Kepler Computing, a San Jose startup that spent seven years quietly rethinking how memory chips get built, has stepped out of stealth with a claim that could reshape the economics of AI infrastructure: high-bandwidth memory without extreme ultraviolet lithography.

The company announced $468 million in total funding and a manufacturing approach that works inside existing fabs rather than requiring the $20 billion to $40 billion investments that new memory facilities demand. In a market where SK Hynix and Micron are racing to construct multibillion-dollar plants to meet insatiable AI demand, Kepler is betting that clever materials science beats brute-force capital expenditure.

Why Memory Has Become AI's Bottleneck

The explosion of large language models since ChatGPT's launch in 2022 turned high-bandwidth memory from a niche component into a critical constraint. HBM sits adjacent to GPUs and accelerators, shuttling massive datasets back and forth during training and inference. Every major AI chipmaker needs it, and the supply chain hasn't kept pace.

Traditional HBM manufacturing relies on stacking DRAM dies and connecting them with through-silicon vias. The process works, but it's expensive, slow to scale, and tied to fabs that cost a fortune to build. Kepler CEO Debo Olaosebikan says the company originally planned to tackle SRAM first, then move to DRAM and HBM. The AI demand surge changed that timeline.

"With the launch of ChatGPT and the explosion of demand for an HBM alternative, we started working on our HBM road map," Olaosebikan says. "Now we're making SRAM and HBM in parallel."

The 3D Stacking Play

Kepler's approach to HBM centers on a novel 3D manufacturing technique that fits more memory dies within a fixed footprint. The core compute logic sits closer to the memory, reducing the energy needed to move data between them. The company's goal is to match SRAM's energy efficiency while retaining HBM's large capacity, a combination that would slash power consumption in data centers running AI workloads.

For SRAM, Kepler uses ferroelectrics, materials that switch polarity at lower voltages than conventional semiconductor storage mechanisms. The key innovation is a proprietary composite material the team developed after 35 iterations. CTO Sasi Manipatruni says the breakthrough came when the team solved the physical limitations of HBM density, then paired that with a material innovation that determines how much memory fits between chips.

The material includes iron, which introduces manufacturing complications. GlobalFoundries senior vice president Ed Kaste notes that iron is a difficult contaminant to manage in a production environment, requiring either dedicated equipment or full encapsulation to prevent it from escaping into other processes. "The art is in keeping that material really well isolated through our production flow," Kaste says.

Working With What Factories Already Have

The most significant aspect of Kepler's approach may be what it doesn't require. EUV lithography machines, made exclusively by ASML, cost upward of $150 million each and have delivery timelines measured in years. By developing a materials-based approach to density improvement, Kepler claims it can achieve the same transistor density as 2-nanometer or 3-nanometer chips without touching EUV equipment.

GlobalFoundries, which invested $50 million and serves as Kepler's manufacturing partner, has been running Kepler's process on its 28-nanometer platform. Kaste calls the work "at the sweet spot of our strategy," noting the multigenerational scaling potential without requiring entirely new fab systems.

In early tests, Kepler converted a GlobalFoundries fab line into a next-generation facility in eight months, compared to the typical 24-month timeline. Testing has taken place primarily in Singapore, where GlobalFoundries has a major facility, and at the company's Burlington, Vermont plant.

Intel Capital managing director Srini Ananth says the investment firm didn't initially position Kepler as a replacement for DRAM or SRAM. "We figured the market would dictate that, and now you're seeing a demand for both," he says.

The Funding and Government Backing

Kepler's investor list reads like a who's who of semiconductor and tech royalty. Beyond GlobalFoundries and Intel Capital, the company counts AMD Ventures, British investment fund Baillie Gifford, and Bill Gates's private Gates Frontier fund among its backers. The U.S. Department of Commerce committed up to $245 million to help Kepler develop high-performance AI memory technology using 3D and ferroelectric approaches on American soil.

The government funding reflects broader U.S. efforts to secure domestic semiconductor manufacturing capacity, particularly for components critical to AI infrastructure. Kepler plans to begin chip production in the United States in 2028.

The Scaling Question

Kepler has tested its technology on roughly 2,000 wafers to date. The company plans to ship first HBM samples later this year and ramp production in Singapore next year. Full-scale U.S. production remains two years away.

That timeline highlights the central challenge. Proving a new materials approach works in a lab or on limited wafer runs is fundamentally different from producing millions of devices at cost, on schedule, with consistent yields. Industry analysts have raised similar concerns about Substrate, another well-funded startup that emerged last year with a new nanoparticle-based lithography approach. Creative Strategies chip analyst Austin Lyons points to contamination management, material consistency, and tooling compatibility as the hurdles that determine whether a semiconductor innovation reaches volume production.

"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," Lyons says.

Kepler's thesis is sound: the world needs more memory bandwidth for AI, existing fabs are underutilized for advanced processes, and materials science can bridge the gap without waiting a decade for new factories. Whether the company can execute on that thesis across millions of wafers, while managing exotic materials in production environments, will determine whether this $468 million bet pays off or becomes another cautionary tale in the unforgiving semiconductor industry.