Here are three things I found interesting in the world of AI in the last week.

Claude is renting GPUs from xAI, and pointing at a profit

Anthropic's investor projections, shared during a fundraise and reported by CNBC, show it expects its first operating profit in the quarter ending June: roughly $559M on $10.9B of revenue, up from $4.8B the quarter before. Yeah, doubling revenue in one quarter. In the same week, SpaceX's IPO filing revealed Anthropic is paying $1.25B a month, commited through 2029, to rent compute at xAI's Colossus data centers. Claude is now paying Grok's parent company about $15B a year for GPUs, with a 90-day exit clause.

The profit is a projection with thin margins but it is a first from a frontier lab and powered by eye watering adoption of Claude in 2026. Lots of commentators are gloating that 'there is no bubble' and Anthropic is overtaking OpenAI on valuations.

Compute has been a massive constraint for Anthropic during this growth, slow responses, tighter token caps, lots of annoyed users. So the deal with xAI makes sense.

My favourite part of this is that xAI has invested heaps in building compute and no one wants to buy Grok (ha ha) so their best option is renting their compute to companies who make models people actually like. I guess don't be a [insert expletive of choice here] is a good business strategy.

But Musk will still probably over subscribe the space X IPO at massive inflated valuations, so what do I know.

The Musk v OpenAI circus ended on a technicality

After about eleven days of trial, an advisory jury found Elon Musk filed his suit too late, and the judge dismissed the whole thing on statute-of-limitations grounds (Axios). There was no ruling on whether OpenAI actually broke its founding promise or unjustly enriched anyone. The roughly $134B in damages Musk was chasing, and his bid to push out Sam Altman, both evaporated. He is appealing to the Ninth Circuit, so the final word is probably a 2027 story.

To me, this sounded like something that should have been caught pre-trial but apparantly the judge denied the arguments for summary judgement on statue-of-limitations grounds. When Musk knew (or should have known) was a disputed fact and the jury decided in OpenAI's favour.

The whole fight was about when the clock started. OpenAI argued Musk should have known something was off years ago: he quietly stopped donating in 2017 (and testified he suspected he was being swindled), and by 2019 there were emails flying about the for-profit arm. Musk's side reckoned he kept taking Altman's reassurances at face value and didn't really clock the breach until 2022, or until the 2023 Microsoft deal stripped out the profit caps. You get two to three years to sue depending on the claim, Musk filed in February 2024, so that gap is the whole case.

So I guess the moral of the story is if you are going to steal a non-profit then just make sure no one notices for a couple of years. I can see the silicon valley tech bros furiously scribbling notes to self already.

Cowork still has the edge, but Codex is a fine alternative

Both camps shipped like crazy this month.

Here is what Claude and Anthropic put out:

• Claude for Excel, PowerPoint and Word (GA), plus Outlook in beta
• Claude Design, with doubled token limits across every plan
• Ready-to-run agents for finance and legal teams
• "Dreaming" in Claude Code: agents leave notes for the next agent working on the same code
• Agent view and parallel `claude agents` sessions in Claude Code
• Self-hosted sandboxes and MCP tunnels for its Managed Agents
• The Claude Platform on AWS, now generally available
• On the corporate side: buying Stainless, rollouts at KPMG and PwC, and a $200M Gates Foundation partnership

And Codex, in the same stretch:

• Codex in Chrome browser plugin
• Goals, from a feature to on by default with progress tracking, plus headless remote control and a `codex doctor` diagnostic
• Codex in the ChatGPT mobile app, so you can monitor and approve agents from your phone
• A "super app" turn: connectors for Slack, Drive, email and calendar, pitched for any task done with a computer
• HIPAA-compliant Codex for local environments, for regulated and healthcare teams
• GPT-5.5, the new flagship, which retook the top of the main coding benchmarks (SWE-bench Verified and Terminal-Bench) at fewer tokens
• Codex on Amazon Bedrock, run on your own AWS credentials and commitments
• The OpenAI and Dell deal putting Codex on on-prem and hybrid enterprise data

For the last year my answer to "which AI should I use" has been to use the AI you have. If you don't have one, get Claude.

Now I'm changing it slightly. Claude Cowork and OpenAI Codex are the two clear leaders which are broadly comparable. Gemini and Microsoft Copilot are second tier, you can use them but Gemini will take a bit more technical learning (getting antigravity installed) and Copilot is much harder to get useful work out of.

Whichever AI you choose, make sure you own your workflows so it's easy to switch and you aren't held hostage