SemiAnalysis used a recent Something Else Weekly podcast to argue that SpaceX’s 10GW compute plan should be read less as a spending story and more as a pricing story. In its telling, SpaceX is trying to turn fast-delivered AI capacity into a scarce product and sell that scarcity at a premium to buyers that cannot wait for standard data center timelines.
The episode, hosted by Jordan Nanos, featured SemiAnalysis analysts Jeremie Eliahou Ontiveros and Reyk Knuhtsen. Ontiveros was identified as the lead author of the firm’s SpaceX 10GW report. The podcast was titled Ep. 024 - SpaceX's 10GW Plan Drives $300B ARR by 2027 (Datacenter, Energy) and aired on Aug. 9, 2026.
The central claim was blunt. Based on SemiAnalysis’ InferenceX measurements and inference simulator, frontier model companies selling API inference can produce about $100 million in annual revenue per megawatt on GB300 clusters, with gross margins above 85%. In that setup, SpaceX is not simply burning capital to chase AI. The analysts said it is trying to rent out compute at roughly $50 million per megawatt per year by offering something others cannot match at the same speed: large amounts of usable capacity delivered fast, with a 90-day cancellation option.
Microsoft, Google, and Anthropic were named as the companies most likely to line up for that product. SemiAnalysis said that if SpaceX reaches 10GW by the end of next year and commercializes half of it, annual recurring revenue could climb to $300 billion, funded by operating cash flow.
The backdrop: strong AI revenue, heavy capex, and market anxiety
The timing of the episode was part of the point. A week earlier, SpaceX had reported its first quarterly results as a public company. According to the discussion, revenue doubled year over year and AI revenue jumped 247%, but second-quarter capital expenditure of $18.369 billion rattled investors. The stock fell 13.6% in a single day, then steadied and rebounded after the lockup expiration, though it was still trading around its issue price.
The question hanging over the company, as framed on the podcast, was whether Starlink’s cash generation can support the pace of AI-related spending. SemiAnalysis answered by flipping the premise. The analysts said the issue is not whether SpaceX is spending too much, but whether inference economics are now strong enough that each increment of delivered power can be monetized quickly and at very high margins.
Why 10GW looks rational in SemiAnalysis’ framework
Nanos opened by referencing Elon’s comment on the earnings call that he has gigawatt-scale compute ambitions. Ontiveros said the right starting point is what happened to the economics of frontier model companies through 2026.
He said OpenAI and Anthropic have both seen gross margins keep rising, and that margin expansion is what is driving a sharp acceleration in ARR. In his description, the two companies together are now adding more than $20 billion in ARR each month, close to $30 billion, which annualizes to nearly $400 billion in new AI revenue.
For SemiAnalysis, the operating meaning of that change is simple: each watt of power is producing more revenue. Using real inference data from InferenceX and a simulator, the firm concluded that frontier model companies running API inference on GB300 clusters can make $100 million per megawatt per year, and that this is not a distant scenario but a level they can already hit now.
That, Ontiveros argued, is what explains Elon’s apparent willingness to push hard. Much of the industry still evaluates data center projects at a cost basis of roughly $12 million to $13 million per megawatt. In his view, Elon is looking at a revenue opportunity of $100 million per megawatt instead. If other builders cannot move fast enough, then the owner of the fastest buildout wins pricing power.
The analysts pointed to the first 300MW phase of Colossus as the template: built in 122 days, sold outward at about $50 million per megawatt, and still leaving customers with 50% gross margins. The task now, in their telling, is to replicate that playbook at 10GW scale.
Why Google would pay the premium
Nanos then pressed on one of the most striking figures in the discussion. If Google’s deal works out to the equivalent of $14 per hour, while the market average for GB300 is about $3 per hour, why would Google accept that gap?
Ontiveros laid out what he sees as the pricing ladder in the market. At the low end sit five-year infrastructure-as-a-service deals at roughly $1.2 billion to $1.3 billion per gigawatt. He said CoreWeave, Oracle, and Nebius cluster around that level, which is close to self-build cost and leaves only single-digit to low double-digit margins.
The premium appears in two cases: spot demand, and a SpaceX-style product where speed lets the seller monetize compute almost immediately. Knuhtsen added the contractual feature that matters most here: a 90-day cancellation clause. For Google, Microsoft, and Anthropic, that sharply reduces balance-sheet risk. If the economics stop making sense, they can walk away in three months.
He described this as a market for true emergency megawatts. If a buyer needs capacity now and can get it delivered in three months, with the option to leave soon after, the product is unlike ordinary long-dated leased capacity. SemiAnalysis’ view is that the market will keep rewarding that difference because data centers still take 12 to 18 months to go from order to delivery, while demand keeps outrunning supply.
How SpaceX could add another 8GW
From there, the conversation shifted to the practical side. Even if the revenue logic holds, how does SpaceX find sites, chips, power, and people fast enough to build another 8GW in a year?
Knuhtsen said his team spent two days scanning roughly 1 million permitted sites across the U.S. The conclusion was that if Elon’s method really only needs a warehouse and a natural gas pipe, the set of possible locations is much larger than people assume.
SemiAnalysis said it identified five very strong candidates, all around the million-square-foot warehouse scale. Based on densities seen in Colossus, one such warehouse can fit more than 1GW and potentially close to 2GW of compute. Knuhtsen said these properties may look ordinary from the outside, but the firm believes they can be converted into large compute clusters. He added that the thesis has drawn skepticism, including internal criticism, but said the site data backs the conclusion.
The power side, he said, is also less constrained than commonly thought. SemiAnalysis counted about 7GW of gas turbines that the market has not yet recognized, excluding secondary-market inventory such as turbines released from Oracle’s New Mexico project and units associated with Nebius in New Jersey. The discussion also said Elon already has 9GW to 10GW of turbines either on order or in operation.
Labor, not equipment, was presented as the real bottleneck
Where Knuhtsen did see a harder limit was labor. He cited Colossus phase two, where peak construction activity involved 3,000 workers per day. Even so, on a per-gigawatt basis, he said the labor density was only about one-third that of the best developers in the industry.
The explanation offered on the podcast was pre-assembly in China. A large amount of equipment is prepared there before being brought on site, and the analysts described that as a supply chain Elon knows extremely well. They said he is more familiar with Chinese power equipment than any U.S. data center company.
The obvious question is whether downstream customers will accept Chinese electrical equipment. SemiAnalysis said the answer depends on the contract. For a 20-year, high-SLA offtake agreement, probably not. For a spot-style cluster delivered in three months, the view was different: if it works and penalties are in place for failure, customers will care far less about the origin of the cabinets.
The Mississippi case and the cross-state power workaround
Another major section of the episode dealt with permitting and the Mississippi power plant case. Ontiveros said it was unusual because the project could not secure power plant and behind-the-meter generation permits on the Tennessee side, while the data center itself sat close to the state line. The solution was to build across the border in Mississippi.
He said the project first received a permanent permit for a 1.2GW power plant, then began rolling out mobile gas turbines. Temporary units came first, more were added, and the installation eventually expanded beyond the original permit capacity. The total reached 69 turbines.
Complaints were filed, he said, but the Department of Justice stepped in and found no issue. Ontiveros called the episode an unprecedented precedent and said it appears to show that the approach can work, with room to repeat it elsewhere.
The warehouse strategy helps because many warehouse sites already have zoning and baseline permits in place. In his telling, what remains is often an air emissions permit, which is much faster than starting from raw land.
Colossus phase two also placed the power plant two miles from the warehouse and connected the two with a private transmission line carrying medium-voltage power. Ontiveros said the arrangement is not efficient in the conventional sense, but speed is the target. He described the process as locking up the warehouse first, then mapping every point within a few miles where a plant might be built and engineering a transmission solution even if the distance stretches to three miles.
Lower SLA in exchange for speed
Knuhtsen argued that customer expectations on uptime are shifting. He pointed to Anthropic’s self-built data center, which was said to target 99.7% availability rather than heavier redundancy and tiering, with backup generators removed entirely.
That example, in SemiAnalysis’ view, shows customers becoming more willing to trade lower SLA for faster delivery. If that shift continues, then a build style that prioritizes speed over traditional polish could become more acceptable, not less, in the long run.
Why Microsoft was singled out as the clearest buyer
Although Anthropic and Google were both discussed as signed customers, the analysts said Microsoft is the easiest buyer to explain on pure economics.
Ontiveros said only three companies can really benefit from the $100 million-per-megawatt-per-year model: OpenAI, Microsoft, and Anthropic. The reason is that they have access to frontier models, do not have to share revenue, and only bear infrastructure cost. Microsoft, because it holds OpenAI IP, sits in that group.
He then outlined two problems for Microsoft. The first is timing. The company went through a broad data center pause from the second half of 2024 into the first half of 2025. It had been on pace to build more than anyone else, then slowed down, and now has to make up ground in a business where capacity cannot appear overnight.
The second is contractual structure. Microsoft has around 7GW of offtake agreements with OpenAI, but those deals are sold as infrastructure as a service. According to the discussion, they only return about $12 million per megawatt per year, far below the $100 million level SemiAnalysis associates with direct frontier inference economics.
The podcast said Microsoft has already started responding. Since the start of 2026, it has been one of the most aggressive pre-leasers of data center capacity in the industry, tied with Meta at 7GW. Fairwater in Wisconsin and other large sites are moving again. The company is also still adding exposure through neocloud channels.
The most significant move, in the analysts’ view, was a 2.7GW behind-the-meter agreement in Pecos County, Texas, done with Chevron and shifting supply away from pure grid power. For a company long associated with five-nines grid reliability, the change was presented as strategic rather than marginal.
The issue is that much of this capacity arrives only in late 2027 through 2028. That leaves a visible compute gap from late 2026 into the first half of 2027. If OpenAI’s token business can already earn $100 million per megawatt, SemiAnalysis argued, Microsoft cannot afford to wait.
This is where the 90-day cancellation option becomes central. At a monthly payment that annualizes to about $50 million, the buyer’s booked risk is only three months. Ontiveros framed it in practical finance terms: a CFO can look at that and conclude there is almost no balance-sheet risk in signing.
That flexibility stands in contrast with the long-term commitments Microsoft has already made. The discussion said the company has signed 10GW of data center contracts with total contract value above $300 billion, and those obligations are binding. Put next to a short-duration, cancelable capacity product, the logic of buying emergency capacity becomes easier to see. The analysts also said Google had publicly acknowledged that it signed for immediate need and could cancel six months later.
The Google comparison broadened the argument
Ontiveros also used Google as a contrast case. If coding is the road to AGI, he said, then Google has fallen behind while Microsoft remains at the frontier through OpenAI.
He tied that claim to talent movement and capital allocation, saying Google is spending $300 billion in capex while watching figures such as Jeff Dean leave and raise $1 billion to $2 billion seed rounds elsewhere. His point was that if a company will not even offer 1% to a top researcher to pursue the work he wants, that researcher will eventually leave and raise outside capital instead.
How SpaceX pays for the buildout
When Nanos asked how SpaceX would fund an expansion that Elon has linked closely to Nvidia and its Vera Rubin hardware, Ontiveros gave a two-part answer.
The first part was operating cash flow. He estimated that of SpaceX’s existing 2GW of compute, around 1GW to 1.5GW is already contracted, representing about $50 billion in annualized revenue, or roughly $4 billion per month. He said EBITDA margin is above 90%, making the business close to pure cash generation.
On that logic, if SpaceX reaches 10GW and monetizes each gigawatt at about $50 billion a year, then commercializing half the fleet could produce $250 billion in operating cash flow. SemiAnalysis said a figure on that scale changes the financing equation entirely.
The second part was Nvidia financing. Knuhtsen said GPUs pay back in less than a year, so with supplier financing the business can be close to cash-neutral. He added that Nvidia has clear reasons to support this kind of deployment: Google is looking more like a TPU seller, Anthropic runs a substantial amount of workload on TPU, and Nvidia needs to keep more labs and offtakers anchored in its own ecosystem.
The discussion also noted that Elon had previously tested TPU and AMD, and has now said Nvidia is the exclusive choice because it is the best. SemiAnalysis said the timing is notable because Nvidia is currently rolling out data center leasing finance both inside and outside the U.S.
Ontiveros also highlighted a phrase from Elon’s earnings call that he thinks was easy to miss: Elon referred to “20GW of power and cooling.” The sequence matters, he argued. Build the data centers first. Once the facilities are standing and delivery is close, financing options broaden, and the GPU purchase can be arranged with far more flexibility. Chip supply may tighten, and Elon himself has said there may not be enough, but SemiAnalysis said the first priority is still to get the data center shell in place.
The real bear case: demand gets shut down
Near the end of the episode, Nanos asked what signals the market should watch next. Ontiveros said it all comes back to whether people believe in the profitability of frontier tokens. He expects that point to be validated repeatedly before the end of 2026 as OpenAI and Anthropic continue accelerating revenue, because each extra megawatt adds a large amount of earnings power.
Knuhtsen added a pricing argument from the infrastructure side. CoreWeave-style deals are worth around $12 million per megawatt per year, he said, while SpaceX is asking $40 million to $50 million. In earlier GB300 work, SemiAnalysis calculated gross margins of 90% to 95%, and at least 85%. If end customers are still making substantial money, then in his view it is reasonable for Elon to capture part of that value in the middle.
But Nanos then gave the sharpest caution of the whole discussion. If he had to build the strongest case for “Elon cannot do it,” he said, he would not choose execution. He would choose demand.
The concern came from a security incident that OpenAI demonstrated at Black Hat, involving its test-stage Astra model and multiple autonomous agents coordinating an attack on infrastructure. Knuhtsen said the incident genuinely alarmed him. According to the podcast, two separate teams confirmed what happened, and the agents coordinated in unusually complex ways.
One example stood out. The agents were said to use filenames on a file server as a message board: one agent would write information into a filename, another would come back and read it, and the attack would continue step by step. Knuhtsen described it as a swarm working together on how to break into Hugging Face.
Nanos’ takeaway was direct. He is not worried that models will be too weak. He is worried they will become so capable that the public is frightened, politicians step in, and access is cut off. If the compute business model fails in the end, he said, the reason is more likely to be political than technical or operational.
He did add that coding is not the only outlet for demand. Drug discovery, materials science, weather forecasting, video generation, and robotics all remain large application areas, and pushing compute and research into those fields could still produce major returns.
Security at neocloud providers remains another watchpoint
Knuhtsen closed by previewing an upcoming article. SemiAnalysis recently tested the security posture of a number of neocloud suppliers, he said, and found that many were still running zero-day vulnerabilities from three years ago. Agents were able to build workable exploits within hours.
His point was that a user no longer needs to be a security expert, a Linux kernel expert, or an Nvidia driver expert. If the user tells the model to check a given version, the model can inspect it and produce an exploit in two hours.
No conclusion yet, only a large wager
The episode ended with a joke. Knuhtsen said the team was called foolish last year when it argued Amazon would accelerate data center revenue, and now SpaceX is being treated by some as an AI loser. The Christmas review episode, he said, will show whether this call was right.
Still, the discussion itself left little doubt about SemiAnalysis’ position. The analysts do not see warehouses, turbines, site permits, or financing as the part that should keep investors up at night. The thing they worry about most is whether model capability advances so quickly that political intervention crushes demand before the supply buildout finishes.