SemiAnalysis: The World Will Build 50 GW of Data Centers in 2027 Alone
SemiAnalysis forecasts that the world will bring roughly 50 GW of data center capacity online in 2027, close to the total global market at the end of 2024. Skeptics such as Brad Gerstner expect only 25 GW, but the analysts point to construction: more than 24 GW is already in an advanced stage, and signed leases total nearly a trillion dollars. Modular assembly and proprietary power plants are accelerating the buildout, which means chips and memory could become the bottleneck in 2027–28.
AI-processed from SemiAnalysis; edited by Hamidun News
Every week someone claims the data center boom will run into a shortage of electricity, workers, or money. SemiAnalysis answers with numbers: in 2027, the world will bring online roughly 50 GW of new AI capacity, by its estimate. For comparison, that is nearly as much as the entire global data center market had accumulated by the end of 2024. For those who build products on models, this is an important signal: there will be much more compute, which means inference will get cheaper and models will evolve faster.
Why the analysts don't buy the skeptics' view
The publication was prompted by a dispute. Investor Brad Gerstner of Altimeter, speaking on the All-In podcast, called SemiAnalysis's earlier forecast of +43 GW for 2027 too aggressive, saying it was based on "intuition" and that about 25 GW would be built. The analysts respond that they track specific facilities rather than going on feel.
Their main argument is construction. More than 24 GW of data centers are already at an advanced stage, and most of that will be delivered in 2027. At the 40 largest campuses alone, many already have more than half of the work done: external cooling and electrical equipment are installed. There are several hundred such facilities at an advanced stage. The source's headline also mentions about 60 GW of capacity under construction.
The second indicator is leasing. Developers typically spend only 5–10% of a project's budget from their own funds when there is no tenant, and the rest they need against "pre-leases": a 15–20 year contract worth $30–50 billion per gigawatt. Such a contract makes it possible to raise debt, and the debt turns into construction.
According to SemiAnalysis, since the beginning of 2026 leases worth almost a trillion dollars have been signed and closed (the report gives both "just under $1 trillion" and "more than a trillion" in different places), and most of the debt backing them has already been raised. The buyer pool has grown substantially: Meta, AWS, Microsoft, Google, Nvidia, AMD, Anthropic, OpenAI, and a range of smaller neoclouds.
How to build 50 GW in a year
The main problem now is not orders but delivery. The industry is responding by simplifying designs. Google replaced part of the outdoor wiring with prefabricated electrical modules. Meta uses lightweight "tent-style" buildings. AWS simplified the building shell and launched the fast SAMDC design, and increasingly assembles machine halls in factories (the Houdini project). AWS's construction timelines are under 10 months, and labor on engineering systems (usually 7–11 months of work) is reduced by up to 63%.
Third-party developers are following the same path. Singapore's DayOne, with roughly 3 GW of leased capacity, builds in about 9 months instead of the usual 18+, and it was that speed that helped it close deals with Firmus, OpenAI, and Nvidia.
On electricity, SemiAnalysis considers the problem largely solved: instead of waiting 5–10 years for a grid connection, companies are installing their own off-grid power stations. When production of turbines at giants like GE Vernova hit a ceiling, new suppliers emerged: fuel cells, aeroderivative engines, and refurbished turbines. Construction is also spreading beyond the US, and in 2027 overseas sites are speeding up markedly, while in 2028, the forecast says, they will take a large share of the market.
What is left out of frame
There are important caveats. This is a forecast from a commercial analytics firm that sells access to its models, and a key part, the question of a possible overbuild of data centers, sits behind a paid subscription. The authors themselves acknowledge delays, labor shortages, and financing problems, and cite the delay at an Oracle site in New Mexico as an example. There is also an oddity: the text gives different estimates of lease volume, so exact sums should be treated with caution.
Still, the conclusion has practical weight. If compute capacity grows as analysts expect, the shortage will shift from sites to memory and logic chips: the authors point to their report on the "great silicon shortage" and call these possibly the main constraint for 2027–28. At the same time, the capital requirement will be enormous, and rates, in the authors' view, will have to rise. Product teams should plan architecture assuming tokens get cheaper and that capacity for training and inference becomes more accessible through large labs. Founders of infrastructure startups should track leasing and construction volumes: these indicators lead the chip market and cloud pricing.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026OpenAI GPT-6 Cyber Found About 35 Flaws in iPhone, Mac and Other Apple Devices
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
Key Questions Answered
Why the analysts don't buy the skeptics' view?
The publication was prompted by a dispute. Investor Brad Gerstner of Altimeter, speaking on the All-In podcast, called SemiAnalysis's earlier forecast of +43 GW for 2027 too aggressive, saying it was based on "intuition" and that about 25 GW would be built. The analysts respond that they track specific facilities rather than going on feel. Their main argument is construction. More than 24 GW of data centers are already at an advanced stage, and most of that will be delivered in 2027.
How to build 50 GW in a year?
The main problem now is not orders but delivery. The industry is responding by simplifying designs. Google replaced part of the outdoor wiring with prefabricated electrical modules. Meta uses lightweight "tent-style" buildings. AWS simplified the building shell and launched the fast SAMDC design, and increasingly assembles machine halls in factories (the Houdini project). AWS's construction timelines are under 10 months, and labor on engineering systems (usually 7–11 months of work) is reduced by up to 63%.
What is left out of frame?
There are important caveats. This is a forecast from a commercial analytics firm that sells access to its models, and a key part, the question of a possible overbuild of data centers, sits behind a paid subscription. The authors themselves acknowledge delays, labor shortages, and financing problems, and cite the delay at an Oracle site in New Mexico as an example. There is also an oddity: the text gives different estimates of lease volume, so exact sums should be treated with caution.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.