SemiAnalysis projects that SpaceX could approach 10 gigawatts of AI computing capacity by the end of 2027, a scale that would place power delivery and construction speed ahead of GPU counts as the central constraints in the race to serve advanced AI models. Its Aug. 7 report argues that the most valuable measure of an AI cluster is increasingly the revenue it can produce per megawatt, particularly from inference workloads where customers pay for model output rather than raw hardware access.
Under the report’s assumptions for frontier-model API inference on NVIDIA GB300 systems, OpenAI and Anthropic could generate more than $100 billion in annual revenue from each gigawatt of deployed inference capacity. SemiAnalysis estimates that renting an equivalent 1GW GB300 cluster would cost roughly $12 billion per year, leaving a wide modeled gap between infrastructure costs and revenue for companies able to sell high-value AI output at scale.
The calculations are scenarios rather than company forecasts, but they illustrate why operators are competing for ready-to-use power capacity. A data center connected to power but lacking customers has limited value; a site supplying low-latency inference to a major AI developer can convert each megawatt into recurring revenue far above conventional infrastructure leasing rates.
Revenue per megawatt replaces the GPU-count race
SemiAnalysis shifts the comparison away from the familiar metrics of chip counts and floating-point operations per second, or FLOPS. It instead models tokens per second, tokens per watt, tokens per megawatt and revenue per megawatt. Tokens are the units of text and other data processed by generative AI systems, making them a closer proxy for the billable output that API providers sell.
The analysis uses an “AgentX Trace” designed around agentic coding tasks, which can involve extended context windows, repeated tool calls and large volumes of generated output. Such workloads place different demands on AI systems than simpler chatbot prompts. Serving configuration, time-to-first-token, input tokens, cache reads and writes, and output-token throughput can all alter how much usable capacity a cluster produces.
In one modeled “Fable 5” scenario, SemiAnalysis estimates NVIDIA’s GB200 NVL72 system could support potential annual revenue of about $73.4 million per megawatt. The newer GB300 NVL72 configuration reaches about $99.7 million per megawatt in the same scenario, an increase of roughly 36%.
Those figures explain the industry’s willingness to spend heavily for newer systems even when their acquisition costs are high. The economic advantage comes from a combination of greater throughput and the ability to serve more valuable workloads within the same power envelope. A marginal improvement in tokens per watt can become substantial when multiplied across hundreds of megawatts.
Fast construction becomes a commercial advantage
SemiAnalysis tracks several projects that it says demonstrate how operators are trying to bypass traditional data-center development timelines. The report describes deployments labeled Colossus 1 and Colossus 2 as reaching about 300MW in 122 days and about 200MW in roughly six months, respectively.
It also follows the Southaven site’s power equipment expansion from 27 turbines generating about 495MW in February 2026 to 69 turbines with about 1.7GW by July 2026. A separate project called MiniHard entered vertical construction in March and is expected to ramp to roughly 450MW to 500MW within around five months, according to SemiAnalysis.
The report identifies three approaches to shortening the wait for computing capacity: retrofitting existing facilities, developing greenfield sites and deploying on-site generation. Each method targets the same bottleneck: conventional grid interconnection can take years, while demand for inference capacity is arriving much faster.
SemiAnalysis estimates that scarce near-term, power-ready capacity could command about $30 million to $50 million per megawatt annually in some circumstances. That pricing would apply to operators supplying infrastructure quickly, rather than necessarily building a full AI API business themselves.
The report places the revenue difference between those models in stark terms. Pure infrastructure provision may yield about $14 million per MW annually in its model, while high-value inference can approach $100 million per MW. The gap gives cloud providers and AI model developers a strong incentive to secure capacity directly instead of depending entirely on third-party leasing arrangements.
Microsoft is modeled as a major buyer of capacity
SemiAnalysis models SpaceX at roughly 2GW by the end of 2026 before an acceleration toward nearly 10GW by the end of 2027. It identifies Microsoft as a likely major offtaker, pointing to a 2025 agreement in which OpenAI committed to purchase an additional $250 billion in Azure services.
The report also cites April disclosures stating that Microsoft no longer receives revenue-share payments from OpenAI, while OpenAI’s payments to Microsoft continue through 2030, subject to a total cap. Microsoft’s license to OpenAI’s model and product intellectual property extends through 2032, according to the disclosures referenced by SemiAnalysis.
Using its data-center model, SemiAnalysis estimates that Microsoft has secured more than 10GW of new capacity in 2026, linked to more than $300 billion in binding commitments. Those totals are analytical estimates, not figures disclosed by Microsoft.
In one scenario, Microsoft obtains 3GW from SpaceX and sells inference capacity at close to $100 million per MW annually. That would imply an exit annual recurring revenue run rate of about $300 billion. Microsoft reported in its FY2026 fourth-quarter disclosure, as cited by SemiAnalysis, that Azure and other cloud-services revenue rose 43% year over year and that Azure annual revenue surpassed $100 billion for the first time.
Financing follows the power buildout
The capital requirements for multi-gigawatt AI campuses are likely to bring infrastructure finance further into the technology supply chain. SemiAnalysis discusses vendor financing as a potential way to fund large deployments, while noting no public confirmation of financing arrangements tied to SpaceX.
NVIDIA announced on Aug. 10 that it had formed AI Compute Infrastructure Financing Platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. NVIDIA said the initiative targets more than $500 billion in long-term third-party capital for dedicated financing pools supporting its customers.
For cryptocurrency markets, the report offers a more grounded lens for evaluating decentralized compute narratives. The growth in AI power demand may create opportunities for networks that coordinate unused hardware, but it does not automatically validate every token associated with distributed computing. The highest-value workloads in SemiAnalysis’ model depend on dense clusters, advanced networking, reliable power, rapid deployment and predictable service quality—requirements that remain difficult for fragmented hardware marketplaces to match.
The pressure point is therefore less about a generic shortage of GPUs than the ability to turn available electricity into reliable, billable AI output. As large operators compete for multi-gigawatt sites, projects claiming to offer compute capacity will face closer scrutiny over whether they control hardware, power access, networking and customer demand rather than simply tokenizing a marketplace for unused machines.
Want to see how AI and blockchain converge in real markets? Explore how AI complements blockchain in real-world trading infrastructure.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
