SpaceX aims for 10 GW of AI computing by 2027, potentially generating $300 billion in annual recurring revenue

Securing 1 GW a year in advance equals $100 billion in revenue—this "war of time" is one where Musk possesses the hardest-to-replicate advantage.

Original report: SemiAnalysis, "SpaceX 10GW in 2027 – Why It’s Real, Will Drive $300B ARR for SpaceX, and Why Microsoft Will Be the Largest Offtaker," August 7, 2026

Compiled and organized by DaiDai, MSX MaiTong

Edited by: Frank, MSX MaTong

With the same 1 GW of AI computing power, one business model might generate only billions of dollars in revenue per year, while another could exceed $100 billion—what would happen?

The answer could be changing the game for the entire AI infrastructure industry.

Over the past two years, the market has measured the AI arms race by the number of GPUs and the scale of training clusters—those with more H100s, GB200s, or larger training clusters were seen as possessing stronger AI infrastructure.

However, SemiAnalysis presents a more aggressive framework in its latest report, stating that the true focus of the future will shift from how many GPUs one owns to how many sellable tokens each megawatt of electricity can ultimately produce, and the revenue this generates.

According to its Tokenomics Model and Inference Simulator, under specific assumptions regarding cutting-edge models, GB300 clusters, real-world Agentic Coding workloads, and API pricing, OpenAI and Anthropic could each generate potential annual revenues exceeding $100 billion per gigawatt of inference computing power. In comparison, the report estimates the annual cost of leasing an equivalent GB300 cluster at approximately $12 billion.

This set of numbers is extremely aggressive and highly dependent on model requirements, token prices, utilization rates, latency demands, and hardware and software efficiency, but it explains an increasingly real issue: why all the major players have suddenly begun competing so fiercely for electricity.

• In February of this year, SpaceX officially acquired xAI, bringing Grok, Colossus, and SpaceX’s infrastructure capabilities under one company;
• Six months later, Musk stunned attendees at SpaceX’s first earnings call, stating that 6–8 GW of new AI computing capacity would be built and delivered by 2027, with potential upside reaching up to 10 GW;
• SemiAnalysis also estimates that by the end of 2027, SpaceX's total computing capacity could approach 10 GW.

If this prediction holds, SpaceX’s next steps may go beyond building rockets, launching satellites, and selling Starlink—it aims to become one of the world’s largest providers of AI computing power.

This also means that as computing power, acquired months or even a year in advance, begins to generate significant economic value, the scarcity of AI infrastructure is shifting from a competition over chips to an industrial war over electricity, engineering, and time.

And compressing complex engineering to its limits is precisely the game Musk knows best.

The most important aspect of the SemiAnalysis report is that it recalculated the revenue per megawatt.

The same GB300 cluster, if rented out solely in the traditional NeoCloud mode, still derives its economic value primarily from equipment cost, depreciation, and rental pricing.

But if these GPUs ultimately host state-of-the-art models and directly sell inference tokens through products like APIs, Coding Agents, and Copilots, the revenue per MW could be an entirely different order of magnitude.

SemiAnalysis estimates that when running cutting-edge model APIs on GB300 clusters, OpenAI and Anthropic could generate over $100 million per MW per year, or more than $100 billion per GW per year.

This estimation does not simply extrapolate based on theoretical GPU FLOPS. SemiAnalysis incorporates variables such as model architecture, GPU memory bandwidth, serving configuration, throughput, TTFT, input tokens, cache reads, cache writes, and output tokens, while utilizing AgentX traces from real-world agentic coding workloads.

This means that the evaluation system for AI infrastructure is undergoing a significant shift.

Initially, the market competed based on the number of GPUs; later, it shifted to FLOPS and training cluster scale. After entering the era of large-scale inference, key metrics may gradually become tokens/sec, tokens/W, tokens/$, and tokens/MW, ultimately culminating in Revenue/MW.

This is why the significance of GB300 goes beyond simply being "faster than GB200."

According to the report model, in the same Fable 5 inference scenario, the GB200 NVL72 generates an estimated annual revenue of approximately $73.4 million per MW, while the GB300 NVL72 increases this to approximately $99.7 million, demonstrating that the true value of the new-generation hardware lies in producing more high-value tokens within the same power budget.

NVIDIA is also using similar language to redefine AI infrastructure.

The DSX platform, released in May this year, has adopted "token performance per megawatt" as one of the core metrics for AI Factory, aiming to extend from chips, networks, and software to power, cooling, and data center operations.

This change becomes even more apparent when reasoning demands expand beyond general chat to include agentic coding, AI workers, multi-agent collaboration, and continuous running tasks.

Training typically follows distinct stages and cycles, whereas inference demand more closely resembles «number of users × number of agents × tokens consumed per task × runtime». Once agents transition from «answering questions» to «working continuously», the upper limit of token consumption is effectively reopened.

From this perspective, the most important variable to watch in the next phase of the AI computing race may no longer be how large training clusters can become, but rather how many tokens—capable of being paid for by users—these computing resources can ultimately generate.

The question arises: if frontier inference truly has such high revenue per megawatt, why don’t OpenAI, Microsoft, and Google build all their own data centers themselves?

Because GPUs can be purchased with money, but electricity and time are not as easily bought.

Large AI data centers, from acquiring land and securing grid connection to constructing substations and transmission facilities, delivering equipment, and finally going live, are becoming increasingly lengthy infrastructure projects.

What SpaceX and xAI have truly demonstrated over the past two years is not the ability to build GPUs no one else has, but rather their capacity to compress these processes to an extreme degree—according to SemiAnalysis, for example, Colossus 1’s approximately 300 MW of computing power was built in just 122 days, while Colossus 2’s roughly 200 MW took about six months.

Even more striking is the power consumption. According to SemiAnalysis’s tracking of the Southaven project, on-site power generation equipment expanded from 27 turbines, approximately 495 MW, in February 2026, to 69 turbines, roughly 1.7 GW, by July. Another MiniHard project, closer to the Greenfield model, reported after entering vertical construction in March that it is expected to reach 450–500 MW in about five months.

This is also what truly matters about SpaceX's current expansion.

Colossus heavily utilized Retrofit in its early stages to quickly convert old industrial buildings into AI data centers; MiniHard has begun validating the Greenfield new-build model; combined with On-site Generation, SpaceX is attempting to simultaneously pursue all three approaches.

Especially for on-site natural gas power generation. If waiting entirely for the traditional utility grid to complete transmission, interconnection, and utility upgrades, GW-scale projects would struggle to come online at the speed Musk desires.

Therefore, SemiAnalysis concludes that to achieve several gigawatts of new capacity by 2027, SpaceX must heavily rely on on-site natural gas power generation to reduce its dependence on the pace of traditional grid connections.

A new business model has emerged here: SpaceX is not just selling GPUs, but providing computing power that others may have to wait one to two years for—power that SpaceX could deliver in just a few months.

SemiAnalysis estimates that this scarcity of "large-scale + short-term availability" could allow SpaceX to charge $30 million to $50 million per MW annually for certain computing power, highlighting the key distinction between value-based pricing and traditional cost-plus pricing: customers are not paying for what a server should be worth, but for the commercial value generated by gaining early access to computing power.

Thus, Time-to-Power has become as important a competitive metric as GPU performance.

SemiAnalysis estimates that SpaceX's computing power will reach approximately 2 GW by the end of 2026, then accelerate significantly in 2027, nearing 10 GW by year-end.

Overall, what will truly be worth watching in the market by 2027 is whether SpaceX can replicate the proven "speed" into an industrial-scale capability.

If SpaceX can truly deliver gigawatts of computing power quickly, the next question is: who will buy it?

SemiAnalysis's answer is very clear: Microsoft, which is also the most important investment implication for publicly traded companies in the entire report:

• In the new agreement signed between Microsoft and OpenAI in 2025, OpenAI additionally committed to purchasing $250 billion in Azure services;
• After entering 2026, both parties further adjusted their partnership. According to the new agreement disclosed by Microsoft in April, Microsoft will no longer pay OpenAI a revenue share, but OpenAI’s revenue share payments to Microsoft will continue at the original rate until 2030, subject to a total cap;
• Meanwhile, Microsoft has extended its license for OpenAI’s models and product IP until 2032;

This presents Microsoft with an interesting question.

It can provide data center capacity to OpenAI as infrastructure, and also leverage its access to OpenAI models to integrate computing power into Azure Foundry, Copilot, API, and Agent products.

SemiAnalysis estimates that the revenue for the first model is approximately $14 million per MW per year, while the latter high-value inference scenario could theoretically approach $100 million per MW per year.

If true, the nearly sevenfold difference in revenue density between the two explains why SemiAnalysis believes Microsoft will aggressively compete for computing power again.

By tracking projects such as leasing, self-built, NeoCloud contracts, PPA, and ESA according to its Datacenter Model, Microsoft has secured over 10 GW of new capacity since 2026, corresponding to more than $300 billion in binding commitments (note that this figure is derived from SemiAnalysis’s model estimates and is not officially disclosed by Microsoft).

More critically, much of this capacity won't be fully operational until the end of 2027 or even 2028.

Ultimately, what Microsoft lacks is not long-term planning, but large-scale computing power that can be deployed right now.

Therefore, SemiAnalysis conducted an extremely aggressive scenario analysis: if Microsoft were to acquire 3 GW of computing power from SpaceX and monetize it at an efficiency of nearly $100 million per MW per year, it could theoretically correspond to an exit ARR of approximately $300 billion, propelling Azure’s growth rate into triple digits.

It is important to emphasize that this is not a Microsoft guideline or a signed SpaceX contract, but rather a scenario model used by SemiAnalysis to illustrate the revenue elasticity of additional computing power. Microsoft’s latest disclosed data for FY2026 Q4 shows that Azure and other cloud services revenue has grown by 43% year-over-year, with Azure’s annual revenue surpassing $100 billion for the first time.

However, the questions raised by this model still warrant the market to reconsider: when evaluating Microsoft’s AI capital expenditures, future assessments may no longer focus solely on “how much was spent,” but also on how much revenue each megawatt of capacity generates once online.

NVIDIA represents another shift.

SemiAnalysis suggests that NVIDIA might assist SpaceX in reducing upfront cash pressure for its massive capital expenditures through vendor financing. There is currently no public evidence that NVIDIA has implemented such financing for SpaceX, so it is more appropriate to view this as a speculative scenario outlined in the report.

However, on August 10, following the report’s release, NVIDIA announced the establishment of AI Compute Infrastructure Financing Platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, aiming to mobilize over $500 billion in third-party capital over the long term and create dedicated financing pools for NVIDIA’s customers.

This at least demonstrates a broader trend: NVIDIA is extending its capabilities from GPUs, CUDA, networking, and rack-scale systems toward the design, construction, and even the capital layer of AI factories.

For AI data centers requiring tens of billions of dollars in capital expenditure, the decision on which technology path customers adopt may no longer be determined solely by “which GPU is faster.” Those who can simultaneously provide chips, networking, systems, software, construction solutions, along with cheaper and more abundant capital, will be better positioned to secure the next wave of AI infrastructure development.

Thus, the AI computing power war is becoming increasingly like an industrial war.

Competition spreads outward from GPUs—beyond chips lie networks and storage; beyond networks lie cooling and power; beyond power lie natural gas, turbines, land, and capital. Ultimately, everything points to one metric: who can convert one megawatt of electricity into tokens that generate continuous revenue at the lowest time cost.

SemiAnalysis's forecast for SpaceX is undoubtedly very aggressive.

10 GW, a $300 billion ARR, Microsoft’s 3 GW offtake, and over $100 million per MW per year in inference revenue are all based on a series of high-growth assumptions.

Any one of these factors—such as agent demand, API pricing, GPU utilization, construction speed, power supply, or even financing terms—若低于 expectations, could lead to significant deviations from the model's outcomes.

Therefore, what truly matters in this report is not whether SpaceX will reach 8GW, 10GW, or some other number by the end of 2027, but rather that it reveals a paradigm shift in the ongoing competition for AI infrastructure:

In the past, the market competed for GPUs, then for electricity; but once all the major players are willing to build power plants, sign PPAs, and purchase natural gas and turbines, the hardest resource to acquire will ultimately become "time."

Securing 1 GW of power a year earlier versus receiving the same 1 GW a year later might have merely been a difference in construction timeline in the traditional data center era, but in a world where 1 GW of cutting-edge inference computing power could theoretically generate annual revenues in the hundreds of billions of dollars, it represents entirely different economic value.

This also gives new meaning to the engineering capabilities Musk has repeatedly demonstrated in the past.

If this logic holds, then SpaceX's "10 GW bet" was never really about the 10 GW itself—it was about turning engineering speed into commercial pricing power.

And in this war over time, Musk may just possess the hardest-to-replicate advantage.