Kepler Computing leaves stealth with $468M to fix AI's memory shortage
Two things landed overnight: a memory startup with serious money behind a claim that matters to every AI data center, and a computer-vision paper that wants to turn ordinary human video into robot skills.
Kepler Computing emerged from stealth with $468 million and a claim the AI industry desperately wants to be true: more HBM and SRAM density without EUV machines. The San Jose company, founded in 2018 by CEO Debo Olaosebikan and CTO Sasi Manipatruni, says a ferroelectric composite material plus a new 3D stacking technique lets it build memory at densities comparable to 2- and 3-nanometer parts on older, cheaper fab equipment. Investors include Intel Capital, AMD Ventures, Gates Frontier and GlobalFoundries, which put in $50 million and is already manufacturing with Kepler in Singapore and Vermont.
The numbers that make this more than a materials-science press release are the unglamorous ones. Kepler says it converted a fab line to next-generation status in eight months against a typical 24, and it claims it can sidestep the $20 billion to $40 billion it costs to build a new memory fab from scratch. That is the actual pitch: not a better memory chip in isolation, but more memory supply from factories that already exist, in a market where, as we reported earlier, AI demand sold out all 2027 DRAM and HBM capacity.
Here is the part worth watching. Kepler has run its process on roughly 2,000 wafers, will ship first HBM samples later this year, ramps in Singapore in 2027, and starts US production in 2028. GlobalFoundries executive Ed Kaste says the fundamental breakthroughs happened and what remains is "good results on thousands of wafers and millions of devices" — and he also notes the composite contains iron, which he calls "a tough contaminant to introduce into a production facility," meaning Kepler's process needs dedicated equipment or very tight containment. The Commerce Department has signed a letter of intent worth up to $245 million in CHIPS funding for the work, which signals belief but is not yet a check. New memory materials fail at scale far more often than they fail in the lab, and 2,000 wafers is a lab result with extra steps.
A new ECCV paper wants to close the gap between video that looks right and a simulation that actually works. HSImul3R, from researchers including Yukang Cao, Haozhe Xie, Liang Pan and Ziwei Liu, reconstructs human-scene interactions from casual captures — sparse photos or a single monocular video — and produces output a physics engine can run without the human sinking through the floor or the chair drifting off into space.
The trick is treating the physics simulator as a supervisor rather than a destination. In the forward pass, scene-targeted reinforcement learning adjusts human motion for both fidelity to the original video and contact stability; in the reverse pass, direct simulation reward optimization uses the simulator's feedback on gravitational stability to fix the scene geometry itself, sorting outcomes into four categories from "objects won't sit still" to "stable, meaningful interaction." The team also released HSIBench, a benchmark of 16-view synchronized captures across objects, subjects and motions.
Why it matters is the last line of the abstract: the reconstructed motions transfer to real humanoid robots. Robotics has a data problem that imitation learning keeps bumping into — there is essentially unlimited video of people doing useful things in real places and almost no paired action data. If reconstruction is physics-stable, YouTube becomes training data rather than just eye candy. That is a big if, and the sim-to-real gap is where confident claims go to die, but this is the most direct attempt yet at making human video deployable rather than merely viewable.
What to watch: whether Kepler's HBM samples show up before year's end, and whether any lab outside the authors' own institution gets HSImul3R to run on a real robot.
Which bet do you think pays off first — memory supply catching up with AI demand, or robots learning from video of people? Tell us in the comments.
Sources: HSImul3R (GitHub)
