M / EI
WorkInsightsAI RadarProfile
Contact
WorkInsightsAI RadarProfileContact

The race to
superintelligence.

01

Race intensity

Composite Intensity Formula

Weighted 0–100 index combining capital allocation speed, cluster compute activations, benchmark leap velocity, and regulatory interventions.

87/ 100Intensity index

Capability, capital and infrastructure are moving in the same direction. The gap between research and deployment is compressing.

Driven by: Capital allocation velocity, cluster compute deployment, frontier release cadence, and regulatory pressure.
02

Superintelligence countdown

Consensus Predictor Methodology

Synthesized from Metaculus AGI/ASI Aggregate (median ~May 2029), Epoch AI Frontier Scaling Trajectories, and prediction market consensus (Kalshi & Manifold). Probability window spans from 2027 (25th percentile) to 2031 (75th percentile).

948Days
:
13Hours
:
29Mins
:
58Secs
Median Target: May 2029Metaculus · Epoch AI · Kalshi Liquidity-Weighted
2027 (25th pct)2029 (Median)2031 (75th pct)
Consensus driver: Test-time inference compute search trees and gigawatt power cluster commissioning accelerate recursive self-improvement timelines.
03

Capability trajectory

Event-Driven Milestone Curve

Points represent pivotal shifts: architecture breakthroughs, 100k+ GPU cluster activations, and autonomous agent cyber discoveries. Hover over any node to inspect details.

FrontierAgentsMultimodalReasoning
Now −12mNow
Frontier velocity
Frontier Velocity (+34%)

Annualised progression rate across frontier reasoning (Artificial Analysis intelligence index, GPQA, Math 500) over the past 12 months.

+34%12m capability gain
Agent reliability
Agent Reliability Multiple (2.8×)

Multi-turn task success multiple without human intervention across SWE-bench Verified, OSWorld, and autonomous developer terminal sandboxes.

2.8×Long-horizon task multiple
Inference pressure
Inference Pressure (Critical)

Surge in test-time compute FLOPs required per answer. Models run hundreds of chain-of-thought search rollouts, stressing real-time GPU cluster occupancy.

CriticalTest-time compute load

Who is moving
the frontier?

Weighted Scoring Formula

30% Reasoning & Benchmarks · 25% Compute & Power · 15% Distribution · 15% Capital · 15% Talent Density. Calibrated against live Artificial Analysis evaluation leaderboards.

01

OpenAI

91+2.4Lead group▼
02

DeepMind

88+3.1Accelerating▼
03

Anthropic

84+1.8Lead group▼
04

xAI

71+4.6Fast mover▼
05

Meta AI

67+0.9Open frontier▼
06

Alibaba (Qwen)

65+2.8Global challenger▼
07

DeepSeek

63+3.5Efficiency pioneer▼
Compute committed
Compute committed

Global hyperscaler cap-ex and sovereign cluster infrastructure committed through 2026.

Source: Hyperscaler Financial Disclosures
$1.4TGlobal hyperscaler cap-ex and sovereign cluster infrastructure committed through 2026.
Frontier labs
Frontier labs

Organisations training models beyond 10^26 training FLOPs under daily observation.

Source: Primary Lab Technical Disclosures
07Organisations training models beyond 10^26 training FLOPs under daily observation.
Critical signals
26 Critical Signals Tracked

8 Capability Breakthroughs, 6 Infrastructure Deals, 4 Capital Allocations, 4 Distribution Moves, 4 Constraints & Regulatory Gates.

2626 pivotal capability, compute, capital, and regulatory milestones logged in active observation window.
Race state
5-Stage Trajectory Gauge

Stalled: Sub-linear benchmark growth and compute deficit.
Plateau: Scaling laws saturate without architecture change.
Cooling down: Cap-ex slowdown or regulatory pauses.
Acceleration: Test-time search scaling and gigawatt cluster additions.
Breakthrough: Autonomous self-improving agents.

STALLEDPLATEAUCOOLING DOWNACCELERATIONBREAKTHROUGH
Test-time inference compute scaling and multimodal synthesis accelerating capability velocity.

What changed
the race?

01/ 04
CapabilityOpenAI ResearchSep 20, 2026 · 05:00 PM UTCOpenAI · Anthropic · DeepMindHigh impact

Reasoning gains are moving from benchmark theatre into useful agent behaviour

The important contest is shifting toward reliability across long, tool-using tasks rather than a single headline score. Frontier labs are transitioning inference compute from static single-pass answers into multi-turn tree-search evaluations.

Why it matters:Frontier models with inference-time compute scaling demonstrate step-function gains in complex multi-step coding, autonomous workflow verification, and scientific synthesis.
Source: OpenAI Research↗
InfrastructurexAI AnnouncementsSep 18, 2026 · 12:00 PM UTCxAI · Microsoft · AlphabetCritical impact

Compute is becoming an energy, capital and deployment race at the same time

Frontier advantage increasingly depends on securing the whole stack, from dedicated sub-station energy grids and liquid-cooled data halls to low-latency interconnects and inference economics.

Why it matters:Access to gigawatt-scale dedicated power infrastructure and specialized silicon interconnects is now the single largest physical bottleneck determining training cluster expansion.
Source: xAI Announcements↗
Power moveAnthropic AnnouncementsSep 19, 2026 · 04:00 PM UTCAnthropic · OpenAI · Meta AIWatch impact

Talent concentration remains a leading indicator of the next capability jump

Small movements in key research teams can matter far more than broad corporate headcount announcements. The race is dominated by dense clusters of reasoning and post-training architects.

Why it matters:Breakthroughs in post-transformer architectures and test-time reasoning algorithms remain concentrated among a few dozen key researchers capable of orchestrating massive compute runs.
Source: Anthropic Announcements↗
ConstraintUS AI Safety InstituteSep 16, 2026 · 03:00 PM UTCUS NIST · UK AISIHigh impact

Evaluation is struggling to keep pace with increasingly agentic systems

The race urgently demands rigorous measures for continuous autonomy, safety boundary adherence, multi-system control, and non-deterministic real-world task completion quality.

Why it matters:Static multiple-choice benchmarks are saturated; global evaluation is rapidly shifting to dynamic cyber, software engineering, and scientific sandbox environments.
Source: US AI Safety Institute↗
EditionPrototype / 001
Static fallback dataset