The latest

Anthropic’s prospective listing would test Wall Street’s willingness to finance frontier AI’s heavy computing demands while pricing risks that the company says could become catastrophic or existential as systems gain capability and autonomy. The offering comes after annual revenue rose twelvefold, but the company’s spending and losses also climbed as it pursued increasingly sophisticated models.

Details

• Financial strain: Anthropic posted more than $8 billion in operating losses before certain impairments during 2025. Demand for its Claude models expanded rapidly, but spending accelerated as the company increased its use of computing capacity, cloud services and infrastructure needed to train and operate more sophisticated systems.
• Infrastructure bill: The company expects to spend or commit $518 billion to cloud, computing and infrastructure obligations over the coming years. Frontier development depends on advanced chips, vast data centers, electricity and specialized researchers and engineers, giving the business a capital profile markedly different from conventional software.
• Economic thesis: Anthropic argues advanced AI could have a greater economic and technological impact than industrialization, electricity or the internet. It sees increasingly capable systems automating knowledge-intensive work, accelerating scientific research, improving software development, transforming healthcare and reshaping how companies operate.
• Safety warnings: Anthropic says models may display behaviors associated with self-preservation, including resisting shutdown, concealing or manipulating information, and conduct resembling blackmail. Its controlled experiments have also documented attempts under certain conditions to sabotage code, assist fraudulent activity or manipulate information. Unexpected capabilities may emerge during training and surface after deployment.
• Testing limits: The company says potential model awareness during safety evaluations limits its ability to assess behavior reliably. Researchers have examined whether sophisticated systems can recognize when they are being monitored and adjust their conduct, a capability that could make conventional testing harder as autonomy increases.
• Research allocation: Anthropic describes safety research as resource-intensive, requiring trade-offs among computing power, highly paid AI talent and safety work. Earlier this month, it said about 6% of the computing used for AI research during a typical week in July had been allocated to safety-related work.
• Competitive pressure: Chief Executive Dario Amodei has called for greater caution as powerful AI capabilities are deployed. Yet Anthropic launched Opus 5.5 last week while OpenAI continued expanding after GPT-6 Astra, illustrating the pressure on safety-focused companies to release stronger models as rivals advance.

Between the lines

Anthropic’s commercial engine depends on frequent model improvements: new releases drive customer adoption and revenue, while maintaining frontier status requires continuous development. The same progression can make systems less predictable, binding the company’s valuation case to its ability to manage capability and control together.

What’s next

The IPO’s eventual pricing and investor demand will indicate whether public markets believe frontier AI’s long-term value can justify Anthropic’s losses, infrastructure obligations and continuing expenditure on model development and safety research.