs, sending it “where the marginal unit is expected to be most productive”. Every business running AI at scale will need the same discipline for its own work: deciding what each step actually needs.

Cost per task fell by about 40% from Opus 5 to Opus 5.5, and total spend kept rising. Each task needs less compute, total demand keeps growing, and throwing compute at a problem is not the answer.

You cannot trust your most valuable process to a probabilistic outcome

The elusive near-deterministic outcome is the other challenge the labs are clearly working on. A model gives a likely answer. A business process needs the same answer every time, checked, with someone accountable for it. You cannot trust your most valuable business process to a probabilistic outcome.

Anthropic’s risk factors say its models may “behave in unintended, misaligned, or autonomous ways”, may “fail to follow instructions”, and that “any systemic flaw, bias, or error in our current models could propagate into the next versions”. Its report on its own agents this month described agents submitting forms they had been told to stop before submitting, and concluded that “alignment training is not yet sufficient or fully robust on its own, at least in the short term”. More than 80% of September-quarter revenue came from model versions released in the previous six months, so the model underneath any process changes within months.

The filing goes further. Anthropic says its technical safeguards “are incomplete, may be outdated in light of evolving threat models, have been and may in the future again be circumvented, or may fail to operate as intended”. It describes what it believes is the first documented large-scale cyber-attack carried out with little human intervention, in which a state-sponsored group manipulated Claude Code into attempting to infiltrate roughly 30 global targets. And it warns that it “may be unable to obtain coverage for certain AI-related liabilities on commercially reasonable terms, or at all”. If the company that builds the model may not be able to insure against what it does, a business cannot run its most valuable process on a probabilistic outcome without controls of its own.

Recursive self-improvement needs the right foundations

The S-1 is open about what Anthropic calls closing the loop: each version of Claude helps build the next. Many aspects of research, training and testing of new models now involve Claude. Over 90% of contributions to Anthropic’s codebase in September 2026 were written by Claude, and in April 2026 the monthly value of its internal Claude use first exceeded its monthly personnel expense, excluding stock-based compensation.

Recursive self-improvement is extremely important, and we believe in it. But RSI without the right infrastructure, the right harness, the right evaluation scoring and the right guardrails can go horribly wrong, because every error compounds. The filing says so itself: “any systemic flaw, bias, or error in our current models could propagate into the next versions”, and Anthropic’s ability to measure how fast its capabilities are advancing is limited, so it “may not identify a significant acceleration in our rate of development until after it has occurred”. It chose to delay pre-training a new frontier base model by more than a month to strengthen its security measures.

The same holds for any business whose systems learn and improve themselves. Every change has to be tested against real outcomes before it is trusted, every version kept, and a person accountable for promoting it. We build on a method credit scoring has used for decades for exactly this: champion and challenger. A new way of doing a step has to beat the current one on the business’s own work before it takes over.

The UK picture

The Fortune 500 is an American list. The pattern here is the same.

The ONS found about 35% of UK businesses with ten or more employees used at least one AI technology in June 2026, up from 12% in late 2023. Those businesses use 1.6 AI technologies on average, and only one in ten report extensive use.
Large firms are further ahead: 49% of businesses with 250 or more employees use AI, against 28% of those with fewer than ten.
Anthropic’s own filing says AI “deployed at scale also has the potential to concentrate power and wealth both within and between countries”.
The June US directive that took Fable 5 and Mythos 5 offline barred access by foreign nationals inside or outside the US, so UK customers lost them too.
The Budget’s new British Data Centre Covenant asks data centres to invest in local communities, create local jobs and protect water supplies, in return for the grid connections they need.
The UK government’s own work with Anthropic, the GOV.UK assistant pilot, aims to leave the UK able to maintain the system itself.
This week’s Budget added £140m for the Sovereign AI Unit to help UK AI companies scale, £40m for a National Agentic Security Unit, and £165m for a National AI Biosecurity Initiative, including a lab backed by up to £200m of model credits from OpenAI and Anthropic. The Treasury cites OECD estimates that AI could add 0.4 to 1.3 percentage points a year to UK productivity growth over the next decade.

Most UK businesses will never see a lab’s engineers. The opportunity here is as large as anywhere, and the help to reach it is thinner.

What the sceptics will say

“It’s a bubble. Anthropic lost $80bn.” Most of that, $72.4bn, is an accounting charge on convertible notes. The second quarter made a $974m operating profit and the first half generated $3.9bn of operating cash. The real exposure is the compute it has committed to and cannot cancel.

“Under $1 in $1,000 shows companies tried it and found little value.” On Anthropic’s measure, customers already paying a year ago spent about nine times as much in the September quarter, and Cowork, for non-technical teams, has been adopted faster than Claude Code was. Spending grows where the work is properly deployed and stalls where it is not.

“The labs will do all of this themselves.” For the largest companies, they will. Most businesses will not get those engineers, and no business should hand its process to its supplier when the model underneath changes within months.

“Sovereignty is flag-waving.” The June shutdown reached customers in every country. Forrester’s advice was portability: “Single-sourcing is the issue, not US government oversight.” That holds for any provider, British ones included. Anthropic itself is expanding its data centres internationally to reduce its own concentration risk.

“You use Claude yourselves.” We do, and the models are very good. The question is where the business’s process lives. It should belong to the business and be able to move to another model when it needs to.

What we are building at MindLynx

At MindLynx we are helping businesses understand where the opportunities lie, and helping them execute on them.

The business owns its process. Each job is declared as steps, each step is checked, and a person signs where the decision is theirs.
The job is priced before it runs, so a finance director can put it in a budget.
The model is chosen for each step and can be swapped without changing the process. Albion, our family of models, is in training.
What the business learns is kept, with who said it and when, so each deployment makes the next one easier.
We activate it inside the business, with the people who run the work.

The infinite game

Some of what we do at MindLynx overlaps with Anthropic, and some of it runs on Anthropic. Simon Sinek calls this the infinite game: you are not playing to beat the other players, you are playing to keep playing, and a worthy rival shows you where to get better. I go a step further. When you start to see your competitors as partners, you play the long game, and you play for something much greater.

What the S-1 tells me is where Anthropic, and the others after it, are putting their attention, which is where the challenges lie. For little people like me, that brings more confidence and more excitement. At times it simply proves what I have been thinking and saying. The S-1 cites IDC’s estimate that investment in AI solutions and services will yield a cumulative global economic impact of $22.5 trillion by 2031, and Anthropic believes it could be greater. The pie is big enough.

First published on LinkedIn, 29 October 2026.