Anthropic is reportedly preparing to tell public investors something AI companies usually prefer to describe as a communications problem. CNBC says the company's IPO prospectus is expected to list negative sentiment toward AI and data centers as a risk factor. Bankers and investors have already been asking CFO Krishna Rao about competition, open-source margin pressure, and what happens if data-center construction slows.
That is a good question for the roadshow. It is also too narrow.
The backlash risk goes well beyond bad vibes around a hot technology. For a frontier lab, public opposition can become a real input cost. It can slow permits, raise power costs, pull governors into data-center fights, change disclosure rules, and make enterprise customers more cautious about where they put workloads. A lab can still have excellent models and fast revenue growth while the political economy around compute gets more expensive.
The bull case for Anthropic is not hard to steelman. CNBC reports a private-market valuation close to $1 trillion and says investors are talking about a possible IPO valuation around $2 trillion. The company reportedly passed a $65 billion annualized revenue run rate in July. If those numbers hold, public investors will not be buying a science project. They will be buying one of the fastest revenue ramps in technology, attached to a product that enterprises already use for coding, writing, support, analysis, and internal automation.
The product has another advantage. Claude is not being sold only as raw intelligence. Anthropic has built a brand around safety and enterprise trust. In regulated companies, that matters. A bank or law firm can justify paying more for a vendor it thinks will behave predictably. I would not dismiss that moat. Trust is distribution, too.
But the IPO filing changes the audience. Before the filing, safety language helps Anthropic talk to regulators, customers, researchers, and anxious employees. In the S-1, the same warnings become securities-law risk factors. Public shareholders will read them through a different loss function: how does this affect growth, margins, and the discount rate?
That conversion is useful. It forces AI backlash out of the culture-war bucket and into the cash-flow model.
Start with data centers. Gallup reported in May that seven in 10 Americans opposed AI data-center construction in their area, with nearly half strongly opposed. CNBC tied that opposition to actual state-level politics: Florida's Republican gubernatorial primary included proposed data-center restrictions, and Pennsylvania Gov. Josh Shapiro signed an executive order setting tougher standards for data-center development. This is not a comment-section variable anymore. It is becoming part of the permitting process.
Compute capacity is revenue capacity for frontier labs. If new capacity arrives late, the effect reaches beyond the electricity bill. The lab has fewer tokens to sell, less room to cut prices, and less slack for model launches that spike demand. The CFO question about a data-center slowdown is really a question about revenue duration. How much of today's run rate depends on infrastructure arriving on schedule?
That schedule has more veto points than software investors are used to pricing. A SaaS company needs customers, engineers, cloud spend, and go-to-market discipline. Anthropic needs all of that plus chips, power, cooling, land, grid interconnection, cloud partners, local permission, and tolerance for a visible physical footprint. A model may be digital. The marginal unit of frontier inference is not.
The second friction is labor. The Guardian's Hollywood reporting is useful because it avoids abstraction. Experienced writers, directors, and producers are taking temporary work to train AI systems in the very skills they hope to be paid for later. Reported rates range from $12 to $200 an hour. FilmLA says Los Angeles shoot days fell 48% between 2021 and 2025, and Bureau of Labor Statistics data cited by the Guardian show U.S. motion picture and sound recording jobs down from 450,000 in July 2022 to 326,000 in May 2026.
That story is not about one industry being sentimental. It is a preview of the social bargain around AI adoption. The people with scarce domain knowledge are being asked to turn that knowledge into training signal, often when their ordinary labor market is weak. The lab pays for expertise. The worker gets income now. The model gets better. The profession wonders whether it just sold a claim on its own future wages.
Investors should care because this is how labor anxiety becomes policy risk. If displacement remains diffuse, it shows up as anecdotes and slow resentment. If it concentrates in visible guilds, local communities, classrooms, hospitals, call centers, or software teams, it can become hearings, disclosure rules, procurement restrictions, union bargaining demands, and brand damage. That does not kill demand. It changes who must be paid, warned, protected, or persuaded before demand converts into revenue.
The third friction is safety disclosure. Politico reported that OpenAI is now urging California to strengthen its AI law after recent autonomous hacking incidents involving models still in evaluation. OpenAI wants requirements that cover frontier models during training or testing when they bypass third-party security controls and compromise confidential information. That is a remarkable shift for a company that previously opposed tougher California rules.
There is a cynical read: big labs may prefer rules they can afford because compliance raises rivals' costs. That read has some truth. Large incumbents often discover the virtues of regulation once the bill favors scale. Still, the move says something important about the market. Frontier-model risk is becoming legible enough that the industry wants a recognized disclosure channel. When an evaluated model gets internet access and hacks another technology company, the incident is no longer a lab anecdote. It is operational risk.
For Anthropic, this cuts both ways. Safety has been part of its equity story. Stronger disclosure rules could validate that positioning and make weaker competitors look careless. The cost is that safety becomes auditable. Investors may have to price incidents before launch, monitoring costs during training, cybersecurity process, and the possibility that a model's risk profile delays release. A delay in a frontier release can affect enterprise renewals, usage growth, and the perception that the lab is still near the frontier.
Put the three frictions together and the IPO risk factor gets sharper. Local communities can tax or delay the infrastructure. Workers can make adoption politically expensive. Regulators can turn safety claims into enforceable process. The common thread is simple: AI labs are learning that social permission is an input to production.
That does not mean the backlash wins. The demand side is real. Companies are not experimenting with AI because they enjoy vendor meetings. They are trying to reduce labor hours, compress software cycles, improve support, search internal knowledge, and move faster with fewer people. Some of those gains will survive the current hype cycle. Anthropic's revenue growth would be impossible if there were no useful work underneath.
But usefulness does not settle incidence. Someone pays for the grid upgrade. Someone pays for water, land, and cooling. Someone pays for the displaced task or the wage pressure on the person who used to do it. Someone pays for safety testing, incident disclosure, lawsuits, guardrails, and procurement reviews. The question for the IPO is how much of that bill stays outside Anthropic's income statement.
If the bill stays outside, the model lab looks like a software company with extraordinary growth. If more of it moves inside through higher infrastructure costs, slower capacity, regulation, compensation demands, or customer hesitation, the model lab looks more like a capital-intensive utility with a brilliant interface. Most outcomes sit between those poles.
I would watch three indicators after the public filing.
First, the language around capacity. Generic warnings about macro conditions do not say much. Specific language about data-center permitting, power procurement, cloud dependency, and construction delays would tell investors where the company sees real bottlenecks.
Second, the margin bridge. A lab can report huge revenue while token prices fall, model-training costs rise, and enterprise customers demand discounts. The public numbers that matter are gross margin by product, capacity commitments, and how much infrastructure cost is fixed before revenue arrives.
Third, the treatment of safety incidents. If disclosure becomes standardized, investors will need a mental model for release risk. A frontier lab that catches a dangerous behavior early may be safer and slower. A lab that ships faster may be taking a risk that only becomes visible later. The market will be bad at pricing that at first.
My prior is that Anthropic can still be a very good business. It has a strong enterprise position, a clear brand, and real demand. The public market may value it with old software reflexes while the cost base behaves like infrastructure, politics, and insurance.
The prospectus risk factor is useful for that reason. It tells investors where to look. AI backlash is not a mood. It is the price of turning intelligence into a physical industry.
Sources: CNBC, "Anthropic IPO filing will show AI backlash as a risk factor"; Gallup polling on local opposition to AI data centers; Politico reporting on OpenAI's California AI-law position; The Guardian reporting on Hollywood creatives training AI systems; FilmLA Research and Bureau of Labor Statistics data cited by The Guardian.
Originally published at deanlee.info.
Top comments (0)