Anthropic CEO warns of “double counting” in AI data center megadeals, urges press not to chase megawatt headlines
A wave of multi‑gigawatt AI data center announcements has electrified the market, but not everyone is convinced the numbers add up. Anthropic CEO Dario Amodei says parts of today’s AI infrastructure boom look “a little bit fishy,” cautioning that some high‑profile deals appear to be double—and even triple—counted in press and partner disclosures.
Amodei, speaking alongside Marc Benioff, didn’t call out specific companies. His point, however, was clear: the industry’s rush to tout multi‑GW facilities and eye‑popping capital commitments can blur what’s actually being built and who is truly paying for it. In many cases, the same data center capacity or investment may show up in multiple announcements from different participants—cloud providers, chip suppliers, and AI labs—creating the impression of several distinct projects when it’s actually one shared effort.
The comments land amid a flurry of partnerships led by major AI players, including OpenAI, whose nonprofit parent oversees its structure, alongside critical suppliers such as NVIDIA and hyperscalers like Microsoft. Markets have reacted enthusiastically to these large-scale infrastructure narratives, but Amodei’s remarks suggest a need for more scrutiny, not less. He emphasized that he’s bullish on the long-term data center buildout; his concern is about how the numbers are being framed and amplified.
What “double” or “triple” counting can look like
– The same facility counted by multiple partners: a cloud operator, a chip vendor, and an AI developer each claim the full value or capacity.
– Repeated tallies across stages: an initial memorandum of understanding, financing commitment, and construction start all get publicized as separate, additive milestones.
– Power versus capacity confusion: total megawatts announced without clarifying timelines, power availability, or how much is reserved versus truly installed.
Why this matters for AI infrastructure
– Realistic capacity planning: training and deploying frontier models depends on dependable power, racks, and supply chains—not just press releases.
– Capital efficiency: investors and partners need to distinguish between unique, incremental data center buildouts and shared investments.
– Policy and grid planning: overstated numbers can distort expectations for power procurement and regional energy strategies.
Despite the skepticism, the direction of travel is unmistakable. Demand for compute is exploding, and the AI data center race is accelerating with multi‑billion‑dollar deals emerging almost weekly. For labs like Anthropic and OpenAI, deep partnerships with chipmakers and cloud platforms remain essential to scale models and deliver reliable AI services.
The takeaway for readers and investors: celebrate genuine progress, but ask the clarifying questions that cut through the hype. What portion of an announced investment is unique versus shared? Who ultimately owns and operates the facility? How much capacity is contracted, funded, and under construction today? What is the delivery timeline and power source? In an era of multi‑GW headlines, transparency will separate real infrastructure from recycled numbers.
