No EUV Required: Kepler Computing Exits Stealth With $468M and a Ferroelectric Fix for the AI Memory Crisis

While the industry’s attention has been locked on GPUs, the sharpest constraint on AI compute in 2026 has quietly become memory. High-bandwidth memory (HBM) stacks are rationed, DRAM contract prices have surged, and every new accelerator design starts with a simple question: where will the memory come from? On September 9, a San Jose startup named Kepler Computing stepped out of seven years of stealth with an unusual answer: don’t build new fabs at all — change the physics of what existing fabs can produce.

Kepler Computing, founded in 2018 by a team of physicists and computer scientists, has raised $468 million to develop a new architecture for high-bandwidth memory that directly targets the supply bottlenecks squeezing the computing market. Its backers read like a semiconductor power list: GlobalFoundries, Intel Capital, AMD Ventures, the British fund Baillie Gifford, and Bill Gates through his private Gates Frontier fund. In July, the US Department of Commerce committed up to $245 million to help Kepler “develop in the US a new class of high-performance AI memory technology, enabled by innovative 3D and ferroelectric technologies.”

The bet: density without EUV

The conventional path to more memory density runs through extreme ultraviolet (EUV) lithography — the staggeringly expensive machines that shrink transistors so more capacity fits in the same footprint. Kepler’s claim is that it can increase density without relying on EUV at all, using two complementary techniques that work with existing semiconductor fabrication plants.

The first is a novel 3D-manufacturing approach that fits more memory chips within a fixed footprint — stacking ferroelectric RAM tiles directly on top of computational logic. Bringing the core compute closer to the memory means data travels a shorter distance, and shorter distances mean less energy per bit moved. Kepler’s ultimate goal is to move data inside HBM at energy levels comparable to SRAM while keeping HBM’s large capacity — collapsing the historic trade-off between the fast-but-tiny cache and the big-but-power-hungry stack.

The second technique attacks SRAM density directly using ferroelectrics, which can read and write data at lower voltages than conventional semiconductor mechanisms. The enabler is a proprietary low-voltage composite material — one the company says took 35 iterations to land on. “Once we found a way to solve the physics problem — as in the physical limitations for HBM — we came up with a material innovation that helps with the amount of memory you can have between chips,” says cofounder and CTO Sasi Manipatruni. CEO Debo Olaosebikan wouldn’t confirm the composite’s exact elements, saying only that “we’re using a small number of materials, and some of them aren’t what you’d typically find in mainstream ferroelectrics.”

The company claims its SRAM achieves density equivalent to 2-nanometer or 3-nanometer chips — without investing in EUV tooling. “It’s a new materials system, with multigenerational scaling potential, without having to build entirely new systems in the fab or invest in very expensive lithography equipment,” says Ed Kaste, senior vice president of GlobalFoundries’ CMOS business.

Why now: ChatGPT rewrote the roadmap

Kepler’s original plan was sequential: SRAM first, then DRAM, then HBM. The launch of ChatGPT in 2022 and the explosion of demand for HBM shredded that plan. “Now we’re making SRAM and HBM in parallel,” Olaosebikan says. The pivot reflects how completely AI has redrawn memory economics — HBM has become the memory “du jour” in a data-center-driven market, and incumbent makers SK Hynix and Micron are racing to build multibillion-dollar fabs on the wager that demand will still be there when those facilities come online.

Kepler’s counter-thesis is that the market shouldn’t have to wait for brand-new megafabs. The company builds what it calls “mini fabs” — in one early build-out with GlobalFoundries, it says a fab was converted into a “next-generation” facility in eight months instead of the typical 24. Much of Kepler’s testing currently happens in Singapore, where manufacturing partner (and $50 million investor) GlobalFoundries operates a facility; tests have also run in GlobalFoundries’ Burlington, Vermont plant, integrated with GlobalFoundries’ 28-nanometer process nodes.

The financial logic is blunt: any additional cost from new materials or retooling existing fabs would still fall far short of the $20–40 billion it costs to build a new fab and outfit it with equipment worth hundreds of millions per tool. “Our goal is to take the fabs and architectures already built, and push them to the limits of physics,” Olaosebikan says.

The hard part: iron, contamination, and 2,000 wafers

Skepticism is warranted, and Wired’s reporting is careful to include it. To date Kepler has run its technology on only around 2,000 wafers — a rounding error against the volume of a production memory line. Its roadmap calls for first HBM samples to ship later this year, production ramping out of Singapore next year, and US-based chip production starting in 2028.

There’s also a materials problem hiding in the composite itself. Kaste of GlobalFoundries notes that Kepler’s material system includes iron — “a tough contaminant to introduce into a production facility.” The solution must run on dedicated equipment or be fully encapsulated so it can’t escape. “The art is in keeping that material really well isolated through our production flow,” he says. He remains confident that “the fundamental breakthroughs have happened. What remains is getting good results on thousands of wafers and millions of devices.”

Independent analysts put the challenge in starker terms. “The question is how to overcome all the limitations of contamination, different materials, and different tooling, in such a way that the resulting innovation can be used at scale and is worth the cost,” says Austin Lyons, a chip analyst at Creative Strategies who was not briefed on Kepler’s work. The caution is earned: the startup Substrate made waves late last year with a nanoparticle lithography alternative, and analysts noted how brutally hard it is to produce huge volumes of chips “that meet incredibly stringent specifications, on time and on budget.”

Why it matters

If Kepler’s ferroelectric 3D stacking works at scale, the implications ripple well beyond one startup. The approach effectively decouples memory supply growth from the EUV oligopoly — a handful of the most expensive machines ever made — and from the megafab construction cycle that takes years and tens of billions of dollars. Legacy 200mm and 28nm-class capacity, of which the world has plenty, could be re-tasked toward AI-class memory. For a market where memory has become the binding constraint on accelerator shipments, that is a supply-side unlock with no obvious equivalent.

It also marks a quiet endorsement of the materials-first thesis: rather than chasing smaller transistors, spend seven years finding the material that makes existing geometry do more. Intel Capital’s Srini Ananth frames the opportunity plainly: “We didn’t go in thinking that this would be a replacement for DRAM or a replacement for SRAM. We figured the market would dictate that, and now you’re seeing a demand for both.”

The US government, meanwhile, is treating domestic AI memory as strategic infrastructure — the $245 million Commerce commitment is explicitly about building this class of technology on American soil, with the 2028 US production target as the delivery vehicle.

None of this is proven yet. Two thousand wafers is a promise, not a product line, and the iron-contamination constraint means Kepler’s economics only work if isolation can be maintained at volume. But in a year where the entire industry’s growth is throttled by memory supply, a startup claiming 2nm-class SRAM density from 28nm fabs with no EUV in sight is exactly the kind of long-shot bet the shortage makes rational. First HBM samples are due before the end of 2026 — that’s when the physics stops being a press release and starts being a supply curve.