The Physics Problem in SpaceX's $1.78 Trillion Orbit

The number that kept rattling around my head this week wasn't $1.78 trillion. It was 3.3 milliseconds.

That's how long it takes light to travel from a satellite in low Earth orbit to the ground, round trip. Speed of light, 300 kilometers per millisecond. LEO altitude around 500 kilometers. The math is simple and it doesn't care about your press release.

SpaceX is a week away from the largest IPO in history. The company filed confidentially, then confirmed plans for a June 2026 listing that would value the combined SpaceX-xAI entity at $1.78 trillion, with the aim of raising $75 billion from public markets. The pitch is audacious: orbital AI data centers, space-based compute clusters, a trillion-parameter Grok running on satellites. Elon Musk called it a step toward a "Kardashev II-level civilization" in a recent investor call.

I don't doubt the ambition. I doubt the physics.

The 3.3ms Problem That Marketing Can't Fix

A user in New York sends a query to a model running on a satellite passing overhead. The signal travels up at the speed of light. The satellite processes it. The response travels back down. Best case, you're looking at 10 to 20 milliseconds of total latency, assuming no routing delays, no queueing, no atmospheric interference.

That's fine for batch training. It's fine for asynchronous model updates. It's fine for workloads that don't care about real-time response.

It's terrible for anything interactive.

Every chatbot, every voice assistant, every code completion tool, every autonomous driving system, every financial trading model, every real-time recommendation engine. These systems need sub-100ms response times. Preferably sub-50ms. The extra 10 to 20 milliseconds from space isn't a dealbreaker on its own. But it's additive. Stack it on top of application latency, network latency, database latency. You chew through your budget fast.

And that's just inference. Training is worse.

Training large models requires constant gradient synchronization across thousands of accelerators. Every training step involves all-to-all communication. The bandwidth requirements are massive and the latency sensitivity is extreme. Move those accelerators into orbit and your synchronization delays go from microseconds to milliseconds. Distributed training across a satellite constellation doesn't accelerate your training loop. It cripples it.

Here's what orbital compute is actually good for:

- Satellite imagery analysis. Process the image where it's captured, send down only the relevant pixels. Reduces data volume by 85 to 95 percent. - Sensor fusion for Earth observation. Thermal, multispectral, radar. Aggregate in orbit, transmit summaries. - Edge inference on cached models. If the model doesn't need real-time updates and the data source is already in space, it makes sense. - Batch scientific computing. Climate modeling, protein folding, radio astronomy data reduction. Things where latency doesn't matter.

That list is real. It has economic value. But it's not "the future of AI." It's a niche.

The $6.4 Billion Hole in the Story

Let's talk about the numbers they don't want you to stare at too long.

xAI lost $6.4 billion in 2025 on just $3.2 billion in revenue, according to SpaceX's own IPO filings. That's a negative 200 percent margin. The company is burning roughly $1 billion per month. Most of that revenue came from X, the former Twitter platform Musk acquired and then sold to his AI company. It's not exactly third-party validation.

The merger in February 2026 valued SpaceX at $1 trillion and xAI at $250 billion, generating a combined enterprise value of $1.25 trillion before the public offering. The IPO seeks to nearly double that to $1.78 trillion. The delta is entirely narrative. There's no revenue growth in between. No product launch. No new customer. Just the promise that putting AI computers in space will unlock something magical.

Analysts are already circling. SpaceNews reports the company plans to use the capital for Starship development, Starlink expansion, and "growth of its artificial intelligence activities obtained through the acquisition of xAI." The prospectus describes plans to scale Grok to "multiple trillions of parameters."

Training a trillion-parameter model in orbit. Let's think about that.

A single H100 GPU draws 700 watts under load. A cluster of 100,000 GPUs draws 70 megawatts. You're not putting 70 megawatts of compute on a satellite. You're not even putting one megawatt. The power available on a typical LEO satellite is a few kilowatts. Solar panels have limits. Radiators have limits. You're not getting around the second law of thermodynamics with better engineering.

Power, Heat, and the Things You Can't Engineer Away

This is where the systems view matters.

A data center on Earth has virtually unlimited power. It has cooling towers. It has backup generators. It has fiber optic connections with microsecond latency. It has technicians who can swap a failed drive in fifteen minutes.

A data center in space has none of those things.

Power is solar. Period. You can deploy larger arrays. You can use more efficient cells. But you're still bounded by the inverse square law and the surface area of your spacecraft. The International Space Station generates about 120 kilowatts from its massive solar array wings. That's enough for a small office building. It's not enough for a single rack of GPUs.

Thermal management in vacuum is a nightmare. On Earth, you dump heat into water or air. In space, you radiate it. Heat rejection scales with surface area and temperature difference. The hotter your radiators, the more heat you reject. But your electronics have temperature limits. You can't just run them at 200 degrees Celsius. So you need bigger radiators. Which means more mass. Which means more launches. Which means more cost.

Launch cost per kilogram to LEO with Starship is projected at roughly $100 per kilogram. That's cheap by historical standards. But a fully loaded data center module with GPUs, networking gear, power systems, and radiators is heavy. A single rack of AI infrastructure on Earth weighs maybe a ton. Getting it to orbit costs $100,000 in launch fees. That's manageable. But you need thousands of racks. And you need to replace them every three to five years, because radiation degrades electronics. And you need to refuel the station-keeping thrusters. And you need to dodge debris, which doubles in probability every five years in LEO.

The economic case gets worse the longer you stare at it.

What the Market Is Actually Pricing

None of this means the IPO will fail. Markets price narratives, not physics. The IPO could soar. Retail investors might chase it. Institutions might allocate for fear of missing the next NVIDIA-level compounder. That's how markets work.

But I want to be clear about what this valuation is betting on.

At $1.78 trillion, SpaceX would be worth more than Meta. More than Tesla. More than every aerospace and defense company combined. It's pricing in a future where orbital data centers become the dominant compute paradigm for AI, where Starlink captures a significant share of global internet traffic, where Starship becomes the dominant launch vehicle for the entire industry, and where xAI's Grok competes head-to-head with OpenAI and Google in the foundation model race.

That's four separate massive bets, each with its own failure modes.

If orbital compute remains niche, the valuation breaks. If Starship hits development delays, the launch cost advantage evaporates. If Starlink faces competition from Amazon's Project Kuiper or terrestrial fiber densification, the revenue base shrinks. If Grok can't catch up to GPT-6 or Gemini, the AI premium disappears.

Philip Johnston, CEO of Starcloud, appeared on Bloomberg Technology this week to discuss these exact challenges. He emphasized that building compute infrastructure in space requires solving problems that don't exist on Earth. Radiation hardening. Thermal cycling. Debris collision avoidance. Autonomous fault recovery. Supply chains for orbital replacement units. The list goes on.

None of these are impossible. They're just expensive and slow. And the IPO valuation assumes they happen fast.

The Real Story Nobody's Telling

Here's what I think is actually happening.

SpaceX needs capital. The Starlink buildout is cash-intensive. Starship development is cash-intensive. The xAI merger added a company that's losing $6.4 billion a year and has no path to profitability without massive additional investment. The IPO isn't primarily about orbital data centers. It's about keeping the whole machine running.

Philip Johnston on Bloomberg: "The challenges of building and maintaining orbital data centers are significant. You're dealing with a harsh environment, limited power, and latency constraints that make most AI workloads impractical in space. The economics work for specific use cases, but not for general-purpose AI compute."

That's the CEO of a company trying to build orbital data centers. If he's tempering expectations, you should listen.

I wrote about this dynamic last month in the context of the broader AI infrastructure spending wave. The pattern repeats: massive capital deployment chasing a narrative that exceeds the underlying technical reality. Data centers on Earth are already constrained by power grid capacity. The solution isn't to move compute into a harsher environment. It's to build more efficient models and more efficient hardware.

What Developers Should Actually Pay Attention To

Let me bring this down to something useful.

If you're building AI applications today, here's what the SpaceX IPO means for your roadmap. Nothing. At least for the next five years. Practical orbital compute is for Earth observation, remote sensing, and batch processing of data that originates in space. Your chatbot, your code assistant, your recommendation engine, your fraud detection pipeline. They run on terrestrial infrastructure. Latency demands it.

The technology to watch isn't orbital data centers. It's model compression. It's on-device inference. It's efficient architectures that can run on edge hardware with limited power. Those are the real constraints that matter. A model that runs on a phone GPU is more consequential for the industry than a model that runs on a satellite.

Space-based compute will have its day. But it'll serve specific workloads, not general AI. The hype cycle is running ahead of the physics cycle. It always does.

The IPO opens next week. The price is reportedly set at $135 per share. At that valuation, SpaceX would raise $75 billion and become the most valuable public company on Earth while losing money on its core AI business and betting on an infrastructure model that violates basic constraints of latency and power.

Maybe I'm wrong. Maybe they've solved thermal management with some breakthrough radiator design. Maybe laser inter-satellite links reduce latency enough. Maybe the power budget scales better than I think.

But I've learned to trust the constraints.

3.3 milliseconds. 120 kilowatts per station. $6.4 billion in annual losses. Those aren't problems you IPO your way out of. They're problems you engineer your way out of, slowly, over years, with hundreds of iterative launches and thousands of engineering hours.

Which is fine. That's how space exploration works. It's just not how a $1.78 trillion valuation works.