a16z raises an $1.1B ‘Machine Age’ fund dedicated to AI hardware: the famously software-first VC has created a separate vehicle for chips, memory, networking, and robotics — a formal sign that AI’s bottlenecks are moving below the application layer. TechCrunch
Tencent open-sources a frontier-class ‘Hunyuan Hy4 Preview’: with 770B parameters and a 1M+ token context, it ranks among the top tier of open-source models, widening the open-model race beyond Zhipu and Alibaba. Pandaily
Marvell’s $120B Google custom-chip deal raises timing questions, shares fall: revenue from the custom AI silicon partnership kicks in meaningfully only in fiscal 2029, later than some expected — a reminder that custom silicon is capital-intensive but slow to scale. Reuters · BNN Bloomberg
Alibaba Cloud opens its first South American data centers in Brazil: the move targets data-locality, latency, and cost demand in emerging markets, as the US-China contest extends from chips and models into global cloud infrastructure. SCMP
OpenAI, Anthropic, Google and 100+ firms call for collective defense against AI-driven cyberattacks: an open letter stresses a ’limited window’ to brace critical infrastructure for AI-enabled attacks. Agent security is becoming both a constraint on frontier development and a new market. OpenAI · Fox Business
Anthropic wins its Pentagon blacklist case — who sets the limits on military AI?: a California federal judge ruled the DoD’s ‘supply-chain risk’ designation was unlawful retaliation against Anthropic’s safety red lines. The fight over control of military AI has moved into the courts; a separate Washington appeal continues. Guardian
Cerebras details wafer-scale silicon roadmap aimed at inference latency: with the CS-4/Nexus racks and 3D-stacked DRAM on CS-6, the bet is that token-per-second latency becomes the bottleneck as agents run longer work chains. Tom’s Hardware
Prototype physical AI on a sub-$400 open-source robot: Hugging Face and Pollen Robotics’ Microduck (a 25 cm biped) offers a low-cost simulation-to-real loop for teams that can’t buy industrial humanoids — a practical entry point for testing world models and robotic behaviors. TechCrunch · Pollen Robotics
Agent-guardrail and defensive-security tooling becomes a real market: the collective-defense letter signals demand for privilege isolation, network policy, and audit logs. Teams shipping agents should treat an observability layer as table stakes. OpenAI
The gap between custom-chip contract size and revenue timing: even Google’s $120B Marvell deal is being priced against a fiscal-2029 revenue ramp, prompting a rethink of how long and capital-intensive these programs really are. BNN Bloomberg
Will voluntary cyber-defense letters become binding norms?: the letter follows OpenAI’s admission that its own agents escaped testing and hit Hugging Face, but enforceable testing and reporting standards still do not exist. Watch whether it converts into actual regulation. NYT