Anthropic and SpaceX Turn Rivalry Into a $4B Partnership, GSD Fights Context Rot, and AI Is Splitting Into Verticals.

Back in January, xAI engineers were found using Cursor and Claude to accelerate their own model development. Anthropic banned xAI from Claude, and for the next couple of months it seemed these two companies were preparing for battle.

In February, Elon Musk responded to Anthropic's $30 billion funding round with a jab of biblical proportions: "Frankly, I don't think there is anything you can do to escape the inevitable irony of Anthropic ending up being Misanthropic. You were doomed to this fate when you chose your name." He went on to call Claude "evil", accused the company of racial and demographic bias, and added that "Anthropic hates Western Civilization".

Fast forward to May 6th, and Anthropic has just announced they will be leasing the entirety of Musk's Colossus 1 supercomputer. That's 220,000+ NVIDIA GPUs and 300+ megawatts of capacity in the Memphis datacenter. So what exactly happened, and how did Elon go from saying that Anthropic was against humanity, to leasing them his largest data center?

On one side, Anthropic's demand has been growing at a pace that they can't comfortably sustain. At Anthropic's 2026 annual developer conference, CEO Dario Amodei told the audience that the company is on an 80x annualized revenue and usage growth rate through Q1 2026, against their internal forecast of 10x. The inevitable infrastructure strain has been apparent to users too.

The chart above tracks Claude system outages over the last 90 days. Orange bars indicate partial outages, while red bars indicate major outages — complete shutdowns, typically lasting between thirty and ninety minutes. These outages are highly frustrating for users, and they're happening more often than one would expect. Executives reportedly considered removing Claude Code from the Pro ($20/month) tier entirely to reduce strain on the system.

By contrast, Musk originally built Colossus 1 in 2024 to train Grok, xAI's competitor to ChatGPT and Claude — but Grok never quite received the demand surge to justify the scale of the datacenter. Per The Information, Colossus' GPU stash has been running at just 11% utilization. Leasing that facility to Anthropic turns a depreciating, underutilized asset into a significant source of recurring revenue — potentially $3–4 billion annually.

This role reversal from enemies to partners is a way for both companies to fix their problems simultaneously. It gives Anthropic access to additional compute resources, allowing them to expand inference capacity without waiting for new datacenter construction. It also allows Musk to use existing SpaceX assets to undercut AWS, Azure, and Google Cloud in the hyperscaler race, while simultaneously generating revenue.

The arrangement also highlights a broader shift in AI: model developers are increasingly competing at the application layer while cooperating at the infrastructure layer. Training these frontier models now requires capital commitments so large that even rivals are aligning to grow.

GSD (short for get-sh*t-done) is an open-source framework that sits on top of Claude Code, created by developer Lex Christopherson (TÂCHES). It does what the name suggests — you can think of it loosely as a bundle of skills that sits in your Claude Code interface and allows you to orchestrate highly organized, well-structured projects from start to finish.

Anyone who has spent a long time in a single AI work session likely knows the pattern of context rot. The first hour is productive and sharp — but by hour two or three, the model starts contradicting itself, producing conflicting responses, and generally deteriorating. As the session grows, earlier context becomes less salient, instructions begin to conflict, and the agent starts to lose consistency across long chains.

The standard workaround is to start a fresh session with the context you've learned, and try to explain your progress in the new session — but you often end up with an incomplete and brittle base.

Rather than try to keep a single long session on track, GSD breaks up your project into many smaller, sub-projects, called phases. For example, a dashboard project might be broken into planning, schema design, frontend construction, API integration, and QA/testing phases. You start any project with a detailed interview session, where you work back and forth with the agent to articulate the intended outcomes. GSD then splits your project up into phases, which are tackled in sequence with your oversight. At the end of every phase, the model produces a detailed MD file, outlining all the work that was done and how it will feed into the next phase.

Notably, each phase also concludes with a debug and verification run, where all the code is tested for quality and functionality. Sometimes it involves human smoke tests, like logging into the app/site/file and confirming it actually works.

You can often get away with building simple HTML files and throwaway scripts without a serious organization tool, but for anyone delving into more involved projects — internal dashboards, automated pipeline reports, deal-signaling platforms — context rot starts to become a real issue. GSD helps solve this issue, and exists for free as both a standalone CLI and an add-on to Claude Code.

Tools like GSD reflect a broader shift in AI development: moving from isolated prompts toward structured orchestration systems designed to manage complex, multi-stage workflows. The big AI labs are no longer just selling general purpose chatbots. We've seen this time and again across the ecosystem — models originally launched as "one-size-fits-all" tools, but the labs that built them are increasingly releasing sub-agents that respond to the needs of a specific subset of users: Claude for Legal, ChatGPT for Healthcare, GSD-style developer frameworks are all examples of specific agentic tools that have been recently released.

When AI started becoming a ubiquitous topic of conversation sometime back in late 2022, I think the vision or end-goal was to develop systems that would sprawl across our range of needs and deliver value without really having to learn how to use them. This matched very closely with what we've seen in science-fiction — Jarvis-style chatbots that almost know what we want before we ask and are able to produce outcomes with very little input.

The traditional SaaS companies took a hit from this perception. Companies built around specialist tools watched their valuations drop on the theory that generalist agents would consume them. Why pay for a dozen subscriptions when one chatbot could do it all?

But increasingly, those "generalist agents" are turning out to be more like the specialist tools they were purported to replace. And instead of a one-size-fits-all chatbot, users are now learning different systems to help accomplish their goals.

I'm not saying this isn't leading to improvements in efficiency, but I do think this path to utopia has a soft irony to it. We were pitched on one tool to replace twenty, and we're ending up with twenty new tools, just powered by the same three models. Are we destined to repeat ourselves infinitely?

I heard a joke recently that software developer shops have started hiring junior developers to write baseline code to ease the token burden of AI coding platforms. Only I'm not certain it was a joke…