Megaport signed $687 million of AI infrastructure deals and raised its forecast, and AMD agreed to acquire Fei-Fei Li's World Labs. Samsung's 2nm yield was reported approaching 60% as China weighed allowing Nvidia chip purchases. Local resistance hardened, from a Tasmanian council move against a Firmus site to residents calling $10,000 checks a bribe.
Samsung is putting money into AI infrastructure outside its own fabs. Samsung Electronics and affiliates committed $1 billion to Helix, an AI infrastructure firm backed by KKR and Nvidia, according to WSJ, Bloomberg and The Korea Herald.
Anthropic's infrastructure ambitions set a new reference point for model-developer spending. CoinDesk reported Anthropic plans to spend $518 billion on AI infrastructure, while pre-IPO perpetual contracts barely moved.
Network provider Megaport booked a large block of AI work. Reuters reported the Australian company signed $687 million of AI infrastructure deals and lifted its annual forecast; other reports said it raised its 2027 earnings forecast.
Samsung is closing in on foundry competitiveness at the leading edge. Its 2-nanometer yield was reported approaching 60%, positioning it against TSMC for orders from large technology customers.
Macro concerns about the build-out gained airtime. IndexBox put US AI data center spending at $10.3 trillion through 2032 and said it could overheat the economy, while Fed Governor Cook said AI infrastructure spending is pushing up inflation first.
Suppliers moved on cooling and power for AI halls. LG Electronics expanded AI data center cooling ties with Microsoft and SK Enmove, Navitas agreed to acquire Claros for data centre power delivery, and Bloom Energy was reported soaring on AI data center orders.
VarioPath (USTC, Alibaba) cut AlltoAllv latency 5.9× for MoE inference on PCIe clusters of up to 256 GPUs, KV-streams (Meta AI, UC Berkeley) reported a 3.8× wall-clock training speedup, and Alibaba's PEARL lifted rollout throughput 2.5×; Nereus reported 2.1× on 8B PPO post-training. None of the eight papers released code or weights. Cambridge and Edinburgh's AgentPerfBench argues existing inference benchmarks mismatch real agentic workloads.
2 charts built from 1 stated figure across 16 stories, the first planned by the composer from the record. Also had the data for when the energy stories landed, pressure and relief by topic. Held back: power committed in deals that state a figure (no deal in this edition states the capacity it commits); power disclosed per edition · last 7 (only 1 of the last 7 editions put a number on it); other figures on the record today (only 1 other figure on the record — the card needs 3).