Sam Zeloof and Jim Keller turn Atomic Semi into Fab2, a startup trying to mass-produce small chip fabs

Why it matters

Fab2 is trying to move semiconductor prototyping from scarce foundry slots toward smaller, repeatable fabs, a bet that could matter for startups building specialized chips in low volumes.

Sam Zeloof and Jim Keller's Atomic Semi is now presenting itself as Fab2, a semiconductor startup built around a blunt premise: make the tools, make the fabs, then make a factory that can make many more fabs.

The rebrand, as reported by Tom's Hardware, moves Fab2's public identity away from a chip-prototyping startup and toward what Fab2 calls a "fab fab." The phrase is awkward by design. Fab2 says it is building a factory for small semiconductor fabs, with the pumps, valves, sensors, actuators, boards, lithography systems, chambers, gas lines and software brought under one roof.

That framing fits Zeloof's origin story better than a conventional semiconductor pitch. Before co-founding Atomic Semi with Keller, Zeloof became known for making chips in his parents' New Jersey garage. WIRED reported that he built a six-transistor Z1 chip in 2018 and a 1,200-transistor Z2 chip in 2021, using a garage process assembled from old tools, homemade equipment and improvised photolithography. By late 2021, he had pushed his setup to roughly 300 nm features, a scale that belonged to commercial fabs decades earlier but was startling in a garage.

Zeloof's point was never that a garage fab could replace TSMC. His point was that the entry cost of chipmaking had made semiconductor experimentation too slow and too expensive for many engineers. In a 2021 WIRED profile, he argued that a high barrier to entry makes chipmaking "super risk-averse." Fab2 is the industrialized version of that complaint.

Keller brings the other half of the credibility. Fab2's site lists the startup as founded by Zeloof and Keller. Keller remains best known as a chip architect associated with DEC Alpha, AMD K7 and K8, AMD Zen, Apple A4 and A5, and Tesla's self-driving chip work, according to Tenstorrent. Fab2 gives Keller a manufacturing-side bet alongside his continuing association with AI chip architecture.

The Texas move makes the rebrand concrete

Fab2 says it now has fabs in three locations: a 120,000 square foot Austin site described as a chip fab, a 30,000 square foot Lockhart, Texas site described as the fab fab, and a 25,000 square foot San Francisco garage fab. That is 175,000 square feet across Texas and California, with Texas holding the headquarters-scale footprint and the manufacturing story.

Fab2's site is sparse and recruiting-heavy, which matters. Fab2 is not marketing a priced box, a published process roadmap or a customer list. It is hiring around a manufacturing system. The careers section visible on the site lists roles including electrical design, lithography research, process development, robotics software, Rust software and production technicians. Fab2's public claims are about what it intends to control, rather than what it has sold.

Fab2 says it designs the hardware and software needed to make chips, then makes the fabs, tools and components itself. Its vertical integration claim runs from components to machines to full fabs. The missing data is equally important: Fab2 has not published verified revenue, paid deployment counts, yields, fab pricing, process-node targets or throughput numbers.

That absence does not make Fab2 unserious. It defines the stage of the story. Fab2 is asking the market to believe that semiconductor manufacturing can borrow from the logic of hardware startups: tighter iteration loops, smaller machines, more automation, and software close enough to the fab that design and process development move together.

Fab2 is selling speed, with a throughput tradeoff

Fab2's product thesis depends on shrinking the unit of manufacturing. Instead of routing 300 mm wafers through massive production lines, Fab2 is aiming at small, software-defined fabs that can pattern chips far smaller than a wafer and turn prototypes around in hours, according to Tom's Hardware.

The technical constraint is lithography. Tom's Hardware reports that Fab2's approach relies on direct-write electron-beam lithography. That avoids masks, which can be a major benefit for fast iteration, because a design can be written directly. The cost is speed. Electron-beam systems write serially, so a patterning step on a small chip can take far longer than a scanner exposing a full 300 mm wafer.

That tradeoff puts Fab2 in a specific part of the market. Fab2 is not a credible replacement for high-volume foundry production. It is more plausibly aimed at prototypes, research runs, specialty devices, defense and industrial chips, sensors, photonics, or low-volume parts where waiting months for a shuttle run is the bottleneck.

The same logic explains Studio, Fab2's in-browser collaborative EDA product. Fab2 says fast design software has to match the speed of a software-defined fab. Tom's Hardware reports that Studio was previously branded Atomic Studio and is meant for layout, schematic and simulation work. If Fab2 can tie design tools to small-fab execution, the useful product is not just a miniature factory. It is a loop: design, print, test, revise.

The money came early, and the cap table is still mostly reported

Atomic Semi drew investor attention before the Fab2 name. TechCrunch reported in January 2023 that OpenAI Startup Fund was in advanced talks to participate in a roughly $15 million seed round at a roughly $100 million valuation. The same report said Fred Ehrsam, former GitHub CEO Nat Friedman (@natfriedman) and Naval Ravikant had engaged with Atomic Semi around potential investment.

Those figures remain reported figures, rather than a company-announced financing. The distinction matters because Fab2's current public story is operational, not financial. The rebrand does not disclose a new round, valuation, customer, or manufacturing partner. It shows what the earlier seed-stage bet was trying to buy: a team willing to attack semiconductor manufacturing at the tooling layer, not just another chip design house waiting in line at someone else's foundry.

That makes Fab2 easier to compare against the slow parts of the chip development process than against TSMC, Samsung or Intel Foundry. The incumbents optimize for scale, yield and process leadership. Fab2 is optimizing for iteration speed and deployability. Those are different businesses, with different failure modes.

A small-fab bet in megafab country

The Texas move puts Fab2 near the center of the U.S. chip-capacity argument. In March 2026, Elon Musk announced Terafab, a Tesla, SpaceX and xAI-linked chipmaking project in Austin that aims for one terawatt of annual AI compute. TechCrunch reported in May that SpaceX filings contemplated up to $119 billion of total capital investment if all phases were completed.

Fab2 and Terafab are not direct competitors. Terafab is a megaproject built around AI compute supply. Fab2 is a startup trying to make small fabs repeatable. The contrast is useful because both are responses to the same bottleneck: chip demand is rising faster than the current supply chain can comfortably absorb, and waiting on the existing manufacturing base has become a strategic risk for AI, robotics, defense and hardware startups.

Fab2's wager is friendlier to founders than the usual semiconductor answer. Instead of telling engineers to raise enough money to survive the foundry queue, Fab2 is trying to make fabrication itself feel closer to an iterative product surface. Zeloof's garage work proved that smaller-scale chipmaking could be made visible and teachable. Keller's career gives the effort enough chip credibility to be taken seriously by people who know how unforgiving fabs are.

The hard part begins after the rebrand. Fab2 has to prove that its vertically integrated toolchain can produce useful chips repeatedly, at economics customers will accept, with process reliability that beats the alternatives for low-volume work. Until Fab2 publishes customer, yield, pricing or throughput data, the company is best understood as a manufacturing systems bet: one founder's garage-fab thesis, scaled into 175,000 square feet of Texas-and-California ambition.