The Board Room

Anthropic ran the classic hidden-profile test on four-agent AI systems, and the groups picked the right answer in only 17–36% of runs, per Exponential View's analysis. One agent given the whole evidence base got it right nearly every time. Separately, twelve frontier models from rival labs in the US, Europe and China returned substantively the same answer to one identical prompt. Any control in your framework that says "ask a second model" is documenting correlated error as verification.

Nvidia is raising prices roughly 17% on Grace Blackwell 300 and Vera Rubin 200 systems for 2027 delivery, and server OEMs have begun notifying customers, The Information reports. The increase adds at least $5 billion to one gigawatt-scale data center, implying chip systems alone already cost $29–30 billion per gigawatt. Any three-year plan built on falling cost per unit of compute is now wrong by double digits. The reporting rests on two anonymous sources and covers only some flagship systems.

Six in ten of about twenty CTOs and VPs of Engineering surveyed by Gergely Orosz are leaving or seriously weighing it. The stated causes are not AI itself but what AI licensed executives to demand: 20–50% cost cuts and transformation timelines nobody believes. Equity has stopped retaining them — one CTO's 2% stake pays nothing below a $210 million exit, because $110 million was raised at a 2x liquidation preference. Flatter orgs leave fewer seats to move into, so the attrition is queued, not avoided.

Exponential View's State of AI work puts token price elasticity at 1.2–1.8, so a 10% price cut lifts usage only 12–18%. The Jevons rebound baked into most AI revenue forecasts has not arrived, which makes a price cut a transfer to the buyer rather than a volume unlock. Adoption is also narrow: since October 2023 the top 1% of firms raised AI spend per employee by $6,542 while the median firm raised it by $9.63. Software engineering leads because its output is already metered.

Harness-Bench ran one model across 106 identical tasks in different harnesses and scored 52.4 to 76.2 — a 23.8-point spread with no change to the weights, per Latent.Space. Roughly half of agent performance is now engineering you own rather than pretraining you rent. The catch is depreciation: trained tool use and context compaction have already migrated into model weights, and tool selection, orchestration and memory are named as next. Harness code is operating expense with a useful life, not IP.