SpaceX's $100 Billion Target Hides the Real Story: A Memphis Factory

SpaceX went public this June as the world's most expensive IPO — $75 billion raised at a $1.75-to-$2 trillion valuation — and has already shed roughly $500 billion from its peak. Then on its first earnings call as a public company, CEO Elon Musk announced the company plans to hit $100 billion in annualized revenue run rate by year-end.

That number is the story everyone is quoting. It's also the wrong number to focus on.

$100 billion in annualized run rate is a velocity target, not an earnings one. It measures how much revenue would flow in if the current quarter's sales repeated twelve times. But velocity without cash flow is just direction. The real question for an investor holding this stock at a $1.95 trillion market cap is whether the plumbing can actually turn that velocity into money — or whether SpaceX will grow itself into a free-cash-flow black hole.

The answer starts with what's driving the growth.

It's not rockets. It's not Starlink. SpaceX's launch business generated $962 million in the second quarter. Starlink connectivity, the company's only historically profitable segment, brought in $4.3 billion. Both are strong. Neither is big enough to triple the company's revenue run rate in four months.

The growth engine is something SpaceX only built this year: an artificial intelligence compute leasing business.

The Company That Builds Data Centers in Four Months

In the summer of 2024, a company called xAI took over a shuttered Electrolux factory in southwest Memphis, Tennessee, and in 122 days filled it with over 200,000 Nvidia GPUs. Nvidia CEO Jensen Huang called it "superhuman." The result was Colossus 1 — one of the largest operational AI training clusters in the world, housing 300 megawatts of compute capacity.

Then xAI merged into SpaceX in February 2026 for $125 billion. Colossus became a SpaceX asset. And what happened next is the critical turn in this whole story.

The AI division found it couldn't use all that capacity running its own Grok model. Grok had roughly 117 million monthly users in March 2026 — impressive, but a fraction of ChatGPT's roughly 1 billion. Musk admitted the system was "being rebuilt from the foundations up."

So instead of letting the cluster sit idle, SpaceX became a landlord.

Anthropic — the Claude company — signed a deal to lease essentially all of Colossus 1 at $1.25 billion per month through 2029. Google agreed to pay $920 million per month for access to 110,000 GPUs starting in October. Reflection AI will pay $150 million monthly for GB300 access through 2029. Together, these three deals commit to a forward run-rate of roughly $26 billion per year.

For context: SpaceX's total revenue across all business lines in all of 2025 was $18.7 billion.

A company that made $18.7 billion last year has now contracted to lease out AI compute at an annualized rate of $26 billion. The Anthropic and Google deals alone exceed the company's entire previous revenue base.

This is the plumbing behind the $100 billion headline. Not rockets. Not satellites. A former appliance factory in Memphis, Tennessee.

What Makes the Compute Scarce

The compute itself isn't the scarce part. Nvidia ships GB300 chips to multiple customers. The scarcity is build speed — and more specifically, power.

Traditional hyperscalers like Amazon, Google, and Microsoft face a fundamental bottleneck: grid connections take five to ten years to approve, permit, and build. The chips can arrive in a year, but the electricity might not be there for a decade. That delay is the single largest constraint on AI infrastructure scaling worldwide.

SpaceX solved it the way it solves everything else: by building its own power.

Colossus 2 and 3 don't connect to the grid at all. They're powered entirely by the Southaven Gas Plant — a 1.2-gigawatt facility SpaceX built across the state line in Mississippi, connected via private transmission infrastructure. By owning both the generation and the load, SpaceX sidestepped the Federal Energy Regulatory Commission's open-access interconnection queue. The strategy is legally creative; the U.S. Department of Justice has already intervened in related litigation, arguing that disrupting the operation would pose "national security risks to classified workloads."

Colossus 1 does connect to the Tennessee Valley Authority grid, but SpaceX hedges that tie with on-site gas turbines and over 240 Tesla Megapack batteries. When the grid demands SpaceX cut 50-80% of its draw — which happens four to five times a year, for two to four hours each — the batteries ride through the event.

The result is a company that can deploy gigawatt-scale, GPU-packed data centers in roughly four months while traditional cloud providers wait years for power. Being able to build the factory floor while your competitors are still applying for permits creates scarcity rent. The question is just how much of it SpaceX actually keeps.

The Cash Burn Behind the Revenue Machine

Here's where the numbers demand attention.

SpaceX spent $18.4 billion on capital expenditures in the second quarter of 2026 alone. Roughly $15.8 billion of that was for AI infrastructure. Management indicated Q3 and Q4 spending may remain at similar levels, implying full-year capex near $65 billion — well above Wall Street's $50 billion estimate. Over the trailing twelve months, capex totals $43.3 billion against operating cash flow of $10 billion, producing $33.4 billion in negative free cash flow.

The AI division — the very segment generating that $26 billion in contracted run-rate — posted a $2.5 billion operating loss in just the first quarter. In fiscal 2025, the AI segment generated $3.2 billion in revenue against $12.7 billion in capex. The gap is enormous.

SpaceX's Starlink business, which produced $4.4 billion in operating profit in 2025, is the cash engine subsidizing this gap. Without Starlink, the AI division would be an unprofitable subsidiary burning investor capital. With it, the whole machine can keep spinning — for now.

The free cash flow margin is -145%. The return on invested capital is -5.2%. The stock trades at 85 times trailing sales.

Compare that to the companies SpaceX is leasing to: Nvidia trades at 27 times earnings and 17 times sales. Alphabet trades at 17 times earnings and 9 times sales. Amazon trades at 20 times earnings and 3.5 times sales. All three are profitable. All three generate positive free cash flow. SpaceX's valuation assumes the company will not just hit its revenue targets but will also close a $40+ billion annual free cash flow gap — and do it fast enough to justify a multiple that exceeds every major tech company on the planet.

The Leases Aren't Long-Term Leases

There's one more detail about those $26 billion in contracts that matters enormously.

Every single one of them includes a 90-day termination clause.

After an initial three-month period, Anthropic can walk away from its $1.25 billion monthly deal with just 90 days' notice. Google can terminate after December 31, 2026 on 90 days' notice — and can walk away immediately if SpaceX fails to deliver the committed GPUs by September 30. Reflection AI has the same 90-day exit right.

This means the $26 billion annualized run-rate is not locked-in revenue. It's a commitment that can evaporate in a quarter if tenants find better pricing, if the hardware underperforms, or if the AI investment cycle softens. In infrastructure leasing terms, these are short-term commitments at premium prices — attractive when demand is desperate, fragile when it isn't.

The contracts also reveal an uncomfortable strategic reality: SpaceX is renting more than half its compute cluster to direct competitors. Anthropic builds AI models that compete with Grok. Google's Gemini competes with Grok. SpaceX is winning the infrastructure game while losing the model race, and it's using the infrastructure profits — when they arrive — to keep funding the race it's losing.

The Clock on the Advantage

So what does an investor actually carry from this?

SpaceX has built something real. The ability to deploy gigawatt-scale AI compute clusters in roughly four months, with self-owned power infrastructure that sidesteps years of regulatory delay, is a genuine bottleneck. The scarcity is real right now. Anthropic's revenue grew from $1 billion to $35 billion in 14 months; neither AWS nor Google Cloud could provision capacity fast enough. They needed Colossus. They paid for it.

But the $100 billion run-rate target is a revenue number, not a profit number. The AI division is losing $6.36 billion annually. The company spent $18.4 billion in one quarter building infrastructure it's hoping to lease back. And those leases can be terminated on 90 days' notice.

The valuation clock runs both ways. At $1.95 trillion, the market is pricing in years of flawless execution: Colossus ramps on schedule, Anthropic and Google stay as tenants, capex efficiency improves, and the AI division crosses from a $2.5 billion quarterly loss into sustained profitability — all while Starlink continues generating the cash to fund the gap.

The capacity cure is already visible. Amazon's Project Kuiper is deploying. Blue Origin's New Glenn begins commercial operations in late 2027. Microsoft and Google are building their own gigawatt-scale clusters with co-located power generation. The 122-day build speed is an advantage only until someone else learns to do it in 180 — or 90. And every quarter that a tenant proves it can build or lease its own capacity, SpaceX's pricing power weakens.

The confirming metric is the AI division's operating margin — the one number that tells you whether the leasing model actually generates profit after all the capex, depreciation, and power costs are accounted for. The normalization signal is a major tenant announcing its own independent compute infrastructure, signaling it no longer needs to rent.

SpaceX is not a rocket company anymore. It's not a satellite internet company anymore. It's an infrastructure landlord in a race where the building speed advantage is real today, the tenants are paying premium prices today, and the clock on how long that lasts is already running.