Anthropic's initial public offering prospectus, filed Monday, reads less like a tech company's growth story and more like a capital-intensive infrastructure play with regulatory headwinds.
The 90-page document spells out operational risks that cut to the heart of large-language-model economics. The most immediate: power and real estate constraints. Anthropic requires an expanding network of data centers to train and operate its Claude models, but lawmakers in 16 states are considering moratoriums on new data center development, according to the National Conference of State Legislatures. Organized opposition has grown across jurisdictions, and Anthropic's public profile may draw specific targeting.
The power problem compounds the economics. The prospectus identifies a national and global shortage of sufficient electrical capacity as a significant risk. Even if Anthropic secures on-site natural gas generation, increased demand will drive up power costs—directly compressing the margins on training and inference at scale.
The filing also discloses risks around AI safety and model reliability. Anthropic states it must prioritize safety protocols, but acknowledges the technology's inherent complexity makes it difficult to anticipate and plan for every failure mode. The company cannot assure its safety measures will be sufficient. Competition and stakeholder pressure complicate safe deployment further.
Anthropically faces political risk from the Trump administration—specific disputes are flagged in the prospectus without elaboration. The company also carries legal exposure around child safety. While Anthropic has avoided the litigation that has dogged OpenAI and Meta (which recently settled child safety claims for approximately 17 billion dollars), its terms of service could expose it to similar claims.
The filing does not disclose specific financial projections or unit economics for model training and inference—a notable omission that leaves open questions about whether the capital intensity of the infrastructure build-out can generate returns sufficient to justify IPO valuations in a sector where compute costs are only rising.
