SpaceX Targets 13GW of AI Data Centers by 2027, Backed by $40B GPU Financing Deal

Goldman Sachs and Apollo Global Management began marketing slices of SpaceX's $40 billion GPU debt package to investors Thursday morning, as a detailed SemiAnalysis analysis published this week projected SpaceX would bring approximately 13 gigawatts of AI data center capacity online by 2027 — a construction timeline that, if achieved, would equal all of Google's accumulated data center capacity built from 1999 through 2026, compressed into roughly 18 months.

The convergence of a live private credit offering and a comprehensive infrastructure roadmap places SpaceX's AI buildout under unusual simultaneous scrutiny from both financial markets and the technology industry. Speaking at SpaceX's third-quarter earnings call on October 20, Elon Musk said he was "highly confident" in 10 gigawatts of capacity online for 2027 alone and "would be surprised" if the total came in below 13 gigawatts.

Five Sites and a Build Pace Without Precedent

SemiAnalysis's analysis maps five distinct sites contributing to the 13-gigawatt projection: the original Colossus 1 campus in Memphis, Tennessee; the Colossus 2 facility in Southaven, Mississippi; a repurposed Starbase facility in Texas; a build code-named Project SPAM that broke ground on October 5; and a Gun Range plot targeting approximately four months from greenfield to operational — which would set the fastest large-scale greenfield build record on record if SpaceX meets that timeline.

The build pace at existing sites already established two records. The Colossus 1 retrofit in Memphis completed in 122 days at 150 megawatts. SpaceX then eclipsed that with the Colossus 2 retrofit in Southaven — 210 megawatts in 91 days, a new record for retrofitted AI cluster buildouts at that scale. Project SPAM, started October 5, is reportedly targeting a sub-91-day completion.

SpaceX's Super Intelligence division generated $7.56 billion in revenue in the third quarter of 2026, an 830 percent increase year-over-year. The company guided fourth-quarter capital expenditure at $35 billion to $40 billion — a figure comparable to what most major hyperscalers spend across an entire year — with approximately $15.8 billion of Q3's $19.6 billion capex flowing to AI infrastructure.

Behind-the-Meter Power: How 8GW Bypasses the Grid Queue

More than 8 gigawatts of SpaceX's planned capacity is tracked as behind-the-meter (BTM) power — on-site generation that connects directly to the data center facility rather than through utility grid interconnection agreements, according to SemiAnalysis. That distinction matters structurally: grid interconnect agreements in the United States typically require two to five years from application to energization, primarily because utilities must plan transmission capacity upgrades and obtain regulatory approvals at each step. BTM power, drawing primarily on mobile gas turbines contracted through FTAI Aviation and Solaris Power Cellular, allows SpaceX to build and power facilities on a timeline that would be physically impossible through the conventional grid interconnect model.

Satellite imagery published in late October confirmed a battery energy storage system (BESS) of approximately 5.9 gigawatt-hours at Colossus 2 in Southaven, a scale that would rank it among the largest single-site BESS installations in the United States. The BESS is not simply backup power. It absorbs excess turbine output during low-demand periods and can discharge during peak GPU utilization, effectively extending the deliverable compute capacity beyond the installed turbine nameplate. Combined with liquid cooling architecture targeting a power usage effectiveness (PUE) of approximately 1.10 to 1.15 — against an industry average closer to 1.5 — SpaceX's power design is engineered to push substantially more compute throughput per installed megawatt than conventional hyperscale data centers.

Colossus 2 is also under construction for a 500-kilovolt gas-insulated switchgear (GIS) substation connecting to the Freeport grid substation, scheduled for completion on December 31, 2026. That grid connection is expected to add approximately 1.15 gigawatts of additional power through two interconnections — Macrohard at 700 megawatts and MiniHard at 450 megawatts — supplementing the BTM generation and creating a layered power architecture not typical of AI data center builds.

The $40B GPU Lease Engineered to Stay Off the Balance Sheet

The GPU financing underpinning the buildout is structured as a master finance lease through a special purpose vehicle named Atlas SP. Under the arrangement, SpaceX creates a separate legal entity that holds title to the NVIDIA chips; the SPV issues debt backed by the GPU assets; SpaceX then leases the hardware from the SPV. The structure keeps $47.3 billion in total obligations — comprising $8.3 billion in senior secured loans plus bond tranches and a $7 billion prepayment provision — off SpaceX's consolidated balance sheet while preserving full operational control of the hardware. Goldman Sachs served as sole structuring agent; Apollo Global Management anchored the transaction alongside BlackRock, Goldman Sachs Asset Management, and KKR.

The hardware being financed is NVIDIA's Vera Rubin architecture. Each Vera Rubin GPU carries 336 billion transistors across a dual-reticle die, 288 gigabytes of HBM4 memory at 22 terabytes per second bandwidth, and NVLink 6 interconnects running at 3.6 terabytes per second per GPU. The NVL72 rack configuration packs 72 Vera Rubin units with 260 terabytes per second of aggregate NVLink bandwidth — a configuration purpose-built for the communication-intensive workloads of large language model training and inference. NVIDIA began full Vera Rubin production in Q1 2026; SpaceX's order under the $40 billion deal is scheduled for delivery in the first half of 2027.

SpaceX as Cloud Provider to Its Own Competitors

SpaceX's current compute tenants define the competitive context as sharply as any benchmark. Anthropic uses all of Colossus 1's roughly 220,000 GPUs at $1.25 billion per month through May 2029, per SpaceX's IPO filing. Google pays approximately $920 million per month since October 2026 for Colossus 2 capacity. Reflection AI runs GB300 GPU workloads at approximately $150 million per month. The pattern positions SpaceX not primarily as a consumer of its own compute — though Grok model development continues — but as a specialized AI cloud provider competing for the same frontier GPU workloads that drive the most profitable segments of AWS, Azure, and Google Cloud, while simultaneously operating as the infrastructure host to some of its chief rivals.

The annualized revenue rate implied by existing tenant deals alone approaches $26 billion, against a quarterly capex guidance of $35 to $40 billion. SpaceX is effectively wagering that the demand for frontier AI compute will continue outpacing the capacity it builds at the rate it is building it.

Turbine Compliance Narrows One Runway

The 13-gigawatt projection carries a regulatory constraint the capacity numbers themselves do not surface. Under an agreed order SpaceX signed with Mississippi's Department of Environmental Quality on July 30, 2026, the company must remove all 69 temporary gas turbines from its Southaven power plant by July 2027 and replace them with a permitted permanent facility rated at approximately 1.2 gigawatts. SemiAnalysis estimates those temporary turbines together provide roughly 2.7 gigawatts of nameplate capacity, creating a potential power shortfall of 300 to 500 megawatts during and after the transition — a gap the company will need to close through accelerated grid connections at the GIS substation, expanded BESS capacity, or new permanent turbine permits within the same window SpaceX is simultaneously trying to maximize total capacity growth.

An active Clean Air Act lawsuit filed by the NAACP and the Southern Environmental Law Center against xAI over unpermitted turbine emissions in Memphis adds further regulatory exposure. The Trump administration's Department of Justice moved to intervene and dismiss the suit citing national security grounds — an unusual intervention that signals both federal backing for SpaceX's buildout and the degree to which the company's infrastructure program operates outside conventional permitting frameworks.

The 13-gigawatt figure remains an analyst projection corroborated by Musk's public statements rather than an independently audited operational target. The first empirical test arrives in the first half of 2027, when NVIDIA begins delivering Vera Rubin GPUs under the $40 billion order and the investors now absorbing Goldman and Apollo's marketed debt tranches receive the first hardware-backed repayment data. Whether SpaceX can physically commission 13 gigawatts of AI compute in 18 months will be answered by build progress and power availability at five specific sites — not by the scale of the ambition that produced the number.