Artificial Intelligence (Foundation Models & AI Infrastructure)

Market size. There is no credible single "AI market size" number, and I decline to invent one. What can be measured: worldwide IT spending is forecast at $6.37 trillion in 2026, up 14.2%, with data center systems — the physical substrate of AI — growing 62.5% to $822 billion and IaaS growing 29.3% to $287 billion (Gartner, July 2026). [Projection, from a vendor that sells research but whose IT-spend series is the closest thing to an industry standard.] Gartner's analyst frames it plainly: "Building the compute capacity required for AI is the largest infrastructure project ever attempted by humanity."

Demand and growth. Real and extraordinary at the model layer. Anthropic's annualized revenue run rate went from $9B (end of 2025) to $47B (May 2026) to $65B (end of July 2026) — $18B of annualized revenue added in two months (CNBC, August 2026; TechCrunch, August 2026). OpenAI reportedly doubled from $20B to **$40B** over the same period. [Estimate — these are company-disclosed run rates shared with investors, not audited revenue, and "annualized run rate" flatters fast-growing businesses by extrapolating the best month.]

Capital intensity. Extreme, and the defining feature. The five largest US hyperscalers plan $660–690 billion in 2026 capex, roughly double 2025's ~$380 billion: Amazon ~$200B, Alphabet $175–185B, Microsoft $120B+, Meta $115–135B, Oracle ~$50B (Futurum Group, 2026). [Projection, built from company guidance.] Foundation-model training is now a capital expenditure category comparable to national infrastructure programs.

Regulatory. Lighter than expected in 2026. The EU agreed a "Digital Omnibus" deferring high-risk AI obligations under the AI Act from August 2026 to December 2, 2027 for stand-alone Annex III systems and to August 2028 for AI in regulated products. Article 50 transparency obligations did take effect August 2, 2026, with a grace period to December 2026 for watermarking existing systems, and new prohibitions on non-consensual intimate imagery and CSAM become enforceable December 2, 2026 (Gibson Dunn, 2026). [Verified.] Copyright remains the live legal risk: US courts have allowed Disney and other studios' infringement claims against AI video generators to proceed (Norton Rose Fulbright AI litigation update, 2026).

Competition. At the frontier: effectively a four-to-six player oligopoly with a capital moat that no new entrant can cross without a hyperscaler sponsor. At the application layer: brutally crowded, minimal defensibility, and exposed to the underlying model providers shipping your feature.

Business models. API/token metering; enterprise seats and committed-capacity contracts; consumer subscription; increasingly, revenue-share and outcome pricing at the application layer.

Investor interest. Maximal. Foundational AI startups raised $178 billion across just 24 deals in Q1 2026 alone — double all of 2025 ($88.9B across 66 deals) and 5.7x 2024 (Crunchbase News, 2026). Median Series D+ AI valuations run $4.25 billion versus $134 million for non-AI companies at the same stage (Venture Monitor) — a ~32x premium that is the cleanest single measure of the dislocation.

Risks. (1) The capex-to-revenue gap: at even the most generous accounting, model-layer revenue is a low-single-digit percentage of hyperscaler capex, and infrastructure precedes revenue by 18–36 months. (2) Depreciation schedules on GPUs that may not have the assumed useful life. (3) Circular financing — see the synthesis section. (4) Commoditization: capability gaps between frontier models have been narrowing, which is good for buyers and bad for pricing power. (5) A single disappointing capex guide from one hyperscaler could reprice the entire complex.

Notable companies. OpenAI, Anthropic, xAI (absorbed by SpaceX in a $250B transaction — the largest VC-backed M&A ever recorded), Mistral, Reflection AI ($2B Series B led by Nvidia), Safe Superintelligence, Advanced Machine Intelligence ($1.03B — Europe's largest seed round on record).

Underserved opportunities. [Analysis] Evaluation and observability for agent systems in regulated industries; inference cost optimization as a service (the arbitrage between what models cost and what applications charge is currently enormous and will compress); data rights, licensing and provenance infrastructure, which the copyright litigation is about to make mandatory rather than optional; and non-English, non-US-regulatory-context deployment tooling. Note that all of these are picks-and-shovels businesses serving the boom rather than bets on the boom — which is the appropriate risk posture for a startup that cannot raise $10B.

---