Delos Data, a Palo Alto startup founded by Intel networking veterans Ed Doe and Dan Daly, raised over $100M in a round led by Matrix Partners and Playground Global — with former Intel CEO Pat Gelsinger joining as an investor — to build a clean-sheet interconnect for AI data centers, spanning a 30 Tbps I/O chiplet, a 10 Tbps near-packaged-optics card, and software to knit them into a single data-movement fabric. The bet is as GPU clusters scale past what existing NICs and switches can feed, the network rather than the accelerator becomes the binding constraint on inference throughput, and Delos is pitching itself as vertically integrated hardware-and-software plumbing rather than another chip competing on FLOPS. (PR Newswire)
Private Companies
TeRAMfundraise
TeRAM emerged from stealth with a $37M seed round led by Primary Venture Partners, B Capital, Hyperion, and SemiAnalysis Capital, with Alumni Ventures and Lightscape Partners also participating, to build 3D SRAM stacked directly onto AI compute chips. CEO Charlie Cheng's team is targeting the memory-bandwidth wall inference workloads are hitting as agentic AI shifts demand toward serving rather than training; the company is targeting initial customer production in 2029. (Semiconductor Digest)
Public Markets
AMD led a sector-wide rebound as chip stocks recovered from Monday's AI-pacing selloff, with investors returning to Instinct GPU and EPYC server exposure after Wall Street reiterated bullish ratings on the company's data-center roadmap. (IBTimes)
Broadcom lagged the sector's rebound, still weighed down by the below-consensus fourth-quarter revenue guidance it issued alongside a record $29.6B third-quarter print, even as its AI accelerator business kept growing. (NAI 500)
Emerging
Private inference gets cheaper: A new paper, OptiPrime: Optimizing Private Inference through Protocol-Hardware Co-design, redesigns the encryption protocol for private neural-network inference to sharply cut the data volume secure computation has to transmit, paired with a 10x reduction in weight-memory traffic. Accepted to MICRO 2026, the 59th IEEE/ACM International Symposium on Microarchitecture, it delivers up to 5.7x faster private inference than the standard Cheetah baseline on CPUs and 4.2x with dedicated accelerator hardware — evidence that confidential AI inference, long bottlenecked by network overhead rather than compute, is closing the gap with plaintext serving. (arXiv)