
Semianalysis: SpaceX to Build 10GW Compute, ARR to Reach $300 Billion, Microsoft to Be the Largest Customer
Author: SemiAnalysis
Compiled by: TechFlow
TechFlow Editor's Note: Musk announced at SpaceX's first earnings call that he plans to add 6–8GW of compute capacity in 2027, implying $300–500 billion in capital expenditure for that year, comparable to AWS and Google. This plan seems crazy, but from three dimensions—inference revenue, Microsoft demand, and SpaceX construction speed—it is entirely feasible—Microsoft is frantically expanding compute capacity to seize the API inference opportunity of $100 million annual revenue per MW, while SpaceX has proven twice that it is the fastest company in the world at building data centers.
Musk shocked the world again at SpaceX's first earnings call, announcing his gigawatt-level ambition: he "conservatively" plans to add 6–8GW of compute capacity in 2027 alone, and the actual number could exceed 10GW. Calculated at $50 billion investment per GW, this equates to $300–500 billion in capital expenditure in 2027, comparable to what we expect AWS and Google to invest—for a company with profitability far below its competitors, this number is incredible.
As we explained in our Meta Compute Deep Dive, large-scale + short delivery cycle compute is an extremely scarce combination, commanding a huge pricing premium—up to $50 billion per GW per year. But AI labs can afford this price and are thriving.
Our Tokenomics model and inference simulator show that at actual performance levels (e.g., tokens/second per GPU), OpenAI and Anthropic selling API inference services on GB300 clusters can both generate over $100 billion in revenue per GW per year. This is far higher than the cost of renting a GB300 cluster for a year at current new cloud provider prices.
For frontier model companies, the profits from providing inference token services are astonishingly high.

We assume approximately $12 billion cost per GW per year, adopt a conservative lease price of $3 per GPU hour, and use our inference simulator based on frontier model architectures and our agent programming benchmark AgentX (part of InferenceX, built on real production programming trajectories) for token output estimation. We calculate the final valuation exceeding $100 billion per GW per year by mixing this token output rate according to input, cache read, cache write, and output token costs as well as real workload proportions.
Our inference simulator is built from the ground up, based on a fundamental understanding of how modern AI accelerators work. We have built performance upper bounds and actual performance models for how frontier models work during inference, timing every operation and outputting real trajectories. This is an end-to-end simulation of executing real workloads on actual silicon. We have verified the simulator's accuracy on various accelerators and workloads and continue to improve its ability to predict future accelerator performance based on design specifications.


Besides OpenAI and Anthropic, there is actually a third company in the world that can achieve such economic benefits per GW: Microsoft. Due to having full access to OpenAI models, they can generate exactly the same revenue per MW and profit margins, while paying no training costs. Nadella's negotiations with OpenAI were very successful: the renegotiated agreement in April 2026 eliminated the old 20% revenue share. In short, Microsoft has huge incentive to procure as much MW compute capacity as possible as fast as possible. Although they currently lease most of their data center capacity to OpenAI at a price of $14 million per MW per year, they have an opportunity to improve this mix. The potential impact is accelerating Microsoft Azure revenue growth rate from about 42% to over 100% next year. This is a once-in-a-lifetime opportunity, and SpaceX is in an excellent position to serve this demand.

Although Microsoft signing 3GW with SpaceX at $50 billion per GW per year sounds crazy, we believe this is possible for two reasons:
- Microsoft is already preparing for large-scale data center expansion. As described below, they have signed 10GW in contracts year-to-date, with total contract value exceeding $300 billion (excluding GPU costs). We expect more to be signed. It should be noted: these contracts contribute capacity in late 2027 and 2028, leaving a gap to be filled in the near term.
- There is a 90-day cancellation policy, similar to SpaceX's agreements with Anthropic and Google, zero risk to the balance sheet. Considering the revenue opportunity, Amy Hood would easily approve this plan.
For SpaceX, the next natural question is financing. Without the balance sheet support of top hyperscalers, how will Elon pay for such high capital expenditure? We expect a combination of the following two methods:
- Support from Nvidia, in the form of vendor financing to reduce upfront cash costs. This might be why Elon announced exclusive use of Nvidia at the earnings call! As our accelerator models have repeatedly explained, xAI/SpaceX actively evaluated alternative solutions like TPU and AMD—so financial considerations may have prompted them to abandon these options and focus on Nvidia.
- Operating cash flow financing supported by the industry's highest pricing, relying on the fastest delivery cycle: SpaceX will continue to sell large-scale compute with a 3–5 month delivery cycle—this is unmatched supply—and price it accordingly at $30–50 million per MW per year. This recoups capital expenditure in less than a year. We explored this in depth in our Meta Compute article.
This means SpaceX is expected to achieve $300 billion ARR by the end of 2027. This only assumes 50% of their new compute capacity added in 2027 is commercialized, with the rest left for the Grok and Cursor teams for training (unmodeled inference revenue).

Now let's dive deep. We start with Microsoft, which has finally awakened amazingly: last year's pause has reversed, and they have signed 10GW in binding contracts year-to-date. We will briefly discuss the economic model of how to achieve $100 million inference revenue per MW per year. Then we turn to SpaceX, analyzing their data center expansion and the feasibility of achieving 10GW+ by the end of 2027.
Microsoft's 10GW Awakening, Seizing the $100 Million Annual Revenue per MW Opportunity
In December 2024, we were the first to point out the significant pause in Microsoft's leasing activity in our data center model. Today, the giant has awakened. Our model tracks leasing activity, new cloud provider signings, self-built construction starts, and large binding PPAs and ESAs quarter by quarter. The chart below shows the output. Microsoft has signed over 10GW across all these areas, equivalent to about $300 billion in new binding commitments.

A key reason for this awakening is their urgent need for compute capacity to seize the $100 million annual revenue per MW opportunity. Microsoft signed a $250 billion agreement with OpenAI in October 2025, which we estimate totals about 7GW in our Tokenomics model—this is the world's best tool for understanding the nuances of the dollar-to-watt mathematics. This massive Infrastructure-as-a-Service deal left Microsoft highly constrained on compute capacity for other use cases. They have been unable to leverage their OpenAI model access to launch the API business Foundry, or for applications like Copilot.
However, the profit margins and revenue per MW for these services are the highest to date. We explained this in depth in our AI Value Capture article.
AI Value Capture—The Shift to Model Labs

A day in the AI field feels like a year in other industries. Model releases, software breakthroughs, and hardware improvements are compressing cycles that take years in other industries into weeks. Just in the past few months, Agentic AI has crossed a real inflection point, driving step-change growth in token value while software and hardware improvements significantly reduce token generation costs.
To arrive at these margin estimates, we must carefully synthesize leaked financial data, InferenceX data, micro-benchmarks of all the latest accelerators in the industry, papers from open-source labs, blogs, tweets, and more. New data points like the leaked DeepSeek investor call (they said the GPU payback period is 10 months) confirm our estimates are in the right range, but we first admit that the lack of granularity is indeed unsatisfactory. Rather than the single number of overall company inference gross margin, what you really want to know is the gross margin for each (model, accelerator) combination across the entire throughput vs. latency Pareto frontier. For example, what are the gross margins for serving Opus 5 Fast on Trainium3 versus serving Fable 5 on TPUv7?
We answer this question with our inference simulator, a tool available only to SemiAnalysis consulting clients.
Our AI Cloud TCO model has already answered the cost side, but the revenue side has historically been unknown. To solve this, we wrote a simulation framework that simulates real model execution on virtual hardware, supported by tuned fine-grained performance models covering various accelerators and operation types. We run each model on simulated XPUs with all possible service configurations, mixing real-world and idealized service conditions. This allows us to accurately estimate the performance of any combination of software, hardware, and workloads assuming a thorough understanding of the model architecture.
Thanks to this simulator, our Tokenomics model now includes high-level revenue per MW figures for running OpenAI/Anthropic flagship models on all relevant chips. Workload shape is obviously a huge factor; we simulate running over $1 million worth of agent trajectories collected from our own usage, while meeting real interactivity and TTFT levels observed from accessing first-party endpoints. As a preview, here are our numbers for serving Fable 5 on GB200 vs GB300:


This is Microsoft's $100 million per MW opportunity. Given the recent surge in Codex demand and corresponding OAI ARR acceleration, we believe Microsoft can commercialize compute capacity at a similar rate by providing OAI model services.
Now, to seize this once-in-a-lifetime opportunity, Microsoft and AI labs need data centers, and they need them fast and large. SpaceX has proven twice that their construction speed is faster than others, but by the end of 2026 their compute capacity is "only" 2GW. Can they really build 10GW+ within the single year of 2027?
SpaceX: Building Data Centers at Astonishing Speed
In our Meta Compute article, we explained in depth why Elon has once again proven himself a business genius. He understands that AI lab profit margins have increased significantly, and accordingly introduced "value-based pricing" for his GPU clusters, rather than the more common "cost-plus".
To keep the machines running, Elon needs to build data centers faster than anyone. We believe he can do it. What gives us this confidence? We have written several articles about Elon's speed, including building 300MW of Colossus 1 in 122 days, 200MW of Colossus 2 in six months, deciding to build an on-site power plant 1 kilometer across the border to avoid permitting approvals, and so on.
There have been more speed demonstrations since. The power plant in Southaven has expanded from 27 turbines (about 495MW) in February 2026 to 69 turbines (1.7GW) in July 2026.


There is also the emergence of "MiniHard", which after starting vertical construction in March 2026, could reach 450–500MW in about 5 months! This does not mean Elon builds better than others. He just adopts a different approach.
How to Achieve All This?
Distribution equipment and high-power transformers sold out two years ago? Buy power modules directly from China, skip high-power transformers, and deliver medium-voltage power from the generation end directly to low-voltage transformers—the latter are much more plentiful. Companies under Elon have the best electrical engineers in the world—no one knows better than them how to trade off speed versus efficiency. Most data center operators globally prioritize efficiency and quality—this is the only way to secure 15–20 year committed hyperscaler contracts. SpaceX will focus on a completely different trade-off: speed above all. In a period of compute scarcity and extremely high AI token profit margins, one that can be delivered within three months with a 90-day cancellation period 500MW cluster is one of the scarcest and most valuable assets in the world. All conventional "quality" metrics do not matter. Evidence? Google, the most vertically integrated infrastructure company globally, ultimately chose to sign with SpaceX.
Gas turbine orders booked out 5 years? That is true for GE Vernova, but there are many other options—our energy model shows that over 30 gas power equipment manufacturers have received large orders to serve data centers. As long as you search hard enough and are willing to work with new suppliers, there is ample available capacity. Turbine secondary market trading volume is surging—for example, all turbines originally planned for delivery to Oracle's New Mexico site are now circulating in the market. Secondary market prices are high, but Elon can afford it.
Labor is the ultimate constraint? Then parallelize operations as much as possible, simplify commissioning processes, and pre-assemble as much as possible. Reportedly, Colossus 2 used about 3,000 construction workers daily at peak, about half less than other gigawatt-scale data centers under construction. Elon has long accomplished feats with labor below industry standards—just look at the history of Tesla and SpaceX.
This leaves enough time to build multiple such shells before 2027. Moreover, this is his first true greenfield project; next time he will likely do even better. Of course, another option is retrofitting existing facilities. As we stated in our deep dive on xAI last year, Colossus 1 and 2 were both built through retrofitting at astonishing speed.
xAI's Colossus 2 — The World's First Gigawatt-Scale Data Center, Unique Reinforcement Learning Approach, Financing Progress

There has been extensive reporting on xAI's Colossus 1. The construction in Memphis is destined for the history books: the world's largest AI training cluster, built from scratch in just 122 days. With about 200,000 H100/H200 and about 30,000 GB200 NVL72, it remains the largest fully operational, single coherent cluster to date (except for Google).
Building 10+GW within a year is a completely different story. SpaceX needs to find suitable land across the country that is easy to permit and has access to natural gas. But we believe there are enough options to support large-scale expansion. This will naturally rely heavily on on-site gas power generation—check out our energy deep dive to understand how it works and its necessity.
Behind the paywall, we will discuss some sites we speculate Elon might choose.













