SAN FRANCISCO, 24 AUG 2026 — Anthropic's forthcoming initial public offering will reportedly list public opposition to artificial intelligence among its risk factors, according to people familiar with the filing. The company confidentially filed in June, and its most recent private valuation was US$965bn in a round that closed in late May.
We reported last week that the US$65bn figure circulating among investors is an annualised run rate and that a prospectus would have to state something more conventional. The risk factors are the other half of that document, and the more telling half.
What a risk factor is for
Risk factors are a legal instrument rather than a forecast, drafted to make later claims of surprise unavailable to an investor who says nobody told them.
That makes them a poor guide to what management fears and an unusually good guide to what management cannot rule out. Everything plausible goes in, including risks the company considers remote, because omitting one is the expensive mistake.
The presence of a risk factor proves little on its own. The category, however, is notable. A frontier AI company disclosing risk from competition, regulation or model performance would be unremarkable. This one is reportedly disclosing risk from the electorate.
The number behind it, and when it was actually taken
The survey being cited in this connection is a Gallup poll finding that roughly seven in ten Americans oppose construction of an AI data centre in their local area, with 48 per cent strongly opposed against 7 per cent strongly in favour.
On the environmental question, 46 per cent said they worry a great deal about the impact of AI data centres and a further 24 per cent a fair amount.
The poll is being cited as a reading of current sentiment, and one detail complicates that. It was published in May, but the interviews were conducted in early March, with a sample of about 1,000 adults across the fifty states and the District of Columbia. A two-month lag between fieldwork and publication is normal. It does mean the figure describes March, in a debate that has moved since, and anyone treating it as this month's temperature is over-reading it.
How local opposition becomes a delivery constraint
The mechanism connecting a planning objection to a technology company's revenue is short, and it is the reason this belongs in a filing rather than a press release.
Frontier models are trained and served on capacity that has to exist somewhere physical. That capacity needs land, a grid connection and water, and each of those is granted locally. An objection does not need to defeat a project to matter; it only needs to delay it, because a delayed connection means compute that arrives later than the roadmap assumed.
Compute arriving late constrains what can be trained, what can be launched and what a service can be sold to do. It also raises operating costs, because the capacity that does exist becomes more expensive to rent when everyone's alternative is delayed too.
That is a supply risk expressed as a civic one, and it is not hypothetical for a company whose commitments are made years before any concrete is poured.
The industry has already moved on this
The risk is demonstrably real, because developers have already changed their behaviour without waiting for a prospectus to name it.
New AI data centre sites in Europe are being planned an average of 175km from major hubs for delivery between 2026 and 2028, against 46km for projects delivered from 2022 to 2025, as we reported this month. Inner-city projects are expected to fall to a twentieth of the pipeline. Of nine proposed gigawatt-scale European sites, one is near a major city.
That shift is usually explained by power availability, which is accurate and incomplete. Land far from population centres is also land far from objectors. Siting a campus at a distance is simultaneously the answer to a grid constraint and to a planning one, and the two are difficult to separate in any individual decision.
The same pressure explains why the buildout has become so interested in jurisdictions that decide these questions faster. Global data centre capital expenditure is on course to pass one trillion dollars, and a meaningful share of that is chasing places where a grid connection is a decision rather than a hearing.
What the investor meetings are actually asking
Chief financial officer Krishna Rao has been leading preliminary test-the-water meetings with bankers and investors in San Francisco. The questions reported from those rooms suggest the risk factor is more than a public-relations gesture.
Investors are asking about competition, about margin pressure from open-weight models, and about what happens if data centre construction slows. The third question is the risk factor, asked by the people whose money is at stake, which suggests it went into the document because it kept coming up rather than because a lawyer was being thorough.
The margin question deserves its own attention. Open-weight releases do not have to win on capability to compress pricing; they only have to be good enough for the tasks that make up the bulk of paid volume.
The valuation gap nobody has confirmed
Reporting has investors working towards a float valuation of around two trillion dollars. That number comes from those investors' own calculations, not from any figure Anthropic has communicated. Rao is reported not to have discussed valuation in the meetings at all.
Set against a US$965bn private mark from late May, a two trillion dollar float would be roughly a doubling in a few months. That may prove right. It is not a company projection, and treating it as one is the most common error in the current coverage.
What remains unconfirmed
The filing is confidential, so its contents are known only through people describing it. The risk factor's actual wording is not public. Anthropic has not commented on the survey, the valuation or the timing.
Morgan Stanley, Goldman Sachs and JPMorgan have been named in earlier reporting as working on the offering. That is reporting rather than an announcement, and the syndicate on a deal this size can change.
What it means from here
For readers across ASEAN the useful part is not the American polling but the dependency it exposes. The compute a regional business rents is now subject to a planning process on another continent, and the industry's response to that is dispersal.
That dispersal is why Johor, Batam and Luzon are in these conversations at all. It is also a warning about the terms on which that investment arrives, because a project that relocated to avoid an objection has demonstrated exactly how much it values a fast approval. The same Gallup finding will eventually be taken locally, and the answers will not be automatically friendlier — Malaysia has already narrowed its own data centre policy, and Singapore rationed capacity for years before releasing it.
The question worth asking of any campus announced in the region is which constraint it moved to escape, and whether the host is being paid for solving it.
