Kepler Computing emerged from eight years of stealth work with a ferroelectric composite material that CEO Debo Olaosebikan and CTO Sasi Manipatruni say lifts compute-in-memory density 15 to 20 times past conventional SRAM and DRAM while cutting power draw, built in partnership with GlobalFoundries. The unveiling follows the Commerce Department's July letter of intent for up to $245M in CHIPS Act funding and sits on top of roughly $663M raised to date, including a $391M round that valued Kepler near $1.9B. It also lands as Western Digital says it is sold out of hard-drive capacity for all of 2026, with new fab supply not arriving before 2027. A material-science claim reaching this stage without a shipping product is a wager that memory, not raw compute, is now AI infrastructure's binding constraint. (Wired)
Private Companies
Alif Semiconductorother
Analog Devices agreed to acquire Alif Semiconductor for $1.35B in cash, with up to $200M in earn-outs. Alif, based in Pleasanton, California, builds microcontrollers and fusion processors for real-time sensor fusion and on-device AI inference, already shipping to industrial and consumer customers. The deal is Analog Devices' second AI acquisition this year after Empower Semi, pairing Alif's on-device processors with ADI's sensing and power-management lines to target systems that sense and react to the physical world in real time. (Reuters)
PsiQuantumother
PsiQuantum executed a definitive $100M award with the U.S. Department of Commerce, converting May's letter of intent into binding CHIPS Act funding for domestic quantum-hardware manufacturing. The award funds R&D on optical switches, single-photon detectors and packaging at PsiQuantum's California facilities, building on its GlobalFoundries partnership. It is the company's second active federal program this year, after its Department of Energy work with Brookhaven National Laboratory, and reflects Washington treating photonic quantum hardware as a supply-chain priority rather than a research grant line. (HPCwire)
Positronfundraise
Positron raised an $875M Series C at a $5B valuation, co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital and Jim Clark, the Silicon Graphics and Netscape founder. The round funds tapeout of Asimov, Positron's next-generation inference chip with up to 2,304GB of memory per chip, and production of Titan, a system pairing up to eight Asimov chips to serve models beyond 16 trillion parameters and context windows past 10 million tokens. Investors are betting memory-dense, CPU-adjacent chips can undercut Nvidia GPUs on inference cost and power. (WSJ)
Emerging
Simulating memory before silicon: A new paper, HBFSim, modifies Nvidia's PTX intermediate code to run real GPU inference workloads against a faithful simulation of high-bandwidth flash memory's timing, capacity and thermal behavior, validated against Qwen3-30B-A3B serving on vLLM. The tool runs roughly 20x faster than cycle-accurate simulation, letting designers evaluate flash-based memory tiers for LLM serving before committing to physical silicon, addressing the same memory-capacity bottleneck now drawing hardware bets from ferroelectric startups to hyperscaler storage teams. (arXiv)