Every metric below has either inflected or accelerated since our Q3 2025 edition. No gauge on this page has materially softened.

For GPT-3.5-class capability on MMLU, $/M tokens has fallen from $20.00 in Nov 2022 to $0.011 in Mar 2026. Nothing in the history of computing compares — not Dennard scaling, not photovoltaics, not bandwidth.

Two years ago this heatmap had deep cold spots. Today every cell is in use. The question has moved from where to how well.

On six of the most-watched capability tests, frontier AI has gone from student to examiner in under two years. The vertical line below is expert human performance. Most benchmarks now retire a year after introduction.

Four compounding forces drive the pace. Each one feeds the next. Together they produce an annual capability multiplier of ~4.3× — and rising.

The 2024 stack had separate text, vision, audio, and video models. The 2026 stack has one multimodal token stream. Latency collapsed; context expanded; reasoning unified.

The fusion of generative creation with agentic execution is the engine of 2026 enterprise value. Systems no longer just draft — they plan, act, monitor, and correct.

Still only 8% of organizations have embedded AI deeply enough to reach 4×+ ROI. The “pilot purgatory” cohort has shrunk — but the middle bulge, not the top, has grown most.

Sovereign AI is the strategic endeavor to develop, deploy, govern, and control AI capabilities aligned with national interests. The dimensions: technology, culture, economics, security.

No single region leads on all dimensions. The US dominates frontier capital; China dominates patents and public investment; the ME has the fastest-growing new entrants; India leads on growth pace from a smaller base.

Europe's sovereign LLM, live since Nov 2025. A concrete proof-point that a values-first, regulator-compatible frontier model is technically and economically viable — if politically funded.

An agentic AI platform operating at scale across 312 enterprises. The playbook behind the 4.1× 12-month ROI — and how the perceive-plan-execute-learn loop actually closes in practice.

The gap between experiment and deployment has closed. 89% of surveyed enterprises now run AI in customer service — up from 62% a year ago. Legal remains the final frontier, but is moving.

NVIDIA still owns ~62% of the frontier training market, but concentration is slowly easing. Google's in-house Ironwood-2 TPU and Huawei's Ascend 920 are the two credible challengers — for different customers.

Quantum advantage for ML remains 3–5 years out. But neuromorphic silicon is already delivering 100× energy efficiency for specific inference workloads — in production at Intel, IBM, and several defense primes.

AI data centers will consume 940 TWh in 2026 — roughly the electricity use of Germany. By 2028, projections put the sector at ~2,000 TWh, or the entire electricity consumption of India today. No credible decarbonization path has been published.

Renewable share varies wildly by region. The EU leads at 84%. The Middle East, despite massive solar resource, remains at 62% because sovereign clusters are running 24/7 on natural gas for reliability. China lags at 42%.

Edge inference has quietly become the dominant form of AI compute by query count. Every new iPhone, Pixel, Galaxy, and Quest ships with frontier-lite models running locally — private by default, latency-free, offline-capable.

Every major IP regime has been forced to take a position on AI training data and AI authorship. The positions are now diverging enough that multinational enterprises need a regional IP playbook, not a single global policy.

The dominant 2024 risks — hallucination, prompt injection — remain. But 2025 added agent autonomy misalignment and CBRN uplift as frontier-class concerns. Mitigation coverage varies wildly.

Every AI ethics framework converges on roughly the same five principles. The test is whether they're operationalized — reported, audited, budgeted — or merely aspirational. Most orgs still fail this test.

Every AI regulator is tuning the same four dials — risk-aversion, innovation, sovereignty, and ethics. Where they land defines compliance, market access, and strategic freedom.

The counter-narrative to 2025's scandals: this year's most important AI applications were not chatbots. AlphaFold-4 mapped ~97% of the known proteome. Grid-optimization AI cut measured emissions by 32% across 142 nations. Early-warning systems for floods, cyclones, and wildfires deployed across 28 regions via UN OCHA.

The gap between 8% (high maturity) and the rest is not a tooling gap — it is a discipline gap. Every high-ROI deployment we studied hews to these five principles.

For all the disruption, 2025–26 produced the largest annual leap in AI-for-good deployment ever measured. The dividend — if captured — could compound faster than the risk.