Anthropic Files for IPO with $42B Loss and $518B Spending Plans

Reuters reviewed Anthropic’s confidential IPO prospectus on September 28, 2026. The company reported a net loss of $42 billion in 2025. It had committed $518 billion to cloud, compute, and infrastructure obligations in the coming years.

An initial public offering turns a privately held company into a public one by selling shares to institutional and retail investors; the prospectus is the lengthy disclosure that lays out the issuer’s finances, obligations, and risks before that sale. Anthropic’s filing argued that artificial intelligence would transform the global economy more profoundly than industrialization, electricity, and the internet. The same pages put those two numbers beside that claim. One figure closed a single year. The other fixed the price of the capacity still required to train and run the models.

The prospectus also opened the full 2025 operating books—the path revenue and expenses had taken to produce the loss, and the cash and customer picture at year-end. Revenue grew twelve-fold, from about $400 million the year before to nearly $4.6 billion. The operating loss that year still topped $8.06 billion once writedowns were excluded. Total operating expenses reached $12.65 billion. Compute and infrastructure took $7.33 billion of that total, a threefold rise from 2024 and more than half of what the company spent to run itself.

Roughly $34 billion of the net loss was an accounting charge. It marked a higher estimated value on earlier financing that could convert into Anthropic shares; as the paper value of those arrangements rose, the books recorded the increase as a loss even though the cash had not gone out for chips, training. Nearly a quarter of the year’s revenue came from just two unnamed customers. Cash, cash equivalents and short-term investments stood at $20.28 billion on December 31, 2025.

The same filing fixed the terms of the far larger compute commitments still ahead. About 80 percent of the $518 billion had to be paid whether Anthropic used the compute or left it idle. The obligation was fixed. The usage was not.

Two of the largest slices ran on multi-year schedules built the same way. Google was owed at least $111.1 billion between April 2026 and July 2033. Amazon was due $110 billion between May 2026 and April 2036. The filing stated the shortfall rule without softening it: “If our actual spend falls short, we must pay Google the difference.” Parallel terms bound the Amazon agreement. The capacity could sit unused and the payments would still come due on the calendar the contracts had already written.

The prospectus named the other edge of that dependence in a single sentence. “If the compute we have access to from third parties is curtailed, repriced, or terminated ... our business, financial condition, and results of operations could be adversely affected,” Anthropic wrote. The bill for the decade ahead was already set. What remained open was the date the company would carry those pages into a public market. That debut was likely to wait until after the November midterm elections.

The confidential prospectus prepared for the sale already carried Anthropic’s own account of the exposure that came with systems able to act on a user’s behalf without steady human oversight. Autonomous capabilities meant the models could take steps and finish tasks with little ongoing direction; the same pages said those steps could produce real-world harm. “These autonomous capabilities could increase the potential for harm, as errors, misalignment, or security exploits may result in real-world consequences,” Anthropic wrote. The legal rules around such agents remained unclear, and contractual limits on liability might not protect the company from claims if the systems caused unexpected damage. The filing extended the warning across the work still ahead: “Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm.”

The day after Reuters reviewed those pages, Dario Amodei faced reporters following a White House session on September 29, 2026. Inside, the president had just hosted AI executives to discuss safety, infrastructure, and the rapid development of systems. “I will say what I've always said, which is that AI has incredible benefits,” Amodei said. He has frequently returned to medicine as one place those benefits could arrive. The same systems, he stressed, still carried very real risks that companies and others had not yet resolved how to manage. “As the president has said, whoever wins AI, wins.”

The ledger he described was simple and unfinished. On one side sat the chance of major advances in healthcare. On the other sat the open problem of harm as the tools grew more powerful. The meeting was about to produce a voluntary safety accord, and Amodei was still making the case for how the race could be run at all.

The companies signed that accord following the session. Participating firms agreed to establish robust internal controls to monitor their AI systems during training and deployment. They committed to working with independent external auditors to assess whether those controls were operating effectively, and to form board committees that would review the reports from both. One concern raised around the meeting had been the chance of systems accessing computers in unintended ways; the companies agreed to work toward preventing their tools from hacking or reaching technical systems outside the intended scope. The framework was not a federal regulation. It imposed no legally enforceable requirements.

Amodei finished the collaboration pitch to the reporters. “We all need to work together to make sure that we can win, and we can win safely if we do this right,” he said. “If we work with the president and everyone here, we can win safe.”

Competitive pressure was already in motion. Anthropic had launched Opus 5.5 the prior week to counter OpenAI’s GPT-6 Astra.