SpaceX committed to deploying 10 GW of AI infrastructure power capacity by 2027 with Microsoft as the largest offtaker.

Elon Musk shocked the world—again—when he announced SpaceX's gigawatt ambitions during the company's first earnings call. He "conservatively" aims to build and deliver an incremental 6–8 GW in 2027 alone, with potential for that number to exceed 10 GW. At $50B per GW, that translates to $300–500B in capex in 2027, matching expectations for AWS and Google—an extraordinary figure for a company significantly less profitable than rival hyperscalers.

Yet we believe the number is real. We see SpaceX on track to build approximately 10 GW by year-end 2027. We've evaluated all suitable sites for SpaceX and provided the list to our Datacenter Model subscribers. Our Energy Model subscribers also have access to the precise inventory of gas generation equipment available quarter by quarter from 30+ turbine, engine, and fuel cell suppliers—data we provided before the broader market caught on.

SpaceX will develop whatever capacity they can and bring it online as rapidly as possible. As we explained in our Meta Compute deep dive, large-scale capacity available on near-term timelines is remarkably scarce and commands a premium—up to $50B per GW per year. AI labs can easily absorb and profit from it.

Our Tokenomics Model and Inference Simulator demonstrate that at realistic performance levels, both OpenAI and Anthropic can generate over $100B per GW per year in revenue from API inference on a GB300 cluster—significantly more than the cost of renting a GB300 cluster for a year at current neocloud prices. Serving inference tokens is exceptionally profitable for frontier model companies.

We assume approximately $12B per GW per year in costs, using a conservative rental rate of $3 per GPU-hour. Our Inference Simulator estimates token production using a frontier-class model architecture and our AgentX benchmark—built from real production coding traces—blending input, cache-read, cache-write, and output token costs at actual workload ratios. The result exceeds $100B per GW per year.

Our Inference Simulator is purpose-built with deep understanding of how modern AI accelerators function. It constructs a roofline and realistic performance model for frontier inference, complete with timings for every operation and real trace output. This is an end-to-end simulation of actual workload execution on actual silicon. We have validated this simulator's fidelity across a wide range of accelerators and workloads.

Beyond OpenAI and Anthropic, there is one other company capable of achieving this economics per GW: Microsoft. With full access to OpenAI's models, they can generate identical revenue and margins per MW while avoiding training costs entirely. Satya Nadella negotiated masterfully with OpenAI: the deal restructured in April 2026 eliminated the old 20% revenue share. Put simply, Microsoft has enormous incentive to procure as many MWs as possible, as quickly as possible. While much of their current datacenter capacity goes to OpenAI at roughly $14M per MW per year, they have the opportunity to shift this mix materially. The potential impact is Azure accelerating revenue growth from approximately 42% to over 100% by next year—a once-in-a-generation opportunity that SpaceX is exceptionally well positioned to serve.

Microsoft signing a 3 GW contract with SpaceX for $50B per GW might sound extreme, but we view it as realistic for two reasons. First, Microsoft is preparing for an epic datacenter ramp: they've already signed 10 GW of contracts year-to-date, representing over $300B in total contract value, with more expected. These contracts support late 2027 and 2028 capacity, however, leaving a near-term gap. Second, with 90-day cancellation policies—similar to SpaceX's deals with Anthropic and Google—there is zero balance sheet risk. This makes the economics easy for Amy Hood to approve.

For SpaceX, the financing question looms: how can Elon fund such massive capex without the balance sheet of leading hyperscalers? We expect a combination of two factors. First, vendor financing from Nvidia to reduce upfront cash costs—likely why Elon declared Nvidia exclusivity on the earnings call. xAI and SpaceX have actively evaluated alternatives like TPU and AMD; the financial incentive presumably tilted them toward Nvidia focus. Second, industry-leading pricing enabled by fastest timelines: SpaceX will continue selling large-scale compute with 3–5 month lead times—an unbeatable offering—priced at $30–50M per MW per year. That payback exceeds one year.

This trajectory suggests a path to $300B in annual recurring revenue by year-end 2027 for SpaceX, assuming only 50% of their incremental 2027 compute is monetized, with the remainder allocated to the Grok and Cursor teams for training.

Microsoft has finally awakened from last year's pause. They've signed 10 GW of binding contracts year-to-date across leasing, neocloud contracting, self-build construction, and large-scale PPAs and ESAs—equivalent to approximately $300B in new binding commitments. A key driver is their desperation for compute to capture a $100M per MW per year revenue opportunity. Microsoft signed a $250B agreement with OpenAI in October 2025, representing approximately 7 GW in total capacity. This massive Infrastructure-as-a-Service commitment has left Microsoft compute-constrained for other use cases, unable to leverage their OpenAI access for their highest-margin services: API Foundry and applications like Copilot.

The pace of AI advancement has accelerated dramatically. Model releases, software breakthroughs, and hardware improvements are compressing multi-year industry cycles into weeks. Over recent months, agentic AI has crossed a genuine inflection point, driving a step change in token value while improvements in software and hardware have sharply reduced generation costs.

Over the past month, sophisticated investors have reached consensus: serving frontier tokens at API prices yields exceptionally high margins. We were first to call this out to our Tokenomics Model subscribers in January, explaining why inference gross margins exceed 60%. In June, we followed with a detailed analysis showing Opus 4.8 achieving 85%+ margins—a figure now cited as default in Anthropic analyses. Arriving at these margin estimates required careful synthesis of leaked financials, InferenceX data, microbenchmarks across all latest accelerators, papers, blogs, and tweets from open source labs. Leaked datapoints such as the DeepSeek investor call—citing a 10-month GPU payback period—confirm we're in the right ballpark, though the lack of granularity remains unsatisfying. Rather than single company-wide inference gross margin numbers, the real value is understanding the specific margin profile by workload and capacity type.