Losing billions, the AI labs now want to slow down for safety

By Takeshi Mori · 2026-09-29

The verdict

Safety-first rhetoric from labs burning at this rate functions as regulatory capture. The rules they helped write raise barriers that only incumbents can clear. Read the financials.

What the UN heard

On 23 September 2026, the United Nations Security Council convened a high-level meeting on artificial intelligence and international security. Executives from Anthropic and OpenAI briefed the Council on what they described as real and imminent threats from runaway AI development.^[6]

Five days earlier, Anthropic CEO Dario Amodei published an essay on his personal blog titled "We Must Pace the Frontier." Amodei argued that AI labs must slow the pace of capability improvements. He warned that AI-driven botnets could cause hundreds of billions of dollars in damage within six to twelve months.^[1] The essay landed while Anthropic's bankers were finalising an IPO prospectus that told a very different story about the company's own trajectory.

Anthropic's numbers

The prospectus, filed 28 September 2026, showed Anthropic lost more than $8 billion in 2025 against $4.6 billion in revenue. A quarter of that revenue came from just two customers. That concentration would concern any credit analyst, let alone a safety-focused mission-driven lab.^[7] Run-rate revenue had surpassed $30 billion by early 2026, up from roughly $9 billion at the end of 2025. The model is scaling fast, but costs are rising too.^[3]

The compute commitments underpin those numbers. Anthropic signed a multi-year agreement with Google and Broadcom for multiple gigawatts of TPU capacity. The infrastructure is expected to come online from 2027.^[4] Anthropic CFO Krishna Rao called this part of a disciplined approach to scaling infrastructure. This is a curious word choice for a company whose losses outrun revenue by nearly two to one.^[4]

A separate agreement with SpaceX gave Anthropic access to all compute capacity at SpaceX's Colossus 1 data centre, over 300 megawatts of new capacity.^[5] That deal, signed in May 2026, followed the Google-Broadcom announcement by less than a month. Every gigawatt locked in becomes another argument against any regulator suggesting the lab should slow down. The capital is already committed.

OpenAI's numbers

OpenAI's situation involves larger sums and steeper trajectories. The company closed its latest funding round in March 2026 with $122 billion in committed capital at a post-money valuation of $852 billion.^[2] Leaked financial projections show revenue scaling from $36 billion to $350 billion between 2025 and 2030. The cash burn keeps pace. Negative free cash flow is projected to reach $278 billion by 2030. The $122 billion raise is expected to be exhausted by 2028 under the company's own model.^[8]

Why the math makes slowing down impossible

Both companies have structured themselves so that deceleration is financially catastrophic. When compute agreements run to hundreds of billions and investors expect returns on an $852 billion valuation, model development does not pause because a rival lab's CEO published a blog post. The play is to ship faster and raise again before 2028.^[8]

The signalling value of safety rhetoric deserves a clear read. Regulations shaped by labs at this scale, covering compute thresholds, safety-evaluation requirements, and licensing regimes, function as moats. A startup burning $10 million a year cannot satisfy the same audit requirements as a company that wrote them. Amodei's essay and the UN briefing are part of the financial plan.

On 28 September 2026, the same day Anthropic's prospectus published, OpenAI shelved the GPT-6.1 Astra model launch and apologised to Australians over a Medicare data exposure incident.^[9] Labs that brief the UN Security Council on existential risk one week and pull a major product launch the next have not resolved the tension between safety and growth. Operators building on their APIs should price that instability into their own planning. The combined compute spend across both companies sits at approximately $1.37 trillion. OpenAI's next raise, required by 2028 under its own projections, will set a new record.^[8]

Takeshi Mori

Takeshi Mori writes about technology and start-ups. He is curious about how products get built and who they are really for, and he would rather see a thing working than hear it described.