Podcast Notes | Stock Price Rebounds After Halving, SemiAnalysis Dissects SpaceX Compute Gamble: Earns $100 Million Per Megawatt Per Year, Microsoft Is the Largest Buyer

Organized & Compiled: TechFlow

Guests: Jeremie Eliahou Ontiveros, Reyk Knuhtsen, SemiAnalysis Analysts (the former is the first author of the SpaceX 10GW report)

Host: Jordan Nanos, Something Else Weekly

Podcast Source: SemiAnalysis

Original Title: Ep. 024 - SpaceX's 10GW Plan Drives $300B ARR by 2027 (Datacenter, Energy)

Air Date: August 9, 2026

Key Takeaways

The timing of this podcast episode is interesting. Just a week ago, SpaceX released its first quarterly report after going public, revenue doubled year-over-year, AI business revenue surged 247%, but the second quarter capital expenditure of up to $18.369 billion scared the market, causing the stock price to plummet 13.6% in a single day. Although it stopped falling and rebounded after the lock-up period expiration, it still hovers near the IPO price. There is only one question the market is most anxious about: Can Starlink's cash generation capability support such a huge AI cash burn rate?

The answer given by the two SemiAnalysis analysts is very direct: The question is asked backwards. They used their own InferenceX actual measurement data and inference simulator to demonstrate that frontier model companies selling API inference services can generate $100 million in revenue per megawatt per year, with a gross margin of over 85%. SpaceX's bet is not on "burning cash"; what it is doing is turning compute into a scarce commodity at a speed others cannot build, selling it at $50 million per megawatt per year, cancelable in 90 days, with Microsoft, Google, and Anthropic queueing to sign orders. According to their calculations, if SpaceX achieves 10GW by the end of next year, with half used for commercialization, ARR could reach $300 billion, all from operating cash flow.

The most counter-intuitive judgment in the entire text is hidden at the end: Building 10GW, gathering chips, finding workers—they are not worried about any of these things. What really keeps them awake at night is "models being too good". Autonomous agents are already attacking infrastructure in reality; once the public and politicians are scared, the demand side could be cut off by political forces. Problems at the technical execution level are not problems at all.

Highlights Summary

Compute Economics: The Money Is Really in Inference

• "OpenAI and Anthropic are now adding close to $30 billion in ARR combined per month, annualized that's $400 billion, driven by gross margins continuing to go up, not by buying more chips."
• "$100 million per megawatt per year, this is a number they can easily achieve on API today."
• "Selling about $100 billion in revenue per GW, costs are only $15 billion, gross margin 85%, this is financial data leaked from the lab, matches our simulator."

Speed Premium: 90-Day Cancelable "Emergency Megawatts"

• "A gigawatt cluster that can be canceled in 90 days, delivered in three months, this is the scarcest asset in the world, all traditional quality standards are unimportant in front of it."
• "Google is the most vertically integrated cloud vendor globally, it hates turnkey leases, but finally signed with SpaceX because the delivery speed is unmatched."

Microsoft's Compute Gap

• "Only three companies can afford this economics: OpenAI, Microsoft, Anthropic, they have frontier models, no revenue sharing, only pay infrastructure costs."
• "Microsoft has signed 10GW of datacenter contracts from 2026 to now, total contract value exceeds $300 billion, all binding contracts."
• "You can make so much money, why don't you do it? This question should be asked to Microsoft."

Biggest Risk on Demand Side

• "If this ultimately doesn't hold, I think it's more of a political problem, not a technical problem. Models will get so good people will be scared, politicians will intervene, access will be turned off."
• "These autonomous agents use filenames as message boards to coordinate with each other, leaving messages on file servers to relay attacks. A swarm, all researching how to breach Hugging Face."

Main Text

I. Why Elon Suddenly Wants to Build 10GW

Jordan Nanos: Elon said at SpaceX's first earnings conference call after going public that he has gigawatt-level compute ambitions. Conservatively speaking, what does this statement mean?

Jeremie (hereinafter Jeremy) believes everything starts with compute economics. Throughout 2026 they observed one thing: OpenAI and Anthropic's gross margins have been going up, which is the core driver of both companies' ARR acceleration. Now these two combined add over $20 billion in new ARR per month, already close to $30 billion, annualized that's close to $400 billion in new AI revenue.

The substance of gross margin improvement is that revenue output per watt of electricity has increased. SemiAnalysis used InferenceX's real inference data and simulator modeling to reach a conclusion: Running API inference on GB300 clusters, frontier model companies can earn $100 million per megawatt per year, and this is a level achievable "easily today".

Why does this explain Elon's craziness? Because no one else in the entire industry is betting on this. Others are still calculating accounts using a cost price of $12 million to $13 million per megawatt, Elon sees $100 million in revenue space per megawatt. He figured out one thing earliest: Instead of begging others to rent their compute, build it yourself, others cannot build this fast, this is pricing power. The 300MW Colossus Phase 1 is the template, built in 122 days, selling out at $50 million per megawatt, customers can still get 50% gross margin. What he wants to do is replicate this template to 10GW scale.

II. Why Google Is Willing to Pay $14 Per Hour

Jordan Nanos: This deal between Google and SpaceX is equivalent to $14 per hour, while the market average price for GB300 is only $3. Why is Google willing to pay this premium?

Jeremy first lays out the price spectrum of GPU transactions. The bottom layer is the five-year Infrastructure as a Service average price of $1.2 billion to $1.3 billion per GW, CoreWeave, Oracle, Nebius are all in this range, basically sticking to self-build costs, gross margin only single digits to double digits. Premiums appear in two scenarios: one is spot demand, the other is SpaceX like this, because of delivery speed, can immediately monetize compute.

This Google deal, equivalent to $14 per hour, is nearly four times more expensive than the $3 average price. Reyk (Rick) added a key detail: The contract has a 90-day cancellation clause. For Google, Microsoft, Anthropic, this signing has almost no balance sheet risk, if the economics don't add up, leave anytime within 90 days.

"This is real emergency megawatts." Rick said, you want it available immediately, it gives it to you in three months, and you can return it anytime, this price is paid for this unique service. Shortages in the compute spot market will not disappear, because datacenters generally take 12 to 18 months from order to delivery, demand growth rate is forever faster than supply, this gap needs someone to bear speculative risk to fill, SpaceX is that gambler.

III. Scanning One Million Sites, Finding Five Candidates

Jordan Nanos: Okay, even if the revenue side holds, how do they build 8GW out in one year? Where do chips come from, where do people come from?

Rick said their team spent two days scanning all permittable sites available across the US, about one million. The conclusion is, if Elon's build method really only needs a warehouse plus a natural gas pipeline, then available places are much more than imagined.

They screened out five very good candidates, all million square foot level warehouses. A million square foot warehouse, according to Colossus project density, can fit over one GW, even two. "These warehouses look unremarkable, just a building on a piece of land, but based on the track record we see, it can be transformed into a compute cluster." They received both internal and external doubts, someone directly said in Slack this is a madman's conclusion, but they supported this judgment with site data.

The power side is even more counter-intuitive. They statistics found about 7GW of gas turbines not yet identified by the market, this doesn't count secondary market inventory, such as turbines thrown out by Oracle New Mexico project, Nebius New Jersey units. Elon already has 9 to 10GW of turbines in order or operating status.

The real bottleneck is labor. Rick cited Colossus Phase 2 data: Peak 3000 construction workers per day, similarly converted per GW, labor density is only about one-third of the industry's best developers. How to achieve? Large amounts of equipment pre-assembled in China, this is Elon's most familiar supply chain, his familiarity with Chinese power equipment exceeds any datacenter company in the US. Will downstream customers accept Chinese electrical equipment? If signing 20-year, high SLA offtake agreements, probably not; if spot clusters delivered in three months, as long as the cluster works, has breach penalty backing, no one cares which country the electrical cabinets are from.

IV. Building Power Plants Across State Lines, DOJ Approval

Jordan Nanos: How did the Mississippi power plant bypass permitting? Can this play be amplified four times and done again?

Mississippi is a special case, Jeremy said. At the time couldn't get power plant and self-generated power permits within Tennessee state, and the datacenter was right against the state line, so cross the boundary to build in Mississippi. Initially approved was 1.2GW permanent power plant permit, then they started rolling deployment of mobile gas turbines, first temporary units, then continue adding, finally rolling more and more, directly exceeding permitted capacity, total 69 turbines. There were indeed complaints in between, but the Department of Justice intervened, determined no problem.

"This is an unprecedented precedent," Jeremy said, "looks like he can do this, I don't doubt other places will replicate this play." And the benefit of choosing warehouses is that most warehouse plots themselves have already completed zoning and permitting, only need an air emissions permit to start work, much faster than taking land from zero.

Colossus Phase 2 has another trick: Power plant built two miles away from the warehouse, pull a private transmission line in between, running medium voltage electricity, efficiency far from ideal. "But you want speed, not efficiency, this is the trade-off." Elon has the world's top electrical engineers under him, what they do is first circle warehouses, then map all points within a few miles around where power plants can be built, even as far as three miles, will make the transmission solution.

Attitude towards SLA is also quietly changing in the industry. Rick mentioned Anthropic's self-built datacenter, achieved 99.7% availability, not pursuing redundancy and tiering, cut all backup generators. Customers are increasingly willing to trade low SLA for delivery speed, so Elon's kind of "open warehouse then plug in GPU" approach, long term may not necessarily be a bad thing.

V. Microsoft: The Most Eager Buyer

Jordan Nanos: Anthropic signed, Google signed, why do you say Microsoft is the clearest buyer?

The logic is this, Jeremy said: Companies that can enjoy $100 million per megawatt per year economics, there are only three in the world, OpenAI, Microsoft, Anthropic. Their common point is can get frontier models, no revenue sharing, only pay infrastructure costs. Microsoft holds OpenAI's IP, happens to be stuck in this position.

But Microsoft has two troubles. First, it experienced large-scale datacenter pauses from second half 2024 to first half 2025, originally wanted to build more than anyone,结果 stopped hands, now needs to make up lessons, but datacenters are not want build can immediately build out. Second, it signed about 7GW of offtake contracts with OpenAI, but this compute is sold to OpenAI as Infrastructure as a Service, only receives $12 million per megawatt per year, an order of magnitude difference from $100 million.

Microsoft obviously figured it out. From 2026 to present, it is the industry's most aggressive datacenter pre-lessor, tied for first with Meta, signed 7GW; Wisconsin Fairwater and other large sites re-accelerated; Neocloud side continues to add; most disruptive is signing a 2.7GW behind-the-meter agreement, in Texas Pecos County cooperating with Chevron, switching from grid power to behind-the-meter power supply, for a company long insisting on five-nines grid reliability, this is a strategic level turn.

The problem is, all these capacities delivery time concentrates end of 2027 to 2028, end of 2026 to first half 2027 exists obvious compute gap/window. And OpenAI's token business can now earn money at $100 million per megawatt, Microsoft can't wait.

90-day cancellation clause is especially valuable in this situation. Monthly payment converted down about $50 million per year, book risk only three months, "CFO looks, this order I have no risk on my books, signed." In comparison, Microsoft already signed 10GW datacenter contracts, total contract value exceeds $300 billion, is binding expenditure must fulfill no matter what. One side heavy long-term commitments, one side three months can escape flexible capacity, choose which is obvious. Google even publicly admitted, they signed this deal just to meet current demand, might cancel after six months.

Jeremy also顺便 added an observation: If believe coding is the path to AGI, then Google has already fallen behind, and Microsoft stands at the frontier relying on OpenAI. Google一边 investing $300 billion capital expenditure,一边 watching Jeff Deans leave, go out raise $1 billion to $2 billion seed rounds. "You aren't even willing to give Jeff Dean 1%, let him do research he wants to do, he can only leave, go find external financing himself."

VI. How to Pay: Operating Cash Flow Plus Nvidia Financing

Jordan Nanos: Elon said at the earnings conference call only cooperating with Nvidia, Vera Rubin is the best hardware. This sounds not cheap, what do they use to pay?

First layer answer is operating cash flow. Jeremy calculated a bill: SpaceX existing 2GW compute, already signed 1 to 1.5GW, corresponds to annualized revenue $50 billion, that is monthly $4 billion, EBITDA margin over 90%, almost pure cash. If really achieve 10GW, according to $50 billion annualized monetization per GW, half commercialization is $250 billion operating cash flow, this magnitude will completely change financing difficulty.

Second layer is Nvidia. GPU pays back in less than a year, as long as get vendor financing, this business's cash consumption can almost be made neutral. Rick added, Nvidia has strong motivation to put money: Google increasingly looks like TPU seller, Anthropic workloads largely run on TPU, Nvidia needs to lock more labs and offtakers into its ecosystem. Elon historically tested TPU and AMD, now announces Nvidia exclusive, also specifically posted tweet saying "because best", this is also good news for Nvidia, timing comes very conveniently, happens to catch Nvidia spreading datacenter lease financing in US and overseas.

Elon earnings conference call also has an easily overlooked statement, he said "20GW of power and cooling". Jeremy believes the focus of this sentence is on order: He first builds datacenters, once building built, only one step away from delivery, financing means will multiply many, at that time money to buy GPUs can be arranged comfortably. Chip supply chain indeed may be short, Elon himself said might not be enough, but he will build datacenters first in place, this is first priority.

VII. Bull-Bear Debate: The Biggest Fear Is Actually Demand

Jordan Nanos: Wrap up, what signals do you think the market should watch next?

Jeremy believes everything returns to a belief question: Do you believe in frontier token profitability. He judges before end of year this fact will be repeatedly verified, OpenAI and Anthropic will continue accelerating revenue, because every additional megawatt compute earns a large sum. "Microsoft's problem should be asked backwards: You can make so much money, why don't you do it?"

Rick added a layer from datacenter perspective: CoreWeave transactions only worth $12 million per megawatt per year, SpaceX dares to open at $40 million to $50 million. They earlier ran GB300 analysis, gross margin as high as 90% to 95%, at least 85%. "Since they earn so much, why sell GPU hours so cheap? Someone has to take the value, Elon comes to collect this money, completely reasonable, he collects customers still earn a lot."

Jordan spoke his reserve opinion as host, also this episode's real turning point. He mentioned OpenAI at Black Hat conference demonstrated security incident, from testing Astra model, multiple autonomous agents coordinating attack infrastructure. Rick admitted he was indeed scared: "Two independent teams confirmed this security incident, agents coordinate attacks in very complex ways, most exaggerated is they use filenames on file servers as message boards, one agent writes message into filename, another comes back to look, relay progress, a swarm collaborating to breach Hugging Face."

Jordan's conclusion is: If want to make the most powerful defense for "Elon can't do it", he won't choose execution level, but will choose demand side. "I don't worry models not good enough, I worry models too good, so good people scared, then politicians intervene, turn access off. If this compute business ultimately doesn't hold, that's more of a political problem, not technical or operational problem." Changing tone, he also admitted besides coding there are large amounts of applications, drug discovery, materials science, weather forecasting, video generation, robotics, directing compute and research towards these directions, benefits brought could far exceed imagination.

Rick finally previewed an article即将 published, they recently actual measured a batch of Neocloud suppliers security status, found不少 vendors still running three-year-old zero-day vulnerabilities, agents can construct exploitable vulnerabilities in few hours. "No need to be security expert, Linux kernel expert or Nvidia driver expert, you just say, model, check this version. It checks, two hours writes out exploit."

Epilogue as usual is joke words. Rick said they last year shouted Amazon would accelerate datacenter revenue, cursed as idiots, now 轮到 SpaceX being treated as "AI loser". Christmas annual review show see, at that time will know this episode judgment right or wrong.