Musk is building a chip fab ten times the size of Giga Texas

Agents grew hands this week. Meta shipped its first coding agent, a startup came out of stealth with one that clicks through the open web, and DeepMind handed humanoids control over their entire frame. On the infrastructure side Musk committed $16.8 billion to making the chips and signed Nvidia up to fly them.

Meta: entering the coding agent war, late and cheap

Meta launched Muse Code in beta, its first coding agent. Terminal-based, macOS and Linux, built to handle complete software engineering tasks across large repos. Planning changes, writing code, validating results. It runs on Muse Spark 1.2, a model co-trained alongside the agent with a dedicated code-management harness.

The pricing is where it gets interesting. Standard pay-as-you-go API rates, or a contributor tier that cuts costs by over 90% if you share your data.

Meta is buying training data with a discount coupon. For agencies running thin margins on internal tooling that is a real decision to make, and the answer depends entirely on whose code you point it at. Client repos under NDA? Probably not.

Alexandr Wang's Superintelligence Labs owns this one. Third serious entrant in a category Anthropic and OpenAI have had mostly to themselves.

Hark: an agent that actually clicks

Hark stepped out of stealth with Handoff, a computer use agent built for long-horizon work on the open web. It drives a dedicated virtual machine, clicks buttons, fills in forms, and figures out sites it has never seen before.

The demos are mundane in a promising way. Ordering dinner end to end on DoorDash and Uber Eats. Comparing and booking flights across United, Delta and American.

Hark claims the top recorded score on the OM2W web-use benchmark, ahead of Anthropic, OpenAI and Google, at under a tenth of the token price. Those are vendor numbers from a launch post, so treat them accordingly until somebody independent runs it.

The founder is Brett Adcock, who started Figure. A man who spent years teaching robots to manipulate the physical world, now pointing the same problem at web UIs. Every booking flow your clients complain about is about to become an API whether they built one or not.

Signups are open, release later this month.

Google DeepMind: robots that finally use their legs

Gemini Robotics 2 landed and the jump is whole-body control. The previous generation mostly drove a robot's upper half. This one coordinates the entire humanoid, top to bottom, while reasoning through multi-step tasks in environments built for people.

Three models ship together. A vision-language-action model, an embodied reasoning VLM, and an on-device VLA for when latency matters more than raw capability.

The number that stands out is transfer. DeepMind says it adapts to an entirely new robot body with a few hours of data, typically under 200 examples. Apptronik's Apollo ran the demos, with Boston Dynamics and Agile Robots on early access.

Also, a 92% success rate at unscrewing a lightbulb. Sounds trivial until you try to describe the required footwork, grip pressure and balance correction as code.

Robotics spent a decade stuck on the hardware. The control layer is now the part moving fastest.

Nvidia and SpaceX: data centres are leaving the planet

SpaceX picked Nvidia to build the compute payload for Starmind AI1, its orbital data centre satellites. Rubin GPUs and Vera CPUs, with Nvidia claiming the space module does up to 25 times the AI work of an H100.

The physical spec is absurd in the good way. 30 metres tall deployed, 75 metre wingspan, wider than a 747. Prototype testing is planned for early 2027, and the long-term pitch is a constellation of up to a million satellites acting as one distributed supercomputer.

The logic is thermal and electrical. Constant solar input, radiative cooling, no grid connection to negotiate with a local municipality. On Earth those three things are currently the hardest part of building anything at this scale.

The Register called it Elon handing Nvidia a monopoly over the stars, which is uncharitable and probably correct. Nobody has shown orbital compute beats a datacentre in Arizona on cost. Radiation, latency and the small matter of repairing a failed rack 500km up are still unsolved.

Ambitious enough that we will be watching the 2027 prototype closely.

Terafab: now they want to make the chips too

Two days after the Nvidia deal, Tesla and SpaceX confirmed a $16.8 billion first phase for Terafab, a semiconductor complex in Grimes County, Texas. It is a joint venture with xAI, which SpaceX absorbed in an all-stock deal back in February.

The scale is hard to picture. Over 100 million square feet of manufacturing space, 100,000 wafer starts per month at launch, scaling toward a million. At least 3,000 people hired from Grimes and Brazos County. Intel joined in April and the fab will run on its 14A process.

That is the first phase. Earlier filings pointed at up to $119 billion for the full buildout, and the stated goal is closing the gap to the terawatt of compute Tesla and SpaceX expect to need.

Put it next to Starmind and the shape is obvious. Musk wants to make the silicon, launch the silicon, and run the workloads on it. Vertical integration from sand to orbit.

Whether one company should own that entire stack is a conversation nobody seems to be having yet.

This roundup first went out to our Dotbite Tech Pulse subscribers. Want it in your inbox, too?