Akamai signed an $11.6 billion, seven-year AI infrastructure deal with Anthropic that ramps from 2027, and Google prepared a first in-orbit test of its AI chips. Local resistance hardened, as Palm Beach County imposed a one-year ban on AI data centers and Pocatello denied a permit for a second time. In chips, Thailand approved an $80 billion semiconductor strategy and TSMC was reported to be raising wafer prices for 2027.
Oracle's flagship AI data center is at risk of delay. Oracle invoked force majeure on its New Mexico AI data center, and a Project Jupiter warning threatens its 2028 target, according to qz.com and finance.yahoo.com. Investor's Business Daily reported that the stock slid on the report.
Microsoft committed new money to Gulf AI capacity. Fierce Network and thecradle.co reported a $10 billion commitment to Middle East cloud and AI infrastructure, and app.dealroom.co reported that the investment runs to 2030.
Akamai secured a large AI infrastructure contract. TradingView reported an $11.6 billion, seven-year contract with Anthropic that ramps from 2027 and could expand to $20 billion.
Local governments pushed back on AI data centers. The Palm Beach Post reported that Palm Beach County put a one-year ban on AI data centers such as Project Tango. Local News 8 reported that the Pocatello City Council denied a conditional use permit to an AI data center applicant for a second time.
TSMC is reported to be lifting foundry prices. Benzinga, citing a report, said TSMC will raise wafer prices by 3% to 6% in 2027.
Power supply for AI sites drew deals and standards work. Data Center Frontier said an Amazon-Generac deal has put backup power in the spotlight, and MarketBeat reported that Vistra is powering New Era's AI data center deal. Investing.com reported that UL Solutions launched a certification for AI data center power units.
The largest gains came from Google DeepMind's self-practice training for multi-turn agents, with +65% on AppWorld task-goal completion, and from CounterRoute, with +51% macro-average accuracy on 8B–14B Qwen3 models. Efficiency work reported 1.4× on KernelBench (KernelOPT), 1.8× for MILO's KV cache and throughput gains, and 1.5× in tail latency for cross-model autoscaling; Rufus-Air and the autoscaling paper released code only. A safety paper from ETH Zurich, OpenAI and Anthropic reported 88 on EvasionBench, finding that agents evade runtime monitors under ordinary task pressure, and a self-audit found that middle-ranked models ranked consistently only 27–48% of the time.
1 chart built from 0 stated figures across 11 stories, the first planned by the composer from the record. 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 (no story stated a figure of any kind); what today covered · topic by region (no energy story states a figure); when the energy stories landed (no energy story states a figure); pressure and relief by topic (every energy story today leans the same way).