Chaos detected. Analysis loading.
Elon Musk just dropped a number that breaks the data center playbook. 10 gigawatts of incremental computing power by end of 2027. From SpaceX. The rocket company. The one that builds Starships. Not Amazon. Not Google. SpaceX.
SemiAnalysis put the feasibility stamp on it. Musk's conservative target: 6-8GW in 2027. Upside? Above 10GW. The capex per GW? Roughly $50 billion. That pencils out to $300-500 billion in 2027 alone. For context, that's half the US federal budget for a single year. Poured into silicon and cooling towers.
Context: Why Now?
This isn't a random pivot. SpaceX has been quietly building a data center division inside the company. Their Starlink network already handles massive data throughput. The next logical step: host compute where the bandwidth lands. The SemiAnalysis report, dated late 2025, reveals the economics. Each GW of GB300 clusters can generate over $100 billion in annual revenue from API inference services. OpenAI and Anthropic are the anchor tenants. At $3 per GPU per hour, the annual operating cost per GW is about $12 billion. Gross margin? 88%. That's not a data center. That's a money printer.
Microsoft already signed a $250 billion infrastructure deal with OpenAI in October 2025. That's roughly 7GW. SemiAnalysis now models a potential 3GW contract between Microsoft and SpaceX. Total value: ~$150 billion. If that lands, SpaceX's annual recurring revenue could hit $300 billion by end of 2027. That's bigger than all of crypto's current market cap. Combined.
Core: The Numbers That Matter
I've been tracking compute infrastructure buildouts since 2021. I watched the hyperscalers gobble up every available watt. But this is different. The capex per GW is $50 billion. That's not a typo. To put it in perspective: a typical hyperscale data center costs $1-2 billion per 100MW. Scaling that to 1GW is $10-20 billion. SpaceX's figure is 2.5-5x higher. Why? Because they're building from scratch, likely leveraging SpaceX's own manufacturing and launch capabilities. The cost includes the GPU clusters, the power infrastructure, the cooling, and the real estate. They're not leasing colo space. They're building cities.
SemiAnalysis's model shows that at $3/GPU/hr, the annual revenue per GW is $100B+. That's a 2x return on capex in year one. But that assumes 100% utilization. The reality? AI inference demand is spiky. OpenAI and Anthropic need massive capacity for training and inference, but they also have competing offers from Google, AWS, and Oracle. The 3GW deal with Microsoft is not a lock. It's a negotiation.
I've seen this playbook before. In 2017, EOS raised $4 billion for a blockchain that never delivered. The narrative was seductive. The numbers were huge. The execution? A mess. Musk's track record with timelines is notorious. FSD was supposed to be solved in 2019. Neuralink's human trials are still in early stages. The pattern is: promise big, underdeliver, recalibrate, eventually produce something. But 10GW by 2027? That's not a stretch. That's a leap.
Contrarian: The Unreported Angle
Everyone is focused on the revenue. The blind spot is the power grid. 10GW of computing power requires 10GW of electricity. That's the output of 10 nuclear reactors. The US grid is already strained. SpaceX plans to build in Texas, likely near the Gulf Coast. But the interconnection queue is years long. Even if they build their own power plants, the permitting alone takes 2-3 years. The timeline is tight.
Another blind spot: the GPU supply chain. GB300 clusters are based on NVIDIA's next-gen architecture. NVIDIA is already sold out through 2026. SpaceX would need to secure allocation. That means competing with every hyperscaler on the planet. Musk's relationship with Jensen Huang is good, but business is business. The 10GW number might be aspirational, not contractual.
Then there's the software stack. The SemiAnalysis report assumes OpenAI and Anthropic will use SpaceX's clusters for inference. But those companies are vertically integrating. OpenAI is building its own chips. Anthropic is working with Google. The 3GW deal with Microsoft might be the only solid anchor. If that falls through, the revenue math collapses.
Let's not ignore the regulatory risk. The US government is increasingly wary of concentrated compute power. The CHIPS Act and export controls are tightening. A single entity controlling 10GW of AI compute could trigger antitrust scrutiny. Musk's own political stance adds another layer of uncertainty.
EOS didn't die; it evolved. Do you? The same applies here. The narrative is evolving from 'rocket company' to 'compute utility'. But the risks are real. The margin of error is zero. Until it isn't.
Takeaway: What to Watch
The next 12 months are critical. Watch for SpaceX's data center construction permits in Texas. Watch for any public compute contract with Microsoft or other hyperscalers. Watch for NVIDIA's allocation announcements. If SpaceX secures 3GW of GB300 supply by mid-2026, the 10GW target becomes plausible. If not, it's vaporware.
I'm not betting against Musk. I'm betting against the timeline. The math works. The physics works. The politics? That's the wildcard. The question is not whether SpaceX can build 10GW of compute. The question is whether the world will let them.