Software copies in a weekend; a launch pad and 13 GW cannot

Pay attention everyone. This is not a joke.

This is an existential crisis for software. Most of the Mag7 is inside it.

Today showed what it looks like. Anthropic launched Claude Sheets, Whiteboards and Meetings, and Microsoft dropped as much as 3%. Adobe, Atlassian and HubSpot apps now open inside Claude. OpenAI gave businesses their own agents inside ChatGPT.

Software was a moat for 30 years because code was hard to write. Now code is the cheap part. My take: an office suite is a weekend project for any lab with enough compute. Anything you can describe in words, AI can now build.

So where does the moat go? Down the stack, to whatever you can't prompt into existence.

First stop is compute. The best models can't run without it, and Anthropic's own S-1 shows who it depends on: at least $121.3B with Google, $110B with Amazon, $44.9B with Microsoft and up to $120.5B with SpaceXAI. The company disrupting everyone's software is a tenant. Microsoft is both its landlord and today's victim.

After compute comes power, land, chips and launch. Amazon, Alphabet, Microsoft and Meta spent $201.9B in one quarter on buildings and equipment. They are not spending that on code.

My take: that splits the Mag7 in two. Amazon and Microsoft carry the most software risk. Meta has users and distribution, which no lab can copy in a weekend.

Apple is the interesting case. I don't think it gets disrupted, because Apple is a hardware company at heart: devices, chips, a supply chain nobody else has. But it has the devices and not the compute. My guess: that is why Apple should work with SpaceX. Both are hardware companies, just different kinds, so they complement each other instead of competing. If Apple doesn't plug into someone who owns compute, the AI labs come for its software layer next.

SpaceX is different. It launches its own rockets, 35 in one quarter. It has 11,100 satellites in orbit, its own spectrum, its own power and 13 GW of compute targeted by the end of 2027. You can clone an app in a weekend. You can't clone a launch pad. My guess: even a decade of head start would not close that gap.

It still takes part in the software shift, from the safe side. Grok Bot is a harness. Elon said it himself: "Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs." Models are swappable, so no model owns the user for long. Grok 5 is its own shot at the top.

In 2027 compute starts moving to orbit. Elon's own guess for the first StarMind computers is Q3 2027. No permit fight, no grid queue.

Software got cheap. Atoms did not. Whoever scales compute fastest decides which labs live, and the company that owns the atoms under all of it does not have to pick a winner.

$SPCX