Kepler raised less than 4% of what one new memory fab costs

Kepler Computing came out of stealth after seven years with a new architecture for high-bandwidth memory, WIRED reports. Its pitch is that the world does not need to wait for new memory fabs, because existing ones can be retooled. Five numbers decide whether that holds.

Five numbers

• $468 million raised privately, from GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford and Bill Gates through his Gates Frontier fund.
• Up to $245 million committed by the US Department of Commerce in July.
• $20 billion to $40 billion, the cost of building and equipping one new memory fab.
• 2,000 wafers run to date.
• 8 months to convert an existing fab, against a typical 24.

Add the private round to the government commitment and Kepler has about $713 million. Set that against the low end of a new fab and it is 3.6%; against the high end, 1.8%. The company is funded at a small fraction of the thing it proposes to make unnecessary, which is either the whole point or the whole problem.

The 2,000 wafers are the number to hold onto. A high-volume fab moves tens of thousands of wafers a month, so Kepler has run roughly what a working plant does in a day or two. Its own manufacturing partner framed the remaining task in the same units.

“The fundamental breakthroughs have happened. What remains is getting good results on thousands of wafers and millions of devices.”

— Ed Kaste, WIRED, 9 September 2026

Ed Kaste, senior vice president of GlobalFoundries' CMOS business. Source: WIRED, 9 September 2026

What the technology claims

Chipmakers normally buy density with extreme ultraviolet lithography, which is the most expensive equipment in the building. Kepler says it reaches the density of 2-nanometre and 3-nanometre parts without EUV at all, using 3D stacking and a proprietary composite, on machines fabs already own.

For SRAM it uses ferroelectrics, which switch at lower voltages than the usual mechanisms. Chief technology officer Sasi Manipatruni says the team went through 35 composite iterations before settling on a material class.

For HBM the goal is stated plainly: move data inside high-bandwidth memory at an energy cost comparable to SRAM while keeping HBM's capacity. Core compute sits closer to the memory, so each transfer costs less, and inference is what pays for those transfers.

The catch its own partner named

Kaste named the catch himself. The composite contains iron, which he calls a tough contaminant to bring into a production facility, so the process has to run on dedicated equipment or stay fully encapsulated. He describes keeping the material isolated through the production flow as the art of it.

Austin Lyons of Creative Strategies, who was not briefed on the technology, put the general problem as contamination, different materials and different tooling, and whether the result works at scale and is worth the cost.

Kepler plans first HBM samples later this year, a production ramp out of Singapore next year, and US chip production in 2028. The memory shortage it is built to relieve will have resolved itself one way or another well before that.

Where this sits

We wrote about a chip designer chasing the energy side of the same problem last week, and about Anthropic booking $517 billion of compute. One company is buying its way through the bottleneck and the other is trying to remove it. The bottleneck is the same.

Watch the wafer count rather than the funding. Money has never been the hard part of semiconductor start-ups, and Kaste has already told you what the test is.

This piece is informational, not a recommendation to buy, sell, or hold any asset.