TLDR Hardware 2026-10-28

TPU capacity will be deployed in OCI's liquid-cooled data centers, tied in through Oracle's Acceleron SmartNIC fabric. Google says its eighth-generation TPU delivers up to 12.6 PFLOPS of FP4 per chip at up to 2x performance per watt, sustains over 97% goodput with optical circuit switching, and cuts network hops by over 50% for up to 80% better low-latency serving performance per dollar.

Big Iron is an 86 m stainless-steel, road-transportable rocket that can carry 15,000+ kg to sun-synchronous orbit on nine 300,000 lbf methalox ‘Clydesdale' engines. The upper stage can stay in orbit as a data center with five racks of Vera Rubin GPUs, with the rocket skin doubling as the radiator and ammonia propellant as coolant. Its first flight is targeted for 2028, with up to 20,000 satellites planned pending FCC approval.

BlueChip runs a Snapdragon X Series chip built at TSMC's Arizona fab, with about 90% of its bill of materials from the US or allies and a roadmap to 100% domestic. The display is one component still made only in China. It's aimed at government and military buyers rather than consumers. The chip is expected to cost about 20% more than typical high-end PCs, and open for preorder in 2027. Anduril hasn't finished contracting its suppliers.

The push for silicon photonics in data centers requires testing infrastructure that matches evolving demands. Automated alignment and integrated optical/electrical testing are critical for overcoming challenges in coupling efficiency and throughput. Collaboration between equipment providers, developers, and manufacturers is essential to create a seamless production system for high-volume manufacturing.

A theory paper proposes reading an Andreev qubit without ancilla qubits: its state shows up as an oscillating supercurrent in a large superconducting loop, and a small coil coupled to a SQUID picks up the magnetic signal. The method still disturbs the qubit, but within standard error-correction thresholds. It's a proposal in Physical Review B, supported by FAPESP, with no experiment yet.

Renesas has launched the RX74M and RX74N, the first 32-bit MCUs using a 22nm process with embedded MRAM, reaching speeds up to 600MHz. These MCUs offer significant performance improvements and support major industrial networks like EtherCAT and TSN with integrated controllers. They enhance security with hardware-based features and support post-quantum cryptography, while a partner ecosystem aids development efficiency.

The fusion industry is close to delivering a roadmap to commercial power plants. The fuel cycle, especially tritium handling, is one of the most important development areas.

Intel has agreed to acquire Classiq, pairing its hardware-agnostic algorithm synthesis platform with Intel's silicon spin-qubit chips and cryogenic control electronics, which Intel says it alone can make at scale on conventional fab processes. The goal is a full-stack platform that adds QPUs alongside Intel's CPUs, GPUs, and ASICs. Intel plans to keep Classiq's multi-platform integrations.

OpenAMRobot is an open-source, modular mobile manipulation platform designed to lower the barrier to entry for embodied AI research and hands-on imitation learning.

The Luxonis OAK 4 FFC is a compact embedded vision board that pairs a 52 TOPS AI processor with five flexible camera ports on a single lightweight carrier.

Amazon FAR and university collaborators collect demonstrations without a robot, using two force-sensing UMI-style handheld grippers plus chest and ankle trackers.

Agility Robotics and GXO Logistics will expand the deployment of Digit humanoid robots to GXO sites in North Carolina and Tennessee.

Micron's roadmap puts wafer bonding in its 1ε DRAM node, with High-NA EUV and new materials in later ζÅ and ηÅ nodes after 1δ ramps in the second half of 2027.