SpaceX's AI Compute Business Is Real. The Moat Is the Question.

SpaceX CFO Bret Johnsen said something that sounded almost too good to be true on the company's first earnings call as a public company: new compute capital is "behaving more like cost of goods than capex". His reasoning was that the payback period on AI infrastructure investments is less than one year. That is the kind of statement that makes a rocket company sound more like a software business — and it's the reason the stock has rallied over 40% from its late-July lows, with Argus Research upgrading it to Buy on Friday with a $160 price target.

But the question isn't whether SpaceX has found a profitable new revenue stream. The question is whether a compute-renting business built on 90-day cancellation clauses, funded by a sixfold jump in quarterly capex, can justify a valuation that briefly topped $2 trillion and is still near $127 per share.

I need to look at the product architecture, the supply chain signal, and the competitive positioning before I decide what kind of allocation this deserves.

The company that isn't what it used to be

SpaceX reported $7.8 billion in Q2 2026 revenue, a 92% year-over-year increase that beat the $6.93 billion consensus. The company IPO'd at $135 per share on June 12 — the largest IPO in history at $85.7 billion raised — and the market initially valued it at over $2 trillion. Since then, the stock has been a rollercoaster, dropping as much as 37% in July before rebounding sharply on the earnings beat and the Argus upgrade.

The revenue mix tells the real story of what's changed. Starlink contributed $1.7 billion in incremental growth, pushing its total and driving 66% revenue growth in the Connectivity segment. But nearly $2 billion of the growth came from the AI division, which more than tripled year-over-year to $2.6 billion in Q2. This AI segment is still loss-making on an operating basis — it ran approximately $6.4 billion in operating losses last year — but the trajectory is the part that matters for valuation.

SpaceX has three reporting segments: Space (launches), Connectivity (Starlink), and AI. Last year, revenue broke down as $11.4 billion from Starlink, $4.1 billion from launches, and $3.2 billion from AI. This quarter, the AI piece grew from a rounding error to a material third of the business in one year.

What this means is that SpaceX is no longer primarily a rocket company in the eyes of the market. It's an AI infrastructure play. And that changes how you evaluate the capital intensity, the competitive moat, and the durability of the growth.

The architecture of the AI compute business

Here's how the AI compute business actually works. SpaceX built massive data center complexes — Colossus and Colossus II near Memphis, now roughly 2 million square feet and providing approximately 1 gigawatt of compute power with capacity for 1 million Nvidia GPUs. The initial cluster went online in 122 days, built by converting an existing factory. That speed of construction is the engineering advantage: SpaceX engineers are used to solving problems that literally involve rocket science, so building data centers was comparatively straightforward.

Then they started renting the unused compute capacity to AI companies. The two deals that anchor the business model are striking in scale:

• Anthropic: ~325,000 Nvidia GPUs for $1.25 billion per month
• Google: ~110,000 Nvidia GPUs for $920 million per month

Those two contracts are projected to generate approximately $26 billion annually — more than SpaceX's total revenue from all segments in the prior year.

And since the end of Q2, the company signed an additional $6.7 billion in cloud computing contracts in just a few weeks, with executives projecting year-end compute capacity exceeding two gigawatts.

The architecture here is straightforward: SpaceX buys Nvidia GPUs, builds the data center infrastructure, and leases compute capacity at a premium during a period of extreme demand. It's essentially an infrastructure-as-a-service business, sitting between chip manufacturers and AI model companies.

This is what separates the SpaceX AI thesis from a typical cloud story. SpaceX isn't building a software platform or a developer ecosystem. It's providing raw compute — the same way a landlord provides square footage. The question isn't whether there's demand. It's whether the margin on that arrangement is durable once competitors figure out how to replicate the model.

The supply chain signal: $18.4 billion in one quarter

This is where the picture gets more complicated. Q2 capital expenditures jumped sixfold to $18.4 billion, of which $15.8 billion went to AI infrastructure. That represents roughly one-fifth of the $85.7 billion raised in the June IPO. And CFO Bret Johnsen signaled that AI spending would remain elevated, with capex likely staying at Q2 levels for the next two quarters.

To put that in perspective: $18.4 billion in one quarter is $73.6 billion annualized. That would make SpaceX one of the highest-capex companies in the world, rivaling the combined infrastructure spending of the three major cloud providers. Against Q2 revenue of $7.8 billion, the capex-to-revenue ratio is over 200%. The company is deeply free cash flow negative.

Johnsen's defense is the payback claim. He said new compute capital has less than one-year payback, which is the basis for his "cost of goods rather than capex" framing. If compute infrastructure earns back its cost in under a year, then the economics do look more like a high-margin service business than a capital-intensive infrastructure play.

But supply commitments surging this aggressively is a dual signal. It confirms demand strength — you don't spend $15.8 billion on a whim — but it also creates leverage risk. If contract cancellations hit, if GPU pricing shifts, or if competition drives down compute rental rates, the company has very little flexibility to unwind this spending cycle.

This is what I check first when I evaluate a company's forward trajectory. Capex of this scale relative to revenue tells me the market is being asked to trust a pipeline that doesn't yet show sustainable unit economics. The revenue growth is real. The question is whether the cost structure to generate that revenue can normalize.

The competitive moat that might not be a moat

Here's the part of the architecture that changes the risk calculus. Both the Anthropic and Google contracts include a 90-day cancellation provision. That means either party can walk away with three months' notice. If cheaper alternatives emerge, customers can cancel. If SpaceX needs the compute back for internal Grok or Starlink use, they can reclaim it.

A 90-day cancellation window on multi-billion-dollar contracts is not a moat. It's an arbitrage.

Sridhar Tayur, a professor at Carnegie Mellon, put it bluntly: the question is whether this infrastructure-as-a-service model will become a permanent main line of business or remain a one-off opportunity. And there's reason to think it could be the latter. Meta is reportedly negotiating similar compute leasing deals with Anthropic, potentially worth up to $10 billion over two years. If Meta, Google, and the other hyperscalers decide to replicate SpaceX's model at scale — and they have the engineering talent and capital to do exactly that — the premium SpaceX charges for compute access could compress.

The current demand outstrips supply. That's what Johnsen said on the call, and it's true right now. But supply chain dynamics change. TSMC is expanding GPU production capacity. Nvidia's newer architectures are coming online. And the hyperscalers that SpaceX is selling to today have no incentive to remain customers forever if they can build their own infrastructure at a comparable cost.

The AI compute business is real. It's generating real revenue with real contracts. But the competitive dynamics look more like a window of opportunity than a structural advantage. And windows close.

What the Argus upgrade gets right and wrong

Argus Research upgraded SpaceX to Buy from Hold on Friday, setting a $160 price target based on 20x estimated 2027 revenue. The analyst, Steven Silver, cited "rapid payback" on AI infrastructure investments and a "robust growth outlook". Argus also noted that the forward multiple has compressed significantly and faster than expected, creating a cheaper entry point.

I agree with Argus on the payback dynamic. If Johnsen's less-than-one-year claim holds, the economics of this business are genuinely attractive on a per-dollar-of-investment basis. And the multiple compression is real — the stock is roughly 50% below its June 15 record high of $225.64, down from a brief peak that briefly made SpaceX the most valuable company on the Nasdaq.

But I have two reservations about the upgrade's framing.

First, Argus explicitly cites Elon Musk's track record at Tesla — noting that a $10,000 investment in Tesla in 2010 is worth $2.5 million today — as part of the basis for the Buy rating. I don't deal with opinions, and I certainly don't deal with founder halo effects as a substitute for current unit economics. Musk's past execution is impressive, but past performance on one company doesn't guarantee the same return profile on another, especially one with a different capital structure and a different competitive landscape.

Second, the 20x 2027 revenue multiple assumes that the $26 billion in annualized contract revenue from Anthropic and Google persists and grows — without accounting for the 90-day cancellation risk, the competitive threat from hyperscaler self-builds, or the sustainability of capex at current levels. Argus projects $110 billion in 2027 revenue. That's a massive number for a company that did $18.7 billion total revenue last year. It requires near-flawless execution on AI infrastructure deployment, zero major contract cancellations, continued Starlink growth, and launch business stability.

The upgrade is directionally reasonable. The stock has pulled back enough from its highs that the entry point is more defensible. But the price target assumes a best-case scenario on the AI compute business that doesn't stress-test the competitive risks.

What would change my view

I believe SpaceX has legitimately diversified beyond rockets into a meaningful AI infrastructure business. The Q2 results prove that the compute-renting model generates substantial revenue, the new $6.7 billion in bookings since quarter-end shows demand isn't cooling, and the engineering advantage in build speed is real.

But here's what would change my thesis, and it's worth being explicit about:

• If Anthropic or Google exercise the 90-day cancellation clause on their contracts, or if new deals are priced significantly lower, the premium-compute narrative breaks.
• If capex remains at $18.4 billion per quarter through 2027 without revenue growing commensurately, the free cash flow negative position becomes unsustainable even for a company that raised $85.7 billion.
• If Meta, Google, or other hyperscalers build their own compute capacity at comparable cost and reduce their reliance on third-party compute leasing, the addressable market for SpaceX's AI business shrinks.

Conversely, what would strengthen the case: evidence that contracts are being signed on longer terms beyond the 90-day cancellation window, proof that Starlink's revenue growth can sustain its trajectory independently, and capex normalization as a percentage of revenue even as absolute spending grows.

Where the capital goes

The debate is not whether SpaceX has built something important. It has. The debate is whether the return profile at current levels — with the stock trading around $127 after bouncing from its lows, still down roughly 44% from its June highs — justifies the allocation.

I believe the long-term thesis for SpaceX as an AI infrastructure company has legs, but much of the durable return is likely back-half weighted. The current setup is a high-conviction, high-risk play on execution. The revenue growth is real, the payback economics are compelling if Johnsen is right, and the multiple has compressed enough from its peak to remove some of the obvious valuation risk.

But the capex signal — sixfold in one quarter, with no guarantee of normalization — and the competitive architecture — 90-day cancellations on the anchor contracts — tell me this isn't a position that deserves a large allocation. It's the kind of opportunity that rewards smaller sizing and patience. The risk/reward improves as the AI compute business proves its durability, as contracts show longer-term commitment, and as the capex-to-revenue ratio comes down.

For me, the question is always opportunity cost. The AI trade is crowded. Nvidia's valuation has its own challenges. Hyperscaler capex is surging. And SpaceX sits in the middle of all of it — a company that's betting its IPO proceeds on becoming the compute landlord for the AI industry, at a time when the landlords might decide to build their own properties.

I'm not saying sell. But I am saying that $18.4 billion in quarterly capex, 90-day cancellation clauses, and a 20x 2027 revenue multiple ask for a lot of faith in execution. Faith has its place in investing, but it shouldn't replace a manageable position size.