Oratomic closed a $475M Series B, bringing its total funding since its March launch to $775M, in a round co-led by ARCH Venture Partners, Spark Capital, Khosla Ventures, Index Ventures, General Catalyst and Bezos Expeditions. The Pasadena company builds fault-tolerant computers from reconfigurable arrays of neutral atoms and says it can reach high fidelity with about 10,000 qubits, against the roughly 1M that has been the industry benchmark. Targets that low move the constraint from qubit volume to error rates and the control systems that drive atom arrays. (Oratomic)
Private Companies
DARPA QBIother
DARPA moved Atom Computing, Diraq, IBM and IonQ into Stage C, the final phase of its Quantum Benchmarking Initiative, where government evaluators test whether each proposed utility-scale system can be built to its design. The four span neutral-atom, silicon spin-qubit, superconducting and trapped-ion approaches and join Microsoft and PsiQuantum in the final phase. IonQ says its agreement is worth up to $300M through 2029, contingent on appropriations. DARPA's Micah Stoutimore said he increasingly expects someone to build a utility-scale machine by 2033. (GovConWire)
Public Markets
Marvell lifted its fiscal 2028 revenue outlook to about $20B against $18.2B expected, raised its fiscal 2029 custom chip target to $12B and set fiscal 2031 revenue at $70B to $90B. The stock rose about 6% on the October 6 investor day. (Reuters)
Samsung guided third-quarter operating profit to about 107.4T won ($80B) on sales of about 195T won, 1.2% above consensus. The divisional split due October 29 will show how much came from HBM. (Samsung)
Emerging
Faulty hardware: A paper from Trevor McCourt, Ila Fiete and Isaac Chuang, Fault-tolerant foundation models, trains language models on simulated faulty digital hardware across 40,000 GPU-hours and finds error resilience rises with model size. The authors conjecture that large models learn to compute inside error-correcting codes whose overhead stays finite. If that holds, inference could run on low-energy, faulty hardware, which would let designers trade reliability for energy savings. (arXiv)
Agent-designed chip: Zettascale Computing, a trade name of Exa Laboratories, unveiled XPU Grasshopper, an AI chip co-designed with autonomous agents, and opened pre-orders for a 100-board FPGA devkit shipping in February 2027. A four-tile build on an AMD FPGA runs gemma-3-1b-it at 51 tokens per second, 816x its June 30 first run and above the 41.3 reported for Nvidia's Jetson Orin Nano Super, though at different quantization. The gain shows how fast agent-driven iteration can move a design, not parity with shipping silicon. The Gen 1 design targets TSMC N7. (Zettascale)