SemiAnalysis Doubles Down: 50GW of Data Centers Due in 2027, With Just Under $1 Trillion in Leases Signed

Every week brings another warning that the AI datacenter boom will stall — moratoriums, grid limits, financing stress. SemiAnalysis, the research firm behind some of the most closely watched capacity models in the industry, published a blunt rebuttal on Monday: its new report, "Datacenters 2027: The Year of Acceleration," forecasts more than 50 gigawatts of new datacenter capacity delivered in a single year — roughly the entire cumulative global datacenter market built through 2024, replicated in twelve months.

The immediate trigger was a public disagreement. On the All-In Podcast, Altimeter's Brad Gerstner cited SemiAnalysis' earlier forecast of 43GW for 2027 and called it "too aggressive," offering his own estimate of about 25GW based, by his own description, on a hunch. SemiAnalysis' answer: raise the number to 50GW and show the construction photos.

The firm's leading indicator is physical, not financial. It counts more than 24GW across 40 campuses already in advanced stages of construction for 2027 delivery — most more than halfway done, with exterior cooling and electrical equipment already installed. Those 40 sites, it notes, are a small share of the several hundred campuses it tracks in advanced construction. Completed proof points include Crusoe's Abilene site, AWS's Indiana build and QTS and Microsoft's giant Georgia campus — the facilities behind models like GPT-6 and Fable 5.5. The same firm has called out delays before, including Oracle's New Mexico force majeure in April, so its model is not simply cheerleading.

The second indicator is money. Developers typically commit only 5-10% of project capital before securing a customer, then fund construction through pre-leases — usually 15-to-20-year contracts worth $30 billion to $50 billion in total contract value per gigawatt. Year to date, SemiAnalysis counts just under $1 trillion in binding, non-cancellable leases signed. Oracle alone closed roughly 4.5GW across three sites between August and October 2025, and Meta, AWS, Microsoft, Google, Nvidia, AMD, Anthropic and OpenAI have all entered the gigawatt-scale leasing market, alongside neoclouds signing hundred-megawatt deals.

The bearish case is not empty. Bernstein analysts recently argued that only about 35% of the US data center pipeline is credible, and SemiAnalysis itself lists labor shortages, supply chain delays and grid constraints as real risks it adjusts for. It is also worth remembering that SemiAnalysis sells access to its Datacenter Industry Model — a commercial interest in being right and being loud.

The report's most interesting warning is about what comes next: the shortage may move from power to silicon. Datacenters and electricity, it argues, may be the tip of the iceberg, with memory and logic chips possibly becoming the number one constraint in 2027-28. The industry is redesigning around the constraint — Google has modularized its electrical architecture with prefabricated modules, Meta has adopted a tent design, and modular builds are cutting construction timelines below ten months.

If even half of the 50GW lands on schedule, the implications compound: materially more compute supply, cheaper and smarter AI, faster model progress, and what the firm calls "unthinkable" capex numbers that will require interest rates to rise to clear the capital market. The 2027 debate, in other words, is no longer whether AI infrastructure gets built — it is whether chips and financing arrive fast enough to fill it.