Global AI — Applications, Geographies, and Market Trajectory
Executive Summary
By May 2026, AI has gone from a defensible technology category to the dominant axis along which the global economy is reorganising itself. Worldwide AI spending will reach roughly $2.5 trillion this year; hyperscaler capex has nearly doubled in twelve months; and the four largest US technology companies will pour somewhere between $660B and $725Binto infrastructure alone — most of it for AI. At the same time, only about 6% of organisations attribute meaningful EBIT impact to AI, 95% of generative AI pilots fail to graduate to production, and the gap between capex and recognised revenue has widened to roughly 10:1.
The defining tension of the moment is the distance between an extraordinary supply-side build-out and a slower, messier, more selective demand-side adoption.
The eight things that matter most
- 1. Anthropic overtook OpenAI on enterprise revenue
$30B+ run-rate by April 2026 (from $1B Jan '25). 40% enterprise spend share vs OpenAI's 27%. Claude Code alone at $2.5B ARR.
- 2. The model layer is bifurcating
Frontier reasoning (GPT-5.5, Opus 4.7, Gemini 3 Pro, Grok 4.1) stays expensive & closed. Open-weights bench (DeepSeek V4, Qwen 3.6, Kimi K2.6, Llama 4) matches them at ~10× lower cost.
- 3. Coding is the only category with proven unit economics
55% of departmental AI spend. Labs now compete with their own customers (Cursor, Cognition). Legal next: Harvey $190M ARR; Legora fastest-ever to $100M.
- 4. Sovereign AI is the defining geopolitical force
UAE, Saudi, India, Korea, Japan, FR, UK, EU, Brazil all building national capacity. Only China attempts full-stack independence — most 'sovereign' efforts run on NVIDIA + US clouds.
- 5. Agent layer is real but earlier than the marketing
Agentforce $1.4B ARR. MCP at 78% enterprise adoption. But only 31% of orgs have an agent in prod. 88% of agent pilots stall.
- 6. Capex has decoupled from revenue
$725B infra spend vs ~$51B direct AI revenue in 2026. Goldman calls it 'elongation' — buildouts stretch, gap widens, no one blinks.
- 7. Pricing is genuinely unsettled
Salesforce runs 3 models in parallel: per-seat, per-conversation ($2), Flex Credits. Intercom Fin's $0.99-per-resolution is the new primitive. IDC: 70% of vendors refactor away from per-seat by 2028.
- 8. The largest underbuilt areas aren't where capital is
Vertical AI for regulated industries, blue-collar/field work, non-English markets, agent identity & governance, AI-native services that sell completed work.
1 · Market Sizing
Gartner projects $2.52T worldwide AI spending in 2026, up 44% YoY, on track for ~$3.3T in 2027. AI infrastructure alone accounts for ~$401B; AI-optimised servers grow 49%; generative AI model spend grows 80.8% — the highest of any line item. IDC forecasts $22.3T cumulative global economic impact from AI by 2030. The spend is overwhelmingly concentrated in five companies: Microsoft, Google, Amazon, Meta, and Oracle, with xAI/SpaceX as the sixth pillar.
2 · Frontier Labs
Anthropic's run-rate exploded from ~$1B (Jan '25) to $30B+ (Apr '26), passing OpenAI in monthly revenue while spending ~4× less on training. The Chinese frontier — DeepSeek V4, Qwen 3.6, Kimi K2.6, Doubao Seed 2.0 — matches Western frontier on many benchmarks at ~10× lower cost. Production architectures are increasingly routers: cheap open-weights for ~80% of work, expensive closed models for the rest.
3 · Infrastructure & The $700B Question
Hyperscalers will spend ~$725B on infrastructure in 2026 against perhaps $51B of direct AI revenue. Goldman calls the dynamic "elongation" — buildouts stretch, costs rise, the gap widens, but commitment holds because no one wants to be the company that blinked. Power is becoming the binding constraint everywhere: 103 GW global AI capacity today, projected to ~200 GW by 2030.
4 · Enterprise B2B SaaS
Enterprise GenAI spend totalled ~$37B in 2025 (3.2× YoY). Coding alone was $4.0B (55% of departmental AI spend). AI-native NRR is dramatically lower than incumbents — median 48% vs 82% for traditional SaaS — meaning product moats matter more than model moats. Salesforce runs three pricing models in parallel; IDC predicts 70% of vendors refactor away from per-seat by 2028.
5 · Verticals
Healthcare ambient documentation is the most-validated AI use case in any vertical. Legal is next: Harvey at $190M ARR / $11B; Legora became the fastest enterprise SaaS ever to reach $100M (18 months). Construction, manufacturing, insurance, and compliance are all crossing the production threshold but with high pilot-to-production drop-off.
6 · Geographic Landscape
US dominates new infrastructure (≥80% of new datacenter construction). China dominates open-weight deployment and humanoid robotics manufacturing. Europe is rebuilding around sovereign infrastructure under the EU AI Act. UAE, Saudi, India, Korea, Japan, France, UK, Brazil run parallel sovereign programmes — though most still run on NVIDIA + US clouds (Bloomberg's "NVIDIA's sovereign AI irony"). True full-stack independence is being attempted only by China.
7 · Predictions
Next 18 months: agent maturation real but uneven; on-device AI inflection (iOS 27 model marketplace); reasoning models do NOT commoditise on price; voice becomes primary interface. 2027–2030: pricing refactors away from per-seat; sovereign fragmentation deepens; agent identity & governance becomes a category; data centre power binding constraint. 2030+: AI-native services replace SaaS in many categories; China achieves full-stack independence (probability ~60%).
8 · Whitespace
Capital is concentrated in foundation models, coding, and enterprise search. The whitespace is in vertical AI for regulated industries (insurance ops, K-12 admin, municipal gov, SMB trades), AI for blue-collar / field work, AI in non-English markets, agent identity & governance, and AI-native services that sell completed work rather than software.
9 · Risks
Top risks: power & grid constraint (82), capex/revenue bubble (78), agent reliability failure (72), China decoupling (70), regulatory fragmentation (65), talent flight (60), model commoditisation in lower tiers (55), antitrust scrutiny on reverse-acquihires (50).
10 · Recommendations
- Build / buy in vertical AI for regulated industries before capital crowds in.
- Position around agent governance, identity, evals, and observability — the layer beneath the headline agents.
- Treat per-seat as a wasting asset; design pricing for outcomes, conversations, or credits.
- For sovereign markets: partner with national champions (G42, HUMAIN, Sea, Mistral) early; do not expect to win them late.
- Keep frontier model exposure optionable — route across Anthropic, OpenAI, Google, and the Chinese open bench.