AI Inference’s $100 Billion Jackpot May Be Mostly Modeled

A new SemiAnalysis report puts extraordinary numbers around AI inference, including more than $100 billion in annual revenue for each gigawatt of frontier compute. The estimate is worth examining closely because its biggest figures depend on several assumptions that have yet to be proven at scale.

Microsoft, OpenAI, Anthropic and SpaceX are all racing to secure more data center capacity. The reason is easy to understand: powerful AI models need enormous amounts of computing power after training, when users send prompts and expect answers in real time. That recurring workload could become a major business.

SemiAnalysis argues that a single gigawatt of advanced AI infrastructure could support more than $100 billion in annual inference revenue. The report also says Microsoft could capture economics similar to OpenAI and Anthropic because it controls Azure capacity and has long-term access to OpenAI technology. SpaceX, meanwhile, is presented as a possible new force in AI cloud infrastructure, with a path toward roughly $300 billion in annual recurring revenue by the end of 2027.

Those are eye-catching numbers. They are also modeled outcomes rather than established operating results. The gap between a profitable inference cluster on a spreadsheet and a profitable inference business in the market can be wide.

SemiAnalysis used its tokenomics and inference models to estimate what frontier AI companies could earn from selling API access on a GB300 cluster. Under the report’s assumptions, OpenAI and Anthropic could each generate more than $100 billion per year for every gigawatt of deployed compute.

The same analysis estimates about $12 billion in annual cost for renting a gigawatt of GB300 capacity, using a rental rate of roughly $3 per GPU-hour. That produces a very large spread between revenue and infrastructure expense.

The calculation depends on several inputs:

A change in any one of those inputs can move the result substantially. A cluster that is technically capable of processing a huge number of tokens still needs enough paying demand to use that capacity. It also needs customers willing to keep paying current prices after cheaper models enter the market.

The revenue estimate should therefore be read as a capacity economics scenario. It describes what a highly utilized, well-priced cluster might produce. It does not establish that every gigawatt will generate $100 billion, or that the resulting revenue will become equivalent profit.

SemiAnalysis’s claim that frontier AI inference can generate more than $100 billion in annual revenue per gigawatt illustrates why cloud providers and model companies are racing to secure power, chips and data centers. But it is a high-end capacity model, not evidence of realized profit: the outcome depends on utilization, token prices, demand, operating costs and buildout timing. Microsoft has a credible advantage through Azure and its revised OpenAI partnership, while SpaceX has shown unusual construction speed. Both opportunities ultimately require the same proof—available capacity must turn into paying usage, sustainable margins and cash flow.

The report’s cost comparison focuses on the price of renting a cluster. That is useful, but it leaves out several costs that sit between infrastructure access and corporate earnings.

AI companies still need engineering teams, model researchers, safety staff, product support, sales operations and customer service. They also need to pay for data preparation, software, networking, electricity and facilities. Large clusters require maintenance and replacement spending as new accelerator generations arrive. Financing costs can become significant when companies commit billions of dollars before customers have fully adopted the service.

Microsoft’s own results show why the distinction matters. Research cited in the background material says Microsoft Cloud gross margin fell to 66% as the company invested heavily in AI infrastructure and served growing usage. OpenAI-related revenue reported by Microsoft also combines multiple pieces of the partnership, including cloud activity, revenue-sharing payments and investment-related economics. It cannot be treated as a clean measure of pure, high-margin Azure sales.

That pressure can persist even when demand is strong. Data center construction, server deployment and depreciation happen before the full revenue stream arrives. A customer contract can secure future demand while leaving the cloud provider with a long period of capital spending and operating costs.

A more useful investor question is the margin after all-in costs. That means looking beyond the price of an API call and asking how much remains after hardware depreciation, energy, labor, financing and idle capacity. The answer could still be attractive. The available evidence does not support treating the full $100 billion estimate as cash profit.

Microsoft has a strong position in the race because it sits on both sides of the AI infrastructure market. Azure supplies the computing capacity, while Microsoft’s partnership with OpenAI gives it access to leading models and a large commercial distribution channel.

The April 2026 agreement changed the economics of that relationship. Microsoft remains OpenAI’s primary cloud partner and retains a license to OpenAI intellectual property through 2032. OpenAI products are expected to ship first on Azure when Azure can support the required capabilities. OpenAI can also distribute its products through other cloud providers under the revised terms.

Microsoft no longer pays OpenAI a share of Microsoft’s revenue for offering OpenAI models through Azure. OpenAI continues making payments to Microsoft, with Reuters reporting a 20% share through 2030 subject to a total cap. OpenAI also retained a minimum commitment to purchase $250 billion of Azure services through 2032.

That structure gives Microsoft a strong reason to build capacity quickly. It also reduces one direct cost in the Azure model distribution arrangement. Yet the opportunity still depends on how much of that capacity gets used and when it becomes available.

SemiAnalysis says Microsoft signed more than 10 GW of data center contracts during 2026, representing more than $300 billion in total contract value. The figure reportedly excludes GPU costs, and much of the capacity is expected to come online in late 2027 and 2028. That timing matters for investors watching Azure growth over the next several quarters.

A contract’s total value spread over multiple years does not equal annual revenue today. It also does not guarantee that every facility will open on schedule, that every customer will use the full allocation or that the contracted price will remain above the cost of delivery. Microsoft can build a powerful long-term backlog while still facing a near-term capacity gap.

The most useful way to read giant AI-infrastructure forecasts is as sensitivity analyses rather than forecasts of inevitable earnings. Securing compute early may prove strategically valuable, but scale alone does not create attractive returns when power constraints, depreciation, financing and falling inference prices are all moving at once. Investors should reward companies that disclose utilization, realized service margins and backlog conversion—not simply the largest announced gigawatt totals.

The analysis suggests Azure growth could accelerate from about 42% year over year to more than 100% by next year. That forecast would require a very large increase in available capacity, customer usage and recognized revenue.

Hardware is one constraint. Power availability, permitting, transmission equipment, cooling systems and construction schedules are others. Microsoft also has to decide how to allocate capacity among OpenAI, enterprise customers, Microsoft’s own products and other model companies.

Customer concentration adds another risk. A major portion of Microsoft’s AI capacity currently goes to OpenAI. That relationship provides scale, but it also ties the economics of Azure expansion to one partner whose products and costs continue to evolve. OpenAI’s ability to use other clouds under the new agreement gives it more flexibility and gives Microsoft less exclusivity.

The partnership remains valuable because Microsoft still has preferred positioning, long-term intellectual property rights and a substantial Azure commitment. The investment case becomes less straightforward when the same capacity must serve customers with different pricing, latency and reliability requirements.

SpaceX is the more speculative part of the story. Elon Musk said the company could deliver 6 GW to 8 GW of additional data center capacity in the following year, with a possible path above 10 GW. Other reporting put SpaceX at 1.4 GW of nameplate compute capacity at the end of June, with more than 2 GW targeted by year-end.

Musk also discussed larger supporting infrastructure targets, including as much as 20 GW of power and cooling capacity and approximately 15 GW at the power plant level. Those figures describe different layers of the system. Compute capacity, electrical capacity, cooling capacity and generation capacity cannot be treated as interchangeable units.

SemiAnalysis believes the 2027 buildout is technically plausible because SpaceX has demonstrated unusually fast construction. The company reportedly built the 300 MW Colossus 1 facility in 122 days. Its Southaven power plant grew from about 495 MW in February 2026 to 1.7 GW in July. A separate MiniHard project was projected to reach 450 MW to 500 MW within roughly five months.

Those examples support the argument that SpaceX can move quickly. They do not confirm that the company will complete 10 GW of usable compute on schedule, secure enough chips, connect the sites to customers or earn the projected revenue.

SemiAnalysis estimates that SpaceX could reach around $300 billion in annual recurring revenue while monetizing only half of its incremental 2027 capacity. The remaining capacity would serve xAI and Cursor for training under the report’s scenario. That projection contains two separate uncertainties: the amount of infrastructure SpaceX builds and the amount of external demand it can sell.

The company may become an important AI cloud provider. It may also use much of its capacity internally for Grok and other projects, which would create strategic value without producing the same external revenue as a cloud contract.

AI inference demand is growing, but the price of inference is under pressure. Model providers compete on speed, accuracy and cost. New hardware can process more tokens per dollar, while smaller models can handle workloads that once required frontier systems.

That creates a difficult tradeoff for infrastructure owners. Better efficiency can increase the number of tokens a gigawatt processes, yet it can also push prices lower. Revenue growth then depends on demand expanding faster than prices decline.

Enterprise customers may use more AI as prices fall, especially for coding, search, customer support and internal analysis. The resulting volume could support large data center deployments. Still, utilization will vary by application. Workloads with strict latency requirements may need reserved capacity, while batch workloads can be shifted to cheaper periods. A cluster’s average economics will reflect that mix.

Competition adds another variable. Microsoft, Google, Amazon, Oracle, CoreWeave, specialized neoclouds and possible SpaceX offerings are all pursuing the same customers. OpenAI and Anthropic may also negotiate lower infrastructure costs as their purchasing power grows. That could benefit model companies while compressing the spread assumed in the $100 billion calculation.

The SemiAnalysis work offers a useful way to understand why companies are signing enormous data center contracts. If inference demand reaches the report’s modeled levels, securing power and accelerators early could create a major advantage. Microsoft has a credible route to that opportunity through Azure and OpenAI. SpaceX has shown construction speed that deserves attention.

The most aggressive claims still require several milestones:

Investors can track those milestones through Azure growth, Microsoft Cloud margins, backlog conversion, capacity online, external customer announcements and the cost of serving AI workloads. For SpaceX, reported compute capacity and signed cloud contracts will matter more than power plant ambitions alone.

The $100 billion per gigawatt figure is best treated as a high-end economic possibility. It helps explain the rush to build AI infrastructure, but it does not remove the execution risk. The companies that build the most capacity may win the next phase of AI, or they may spend years carrying underused hardware purchased during a period of unusually strong expectations.

The market will eventually judge the difference through utilization, pricing and cash flow. Until those figures arrive, the jackpot remains mostly modeled.