Borrowed Compute

More and more, vulnerability news reaches me on X before it shows up anywhere else.

CVE numbers, fragments of repro code, the reactions from each direction. Grok pulls the context together. Who's responding how, where the PoC lives, what the vendor is saying. With X search behind it, the whole picture lands in one shot.

When you run a talent agency, what happens on X feeds straight into the work. How often a name is being searched, the heat in each fandom, the smell of a brewing flare-up. Web-search-based AIs don't reach that granularity. The staff already has Gemini, but it can't quite catch the real-time pulse.

I wanted to give them Grok too. Heavy isn't needed — plain SuperGrok is enough. But I was already weighing whether to roll out GPT for image work. Layering Grok on top, even at $30 a month, the per-head bill starts to add up. I dropped it.

Fortunately, xAI has a usage-based API. If an in-house tool calls it, I don't have to stack monthly seats by headcount. So I wrote grok-mcp. An MCP server that lets Claude Code, Codex CLI, and other MCP clients talk to Grok. Send the time-sensitive questions to Grok. Send the code questions back to Claude. Now that each AI is settling into its own specialty, you need a router.

While I was writing it, Anthropic struck a compute deal with xAI/SpaceX. Colossus 1 — xAI's giant GPU cluster in Memphis. Claude would run on it, and in return, Claude Code's rate limits got doubled. I seem to remember Musk and Anthropic publicly trading jabs not long before that. The map of the AI war gets repainted daily.

In the same month, xAI shipped Grok Build CLI. A coding agent against Claude Code, running from the CLI. For now it's limited to SuperGrok Heavy — the $300-a-month plan.

When I line up what I'm actually doing, it gets funny. I'm using a Claude that runs on borrowed xAI infrastructure to write an MCP that hits Grok. To put Grok in front of my staff, I'm hitting Grok's rival CLI to generate the code that connects to Grok.

Up above, the giants trade massive compute. I'm writing a small router that ties one AI to another. I can't ride the main current of the AI war. But router-sized work is still left in my hands.