The race for artificial intelligence is almost always framed around processors, but in modern data centers, the problem is not just how fast a GPU can compute. Increasingly, the limit is the speed and energy required to move data between memory and compute units. That is why Kepler Computing’s emergence from stealth deserves attention: the startup is not promising a new accelerator, but a new memory architecture designed for one of AI’s toughest bottlenecks.
Kepler, founded in 2018 in San Jose, has raised $468 million from investors including Intel Capital, AMD Ventures, GlobalFoundries, Baillie Gifford, and Bill Gates's Gates Frontier fund. In July, the US Department of Commerce also announced a commitment of up to $245 million to support the development of a new class of high-performance AI memory based on 3D and ferroelectric technologies in the United States.
The problem is not just building more chips, but moving less data
In AI systems, high-bandwidth memory, HBM, has become nearly as important as the accelerators themselves. Large models must continuously transfer massive amounts of weights and activations; when data is far from where it is processed, energy and performance are wasted in transit. That is why the industry is pouring billions into memory stacks placed ever closer to processors.
Kepler claims it can increase memory density using three-dimensional stacking techniques and proprietary materials without relying entirely on the most advanced EUV lithography. For SRAM, the fast memory integrated near the core, the company uses ferroelectric materials capable of operating at lower voltages. The goal is simple to state and hard to achieve: more memory in the same footprint, less power to read and write it.
The interesting part is the use of existing fabs
The cost of a new advanced fab is enormous, and construction timelines can stretch over several years. Kepler aims to step in right here, claiming its technology can be adopted in existing facilities without requiring an entire new generation of EUV equipment. GlobalFoundries is collaborating with the startup and highlighted this specific aspect as one of the technology’s key points of interest.
If the approach works at industrial scale, the advantage would be not just technical, but strategic: expanding memory capacity without waiting for the entire industry to build new fabs. It is a particularly crucial promise at a time when HBM has become one of the most fiercely contested components in the AI supply chain.
The caveat is enormous: physics must become manufacturing
Semiconductor startups can demonstrate convincing performance on experimental wafers and still fail in the transition to production. Kepler has tested its technology on thousands of wafers, but it will need to prove uniformity, yield, reliability, and compatibility with industrial processes running across millions of devices. Introducing new materials into fabs also creates contamination and isolation challenges that must be managed with extreme precision.
The company itself expects initial HBM samples by the end of the year and aims for US production in 2028. These timelines show that the revolution, if it comes, will not be immediate.
Why Kepler matters even if it won't replace DRAM and SRAM
Kepler’s proposition does not necessarily have to wipe out existing technologies to succeed. In the AI market, improving just one tier of the memory hierarchy is enough to generate immense value, because every watt saved moving data can be converted into more computing power.
The most compelling takeaway, then, is that AI is reopening problems that for years seemed secondary to the processor. Memory, packaging, and the physical proximity between data and compute are becoming just as decisive as transistor counts. If Kepler succeeds in turning its material into an industrial product, it could prove that AI’s next leap will not come from a more powerful GPU, but from data having far less ground to cover.
