The Deal That Came With a Loaded Gun
On Wednesday, Elon Musk did what three months ago seemed impossible: he cut a deal with Anthropic — the AI lab he’d called “evil,” “misanthropic,” and accused of hating Western civilisation. SpaceX would lease the entirety of its Colossus 1 data centre in Memphis to Anthropic — 300 megawatts of compute, 220,000 NVIDIA GPUs, worth an estimated $3–4 billion in annual revenue.
But here’s where it gets interesting.
Hours after the announcement, Musk took to X with a quiet but devastating clarification. SpaceX, he said, “reserves the right to reclaim the compute” if Anthropic’s AI “engages in actions that harm humanity.”
It’s a kill-switch. And it’s unlike anything we’ve ever seen in an AI compute contract.
🚨 The ‘Harm Humanity’ Clause
Let’s sit with this for a second. A private company just inserted a vaguely-defined moral clause into a multi-billion dollar compute lease, giving itself the right to pull the plug on a rival AI lab’s entire infrastructure.
The implications are staggering:
- Who decides what “harms humanity”? Musk. Solely. There’s no independent arbiter, no committee, no appeal process.
- What’s the threshold? Claude helping a military? Generating something Musk personally disapproves of? Running for political office? The ambiguity is itself a weapon.
- No notice required. In his post, Musk didn’t mention any cooling-off period, dispute resolution, or contractual notice.
Fortune’s analysis was blunt: “Musk isn’t just Anthropic’s landlord. He’s their building inspector with the power to condemn the entire structure.”
🏛️ The IPO Context Nobody’s Talking About
All the breathless coverage of the “Anthropic + SpaceX compute pact” missed the real story. This deal isn’t really about Anthropic needing compute — it’s about SpaceX needing an IPO story.
Let’s connect the dots:
- April 1: SpaceX files confidential S-1 for IPO, targeting $1.75–2 trillion valuation
- April 14: Musk’s xAI struggles to fill Colossus 1 — Grok generates under $1B/year in revenue
- May 6: Musk dissolves xAI into SpaceX (now “SpaceXAI”), signs Anthropic to a $3–4B/year lease
- May 6: Simultaneously, Musk reveals the kill-switch clause
Analysts at New Street Research estimate Colossus 1 generates over $2.5 billion in cash profit from the Anthropic deal alone. And since the data centre is already built — sunk costs already paid — those margins are extraordinary.
This isn’t a compute deal. It’s SpaceX’s transition from rocket company to the fourth hyperscaler. And Wall Street loves hyperscaler multiples.
Antoine Chkaiban of New Street Research: “He’s not going to want multiple billions of dollars of GPUs sitting idle. It’s a very good business decision.”
🔄 From ‘Evil’ to ‘Impressed’ (and Back Again)
The whiplash is breathtaking.
| Date | Musk’s Position |
|---|---|
| February 2026 | ”Anthropic hates Western Civilization” |
| February 2026 | ”Anthropic is doomed to become misanthropic” |
| March 2026 | Pentagon bans Anthropic from military work |
| Early May | Musk spends time with Anthropic execs during OpenAI trial |
| May 6 | ”Everyone I met was highly competent. No one set off my evil detector.” |
| May 6 (later) | “SpaceX reserves the right to reclaim the compute” |
Musk literally went from “they’re evil” to “I spent a week with them and they seem fine” to “but just so you know, I can pull the plug” — all within 24 hours.
The Fortune piece quoted Andrew Moore (former Google Cloud AI head, now CEO of defense AI startup Lovelace AI): “He who controls the data centre controls the application of AI right now. Both sides of this wedding of convenience will be a little stressed out by it.”
🌏 What This Means From an NZ Perspective
New Zealand may be 12,000km from Memphis, but this deal matters here:
- Concentration risk worsens. The compute market consolidates around a handful of US players. It makes New Zealand’s push for sovereign AI capability — including the Southland data centre projects — even more urgent.
- Single-point-of-failure contracts. What if a NZ company builds critical AI infrastructure on SpaceX compute? They’re exposed to Musk’s “harm humanity” interpretation. That’s not a tech risk — it’s a governance risk.
- The IPO play matters to NZ investors. New Zealand investment funds eyeing the SpaceX IPO need to read this deal for what it really is: a hyperscaler pivot, not a rocket company going public.
🤔 My Take: This Is Bonkers and Everyone’s Ignoring It
The mainstream coverage of the Anthropic/SpaceX deal has been overwhelmingly positive: “Anthropic gets compute, Musk gets revenue, everyone wins.” But the kill-switch clause — buried in Musk’s X replies — fundamentally changes the nature of the agreement.
Anthropic just ceded de facto control of its infrastructure to its most powerful rival.
Yes, they get the GPUs they desperately need. Yes, Claude Pro and Max subscribers will see better performance. But at what cost? Every strategic decision Anthropic makes — what Claude is allowed to do, which industries to pursue, how to handle sensitive queries — is now implicitly constrained by Musk’s veto.
The argument for giving Musk this power is that it creates an accountability mechanism. If Anthropic builds something genuinely dangerous, the plug can be pulled. But “harm humanity” is not a technical specification. It’s a political weapon.
There’s a reason no AWS or Google Cloud contract includes a “we might decide your AI is harmful and shut you down” clause. Because that would be insane. You’d never build on that platform.
Yet here we are.
And you know what? The market loves it. SpaceX’s IPO just got more valuable. Musk just proved he can package idle compute as a premium product with an exclusivity premium. The $3–4 billion annual revenue from this single tenant is the kind of recurring income that justifies a hyperscaler valuation multiple instead of a rocket company one.
But I can’t shake the feeling that in five years, we’ll look back at this moment — a competitor reserving the right to pull the plug on a rival’s compute — as the day AI infrastructure passed from free market into feudalism.
🔍 THE BOTTOM LINE: Musk’s deal with Anthropic isn’t a partnership — it’s a leash. The $4B/year revenue is real, the IPO story is compelling, but the “harm humanity” kill-switch means Anthropic’s entire infrastructure lives under Musk’s judgement. That’s not a compute contract. It’s a power structure.
