The data shows a capital expenditure projection that would make most sovereign treasuries blush. SpaceX, according to a SemiAnalysis report, is targeting over 10 gigawatts of incremental computing power by the end of 2027. Elon Musk’s conservative estimate sits at 6-8GW, with upside pushing beyond 10GW. At roughly $50 billion per gigawatt, the total capex for 2027 alone could land between $300 and $500 billion.
Ledgers don’t lie. But the assumptions behind those ledgers require forensic unpacking. I’ve spent the last decade auditing tokenomics and capital deployment models—from the 2017 ICO boom to the 2024 ETF institutional flows. This projection is not a whitepaper. It’s a real-world infrastructure buildout that, if executed, would reshape the global compute landscape. But the numbers demand a rigorous, bear-case-first examination.
Context: The Compute Arms Race
SpaceX is not a cloud provider. It’s a rocket company that also operates Starlink. But the SemiAnalysis report paints a different future: SpaceX as a hyperscale AI compute provider, leveraging its manufacturing prowess and access to cheap energy (via methane and potentially renewable sources) to build massive data centers. The underlying driver is the insatiable demand for AI training and inference. Microsoft’s $250 billion infrastructure agreement with OpenAI, signed in October 2025, corresponds to about 7GW of compute capacity. SemiAnalysis estimates a possible SpaceX contract with Microsoft for roughly 3GW, valued at approximately $150 billion. That’s a single client deal larger than the GDP of many nations.
To put this in perspective, the entire Bitcoin network currently consumes about 0.15GW. One GW of compute running NVIDIA GB300 clusters can generate over $100 billion in revenue per year when providing API inference services for OpenAI or Anthropic, according to the SemiAnalysis model. The annual cost at $3 per GPU per hour is about $12 billion per GW. That’s an 88% gross margin, assuming full utilization. These are not numbers from a crypto white paper; they are derived from public AI pricing and hardware costs. But the pattern here is familiar: high-capital, high-margin infrastructure plays that promise exponential returns. The blockchain remembers every step; do you? The 2021 DeFi summer saw similar margin projections—until the liquidity drained.
Core: Breaking Down the 10GW Thesis
Let’s build the evidence chain. First, the capex. $50 billion per GW includes land, power infrastructure, cooling, and the clusters themselves. For 10GW, that’s $500 billion. SpaceX’s total funding to date is a fraction of that. The capital must come from somewhere—likely a combination of debt, equity, prepaid service contracts (like Microsoft’s), and government subsidies. The SemiAnalysis report claims SpaceX’s annual recurring revenue could reach $300 billion by end of 2027. That implies a revenue-to-capex ratio of 0.6, which is low for a capital-intensive business but plausible if the compute is monetized immediately.
To verify, I modeled the cash flows. Assume 10GW online by year-end 2027, with a ramp starting in 2025. The report suggests SpaceX can deliver 6-8GW in 2027 alone. That means the first GWs come online in 2026, generating revenue that partially funds subsequent builds. The $3/GPU/hour cost is the operating expense; the $100B/GW revenue is the top line. Net margin after opex: $88B per GW. But that assumes 100% utilization at 100% of the contracted price. In reality, AI inference demand is volatile. OpenAI and Anthropic might not need peak capacity 24/7. Over-provisioning is common. My 2020 DeFi smart contract verification taught me that lockups don’t guarantee usage. The liquidity can be locked, but the swaps might not happen.
Patterns emerge only when chaos is organized. Let’s organize the chaos. The SemiAnalysis model shows that at $3 per GPU hour, a GW cluster costs $12B annually. That’s the electricity, cooling, maintenance, and labor. But what about the cost of capital? If SpaceX raises $500B at 5% interest, that’s $25B per year in interest alone. That eats into the $88B margin. The net margin drops to $63B per GW. Still high, but sensitive to interest rates and utilization. If utilization falls to 70%, revenue drops to $70B, net margin to $58B. The breakeven on a $50B GW is less than one year. That’s the kind of arithmetic that makes institutional investors salivate.
But there’s a second layer: the supply chain. NVIDIA’s GB300 clusters are not commodities. The lead time for a single GW of GB300s is likely 12-18 months. SpaceX would need to pre-order $400B+ worth of chips. That’s a bet on NVIDIA’s manufacturing capacity and continued dominance. SemiAnalysis’s model assumes no supply chain disruption. In my 2017 ICO audit, I saw projects assume infinite token demand. The market taught them otherwise. Here, the demand is real—AI companies are spending billions—but the hardware bottleneck is real.
Contrarian: The Bear Case That Demands Attention
I started with the bear case because the narrative hype is deafening. The contrarian angle here is not that SpaceX can’t do it, but that the numbers are too clean. The $100B per GW revenue assumes that API inference pricing remains at current levels. Competition from other compute providers (Google, Amazon, Microsoft, another startup) could compress margins. The $3 per GPU hour is a midpoint; hyperscalers often negotiate discounts for volume. If SpaceX’s effective pricing drops to $2 per hour, revenue drops to $66B, margin to $54B per GW. Still attractive, but the capex payback extends.
More importantly, the 10GW target requires a massive energy infrastructure. SpaceX’s Starbase facility in Texas might host some of this, but 10GW is equivalent to ten nuclear power plants. The grid interconnection timelines are years. Renewable energy at that scale requires land and storage. The SemiAnalysis report assumes SpaceX can solve this, but physical constraints are not solvable by code. Code is law, but intent is the evidence. The intent is there—Musk’s track record with Tesla and SpaceX shows he overcomes physical constraints—but the evidence of a 10GW build in 3 years is thin.
Another blind spot: the demand itself. If AI progress slows, or if open-source models reduce the need for massive inference clusters, the compute glut could crash prices. In 2022, I analyzed the liquidity drain from Celsius and Three Arrows Capital. The market had assumed perpetual growth. The data showed otherwise. The same pattern could repeat here. Compute capacity is a commodity; only the lowest-cost producer survives. SpaceX’s cost structure is supposedly low, but the capex debt load could make it fragile.
Due diligence is the armor against narrative hype. The SemiAnalysis report is thorough, but it relies on public statements and modeled projections. The actual construction timelines, regulatory approvals, and chip availability are unknown. The $250 billion Microsoft-OpenAI deal is signed, but the terms are not public. Is the 3GW SpaceX contract firm or an option? The difference is material. In my 2024 ETF institutional flow analysis, I tracked real money flows, not promises. Here, we have promises. The blockchain of financial records will eventually show the transactions. Until then, skepticism is warranted.
Takeaway: The Next Signal to Watch
By the end of 2027, we will know if SpaceX has delivered. But the leading indicators are earlier: chip orders, construction permits, energy contracts. If SpaceX secures 5GW of pre-paid contracts from Microsoft alone, that’s $150B in guaranteed revenue, reducing the risk. If they raise debt at favorable rates, the cost of capital drops. The key metric to monitor is the ratio of contracted revenue to total capex. A ratio above 0.5 suggests the project is bankable. Below 0.3, and it’s speculative.
For the crypto world, this compute buildout has implications. If hyperscale AI compute becomes a rent-seeking monopoly, the need for decentralized compute networks (like Render, Akash, or io.net) becomes more acute. But also, if SpaceX succeeds, the capital costs for new entrants will rise, centralizing power further. The data will tell the story. Until then, we organize the chaos, follow the numbers, and let the ledgers speak.