On November 4, 2026, industry research firm SemiAnalysis published a report stating that SpaceX plans to build 13 gigawatts of AI data center capacity by the end of 2027—a construction pace equivalent to replicating in 18 months the total data center volume Google accumulated over 27 years since 1999. This figure is 30% to 80% higher than the approximately 7 to 10 gigawatts planned by the Stargate alliance led by OpenAI and SoftBank.
SemiAnalysis noted in its report that it has identified over 9 gigawatts of specific capacity across sites, of which 3.6 gigawatts is already under construction. The main projects are distributed across three locations: the Colossus 2 complex near Memphis, Mississippi; the twin-building cluster at Starbase in Texas; and a new project codenamed "TeraFab."
Just two months ago, this target was only 10 gigawatts. The pace at which SpaceX has successively raised its ceiling illustrates the speed of construction progress.
SemiAnalysis researchers wrote: "We now believe that SpaceX may be the fastest data center builder in the world, even surpassing its Chinese counterparts."
The Real Source of Construction Speed: Energy Storage Buys Time
The key variable behind SpaceX's speed lies in the method of power access.
The bottleneck in traditional data center construction is often not civil engineering but waiting for power: coordinating grid expansion, substation approvals, and high-voltage line construction typically takes more than 12 months. SpaceX bypasses this path. The Colossus complex has deployed up to 5.9 gigawatt-hours of Tesla Megapack battery energy storage systems, which can buffer compute demand using batteries when grid capacity is insufficient, thereby compressing the construction cycle from 12 months to 3 to 4 months.
This is precisely the embodiment of the vertical integration advantage formed jointly by Tesla and SpaceX. Tesla manufactures the Megapacks; SpaceX builds the data centers; both belong to the industrial group under Elon Musk. This relationship allows SpaceX to avoid queuing for external suppliers when procuring critical hardware, while traditional cloud providers building equivalent capacity face simultaneous congestion across multiple supply chains—GPUs, energy storage equipment, transformers, and more.
Specific construction records can already be verified: Colossus 1's first retrofit took 122 days, building 150 megawatts of capacity; Colossus 2 took 91 days, building 210 megawatts. The "Gun Range" facility in Southaven, Mississippi, targets completion in 4 months, and the "MiniHard" expansion in Memphis is expected to complete in about 5 months. SemiAnalysis described "Gun Range" as "to our knowledge, this will be the fastest greenfield data center ever built at scale."
A new project codenamed "Project SPAM" launched in October 2026, with plans to bring the first batch of compute systems online by December. If successful, it will break Colossus 2's 91-day record.
Who Is Paying for SpaceX's Compute
Beyond construction scale, more persuasive data comes from the customer side. The SemiAnalysis report disclosed that enterprise cloud agreements signed by Google and Anthropic with SpaceX have a combined annual contract value of $26 billion. The breakdown is even more striking: Anthropic has taken all of Colossus 1's capacity, securing over 300 megawatts of compute and approximately 220,000 Nvidia GPUs; Google has been allocated 110,000 GPUs at a price of $920 million per month through June 2029.
Anthropic is OpenAI's most direct competitor at the model level, and Google is the largest potential external pressure source for Stargate—yet both companies have chosen to bet substantial compute on infrastructure under Musk. This shows that at least at the engineering delivery level, SpaceX has established reliability that top AI labs can trust.
Extrapolating from business logic, this trust translates into equally striking revenue scale. In an earlier report from August 2026, SemiAnalysis estimated that if SpaceX leases half of its additional compute to external customers, it could support approximately $300 billion in annual recurring revenue by the end of 2027. This figure exceeds the annual revenue of any of the current top three global cloud service providers.
What Stargate's 7 to 10 Gigawatts Means
Stargate is a consortium jointly funded by OpenAI, SoftBank, Oracle, and Abu Dhabi sovereign fund MGX, with a total investment of $500 billion spanning four years. As of October 2026, Stargate's planned capacity is close to 7 gigawatts, with cumulative committed investment exceeding $400 billion. Its first site in Abilene, Texas, has achieved approximately 0.3 gigawatts of operational capacity. At the current pace, Stargate's 10-gigawatt target is expected to be reached around 2029.
If SpaceX reaches 13 gigawatts by the end of 2027 as planned, it will surpass the entire Stargate alliance's planned ceiling on a single-entity basis. The two are not competing on the same dimension—Stargate is a multi-party interest coordination body, while SpaceX is an engineering execution machine with a unified command chain. This difference in organizational structure will be amplified into an order-of-magnitude gap in decision-making speed and engineering response speed in infrastructure construction.
Power: The Only Real Bottleneck
The SemiAnalysis report notes that Colossus 2 has a potential power shortfall exceeding 300 megawatts, although energy storage systems and grid access may partially compensate. This detail is a key variable in assessing the credibility of the 13-gigawatt target.
Power is constrained by both physical infrastructure and regulatory approval. While SpaceX's energy storage strategy effectively reduces the time dependency on grid access, it does not equal unlimited power supply. As capacity under construction advances from 3.6 gigawatts toward 13 gigawatts, the cumulative effect of the power bottleneck will become increasingly difficult to cover through battery buffering.
This is why both parameters—"13 gigawatts" and "end of 2027"—must be met simultaneously to constitute a true milestone. Construction progress can be bypassed by energy storage; power capacity cannot.
The Deeper Impact of Non-Tech Companies Entering
SpaceX's entry into the AI compute arena represents something that has never happened in this industry before: a company known for manufacturing and engineering execution, with almost no reliance on software barriers, is entering the AI value chain from the infrastructure layer.
Over the past five years, control of AI infrastructure has been mainly held by three types of entities: hyperscale cloud providers (AWS, Azure, Google Cloud), GPU manufacturers (Nvidia), and model labs represented by OpenAI. SpaceX represents a fourth type—an infrastructure operator with manufacturing origins. The core competitiveness of such entities is not API calls or model training, but building larger-scale physical facilities at lower cost and in shorter timeframes.
When such a new entity appears on the supply side of compute, it directly affects the medium- to long-term trajectory of model usage costs. Large-scale competitive supply entering the market has historically never failed to compress prices. If SpaceX indeed builds 13 gigawatts of compute by 2027 and leases it externally at competitive prices, the current cloud providers' compute pricing logic based on scarcity will come under direct pressure.
The significance for the entire AI ecosystem is this: the evaluation and comparison of model capabilities is increasingly dependent on the precondition of "who can obtain how much compute at what price." When control of compute disperses to non-traditional entities like SpaceX, this precondition begins to become unstable.
Judgment
SpaceX's 13-gigawatt target, along with the construction speed described by SemiAnalysis, is internally coherent in its core logic: manufacturing supply chain advantage + energy storage replacing grid wait times = halved construction periods. This mechanism is supported by data from delivered projects; it is not a concept demonstration.
More importantly: a company that does not win through algorithms is exerting an asymmetric influence on the entire AI industry structure by controlling the supply speed at the physical layer. The infrastructure moat Google built over 27 years may be being filled in within 18 months by a rocket company.
If the power bottleneck is effectively managed during scale expansion, the AI compute landscape in 2027 will look very different from today. If the power bottleneck cannot be resolved, 13 gigawatts will remain on the engineering blueprints.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接