SemiAnalysis Breaks Down SpaceX's Computing Power Bet: $100 Million Per Megawatt Per Year, Microsoft Is the Biggest Buyer

Compiled & Edited by: Deep Tide TechFlow

Guests: Jeremie Eliahou Ontiveros, Reyk Knuhtsen, Analysts at SemiAnalysis (The former is the lead 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 particularly interesting. Just a week prior, SpaceX delivered its first quarterly report since going public, with revenue doubling year-over-year and AI business income surging 247%. However, the massive $18.369 billion in capital expenditures for Q2 spooked the market, causing the stock to plummet 13.6% in a single day. Although it stabilized and rebounded after the lock-up period expiry, the stock still hovers near its IPO price. The market's most pressing concern is simple: Can Starlink's cash generation support such a massive pace of AI spending?

The two SemiAnalysis analysts' answer is straightforward: The question is framed incorrectly. Using their proprietary InferenceX real-world data and inference simulators, they argue that frontier model companies selling API inference services can generate $100 million in revenue per megawatt annually, with gross margins exceeding 85%. SpaceX's bet isn't about "burning cash." Instead, they are using their unparalleled speed to turn compute into a scarce commodity, selling it at $50 million per megawatt annually with 90-day cancellation terms, with Microsoft, Google, and Anthropic queuing up to sign. According to their calculations, if SpaceX reaches 10GW by the end of next year and commercializes half of it, they could hit $300 billion in ARR, all funded by operating cash flow.

The most counterintuitive judgment in the entire article is tucked away at the end: Building 10GW, sourcing chips, and finding workers are all things they don't worry about. What truly keeps them up at night is "models being too good." Autonomous agents are already attacking critical infrastructure in the real world. Once the public and politicians get scared, demand could be choked off by political forces. Execution-level technical problems, on the other hand, don't seem like problems at all.

Highlights & Key Quotes

The Economics of Compute: The Real Money is in Inference

• "OpenAI and Anthropic are now adding nearly $30 billion in combined ARR monthly, which annualizes to $400 billion. The driver is expanding gross margins, not buying more chips."
• "$100 million per megawatt annually – that's a number they can easily achieve today on their APIs."
• "Each GW sold generates roughly $100 billion in revenue against a cost of only $15 billion – an 85% gross margin. This comes from financial data leaked from the labs, and it aligns with our simulators."

Speed Premium: The "Emergency Megawatt" with 90-Day Cancellation

• "A gigawatt-scale cluster that can be cancelled in 90 days and delivered in three months is the most scarce asset in the world. All traditional quality standards become irrelevant in its presence."
• "Google, the world's most vertically integrated cloud provider, hates turnkey leasing, yet they ultimately signed with SpaceX because the delivery speed is unmatched."

Microsoft's Compute Gap

• "Only three companies can afford this kind of economics: OpenAI, Microsoft, and Anthropic. They have frontier models, no revenue sharing, and only pay infrastructure costs."
• "Microsoft has signed 10GW of data center contracts so far in 2026, with a total contract value exceeding $300 billion, all of it binding."
• "If you can make that much money, why wouldn't you do it? That's the question to ask Microsoft."

The Biggest Risk is on the Demand Side

• "If this doesn't ultimately work out, I think it's more of a political problem than a technical one. Models will become so good they scare people, politicians will intervene, and access will be shut off."
• "These autonomous agents use filenames as a message board to coordinate with each other, leaving messages on file servers to relay attacks. A whole swarm, all figuring out how to breach Hugging Face."

Full Transcript

1. Why Elon Suddenly Wants to Build 10GW

Jordan Nanos: Elon mentioned his gigawatt-scale compute ambitions on SpaceX's first earnings call post-IPO. What does this ambition mean, at a conservative estimate?

Jeremie (referred to as Jeremy) believes it all comes down to the economics of compute. Throughout 2026, they observed one key trend: the gross margins of OpenAI and Anthropic have been consistently climbing, which is the core driver behind the accelerating ARR of these two companies. Combined, they are now adding over $20 billion in monthly ARR, approaching $30 billion, which annualizes to nearly $400 billion in new AI revenue.

The essence of improving gross margins is that the revenue generated per watt of power has increased. SemiAnalysis, using real inference data from InferenceX and their simulators, reached a conclusion: Running API inference on GB300 clusters, a frontier model company can earn $100 million per megawatt annually, a level achievable "easily today."

Why does this explain Elon's audacity? Because no one else in the industry is betting on this. Others are still calculating based on a cost of $12-13 million per megawatt, but Elon sees the potential for $100 million in revenue per megawatt. He was the first to realize a key point: rather than begging others to rent his compute, he should build it himself. No one else can build it this fast, and that's pricing power. The 300MW Colossus Phase 1 is the template – built in 122 days, sold at $50 million per megawatt, still allowing customers a 50% gross margin. All he needs to do is replicate this template at a 10GW scale.

2. Why Google is Willing to Pay $14 an Hour

Jordan Nanos: The Google-SpaceX deal works out to an equivalent of $14 per hour, while the market average for GB300 is only $3. Why is Google willing to pay this premium?

Jeremy lays out the spectrum of GPU deal pricing. At the bottom, there's the five-year Infrastructure-as-a-Service average of $1.2-1.3 billion per GW, where CoreWeave, Oracle, and Nebius all operate, basically matching self-build costs with single-digit to low-double-digit gross margins. The premium appears in two scenarios: spot market demand, or situations like SpaceX, where delivery speed allows for immediate monetization of compute.

The Google deal, at an equivalent of $14 per hour, is nearly four times more expensive than the $3 average. Reyk (Rick) adds a critical detail: the contract includes a 90-day cancellation clause. For Google, Microsoft, and Anthropic, this signing carries almost no balance sheet risk. If the economics don't work out, they can walk away within 90 days.

"This is truly emergency megawatts," Rick says. "You need it right now, they deliver in three months, and you can cancel anytime. This price is paying for that unique service." The shortage in the compute spot market won't disappear because data centers typically take 12 to 18 months from order to delivery. Demand growth always outpaces supply, and this gap needs someone to take on speculative risk to fill it. SpaceX is that gambler.

3. Scanning a Million Sites to Find Five Candidates

Jordan Nanos: Okay, even if the revenue side holds up, how can they build 8GW in a year? Where do the chips and the people come from?

Rick says their team spent two days scanning all the permitted sites they could find across the US, roughly a million. The conclusion is that if Elon's construction method genuinely only requires a warehouse and a natural gas pipeline, the available options are far more numerous than one might think.

They shortlisted five excellent candidates, all warehouses with million-square-foot footprints. A million-square-foot warehouse, based on Colossus density, can house over a GW, sometimes even two. "These warehouses look unassuming, just a building on a plot of land, but based on the track record we've seen, it can be converted into a compute cluster." They've received skepticism from both inside and outside the company – someone on Slack even called it a madman's conclusion – but they backed their judgment with site data.

The power side is even more counterintuitive. They counted roughly 7GW of gas turbines that the market hadn't yet identified, not including existing inventory in the secondary market, like turbines from Oracle's New Mexico project or Nebius's New Jersey units. Elon already has 9-10GW of turbines on order or in operation.

The real bottleneck is labor. Rick cites data from Colossus Phase 2: a peak of 3,000 construction workers daily. Normalized per GW, their labor density is only about one-third that of the industry's best developers. How? Heavy pre-assembly of equipment from China – this is Elon's most familiar supply chain. His familiarity with Chinese power equipment surpasses that of any US data center company. Will downstream customers accept Chinese electrical equipment? Probably not for a 20-year, high-SLA offtake agreement. But for a spot cluster delivered in three months, as long as the cluster works and there are breach penalties as a backstop, no one cares where the electrical cabinets are made.

4. Building a Power Plant Across State Lines, DOJ Gives Green Light

Jordan Nanos: How did the Mississippi power plant bypass permitting? Can this playbook be scaled up four times over?

Mississippi is a special case, Jeremy says. At the time, they couldn't get permits for a power plant and self-generation within Tennessee, but the data center was right on the state border, so they crossed over to Mississippi. Initially permitted for a 1.2GW permanent plant, they then began rolling out mobile gas turbines – first temporary units, then adding more, until it snowballed past the permitted capacity, totaling 69 turbines. There were complaints along the way, but the DOJ got involved and cleared it.

"This is an unprecedented precedent," Jeremy says. "It looks like he can simply do this, and I have no doubt this playbook will be replicated elsewhere." The advantage of choosing warehouses is that most warehouse sites are already zoned and permitted. Only an air emissions permit is needed to start construction, which is much faster than acquiring land from scratch.

Colossus Phase 2 also deployed another tactic: building the power plant two miles from the warehouse and running a private transmission line using medium voltage, which is far from efficient. "But you want speed, not efficiency – that's the trade-off." Elon has some of the world's best electrical engineers. Their process is to secure warehouses first, then map out all viable power plant locations within a few miles, even up to three miles out, and design the transmission solution.

The industry's attitude towards SLAs is also quietly shifting. Rick mentions Anthropic's self-built data center, aiming for 99.7% availability, forgoing redundancy and tiering, and cutting all backup generators. Customers are increasingly willing to trade lower SLAs for faster delivery. So Elon's "rip open a warehouse and plug in GPUs" approach might not be a bad thing in the long run.

5. Microsoft: The Most Eager Buyer

Jordan Nanos: Anthropic signed, Google signed. Why do you say Microsoft is the most obvious buyer?

Here's the logic, Jeremy says: Only three companies in the world can enjoy the economics of $100 million per megawatt annually: OpenAI, Microsoft, and Anthropic. Their commonality is access to frontier models, no revenue sharing, and paying only infrastructure costs. Microsoft holds OpenAI's IP, putting it squarely in this position.

But Microsoft has two problems. First, it experienced massive data center pauses from H2 2024 to H1 2025. It was supposed to build more than anyone, but it stopped. Now it needs to catch up, and data centers can't just be conjured up overnight. Second, it signed an offtake agreement with OpenAI for roughly 7GW, but this compute is sold to OpenAI as Infrastructure-as-a-Service, generating only $12 million per megawatt annually – an order of magnitude less than $100 million.

Microsoft seems to have figured this out. So far in 2026, it is the industry's most aggressive data center pre-lessor, tied with Meta for first place, having signed 7GW. Major sites like Fairwater in Wisconsin have accelerated. Neocloud investments continue to increase. Most disruptively, it signed a 2.7GW behind-the-meter agreement in Pecos County, Texas, partnering with Chevron, shifting from grid power to behind-the-meter supply. For a company that has long insisted on five-nines grid reliability, this is a strategic pivot.

The problem is that all these capacities are scheduled for delivery between late 2027 and 2028, leaving a significant compute gap in late 2026 and H1 2027. Meanwhile, OpenAI's token business can generate $100 million per megawatt today. Microsoft can't afford to wait.

The 90-day cancellation clause is particularly valuable in this situation. With monthly payments annualizing to about $50 million, the balance sheet exposure is only three months. "The CFO looks at it and says, 'There's no risk on my books for this deal. Sign it.'" In contrast, Microsoft's already-signed 10GW of data center contracts, totaling over $300 billion, are binding commitments that must be fulfilled. On one hand, heavy long-term commitments; on the other, flexible capacity you can walk away from in three months. The choice is obvious. Google has even openly admitted they signed this deal to meet immediate demand and might cancel in six months.

Jeremy also adds an observation: If you believe coding is the path to AGI, Google has already fallen behind, while Microsoft, through OpenAI, is at the forefront. Google is pouring $300 billion into capex while watching Jeff Dean and others leave to raise $1-2 billion seed rounds externally. "You won't even give Jeff Dean 1% of that to do the research he wants to do, so he leaves to find external funding himself."

6. How to Pay for It: Operating Cash Flow Plus Nvidia Financing

Jordan Nanos: Elon said on the earnings call that they're only working with Nvidia and that Vera Rubin is the best hardware. That doesn't sound cheap. How will they pay for it?

The first layer of the answer is operating cash flow. Jeremy runs the numbers: Of SpaceX's existing 2GW of compute, 1-1.5GW is already contracted, corresponding to $50 billion in annualized revenue, or $4 billion monthly, with EBITDA margins over 90% – nearly pure cash. If they truly reach