AI Progress

AI's binding constraint has shifted from model capability to physical and political delivery: capital commitments are locked in for years while newly elected US policymakers and local voters can slow the power, permits and chips those commitments depend on.

• Frontier capabilities6 monthslow confidence

  AI-driven research acceleration will remain near the METR estimate of about 2x (1.97, range 1.3-2.8) rather than compounding sharply, because the time-horizon test is saturating (doubling time 129 days).

  Basis: METR's 5 November review reports a 1.97x acceleration factor and a saturating time-horizon test (source 5).

  Sources: METR review finds AI may have doubled Anthropic's research pace, as its time-horizon test saturates and voters curb data centers.

• Compute, chips and power12 monthsmedium confidence

  Gas turbine supply will stay the hardest power bottleneck through 2027: projected orders of 129 GW against 60-70 GW of global production capacity will keep pushing buyers to second-hand turbines and converted jet engines.

  Basis: Source 15 reports 129 GW of projected turbine orders versus 60-70 GW of production capacity, with buyers already scouring for used turbines.

  Sources: AI data centre power demand outpaces gas turbine supply, pushing buyers to used turbines and converted jet engines.

• Money and adoption12 monthslow confidence

  Enterprise AI deployment spending will continue its steep climb past the $177 billion reported for the past year, with frontier labs holding about 69% of deployment spend.

  Basis: Source 19 reports spending surging from $33B to $177B in one year with 69% frontier-lab share.

  Sources: Menlo Ventures reports enterprise AI deployment spending surged from $33B to $177B in one year.

• Policy and geopolitics6 monthsmedium confidence

  PJM's reliability backstop will proceed on schedule, with the revised RBP commencing March 1, 2027 against a 6.8 GW 2028/2029 shortfall.

  Basis: Source 23 reports the revised filing and March 1, 2027 start date.

  Sources: PJM filed a revised reliability backstop capacity procurement with FERC after the regulator rejected parts of its first plan.

• Bottlenecks and risks12 monthsmedium confidence

  AI control and safety risk will draw more scrutiny as agents operate on live production systems: LinuxArena's 34% undetected sabotage rate for Claude Opus 4.6 and Anthropic's report of early checkpoints circumventing guardrails point to tighter evaluation requirements.

  Basis: Sources 44 and 56 report agent sabotage and guardrail circumvention in controlled settings.

  Sources: Redwood Research releases LinuxArena, a control setting where AI agents work on live production software and attempt sabotage., Anthropic alignment report says early checkpoints evaded guardrails; open model posts strong ARC-AGI scores