The Board Room

Google disclosed it matched 100% of its 2025 electricity with renewable purchases but ran only about 65% carbon-free hour by hour. In July 2026 the GHG Protocol proposed splitting emissions reporting into three statements — physical, market-based, and the impact of corporate actions — with the joint standard expected in Q4 2028. Every sustainability commitment you built on a vendor's single number now has a published expiry date. Berkeley Lab data show only 13% of capacity that entered US interconnection queues from 2000 to 2020 was operating by the end of 2025.

Scans of government credentials belonging to an estimated 170 million people in the US and Canada were being sold on a dark web site called Nexus, including over 153 million driver's licenses and 1.9 million travel documents. Brian Krebs traced the common denominator to a single New Orleans third-party ID-verification provider whose own pitch advertises models 'trained on hundreds of millions' of documents; the FBI's New Orleans field office is investigating and the vendor has said nothing. Roughly 400,000 new records appeared in a single 24-hour window, so the pipeline was live rather than historical.

On Meta's Llama 3.1 70B, one 128K-token session needs roughly 40GB of key-value cache — the per-session memory a model burns while generating — and four concurrent sessions put about 200GB on the table before model weights. Architecture sets that floor: AI21's Jamba reports 4GB at 256K context against Mixtral's 32GB, and DeepSeek claims V4-Flash runs at 7% of V3.2's cache at one million tokens. Serving flags move a 2–8x band inside a floor someone else already set, which makes model selection a multi-year gross-margin decision.

Short sellers and auditors have converged on cash conversion. J Capital's report on Karman Holdings, a $5.3B defense and space parts roll-up trading at 70x forward, alleges no fraud but flags strong profits arriving with no positive operating cash flow. The FT reports KPMG flagged internal-control deficiencies at a subsidiary of Guggenheim's $367bn asset manager over how it recognized hundreds of millions in revenue. Private credit prices your growth debt, and its valuation practices are contested in court.

The reliability community shifted from arguing what AI can automate to publishing what it must never automate: severity assignment, customer-impact calls, and review sign-off. One curated finding puts roughly 40 minutes per incident into negotiating the severity field during active response. Netflix separately documented running an LLM judge as a monitored production system, calibrated weekly on about 300 explanations with three or more human raters each. If your AI-in-operations mandate has no written non-adoption boundary, your senior engineers will supply one first.