The compute race just left the planet, and it is no longer a thought experiment.

SpaceX has filed with the FCC to deploy up to one million orbital data center satellites, describing a constellation with unprecedented computing capacity built to run advanced AI models. Blue Origin filed in March for 51,600 of its own. Google is launching test satellites carrying its TPU chips next year. An Nvidia-backed startup already has an H100 running a language model in orbit right now.

And three days from now, on August 7, xAI ships Grok 4.6, from a company that is now, structurally, a rocket company.

The reason all of this is happening at once is not that space is cool. It is that AI has an energy problem on Earth that nobody has solved. Here is the real picture, including the part the hype skips.

Starting with the filing, because it is the part that turned this from a Musk tweet into a regulatory record. SpaceX asked the FCC for permission to operate up to one million satellites between 500km and 2,000km altitude, arranged in clusters roughly 50km apart to serve different latency profiles. The stated purpose is powering advanced AI models and the applications that depend on them.

The financing question got answered on June 12, when SpaceX went public and raised $85.7 billion, valuing the company in the trillions and making Musk the world's first trillionaire. That money is explicitly earmarked in part for orbital compute. Musk's own compensation package is tied to targets including data centers in space drawing 100 terawatts of power, which is more than a thousand times what every data center on Earth consumes today.

The physics is the actual argument. In orbit you get uninterrupted solar power with no atmosphere in the way and no night, and you get passive cooling by radiating heat into vacuum. Power and cooling are the two constraints strangling terrestrial AI buildouts right now, and both get cheaper in space. That is why this is not only Musk. Blue Origin filed in March for 51,600 data center satellites under Project Sunrise, targeting deployment in late 2027. Google is running Project Suncatcher with Planet Labs, launching two satellites carrying its TPU chips in 2027, with SpaceX as a likely launch partner despite Google owning 6.1 percent of it. Nvidia-backed StarCloud already put a satellite up carrying an H100 that is currently running Google's open Gemma model.

Now the part that most coverage skips, and the reason your timeline should be measured in years rather than months.

There is no operational orbital data center. There will not be one soon. Researchers at Northeastern who study satellite systems expect to see building blocks tested over the next couple of years, not a working facility. Google's own Project Suncatcher paper is candid about the economics: launch costs have historically made large space systems impractical, and the projection is that prices fall below $200 per kilogram by the mid-2030s, which is when the math starts working.

SpaceX's own messaging is more aggressive, putting cost-competitiveness with terrestrial data centers within two to three years. Take that as a target rather than a forecast, since it depends entirely on Starship achieving reliable reusability and a launch cadence nobody has demonstrated yet.

There are unsolved problems beyond cost, too. Servicing hardware you cannot physically reach, radiation damage to chips, orbital debris, and light pollution affecting astronomy are all real and largely unanswered. The honest read is that this is a serious, funded, regulator-engaged effort that will produce test payloads soon and useful capacity much later.

The same company betting on orbital compute ships a model in three days. Musk announced on July 28 that Grok 4.6 arrives around August 7, describing it as a 1.5-trillion-parameter model with significantly improved supervised fine-tuning and reinforcement learning. Grok 4.7, at 2.1 trillion parameters, follows a few weeks later and will reportedly be more capable but slower to serve.

There are no official specs, pricing, or benchmarks yet, and Musk's timelines slip regularly, so treat August 7 as a target. What makes it worth watching is the baseline. Grok 4.5, out since July 8, sits around 54 on the Artificial Analysis Intelligence Index at roughly fourth overall, scores about 83.3 percent on Terminal-Bench 2.1 and 64.7 percent on SWE-Bench Pro, and costs $2 and $6 per million tokens, well under most rival flagships. Vercel's CEO called it the best cybersecurity model relative to price in their tests, ten times cheaper than GPT-5.6 Sol and 5.7 times cheaper than Opus 5.

Arena.ai confirmed the August 7 date and will run 4.6 on its Agent Arena leaderboard, which scores models on live agentic sessions rather than static benchmarks. That is the number to watch, since Grok 4.5 debuted there at 13th, up sharply from 4.3 at 29th. If the trajectory holds, agentic capability will be 4.6's headline.

Connect the two stories and the strategy is coherent in a way it was not eighteen months ago. xAI merged into a launch company. That launch company went public and raised $85.7 billion. It filed to put a million compute satellites in orbit. And it is shipping frontier models on a monthly cadence, priced to undercut everyone, while it builds toward energy that does not depend on anyone's power grid.

For you, nothing about orbital compute changes what you build this year or next. What changes sooner is the cadence and the pricing on the ground. If Grok 4.6 lands Friday near 4.5's price with meaningfully better agentic performance, that is a real reason to re-run your evals, particularly if you are paying flagship rates for work a cheaper model now finishes. Watch the Agent Arena placement rather than xAI's own charts, check whether the $2 and $6 pricing holds, and as always, the only benchmark that settles anything is your own workload.

The stories moving fast beyond today's headlines:

SpaceX has filed with the FCC to deploy up to one million orbital data center satellites between 500km and 2,000km altitude, explicitly to run advanced AI models, funded in part by its June 12 IPO that raised $85.7 billion and made Elon Musk the world's first trillionaire. His compensation package includes a target of data centers in space drawing 100 terawatts, over a thousand times current global data center consumption. The physics is the argument: orbital solar delivers uninterrupted power with no atmosphere or night, and vacuum enables passive cooling, addressing the two constraints throttling terrestrial AI buildouts. It is a genuine race, with Blue Origin filing in March for 51,600 satellites under Project Sunrise targeting late 2027, Google running Project Suncatcher with Planet Labs to launch TPU-carrying test satellites in 2027, and Nvidia-backed StarCloud already operating a satellite running Gemma on an H100. But no operational orbital data center exists, researchers expect only building blocks tested over the next few years, and Google's own paper puts cost-competitiveness in the mid-2030s when launch costs fall below $200 per kilogram, against SpaceX's more aggressive two-to-three-year claim. Closer to home, xAI ships Grok 4.6 around August 7, a 1.5-trillion-parameter model with improved SFT and RL, following Grok 4.5 which sits around fourth on the Artificial Analysis index at $2 and $6 per million tokens, with Grok 4.7 at 2.1 trillion parameters a few weeks behind. Watch Arena's Agent leaderboard rather than vendor charts, and check whether the pricing holds.

Stay building. 🤖