AI Supercycle Economics Part 1: Mapping the AI Value‑Chain with an A‑Shaped Framework

The article follows the first lecture of Stanford’s MS&E 435 course, where Altimeter Capital partner Apoorv Agrawal presents an “AI value‑chain framework” – an A‑shaped three‑layer model (Semi, Infra, Apps) that answers the core question “Where’s the money?”.

Physical Laws of the AI Industry: A‑shaped vs. V‑shaped

In cloud computing the value distribution formed a V‑shape (apps largest, infra middle, chips smallest). In the AI era the shape is inverted: the semiconductor layer now captures the bulk of revenue and profit, while the application layer is the smallest slice. The contrast is illustrated by two charts:

Revenue Growth and Structure (2024 → 2026)

In April 2024 the AI ecosystem’s annualized revenue was about $90 B, split roughly as:

Semi ≈ $75 B – NVIDIA data‑center Q1 FY2027 revenue $18 B (≈$72 B annualized) plus other chip makers.

Infra ≈ $10 B – major cloud providers and inference‑cloud services.

Apps ≈ $5 B – OpenAI $2 B ARR plus Anthropic, Midjourney, etc.

Two years later the total annualized revenue exceeds $435 B, but the layer proportions remain similar:

Semi ≈ $300 B – NVIDIA $193.7 B data‑center, Broadcom AI chips $34 B, HBM $25 B.

Infra ≈ $75 B – Azure, AWS, GCP, Oracle each $10‑20 B, CoreWeave $6 B.

Apps ≈ $60 B – OpenAI ~$25 B, Anthropic ~$30 B, plus smaller players.

The semi layer now accounts for about 70 % of revenue and 79 % of AI‑ecosystem gross profit.

Micro‑level Money Flow ($20 Subscription)

A Day1Global podcast breaks down a $20 monthly subscription to Claude:

Anthropic retains $4 gross profit (≈$0.20 net after operating costs).

$6 flows to compute providers; of that $1.2 goes to electricity, $0.9 to operations, $0.9 stays with the compute provider, and $3 goes to NVIDIA.

NVIDIA’s $3 yields $2.25 gross profit and $1.5 net profit (≈50 % net margin).

Across the ecosystem, the eight‑layer panorama shows how the $20 is split among chips, power, fabs, optics, etc.

Why the A‑shaped Structure Persists

Three mechanisms lock the shape:

Model‑capacity race fuels continuous compute arm‑ament. Frontier model FLOPS grow ~4× per year (GPT‑4 ≈2×10²⁵ → GPT‑5 ≈1×10²⁶) while chip cost per FLOP falls only 30‑40 % annually, creating a 3‑4× supply‑demand gap.

Jevons paradox in inference demand. Even with 90 % per‑token price cuts, agent architectures cause token consumption to rise 10‑100× per interaction, so total spend grows 320 % YoY (Menlo Ventures, 2025).

NVIDIA’s pricing power stems from ecosystem lock‑in. CUDA’s dominance and generational performance jumps let NVIDIA keep ~75 % gross margin despite rising supply.

The only path to a flip is demand saturation – when AI inference growth slows below efficiency gains. Current data show no sign of saturation; agent‑driven usage continues to expand.

When Could the A‑shape Reverse?

Two scenarios could break the current order:

Demand saturation. If enterprise AI spend growth falls below per‑token cost reductions, the “compute tax” would shrink and apps could capture more profit.

Structural break. Large‑scale ASIC success in training (e.g., Google TPU or custom ASICs winning frontier‑model training) or a shift to ad‑based app revenue could erode NVIDIA’s pricing power.

New Phenomena Since 2024

1. Extreme Apps‑layer concentration

OpenAI and Anthropic now own ~75 % of app‑layer revenue, far higher than SaaS concentration in the cloud era.

2. ASIC competition blurs layer boundaries

Custom ASICs (Google TPU Ironwood, AWS Trainium2, Broadcom‑designed chips for OpenAI) dominate inference, while GPUs remain dominant for training. This splits the “Semi” layer into general‑purpose GPU and specialized ASIC sub‑markets, challenging the clean three‑layer taxonomy.

3. Jevons paradox intensifies

Inference cost per token fell two orders of magnitude since 2023, yet enterprise AI spend rose 320 % because agents multiply token usage. The paradox means profit improvements in the app layer are offset by exploding compute consumption.

Key Takeaways

A‑shaped flip requires demand saturation; as long as AI spend outpaces per‑token cost declines, the shape stays.

ASICs have won inference economics, but training dominance still rests with GPUs; a full‑ASIC training breakthrough could challenge NVIDIA.

App‑layer value capture is limited by extreme oligopoly (OpenAI/Anthropic) rather than a broad SaaS market.

Agent architectures decouple compute consumption from user count, allowing total AI spend to keep rising even after user penetration peaks.

Three‑layer analysis works for independent firms, but vertically integrated giants (Google, Amazon, Meta) require a cross‑layer view.

When picking AI‑stack investments, prioritize indispensability > margin > growth – e.g., fabs, GPU/ASIC fabs, power providers.

References are listed at the end of the original article and include Agrawal’s 2024 and 2026 Substack posts, Day1Global podcast, FactSet/ Apollo data, and various public financial disclosures.