IGNITION: Hyperscaler Game of Thrones; AI Coding (and more)

Welcome to Ignition, Catalyst Investors’ briefing on what we’re seeing at the intersection of AI, robotics, software, and growth equity. In this issue:

SpaceX joins the hyperscaler Game of Thrones
AI coding productivity at growth-stage companies
The product and engineering AI market map

01 — THE SIGNAL

There is a version of the Elon Musk AI story that the benchmark-watchers are telling. Grok trails ChatGPT and Claude on evals. xAI’s consumer product has failed to dent ChatGPT’s lead. The model race, by conventional scoring, is going to OpenAI and Anthropic.

That is not the whole story.

While the AI industry has been obsessing over model leaderboard positions, Musk has been playing a more complex game. He has been building the physical substrate that every model in the race depends on and signing his competitors as customers.

The evidence arrived in a single week. Anthropic, a company whose CEO Musk had called misanthropic and accused of hating Western civilization just months earlier, announced it was purchasing all 300 megawatts of compute capacity at Colossus 1, SpaceX’s Memphis data center. The reason was simple: Claude’s massive demand growth had outstripped Anthropic’s infrastructure, creating “inevitable strain” on reliability for paying subscribers. Anthropic needed capacity immediately…and Colossus had it.

Separately, SpaceX now holds a $60 billion conditional option to acquire Cursor, the fastest-growing AI coding assistant, with a $10 billion break-up fee if it walks. Cursor’s compute costs have been weighing on its gross margins, and it needed capacity. Cursor now gets Colossus capacity immediately, before any acquisition closes. The option gives SpaceXAI a potential beachhead in the developer tools war, a category where Anthropic’s Claude Code and OpenAI’s Codex are flexing their muscles, and where Cursor has more daily active developers than any rival.

These are not isolated deals. They are the first visible transactions of a strategy that has been hiding in plain sight.

Yes, Musk has a horse (Grok) in the model race. He also has the infrastructure track that the model race runs on, and he reserves the right to enter the application race at any time. Colossus is the opening move of Elon the Builder. The Anthropic and Cursor deals are the first proof points that the strategy is working. And the deeper you look, the larger the infrastructure ambition becomes.

Colossus 1 is a 300-megawatt facility in Memphis, built faster than any comparable data center in history. But Colossus is only the terrestrial layer. Anthropic’s announcement included a detail that received almost no coverage: the company also “expressed interest” in working with SpaceX to develop multiple gigawatts of compute capacity in space. Orbital compute infrastructure, powered by SpaceX’s launch capabilities, connected by Starlink’s low-latency global network, and outside the jurisdictional reach of any single regulator, is a category of hyperscaler that does not yet exist. If it does, SpaceX is the only company on earth (in the solar system?) positioned to build it. The hyperscalers can beat Musk on ground-based data centers. None of them have a rocket. (Blue Origin doesn’t yet count.)

It is an attempt to own the physical infrastructure of the AI era the way Standard Oil owned the refineries of the oil era, with the added wrinkle that some of those refineries are in orbit.

The counterargument is real and worth stating. Colossus 1’s utilization problem is precisely what forced Musk to sign competitors as customers rather than reserving the resource for Grok. With the SpaceX IPO potentially imminent, signed contracts with Anthropic and Cursor make the Colossus asset look very different on a roadshow than empty racks do.

The Cursor situation carries its own tension. Cursor currently runs on Claude and OpenAI’s Codex, not on Grok. Its researchers are building their own model, Composer, partly powered by Chinese model Kimi. There is no current plan to co-develop models with xAI, and no plan to steer users toward Grok. If SpaceX exercises the option and acquires Cursor, it inherits a product whose competitive position depends on its neutrality across AI providers. Plus, Cursor’s best researchers will have opinions about joining a rocket company run by the world’s most polarizing CEO, and retention bonuses only go so far.

None of this diminishes SpaceX’s broader AI strategy. It’s just more complicated than the clean narrative suggests. That said, no one said the narratives of the other hyperscalers (Amazon, Microsoft and Google) are particularly clean either. Complexity is the name of the game in the AI Game of Thrones.

Note: The Cursor acquisition has not closed. SpaceX holds a conditional option; terms and timing remain unconfirmed. The Anthropic-SpaceX compute agreement was announced May 6, 2026. Sources: The Information, CNBC.

02 — OPERATOR’S EDGE

The AI coding productivity gap: what the data says

Devessence published one of the most honest assessments of AI in software development we’ve read this year, cutting through the hype to what the data shows at each stage of the software development lifecycle (SDLC). Worth reading in full. Read it here.

The takeaways every Series B-D engineering leader should internalize:

Adoption is effectively universal. 97% of software organizations are using or evaluating AI tools. The question is no longer whether, it’s how well.
AI helps most where work is mechanical. Boilerplate, scaffolding, test generation, documentation. Developers report saving 30 to 60% of their time on this category. Planning, architecture, and complex debugging still require human judgment AI cannot replicate.
Experienced developers can get slower. The METR study found senior developers working in their core domain were 19% slower with AI tools. Prompting and evaluating suggestions creates friction that outweighs the help when you already have fast mental models. AI is most valuable at the edges of expertise, not the center.
The review bottleneck is the real problem. Faros AI found high-adoption teams merged 98% more pull requests, but review time increased 91%. AI accelerates code generation. It does not accelerate code review. Teams that don’t rebuild their review process alongside AI adoption generate more activity without more output.
AI-generated tests validate what code does, not what it should do. If the underlying logic is wrong, the AI writes a test that confirms the wrong behavior. More tests do not automatically mean better software.
Total adoption cost is 1.5 to 2.5 times the license fee in year one. Onboarding friction, codebase cleanup, review infrastructure, and governance overhead rarely appear on the invoice. Model them honestly before presenting the productivity case to your board.
Agentic AI is the next inflection. Gartner projects 60% of enterprise AI rollouts will include agentic capabilities by end of 2026, meaning systems that don’t just complete code but execute entire tasks. The teams building review discipline now are the ones positioned to absorb that shift without losing control of their codebase.

At Catalyst, we’ve started tracking AI code adoption rate and review infrastructure maturity as diligence inputs alongside NRR and CAC payback. A growth stage company with a systematized AI program is a fundamentally different asset than one with only ad hoc usage.