SemiAnalysis report predicts that in 2027, global AI data center new capacity excluding mainland China will reach 51.2GW

Recently, the semiconductor research firm SemiAnalysis published a report titled "Datacenters 2027: The Year of Acceleration," predicting that new data center IT capacity added in markets outside mainland China will reach approximately 51.2 GW in 2027, more than doubling from 25 GW in 2026 and more than tripling from 15.6 GW in 2025. The new capacity in 2027 is nearly equivalent to the total cumulative data center capacity built worldwide (excluding mainland China) over all years through the end of 2024.

In this report, SemiAnalysis directly characterizes 2027 as the "year of acceleration" for AI data center construction, and uses three indicators—satellite imagery, lease contract values, and self-build construction starts—to directly rebut market doubts about whether "AI infrastructure can be delivered on schedule."

Satellite imagery and trillion-dollar leases: where does 51 GW come from

Faced with the doubt that "planning does not equal delivery," the first piece of evidence SemiAnalysis offers comes from satellite imagery. Satellite images cited in the report show that more than 24.3 GW of data center capacity worldwide is distributed across about 40 campuses in the "late stages of construction"—most projects are more than half complete, external cooling and electrical equipment has been installed, and delivery is expected within 2027. The report stresses that these 40 campuses are only a small fraction of the hundreds of sites that have entered late-stage construction.

The second piece of evidence comes at the data center contract level. Data tracked by SemiAnalysis shows that from the start of 2026 to date, the value of legally binding, non-cancelable data center leases signed worldwide is slightly below $1 trillion. Developers typically invest only 5% to 10% of a project budget before locking in customers, so the signing of large-scale pre-lease agreements is seen as a key leading signal that debt financing is in place and construction is advancing.

The report notes that Meta, AWS, Microsoft, Google, Nvidia, AMD, Anthropic and OpenAI are all seeking gigawatt-scale leases in the market, and that Oracle also completed lease transactions of about 4.5 GW in the second half of 2025.

The third piece of evidence is the volume of self-built data center construction starts by hyperscale companies. The quarterly megawatts of construction starts by U.S. hyperscale cloud providers tracked by SemiAnalysis show "parabolic growth," and the report states bluntly that "U.S. hyperscale cloud providers have never built data centers at this scale."

Focus of disagreement: 51 GW or 25 GW

The release of SemiAnalysis's report comes as the debate over whether the AI infrastructure boom can continue has reached a fever pitch.

At the All-In Summit in September 2026, Altimeter Capital founder Brad Gerstner publicly questioned SemiAnalysis's earlier forecast of 43 GW, calling it "too aggressive," and based on his personal judgment said actual deliveries in 2027 would be close to 25 GW, flat with that year. Gerstner's comments focused mainly on the U.S. market, while SemiAnalysis's latest forecast covers a broader geographic scope.

SemiAnalysis responded to such doubts in somewhat pointed language in the report: "We focus on facts, not narratives." The report also acknowledges that constraints from power, labor, supply chains and capital genuinely exist, but stresses that "mega data centers really are under construction, and many projects are proceeding on schedule." Citing Crusoe's project in Abilene, Texas, AWS's project in Indiana, and the QTS/Microsoft mega-campus in Georgia as examples, the report says these are precisely the actual compute facilities supporting the training of frontier models such as GPT-6.

Modularization cuts construction time to 9 months, but the power grid remains the key bottleneck

Part of the reason doubling data center capacity in 2027 is engineering-feasible is a fundamental change in how data centers are built.

SemiAnalysis calls the current data center construction model "Lego data centers": electrical rooms and liquid cooling systems are prefabricated in factories as modules that can weigh up to 50,000 pounds, advancing in parallel with on-site civil works. Taking a 50 MW AI data hall as an example, traditional construction takes 22 months, while a fully modular approach can compress that to 14 months (excluding permitting); the eight months saved are worth about $200 million to the chip owner—GPUs generate rent as soon as they are powered on, while empty machine rooms continue to depreciate and accrue interest.

The report cites the aggressive case of Singapore-based operator DayOne, saying its construction time has been compressed from more than 18 months to about 9 months. SemiAnalysis estimates that modular construction can reduce on-site labor needs by up to 63%.

However, modularization compresses the construction time of the data hall shell, not the speed of grid interconnection. Around the time the report was published, Microsoft confirmed it had abandoned a 50 MW data center lease near London because the local grid operator said power might not be available until the 2030s. This case illustrates precisely that a built data center truly counts as "online" only when the building, the power and the chips are all in place.

SemiAnalysis's judgment on this is that the power generation problem has been "largely solved." Developers no longer wait five to ten years in grid interconnection queues, but instead accelerate the deployment of on-site power generation equipment, including fuel cells, refurbished gas turbines and even engines originally designed for trucks. These "behind-the-meter generation" solutions bypass the grid connection delays that caused Microsoft's UK project to stall, but the report also acknowledges that they may not be suitable for all sites.

The SemiAnalysis report shows that, to cope with grid power supply bottlenecks, North American data centers are frantically snapping up "self-built generation" equipment, and cumulative bookings for "on-site generation" (behind-the-meter generation) have soared from almost zero in early 2024 to an astonishing 75 GW by August 2026, showing a typical exponential explosion curve. The number of equipment suppliers participating in supply has grown in tandem with the order surge, rapidly expanding from only a very small handful at first to 22.

Facing the huge power gap, data centers have adopted a variety of power generation solutions. Among them, gas turbines (yellow) are the absolute mainstay; reciprocating engines (blue) and fuel cells (orange-red) account for a significantly higher share of the 2026 increment.

Data center increments mainly come from the United States

Looking at the geographic distribution of new data centers, SemiAnalysis breaks down in detail the net new additions of global (excluding China) data center IT capacity from 2023 to 2028, and focuses on comparing the growth momentum and market share changes of the United States (US) versus non-US regions (ex-US).

From the chart above, we can see that the blue bars (non-US data center capacity) maintained steady growth from 2023 to 2025. Explosive growth begins in 2027, rising from 6.9 GW in 2026 to 13.2 GW in 2027. Likewise, global data center capacity (yellow bars, excluding mainland China) will also leap to 38 GW in 2027. Entering 2028, due to domestic infrastructure and power constraints in the United States, its spillover effect will also significantly drive an explosion in non-US regions (share rebounding from 26% to 35%).

The next bottleneck: from power back to chips

The report also points out that the completion of data center construction does not mean GPUs can be filled on day one. As power and construction constraints gradually ease, SemiAnalysis expects that in 2027-2028, semiconductor chips and high-bandwidth memory (HBM) will again become the primary bottleneck. Clearly, this forecast not only responds to the market's concerns about future "compute surplus," but also responds to the market's concerns about future "memory chip surplus."

SemiAnalysis's logic is: an increase in compute supply does not equal demand saturation. On the contrary, falling AI inference costs will accelerate the adoption of autonomous agents, enterprise applications, robots and physical AI; lower costs will unlock new application scenarios that were previously not economically viable, thereby spawning a new round of compute demand and memory chip demand.

Summary:

If the single-year delivery volume of 51.2 GW of global (excluding mainland China) data center capacity in 2027 can be realized on schedule, it means that the demand for GPUs, HBM, network switching equipment, liquid cooling systems, power supply equipment and power infrastructure corresponding to each additional 1 GW of capacity will be released in a concentrated manner in 2027. SemiAnalysis states bluntly in the report that supply chain companies providing compute, memory, networking and power supply will become the main beneficiaries in the coming years.

But the report also plants an observation point that needs continuous tracking: whether the 24 GW of capacity in late-stage construction can truly be converted into "powered and connected" usable compute in 2027, and whether compute chip and memory supply can keep up with this pace without becoming a new chokepoint.