Jeff Dean, Google's chief scientist and 30th employee, is leaving after 27 years to co-found Discovery Loop with three other senior Google AI researchers — Sanjay Ghemawat, Quoc Le, and Oriol Vinyals — betting that automating the experimental loop of scientific research, running thousands of algorithmic experiments in parallel instead of one at a time, is the next unlock for AI progress. Radical Ventures and Khosla Ventures are co-leading the seed round, and Google itself is in as a founding investor and cloud partner, meaning the compute the new lab needs gets bought on Google's own infrastructure rather than a rival's. It's the same shape as Ilya Sutskever's Safe Superintelligence and Mira Murati's Thinking Machines — senior research talent unbundling from a hyperscaler's org chart while the hyperscaler stays inside as investor and vendor, adding a new, well-financed buyer to a compute market where supply is already the binding constraint. For Dean's own case that efficiency, not scale, is where AI hardware goes next, his February conversation on energy-bound inference is worth the hour. (TechCrunch)
Private Companies
Lumilensfundraise
Lumilens raised over $700M in a Series C led by Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures, and Spark Capital, valuing the two-year-old optical-interconnect startup at $5.51B and pushing its total raised past $900M. Founder and CEO Ankur Singla — previously of Contrail and Volterra — says Lumilens is already shipping into a hyperscaler's live data centers under an unaudited multi-billion-dollar deal, replacing the copper wiring that caps chip-to-chip signaling at roughly 1.5 meters with light instead. (WSJ)
Public Markets
Michael Burry has deepened his short against Micron even as buyers push back on record memory pricing and Chinese suppliers CXMT and YMTC signal new capacity — a live test of whether the AI memory cycle has already peaked. (Yahoo Finance)
Record Q2 revenue of $3.3B, up 40% year-over-year, and Q3 guidance both cleared consensus, with CEO Jayshree Ullal citing 800G switching demand as hyperscalers rebuild networks around AI clusters — evidence networking capex is keeping pace with compute. (The Motley Fool)
Emerging
Efficient inference: A paper posted July 29 introduces LLMET, a framework for evaluating monolithic 3D (M3D) memory as an on-chip substitute for the DRAM lookups that drive LLM-serving energy costs. Swapping a 40MB L2 cache for a 1GB M3D cache cut chip energy 44% during Llama 3.1-70B prefill on paired Nvidia A100s, and a modeled 8-GPU B200-class cluster still saw up to 24% prefill savings from the same swap — a direct challenge to the assumption that HBM is the only lever left for serving efficiency, arriving the same week rising memory prices became a flashpoint for AI infrastructure investors. (arXiv)
Power siting: Amazon has formally withdrawn its plan for a roughly 500MW data center campus next to Maryland's Calvert Cliffs nuclear plant, abandoning a project that would have paired hyperscale compute directly with dedicated nuclear power from Constellation Energy. The reversal follows a June primary in which Calvert County voters ousted the three commissioners who had opposed a construction moratorium — a sign that community and political pushback, not grid queues or turbine lead times, is becoming the binding constraint on where AI power gets sited next. (Maryland Matters)