Musk won't rule out a Tesla-SpaceX merger to feed the AI buildout
Musk spent the All-In Summit this week sitting in a trailer next to his biggest data center, saying out loud what the capital markets have been pricing for months. Then Browser Use open-sourced an agent that deletes the screenshot step entirely.
Asked why Tesla and SpaceX are still two separate companies, Musk did not say they will stay that way. Speaking with SpaceX president Gwynne Shotwell at the All-In Summit in Los Angeles, he said the two firms already collaborate so closely that "who knows what actions people will take" — no timetable, no deal structure, no promise, but no denial either. The setup for the question was Terafab, the jointly built chip plant in Texas that Musk now describes as mandatory rather than strategic: "It's either build TeraFab or fail to scale," he said, adding the idea "sort of did come to me in a dream." Tesla and SpaceX have ordered the equipment and hope to have something useful out the other end by the end of next year, against a stated target of more than a terawatt of computing hardware a year. Shotwell noted SpaceX is already doing its own chip packaging, one of the bottlenecks hosts flag as scarce.
The scale of that ambition is the part worth pricing. Terafab broke ground in Grimes County in August on $16.8 billion of committed capital, and analysts put the cost of a plant capable of 100,000 leading-edge logic wafers a month at $30 billion to $45 billion — Morgan Stanley calls it a "Herculean task," noting Micron's Boise memory fab took from late 2022 to mid-2027 to reach shipments. Musk's answer to the capital question is the merger question: an empire whose compute, chips, power and models sit inside one balance sheet is cheaper to finance than three that have to negotiate with each other. He is also living at the site — an Airstream trailer parked in Memphis while he oversees the next xAI data center build, the same instinct that had him on the Model 3 factory floor in 2017 and in Twitter's library in 2022. Colossus went from zero to 100,000 GPUs in 122 days and doubled to 200,000 in another 92, per QbitAI, and it now rents capacity to Google and Anthropic under multibillion-dollar contracts — the revenue line that makes SpaceX's IPO story about AI infrastructure rather than rockets. If Terafab slips, none of that matters; if it lands, Tesla and SpaceX become one company in everything but paperwork, and the paperwork is plainly next.
Browser Use open-sourced a browser agent that never looks at a screenshot, and it is measurably faster for it. Jev Ultrafast replaces the usual capture-describe-translate loop with a single DOM snapshot per decision cycle: the page is read once into a table of visible controls, and TypeSafe's Jev model returns two probability distributions in parallel — which operation to run, and which indexed element to run it on. The published benchmark is a real Google Flights search, one-way Zürich to London for September 20, which ran in 7.1 seconds against 9.45 seconds for the previous loop, with browser protocol calls falling from 1,092 to 101, a 91% cut. A Wikipedia article open finished in 2.798 seconds and a local hotel filter search in 1.896 seconds. The honest part is the limits the authors list themselves: no shadow DOM, iframes, canvas, file uploads, new tabs or arbitrary keyboard widgets, and a warning that "DONE is never independent evidence of success," so the landing has to be verified outside the agent. Mercury 2.5 over OpenRouter handles the decisions with reasoning off, and a small text model only wakes up when something needs typing — which is the architectural point, since it makes most interaction cycles free of language-model text generation. We flagged TypeSafe's calibrated-confidence pitch when it exited stealth with $40 million — TypeSafe exits stealth with $40M for models that don't chat — and this is the first time that architecture has a real, open, measured task behind it.
What to watch: whether Terafab's end-of-2027 output target survives contact with equipment lead times, and whether Jev's constraint-based action space holds up past the structured forms it was benchmarked on.
If a browser agent that can only click and type beats one that can see, how much of what we call computer use was ever about vision? Tell us in the comments.
Sources: Jev Ultrafast (GitHub) · Jev Ultrafast performance notes (GitHub)
