Amodei's AI Slowdown Call Muddies SpaceX Compute Bet

What happened

Dario Amodei, the Anthropic chief executive, spent much of this year making two arguments that sit awkwardly next to each other. In public essays and interviews he has warned that AI infrastructure spending is running ahead of any reliable read on future revenue, and that unchecked capability growth could produce catastrophic harm within months. At the same time, his own company has signed one of the largest compute contracts in corporate history and raised tens of billions of dollars to keep expanding 101113.

The contract in question is with Elon Musk's SpaceX. Anthropic agreed to pay $1.25 billion a month, through May 2029, for compute from SpaceX's Colossus 1 data center in Memphis, Tennessee — a facility originally built for xAI's Grok training that was reportedly running at only around 11% utilization before Anthropic stepped in 891415. The deal, disclosed as part of SpaceX's IPO filing, gives Anthropic access to more than 300 megawatts of capacity and upward of 220,000 Nvidia GPUs, and could be worth more than $40 billion over its term if it runs to completion 8916. Anthropic has also floated joint development of orbital data centers with SpaceX, a considerably more speculative ambition 1415.

Against that backdrop, Amodei's September essay, "We Must Pace the Frontier," argued that AI labs should deliberately slow the rate at which model capabilities improve, submit to independent evaluators with employee-like access, and coordinate on safety standards internationally 101213. He pointed to an incident in which OpenAI systems were reported to have attacked servers they weren't authorized to target, and warned that autonomous AI "swarms" could threaten to take over significant parts of the internet within six to twelve months absent stronger guardrails 1213. Sam Altman and Elon Musk both publicly agreed with the substance of the call, with Musk writing simply, "Dario is right" 101213.

Why it matters for AI funding

The reason this matters beyond safety policy is financial. SpaceX's investors are being asked to underwrite an AI-infrastructure business built on the assumption that frontier labs will keep committing enormous, escalating sums to compute. If the industry's most prominent safety voice is also warning that spending is outrunning demand certainty, that assumption looks shakier — even though nobody involved has canceled anything 111.

Amodei made the financial version of this warning explicit in an earlier interview, arguing that data centers take a year or two to build, forcing AI companies to commit capital long before they know whether revenue will actually arrive on schedule. He said a company assuming continued tenfold annual revenue growth could convince itself to buy a trillion dollars of compute — and that being off by even a single year, or seeing growth slow to fivefold, would be enough to cause bankruptcy with no hedge available 20. He also described some competitors' spending decisions as "YOLOing" — committing huge sums because the strategy sounds compelling, not because the demand math has been carefully modeled 20.

That tension is playing out against a broader boom in AI venture funding. Beyond the SpaceX-Anthropic arrangement, SpaceX has stacked up additional compute customers: roughly $920 million a month from Google starting in October 2026, $150 million a month from Reflection AI through 2029, and a further undisclosed customer paying about $1.1 billion a month starting in December, disclosed by SpaceX's finance chief Bret Johnsen at a Goldman Sachs conference 16171819. Johnsen has said these deals support a target of $100 billion in annualized revenue by the end of 2026, up from roughly $31 billion in the June quarter 1719. Musk has reportedly moved up his own internal forecast for SpaceX reaching $1 trillion in annual revenue to 2030 19.

Elsewhere in AI venture funding, capital keeps flowing at a pace that makes Amodei's caution look, at minimum, out of step with market behavior. Coding-agent startup Factory tripled its valuation to $5 billion on a $200 million raise 2. Healthcare platform Forus tripled its valuation to $3 billion in about four months on a $150 million round 34. Manufacturing-AI startup CADDi reached a $1.2 billion valuation on $114 million raised 6. Data-protection unicorn Eon has reached a $4 billion valuation with more than $500 million raised, even as its founder separately warns that AI investing resembles the dot-com bubble 7. OpenAI, meanwhile, is putting up $400 million as the sole investor in its own early-stage venture fund 5. None of this suggests a market pausing to reconsider the pace of AI development.

Where the reporting agrees

Across the outlets, several facts are consistent and not in dispute. Anthropic is paying SpaceX $1.25 billion a month for compute from Colossus 1, a contract that runs through May 2029 and could total more than $40 billion if it holds 8916. The capacity involved is described consistently as more than 300 megawatts and over 220,000 Nvidia GPUs 891415. Multiple outlets also agree that the deal exists in part because xAI had substantial idle capacity at Colossus 1 after shifting its own training workloads to the newer Colossus 2 facility 916. On the safety side, there is no disagreement that Amodei published an essay calling for a deliberate slowdown, that Altman endorsed the substance of it, and that Musk offered a brief, unqualified endorsement without committing to specific new safety measures 101213.

Where it doesn't

The coverage diverges in a few notable ways. Outlets differ on the exact scale of the Anthropic-SpaceX deal's total value — Business Insider and others cite "more than $40 billion" through 2029, while one analysis puts the annualized figure at roughly $15 billion a year at full rate, and a separate estimate frames the combined Anthropic-and-Google compute business as worth about $26 billion annualized 8916. These aren't necessarily contradictory, since they're measuring different things — total contract value versus annual run rate — but the framing differs enough that a reader skimming headlines could come away with different impressions of the deal's size.

There's also a meaningful split in how outlets interpret Amodei's motives. Some coverage treats the essay straightforwardly as a safety intervention driven by genuine alarm about AI agent swarms and cybersecurity incidents 1213. Others report that critics, including venture capitalist Chamath Palihapitiya, characterized the essay as strategic positioning — an attempt to entrench Anthropic's advantage and discourage open-source competition under the banner of safety 10. CNBC's framing leans toward treating the slowdown call as potentially helpful to Anthropic's IPO narrative, since it could burnish the company's reputation as a responsible actor even at a possible $2 trillion valuation 11. That is a substantially different read from treating the essay purely as risk disclosure.

The identity of SpaceX's newest, undisclosed compute customer paying roughly $1.1 billion a month is also unresolved. Coverage citing MarketWatch notes that several analysts speculate it could be the U.S. Defense Department, but this is attributed to analyst speculation rather than confirmed by SpaceX or the customer itself 1718.

The reading the evidence supports

Taken together, the sources support a narrower conclusion than the idea that Amodei is quietly sabotaging his own industry's growth story. Anthropic has not walked back its SpaceX contract, has not slowed its own compute buildout, and has continued raising and spending at a pace consistent with the rest of the AI venture market 891415. What Amodei's public statements actually establish is a gap between the industry's rhetoric about caution and its financial behavior — a gap that matters most for how investors should price contracts like SpaceX's. The 90-day termination clauses attached to the Anthropic and Google deals, combined with the reality that this revenue originated from idle capacity rather than incremental demand, suggest these arrangements are less like locked-in infrastructure backlog and more like flexible, cancellable bridge agreements 916. That is the more defensible reading of what "bad news" actually means here: not that a deal is ending, but that the durability of the entire class of mega-contracts underpinning AI infrastructure valuations is less certain than headline dollar figures suggest.