When Elon Musk told investors that SpaceX could expand its AI compute capacity from roughly 1.4 gigawatts today to potentially more than 10 gigawatts by the end of 2027, the number sounded almost impossible.
Not ambitious.
Not aggressive.
Almost physically absurd.
Building that much AI infrastructure in such a short period would require enormous amounts of land, power, cooling equipment, networking hardware, transformers, construction labor and—most importantly—advanced GPUs.
It would also require something else.
Hundreds of billions of dollars.
Independent research firm SemiAnalysis estimates that reaching roughly 10GW of compute capacity could require somewhere between $300 billion and $500 billion in capital expenditure by the end of 2027.
That is the kind of number that normally ends the conversation.
Except SemiAnalysis kept looking.
And the deeper its analysts went, the more they began to believe Musk’s plan might not be as irrational as it first appeared.
The reason had surprisingly little to do with SpaceX.
It had to do with Microsoft.
During SpaceX’s first earnings call as a public company, Musk described the company’s AI infrastructure plans in language that immediately caught the attention of anyone familiar with data-center construction.
SpaceX currently operates around 1.4GW of AI compute.
Musk suggested the company could add roughly 6GW to 8GW next year, potentially taking total capacity above 10GW.
And he called that outlook, remarkably, “conservative.”
For most companies, adding even a fraction of a gigawatt is a multiyear infrastructure challenge.
A gigawatt-scale AI data center is not a building filled with computers.
It is closer to an industrial ecosystem.
The electrical infrastructure alone can resemble what is required to power a city.
Transmission lines must be available.
Substations must be built.
Transformers have to be sourced.
Cooling systems need to handle extraordinary thermal loads.
Thousands upon thousands of GPUs have to arrive on schedule.
Then the entire system has to operate reliably enough that expensive computing hardware is not sitting idle.
Traditional hyperscalers have spent years learning how to do this.
Musk was talking about multiplying SpaceX’s existing capacity several times over in roughly a year and a half.
The initial reaction was predictable.
How?
SemiAnalysis, a San Francisco-based research and advisory firm focused heavily on semiconductors and AI infrastructure, decided to work through that question.
Researchers Jeremie Eliahou Ontiveros, Reyk Knuhtsen and Jordan Nanos examined Musk’s proposal in an August 7 analysis.
Their conclusion was not that SpaceX definitely would reach 10GW.
It was more interesting than that.
They concluded that the plan was plausible enough to be taken seriously.
And one of the biggest reasons was that Musk may not need SpaceX itself to consume all that compute.
There may already be customers desperate to buy it.
The most obvious one could be Microsoft.
That sounds strange at first.
Microsoft is one of the largest cloud-computing companies on Earth.
Azure operates data centers across the world.
The company has spent tens of billions of dollars building AI infrastructure.
If anyone should have enough compute, surely it would be Microsoft.
But the AI boom has created a problem even companies of Microsoft’s scale are struggling to solve.
Demand is increasing faster than infrastructure can be delivered.
Microsoft has enormous obligations associated with AI models, including workloads connected to OpenAI.
Those models require increasingly large amounts of compute not only for training but for inference—the ongoing process of actually serving AI responses to users.
Every time millions more people begin using AI assistants, coding tools, enterprise agents and generative applications, the infrastructure burden expands.
And unlike software, physical compute capacity cannot be downloaded overnight.
A data center takes time.
Power takes time.
Grid connections take time.
Transformers take time.
Permits take time.
GPUs can be ordered quickly compared with constructing the electrical system required to operate them.
That is how one of the world’s richest technology companies can still end up with a compute shortage.
SemiAnalysis believes Microsoft may be facing exactly that problem.
And that changes the economics of SpaceX’s plan.
Because scarce compute isn’t priced like an ordinary commodity.
It carries a premium.
According to SemiAnalysis, large-scale, near-term AI compute could potentially command pricing equivalent to roughly $50 billion per gigawatt per year.
Read that number again.
$50 billion.
Per gigawatt.
Per year.
That is why a $300 billion or even $500 billion infrastructure program begins to look different.
If SpaceX can build compute faster than traditional providers can—and if customers are willing to sign enormous long-term contracts for that capacity—then the question is no longer simply:
“How could anyone afford to build 10GW?”
It becomes:
“How valuable is 10GW if almost no one else can deliver it when customers need it?”
That is the twist.
Musk’s advantage may not be that he can build data centers more cheaply.
It may be that he believes SpaceX can build them faster.
And in a market facing extreme scarcity, speed can be worth tens of billions of dollars.
SemiAnalysis went even further.
Using its own inference simulations and coding benchmarks, the firm estimated that frontier AI laboratories could generate around $100 billion in annual recurring revenue per gigawatt of productive compute.
That kind of economics explains why companies such as OpenAI and Anthropic continue spending enormous amounts on infrastructure.
At first glance, paying tens of billions for compute sounds reckless.
But if that compute supports services capable of generating even greater revenue, the spending becomes rational.
The machine is expensive.
What the machine produces may be worth more.
SemiAnalysis believes Microsoft belongs in the same category.
And Microsoft’s position is unusual.
The company has a deep relationship with OpenAI and an economic stake in its success.
That gives Microsoft access to frontier AI capabilities without necessarily bearing exactly the same training economics as an independent AI laboratory building everything alone.
If Microsoft can combine OpenAI-powered models with Azure distribution, enterprise customers, Office, GitHub and its enormous global software footprint, every additional unit of compute can be monetized across multiple businesses.
That makes compute not merely an operating expense.
It becomes inventory.
And right now, there may not be enough inventory.
This is why the analysts began looking at SpaceX as a potential supplier.
Suppose Musk actually builds several additional gigawatts.
SpaceX does not necessarily need Grok, Starlink or internal AI operations to consume all of it immediately.
It can sell excess capacity.
And a customer such as Microsoft might be willing to absorb enormous amounts of it.
Not because Microsoft suddenly wants to depend on Elon Musk.
Because when compute is scarce enough, ideology becomes secondary to capacity.
A GPU running a profitable workload does not care who owns the building.
If SpaceX can deliver usable AI capacity before competing infrastructure comes online, Microsoft could have an economic reason to become one of its largest customers.
That possibility turns SpaceX’s AI expansion into something much bigger than an internal infrastructure project.
It could become a cloud business measured in gigawatts rather than servers.
And Musk already understands the value of selling scarce infrastructure.
SpaceX’s rocket business grew partly because it solved a problem that customers could not easily solve themselves.
Getting payloads into orbit was expensive.
Launch availability was limited.
SpaceX developed reusable rockets, increased launch cadence and drove down costs.
Then it became the company everyone else had to call.
The same pattern appeared with Starlink.
Building a global satellite network required extraordinary upfront capital.
SpaceX spent the money.
Once the constellation existed, customers could buy connectivity instead of building satellites themselves.
Now Musk may be attempting the same play in AI.
Spend first.
Build capacity that looks excessive.
Then wait for everyone else to discover they need it.
That strategy is enormously risky.
But it is not unfamiliar.
Traditional companies generally build infrastructure around confirmed demand.
Musk often tries to build infrastructure around demand he believes is inevitable.
Gigafactories came before electric vehicle demand was proven at today’s scale.
Superchargers appeared before millions of Teslas were on the road.
Reusable rockets required enormous investment before SpaceX could dominate launch cadence.
Starlink required launching thousands of satellites before subscription revenue could justify the network.
In each case, the early stage looked expensive because the infrastructure existed before the mature market did.
The AI buildout could follow the same pattern.
Except this time the numbers are far larger.
At $300 billion to $500 billion of estimated capital requirements, getting the timing wrong would be catastrophic.
If AI efficiency improves faster than expected, customers may need less compute.
If competitors bring massive amounts of capacity online, scarcity premiums could disappear.
If power infrastructure delays deployments, expensive GPUs could sit waiting for electricity.
If model economics fail to justify today’s spending, hyperscalers could pull back.
And if SpaceX builds too far ahead of demand, tens or hundreds of billions of dollars could be trapped in infrastructure earning inadequate returns.
Those risks are real.
But there is another possibility.
What if the bigger risk is building too slowly?
That is the scenario SemiAnalysis appears to be taking seriously.
AI models are increasingly moving from occasional tools to permanent digital workers.
Coding agents can operate for hours.
Research agents can search, analyze and synthesize continuously.
Enterprise AI can process documents, calls, emails and transactions around the clock.
Robotics adds another layer.
Autonomous systems do not just need training compute.
They need continuous inference.
Every successful AI product creates more computational demand.
Better models often generate more usage rather than less.
That phenomenon is important.
Historically, technological efficiency does not always reduce total resource consumption.
Sometimes lower costs create dramatically more demand.
Computing became cheaper, and the world consumed vastly more computing.
Bandwidth became cheaper, and internet traffic exploded.
Storage became cheaper, and companies began retaining enormous datasets.
If AI inference becomes cheaper, humanity may not simply spend less on AI.
It may deploy AI everywhere.
Then 10GW stops sounding like an endpoint.
It starts sounding like an early phase.
This is what Musk appears to be betting on.
And SpaceX may possess advantages traditional data-center developers do not.
The company is famous for compressing industrial timelines.
It designs hardware internally.
It builds manufacturing systems aggressively.
It tolerates iteration and failure rates that would be uncomfortable inside traditional infrastructure companies.
It often treats regulatory and engineering constraints as problems to be attacked in parallel rather than sequentially.
SemiAnalysis believes those traits could help SpaceX bypass some of the bottlenecks that normally slow data-center construction.
That does not mean SpaceX can ignore physics.
A transformer still has to be manufactured.
Electricity still has to come from somewhere.
Cooling still has to work.
Networks still have latency.
But organizational speed matters.
A company capable of building infrastructure six months earlier than a competitor can capture six months of revenue when that capacity is desperately needed.
At tens of billions of dollars per gigawatt, six months is no longer a scheduling detail.
It can be an economic weapon.
This is where Microsoft becomes the clue.
The software giant may be one of the clearest examples of what AI companies are facing.
Microsoft has money.
It has customers.
It has AI models.
It has cloud infrastructure.
It has technical talent.
What it cannot manufacture instantly is time.
If it needs additional capacity this year or next year, a data center completing three years from now does not solve the problem.
SpaceX’s proposition could effectively be:
We will absorb the construction risk.
You buy the compute when it comes online.
For Microsoft, that could be attractive.
For SpaceX, a major off-take agreement could transform the financing logic of the entire expansion.
Infrastructure becomes much easier to fund when future capacity is already contracted to investment-grade customers.
Banks understand contracts.
Investors understand recurring revenue.
Suppliers understand committed volumes.
A data center with no customer is speculation.
A data center with a hyperscaler committed to buying the output starts to resemble contracted infrastructure.
And this is where Musk’s impossible-looking 10GW target begins to look more strategic.
He may not need to finance all of it purely on faith in Grok.
He may be able to pre-sell significant portions of the capacity.
That cash flow could help finance further expansion.
More capacity creates more contracts.
More contracts improve financing.
Improved financing funds more capacity.
A flywheel begins.
And suddenly SpaceX’s role changes.
The company most people associate with rockets could become one of the largest AI infrastructure suppliers on Earth.
That possibility also explains why investors are paying such close attention to Musk’s comments about capacity.
When CEOs announce AI plans, they often speak in vague terms.
More GPUs.
More data centers.
Higher capital spending.
Musk gave a physical target.
Gigawatts.
That forces analysts to calculate what the statement actually means.
How many chips?
How much power?
How much capital?
How much revenue?
And who would buy it?
When SemiAnalysis worked through those questions, Microsoft emerged as a possible answer to the last one.
That is the moment the narrative changes.
At first, Musk’s target looked unbelievable because the cost looked unbelievable.
Then analysts compared the cost with the economic value of scarce AI compute.
A $500 billion investment is enormous.
But 10GW priced at approximately $50 billion per gigawatt per year implies a theoretical revenue opportunity approaching $500 billion annually if utilization and pricing hold.
That does not mean SpaceX will generate $500 billion a year from compute.
It does not mean 10GW will arrive on time.
And it certainly does not mean current scarcity pricing will last indefinitely.
But it demonstrates why sophisticated analysts are unwilling to dismiss the plan merely because the capital requirement sounds absurd.
The potential revenue sounds absurd too.
That is the defining feature of the current AI infrastructure race.
Everything is happening at numbers that would have looked ridiculous a few years ago.
Hundreds of billions in capital spending.
Gigawatts of power.
Millions of GPUs.
Data centers with electricity consumption comparable to cities.
Companies discussing infrastructure commitments on a scale once associated with national governments.
SpaceX is entering that race with one advantage Musk has spent decades building:
a corporate culture designed around attempting things that initially look uneconomic.
Sometimes those attempts fail.
Sometimes deadlines slip dramatically.
Sometimes the promises prove far more difficult than the announcement suggested.
But occasionally the infrastructure works.
And once it works, the market reorganizes around it.
Very few people initially believed reusable orbital rockets could become routine.
Then Falcon boosters started landing.
Few expected one company to deploy thousands of broadband satellites.
Then Starlink became one of the largest satellite networks ever built.
Now Musk is asking investors to believe SpaceX can build AI compute infrastructure measured in double-digit gigawatts.
The smart response is not blind belief.
It is not automatic dismissal either.
It is to ask whether the economics exist to justify the attempt.
SemiAnalysis believes they might.
And Microsoft may be the clearest reason why.
Because the hidden story of the AI boom is that even the biggest technology companies cannot build physical infrastructure as quickly as software demand is growing.
They can write checks.
They can place GPU orders.
They can sign power contracts.
But they cannot instantly create electricity, substations and cooling capacity.
That creates an opening for anyone capable of delivering compute sooner.
Musk appears to think SpaceX can be that supplier.
And if Microsoft really does need enormous amounts of near-term capacity to support OpenAI workloads and its broader AI ecosystem, the strangest partnership in technology may suddenly make economic sense.
Microsoft—the company that helped fuel OpenAI’s rise.
Buying billions of dollars of AI compute.
From SpaceX.
Owned and controlled by Elon Musk.
Whose xAI is simultaneously competing in the same artificial-intelligence race.
Competitors on the model layer could become partners on the infrastructure layer.
That is how severe the compute shortage may be.
And that may be the biggest sign that Musk’s plan deserves attention.
Not because Elon Musk said 10GW.
Not because analysts assume every Musk target will be reached.
But because when they modeled what that capacity could be worth, they found customers wealthy enough—and possibly desperate enough—to pay for it.
SpaceX spent two decades proving that scarce access to orbit could become an enormous business.
Now Musk may be trying to prove the same thing about scarce access to intelligence.
And if Microsoft eventually becomes one of the companies paying SpaceX for that access, the moment everyone laughed at a 10GW target may look very different in hindsight.
The most dangerous assumption in the AI race may no longer be believing Elon Musk can build too much compute—it may be assuming the world won’t find a way to use all of it.
