1. Claude Starts Building Claude (AI)
Anthropic published a report saying Claude wrote more than 80% of the code merged into Anthropic's production codebase in May. Before Claude Code entered research preview in February 2025, the figure was in the low single digits. 1
The same report said a typical Anthropic engineer now merges about eight times as much code per quarter as engineers did from 2021 to 2025. That matters more than the headline. This is not AI replacing the company. It is the company changing shape around AI.
Anthropic paired the numbers with a warning about recursive self-improvement. It said AI has not yet reached the point where it can design and build its own successor. But it argued the world should have the option to slow frontier development if the systems start moving faster than safety work.
The timing was awkward. Anthropic had just confidentially filed to go public. 2
Why it matters
The 80% number is the clearest public sign yet that coding agents are not just helping engineers. They are becoming part of the production line that builds the next systems.
Reality check
This is still a human process. People ask, review, test, and merge. The numbers are also self-reported. The pause proposal is real governance work, but it also arrives at the perfect time for a company about to sell its story to public markets.
2. SpaceX Becomes an AI IPO (Space)
SpaceX set IPO terms at 555.6 million shares for $135 each. That would raise about $75 billion and value the company near $1.75 trillion. Trading is expected under SPCX on Nasdaq. 3
The numbers behind the deal are stranger than a normal space listing. SpaceX posted $18.7 billion of 2025 revenue and a $4.9 billion net loss. Starlink was profitable. The AI division absorbed through xAI was not. 4
That turns the IPO into something cleaner and weirder. Starlink is the cash engine. Launch is the industrial base. AI is the giant option investors are being asked to fund.
Why it matters
The biggest IPO ever is not only a rocket story. It is another AI infrastructure story. Even SpaceX is now being valued partly on compute, data centers, and the idea that AI can move off Earth.
Reality check
A loss is not automatically weakness. SpaceX is choosing to spend. But public investors will not be buying a clean Starlink utility. They will be buying Musk's whole stack.
3. Apple Rents the Frontier (AI)
Apple finally showed the Siri rebuild at WWDC. The new system is meant to understand context, read what is on screen, work across apps, and act more like an agent than a voice menu.
The important part is the model layer. Apple says the features are powered by Apple Foundation Models, but those models were developed with help from Google and Gemini. Apple says requests run on-device when possible and through Private Cloud Compute when needed. 5
That is not the same as handing Siri to Google. But it is a concession. Apple still owns the device, the operating system, and the privacy wrapper. It does not fully own the frontier model path.
Why it matters
Apple is the best distribution machine in consumer tech. If even Apple needs outside model help, the frontier layer has become a harder place to compete than the product layer.
Reality check
This may be the right move. Apple can still win by making AI feel native, private, and boringly useful. But the old Apple story was control from chip to software. Siri AI is more hybrid than that.
4. Microsoft Builds the Agent Stack (AI)
Microsoft used Build 2026 to make agents feel less like demos and more like plumbing. It introduced Microsoft IQ as a context layer across GitHub Copilot, Foundry, and Copilot Studio. It also put the GitHub Copilot app into preview. 6
The company also showed Project Rayfin, Project Solara devices, and seven MAI models, including MAI-Thinking-1 and MAI-Code-1. The message was simple: agents need memory, local compute, enterprise data, models, and a place to run.
This is Microsoft trying to own the work layer. Not the chatbot. The workflow.
Why it matters
The next AI fight is not only model quality. It is where agents live. Microsoft is pushing them into the tools companies already use every day.
Reality check
Build keynotes always make stacks look cleaner than they are. Enterprises still have messy permissions, bad data, compliance limits, and users who do not want surprise automation. The strategy is clear. The operating reality is harder.