The State of AI

Executive Summary

Strategic Research Report — February 2026

The State of AI

Intelligence per dollar has improved 1,000x since 2020. Infrastructure will commoditize. Value will migrate to applications. The window to act is 3–5 years.

1,000x

Intelligence yield improvement since GPT-3

3–5 yr

Strategic window before commoditization

$600B+

Enterprise software exposed to disruption

Chapters across 6 parts with interactive data

This report presents a single, evidence-based thesis: the economics of artificial intelligence are following the same trajectory as electricity, telecom, and silicon before it — and the strategic implications for enterprises are both urgent and specific.

The Core Thesis

The effective intelligence delivered per dollar of AI inference has improved by approximately 1,000x since GPT-3’s launch in 2020. This is not a temporary phenomenon. Three reinforcing vectors — architectural efficiency, market price competition, and algorithmic amplifiers — compound at different rates to produce a super-linear yield curve. Capability density is doubling every 3.5 months, a pace we call the “Densing Law.”

At the same time, the gap between frontier and open-source models has collapsed from 24 months to under 6 months. MMLU scores have converged from a 17.5-point gap to just 0.3 points. This convergence eliminates vendor lock-in and establishes a permanent price ceiling on intelligence.

The result is a structural force that history has demonstrated four times in 140 years: when infrastructure commoditizes, value migrates to the application layer. Edison controlled the entire electrical stack; today, utilities capture 1.8% of GDP while the enabled economy is worth $28 trillion. AT&T spent $133.5 billion in acquisitions trying to escape the “dumb pipe” fate and failed. Intel and Microsoft were valued equally in 2000; by 2026, software is worth 21x silicon. NVIDIA — at $4.3 trillion in market cap — is in exactly the position that Edison, AT&T, and Intel once occupied.

For enterprise strategists, the implication is that the 3–5 year window during which AI infrastructure commands premium economics is also the window during which application-layer positions must be established. Organizations that master intelligence orchestration — routing the right model to the right task at the right cost — will capture the value migration. Those that default to “best model for everything” will overspend by 3x to 24x on a cost base that is itself declining 40–50% per year.

Ten Key Findings

Finding 1 — Part I: The Intelligence Layer

1,000x

Intelligence yield has improved 1,000x since GPT-3

Capability density doubles every 3.5 months. Cost per 1M tokens dropped from $20 (GPT-3 Davinci) to $0.15 (GPT-4o-mini) — a 133x reduction in API pricing alone, with far superior quality. A well-designed system routes 70% of queries to commodity models. Ch 1–2

Finding 2 — Part I: The Intelligence Layer

0.3 pts

Open-source gap collapsed from 24 months to under 6

MMLU gap between frontier and open-source models closed from 17.5 points (2023) to 0.3 points (2025). Open-source has pulled ahead on some benchmarks entirely. Vendor lock-in is dissolving. Ch 3–4

Finding 3 — Part II: The Hypothesis

140 years

Value always migrates from infrastructure to applications

91% of AI value sits in infrastructure today. History predicts inversion to 20/80 within a decade. NVIDIA is in Edison’s, AT&T’s, and Intel’s exact position. The projected crossover year is ~2031. Ch 5–6

Finding 4 — Part III: Historical Analogies

$133.5B

The pattern has never failed across electricity, telecom, and silicon

AT&T spent $133.5B in acquisitions trying to escape commoditization. Intel went from near-parity with Microsoft to 1/13th of its value. Utilities now capture just 1.6% of the $30.6T economy they enabled. Ch 7–10

Finding 5 — Part IV: The Model Landscape

85%

Fine-tuned 7B models beat GPT-4 on their target task

Across 310 fine-tuned models (LoRA Land, Predibase), 85% of QLoRA fine-tuned small models outperform GPT-4 on targeted tasks. The sweet spot is 14B parameters: 94% of frontier capability at a fraction of the cost. Self-hosting at $0.03/M tokens. Ch 11–12

Finding 6 — Part IV: The Model Landscape

60,000x

Over-training ratios have shattered Chinchilla scaling laws

Qwen3 0.6B trains at 60,000 tokens per parameter — 3,000x above Chinchilla-optimal. Inference cost, not training cost, now dominates total cost of ownership. Smaller + massively over-trained = cheaper to serve, comparable quality. Ch 14

Finding 7 — Part IV: The Model Landscape

$72B+/yr

Open-source is a strategic weapon, not charity

Meta spent $72B on AI in 2025 (with $115–135B budgeted for 2026) and gives models away. DeepSeek trained V3 for $5.6M in compute costs (total R&D estimated at $1.3B+) and released it open-weight. Alibaba committed $53B over three years while open-sourcing Qwen. Open-sourcing commoditizes the model layer, shifting value to where these firms dominate. Ch 15

Finding 8 — Part V: Infrastructure Arms Race

$1.7T

Largest infrastructure buildout since electrification

$1.7 trillion compute stack projected by 2030. Inference demand reaches 296 trillion tokens per day, built bottom-up from 7 consumer platforms, 5 enterprise copilots, and 8 industry verticals. Ch 16–18

Finding 9 — Part VI: Enterprise Disruption

$600B+

SaaS and enterprise services exposed to AI disruption

$40–50B in annual SaaS revenue displaced by 2028. Customer support hits near-total disruption earliest. CRM & Sales represent the largest revenue at risk ($25B). ERP and Security protected by integration moats. Ch 19–20

Finding 10 — Part VI: Enterprise Disruption

70 / 20 / 10

Intelligent routing captures 5–10x cost advantage

The optimal architecture: 70% tiny agents, 20% mid-tier specialists, 10% frontier reasoning. 9 of top 10 highest-volume enterprise tasks are served by fine-tuned models at fractions of a cent per invocation. Ch 22

The Strategic Window

The data in this report converges on a single strategic conclusion: organizations have a 3–5 year window to establish positions in the application layer before AI infrastructure fully commoditizes. The actions required are specific and measurable:

Map every enterprise task to the right model tier. The 70/20/10 routing architecture — 70% commodity models, 20% mid-tier, 10% frontier reasoning — reduces AI costs by 5–10x without sacrificing quality.
Build vertical model capability. Fine-tuned 14B models achieve 94% of frontier performance at $0.03/M tokens. For organizations processing 2M+ tokens/day, self-hosting pays back in weeks.
Invest in the application layer, not the infrastructure layer. History’s lesson is unambiguous: the utility never becomes the appliance company. Position where value is migrating to, not where it sits today.
Treat intelligence as an operational cost curve, not a capital investment. API prices are declining 40–50% per year. By 2028, 68% of enterprise AI tasks will cost less than $0.001 per invocation.
Act now. The convergence window is closing. Open-source parity means switching costs approach zero. First movers in vertical intelligence capture compounding data advantages that late entrants cannot replicate.

Report Road Map

The Intelligence Layer

How intelligence per token has grown 1,000x since GPT-3, the cost curve, and the convergence between open and closed models.

Ch 1 · Ch 2 · Ch 3 · Ch 4

The Hypothesis

Value migrates from infrastructure to applications. The thesis that drives the entire report, drawn from 140 years of precedent.

Ch 5 · Ch 6

III

Historical Analogies

Four case studies proving the pattern: electricity, telecom, silicon, and the Britannica Problem of redundant knowledge.

Ch 7 · Ch 8 · Ch 9 · Ch 10

The Model Landscape

Taxonomy of 7 model tiers, the small-model thesis, training economics, and the open-source paradox.

Ch 11 · Ch 12 · Ch 13 · Ch 14 · Ch 15

The Infrastructure Arms Race

$1.7T compute buildout, NVIDIA dominance analysis, and bottom-up inference demand reaching 296T tokens/day.

Ch 16 · Ch 17 · Ch 18

The Enterprise Disruption

$600B disruption map, SaaS displacement timeline, task-level economics, and the 70/20/10 routing architecture.

Ch 19 · Ch 20 · Ch 21 · Ch 22

← Report Index Chapter 1: Intelligence Yield →