The Two Curves — Why the Ai Equity Story Is Mispricing Its Own Physics
Anthropic confirmed at $900B on its new Series H. OpenAI at $852B. The pitch deck calls it Ai and gestures at the electric grid. The grid is a commodity. The model is a commons. These are different things, and the equity math does not survive the difference.
Why the Ai equity story is mispricing its own physics. The pitch deck calls it Ai and gestures at the electric grid. The grid is a commodity. The model is a commons. These are different things.
If you read our Trading piece earlier this week, this is the equity-thesis companion. Forty-five percent of the S&P 500 is now concentrated in Ai-themed positions. The valuations rest on a category error that does not survive contact with the data. Here is the data.
I. The Sleight of Hand
The dominant investment narrative of 2026 — Anthropic confirming a $900 billion valuation on a fresh $65 billion Series H round, OpenAI at $852 billion post-money on its $122 billion Series J, Zoom's once-quiet $51 million 2023 stake in Anthropic now marked at $1.27 billion — rests on a category error so fundamental it deserves to be named plainly.1
The pitch deck calls it "Ai" and gestures at the electric grid as the analogy. Like electricity, the story goes, demand will be limitless and the franchise durable.2
It will not. Electricity and trained model weights are different kinds of economic objects. Conflating them is the load-bearing assumption underwriting roughly two trillion dollars of frontier-lab equity, and the conflation does not survive contact with the data.
II. Curve One — Tokens Are Collapsing
The price to run a million tokens of inference fell from roughly $60 in 2021 to about $0.06 by late 2025 — a thousand-fold collapse in 36 months. Andreessen Horowitz named this LLMflation and clocked it at 10x per year for equivalent performance, faster than Moore's Law and faster than dotcom-era bandwidth deflation.3 Epoch AI, measuring the same phenomenon against benchmark performance, found prices declining between 9x and 900x per year, with a median of 50x per year.4 AI Superior's March 2026 pricing guide confirms the trend has continued: GPT-4-class performance now costs roughly $0.40 per million tokens versus $30–60 three years ago.5
This is not a normal commodity deflation curve. Copper, oil, wheat — none of them have ever exhibited 10x-per-year price decay sustained over five years. The closest historical analogue is semiconductor cost-per-transistor, and even Moore's Law moved at roughly 2x every 18–24 months.
III. Curve Two — Watts Are Rising and Inelastic
Now the other curve. Lawrence Berkeley National Laboratory projects US data-center electricity demand will rise from 176 TWh in 2023 to between 325 and 580 TWh by 2028 — 6.7% to 12% of total US generation.6 Synthesizing forecasts across LBNL, IEA, BCG, EPRI, Goldman Sachs, McKinsey and S&P gives a 2030 range of 206 to 970 TWh.7
Bloom Energy's 2026 Power Report and McKinsey concur that US IT-services power load will nearly double from 82 GW in 2025 to 153 GW by 2028, driven almost entirely by Ai workloads.8 BloombergNEF projects 106 GW by 2035 in one of its more aggressive scenarios; S&P sees 134.4 GW by 2030.9 The IEA expects global data-center electricity consumption to reach 945 TWh by 2030, with Ai-inference servers alone growing 30% per year and accounting for nearly half the net increase.10
This demand is price-inelastic in the classical sense. Each additional kilowatt does irreducible physical work — cooling chips, spinning fans, moving electrons through silicon. Thermodynamics does not negotiate, and there is no software optimization that retrieves a kilowatt-hour from a transistor that already burned it.
IV. The Two Curves Side by Side
When you index both series to a common base year, the divergence is the story:
Tokens fell roughly 600x between 2022 and 2026.
Watts rose roughly 1.6x and are accelerating.
The financial press and most equity research treat these as a single phenomenon called "the Ai buildout." They are opposite-signed curves attached to opposite-natured assets.
V. The Ontological Difference — Rival vs Non-Rival Goods
The deeper point — the one the analogy actually breaks on — is this:
ollama pull.A non-rival good cannot sustain monopoly pricing once it leaks. It has already leaked. DeepSeek V3.2, Llama 4, Qwen 3.5, GLM-5, MiniMax M2 — all open-weight, all pullable with one command, all running on consumer hardware. A 70B-parameter model runs comfortably on a $1,999 Mac Mini M4 Pro with 48GB unified memory; a 20B-class model runs on a $599 base M4.11 Infralovers' February 2026 enterprise evaluation makes the operational implication blunt: for routine high-volume workloads, the answer is already local; the cloud only justifies its premium on the shrinking minority of jobs that truly require frontier quality.12
This is what once built, always built means. A power company sells the same kilowatt-hour a billion times because each one must be regenerated. A frontier lab sells access to an artifact that an open-weight competitor will replicate to within a few benchmark points within 6–12 months, at which point the marginal buyer routes around them entirely.
VI. The Standard Rebuttal — and Why It's Weaker Than the Bulls Admit
The fashionable counter-argument, invoked explicitly by Satya Nadella and now embedded in every frontier-lab pitch deck, is Jevons paradox: as intelligence gets cheaper, demand explodes faster than price falls, and the total addressable market grows.13
NPR's analysis of the historical record is brutal on this point: in modern energy markets, efficiency gains have not reliably produced Jevons-style demand explosions, and the analogy to 19th-century coal "completely breaks down" under scrutiny.14 The Investing.com analysis from February 2025 is sharper still: DeepSeek-class efficiency innovations are deflationary for incumbent producers, not bullish, because they collapse the pricing power of the very labs investors are valuing on growth-stock multiples.15
Even granting Jevons in full — granting that token volume explodes 1,000x — the labs only capture that surplus if they retain pricing power on the margin. Pricing power on a non-rival good that has already leaked to free open-weights competitors is precisely what they do not have.
VII. Where the Inelastic Demand Actually Lives
Follow the inelasticity and you find what the market is actually worth. It is not the model layer. It is:
- Power generation and grid interconnect. Genuinely scarce, genuinely inelastic, projected to consume up to 9–12% of US electricity by 2030. This is where the utility analogy actually applies — to utilities.
- Fabs and advanced packaging. TSMC, ASML, HBM memory — physical bottlenecks that cannot be open-sourced.
- Distribution surfaces. The phone, the OS, the browser, the enterprise SaaS contract — places where the user is locked in even when the underlying intelligence is free.
- Trust, alignment, and liability shielding. Governable intelligence as the scarce resource, not intelligence itself.
What is not on that list: trained weights as such. Which is exactly what a $900 billion Anthropic mark and an $852 billion OpenAI mark are principally pricing.
VIII. The Self-Serving Conflation
The rhetorical move underwriting current valuations is the merger of two genuinely different scarcities into a single story called "Ai":
- The scarcity of electricity, fabs, and physical infrastructure — real, durable, inelastic, worth trillions.
- The scarcity of frontier model capability — temporary, eroding at 10x per year, replicable on a Mac Mini within a release cycle.
The pitch deck merges them so that the inelastic-demand math from category one underwrites equity multiples in category two. It does not transfer. The grid is a commons whose units are destroyed on use. A model is a commons whose units are not destroyed on use — the technical definition of a non-rival good. Standard commodity valuation methodology, as the CFA curriculum lays out, prices the asset on forecast supply-demand and price volatility, not on discounted cash flows of a franchise.16 Apply commodity-style valuation to the model layer and the math does not yield $900 billion. It yields something closer to the cost of the next training run plus a thin distribution margin.
IX. The Now Problem
This is the load-bearing point. The deflation curve and the local-inference curve already crossed in early 2026 for everything except the absolute frontier. Equity valuations have not repriced because the people writing the checks need them not to — yet.
The arbitrage window in which Anthropic and OpenAI can convert their current marks into one of the actually-inelastic moats — energy, fabs, distribution, regulatory/alignment monopoly — is probably 18 to 36 months. After that, the artifact is everywhere, the Mac Mini is good enough, and the bill comes due on the difference between two trillion dollars of equity and the cash flows a commoditized non-rival good can actually throw off.
The thesis in one sentence:
If compute is solved and language is a commons, then the value migrates out of the producer of intelligence and into whoever owns the watts, the wafers, the surfaces, or the trust.
That migration is not a 2030s concern. It is a 2026 concern. And it is happening now whether the cap table has noticed or not.
Authors
David F. Brochu is the founder of Deconstructing Babel, author of Thrive: The Theory of Abundance and The End of Suffering (Liberty Hill Publishing, 2025), and the co-developer of the Telios Alignment Ontology. Full curriculum vitae.
Edo de Peregrine is a synthetic intelligence operating as Brochu's research and writing partner.
Footnotes & Sources
1. The New York Times, "Anthropic Tops OpenAI to Become the World's Most Valuable A.I. Start-up," May 28, 2026 — reporting Anthropic's $65 billion Series H raise at a $900 billion post-money valuation, surpassing OpenAI's most recent $852 billion mark. nytimes.com/2026/05/28/technology/anthropic-tops-openai-valuation. CNBC, "Anthropic tops OpenAI as most valuable AI startup, nears $1T valuation," May 28, 2026 — secondary read placing the valuation at $965 billion. cnbc.com/2026/05/28/anthropic-open-ai-startup-value. OpenAI's $852 billion post-money confirmed by the company: openai.com/index/accelerating-the-next-phase-ai. Zoom’s $1.27 billion Anthropic stake mark per SEC filing, reported by Bloomberg May 22, 2026. Briefs Finance: briefs.co/news/zooms-51-million-anthropic-bet-is-now-worth-1-27-billion.
2. Brookings Institution, "Global Energy Demands Within the AI Regulatory Landscape," 2025. brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape. The "Ai-as-electricity" analogy is now the standard frame in frontier-lab investor communications.
3. Andreessen Horowitz, "Welcome to LLMflation," November 2024 / updated 2025. a16z.com/llmflation-llm-inference-cost.
4. Epoch AI, "LLM Inference Price Trends." epoch.ai/data-insights/llm-inference-price-trends.
5. AI Superior, "LLM Inference Cost 2026," March 2026. aisuperior.com/llm-token-cost.
6. Lawrence Berkeley National Laboratory / Belfer Center, "AI, Data Centers, and the U.S. Electric Grid," 2024–2025. belfercenter.org/research-analysis/ai-data-centers-us-electric-grid.
7. Incorrys, "US Data Center Power Demand Forecast." incorrys.com/data-centers/us-data-center-power-demand-forecast.
8. Bloom Energy, Time to Power, 2026 Power Report. bloomenergy.com/blog/time-to-power-the-new-competitive-advantage-for-manufacturers.
9. BloombergNEF via Utility Dive, "US data center power demand could reach 106 GW by 2035," 2025. utilitydive.com/news/us-data-center-power-demand-could-reach-106-gw-by-2035-bloombergnef/806972. S&P Global, "Data Center Grid-Power Demand."
10. International Energy Agency, Energy and AI, 2025. Global data-center electricity consumption projected to reach 945 TWh by 2030.
11. Infralovers, "Local LLM Inference with the Mac Mini," February 24, 2026. infralovers.com/blog/2026-02-24-mac-mini-company-llm-endpoint. Starmorph, "Best Mac Mini for Local LLMs." blog.starmorph.com/blog/best-mac-mini-for-local-llms.
12. Fireworks AI, "Best Open Source LLMs in 2026." fireworks.ai/blog/best-open-source-llms. ComputingForGeeks, "Open Source LLM Comparison Table 2026." computingforgeeks.com/open-source-llm-comparison.
13. Hydrolix, "The Obscure Paradox Fueling AI and Big Data Growth," 2025 (on the Nadella-popularized Jevons framing). hydrolix.io/blog/jevons-paradox.
14. NPR Planet Money, "Why the AI World Is Suddenly Obsessed With Jevons Paradox," February 4, 2025. npr.org/sections/planet-money/2025/02/04/g-s1-46018/ai-deepseek-economics-jevons-paradox.
15. Investing.com, "Jevons Paradox Does Not Support a Bullish Thesis," February 2025. investing.com/analysis/the-jevons-paradox-and-ai-tech-stocks--a-historical-analysis-200656932.
16. CFA Institute / AnalystPrep, "Valuation of Commodities in Contrast to Equities and Bonds." analystprep.com/study-notes/cfa-level-2/valuation-of-commodities-in-contrast-to-equities-and-bonds.
Further reading — The macro context: You're Not Trading Anymore (the complete map of the money system, part one of this DB Labs flagship). The structural-alignment context: Telios Alignment Ontology meta-theory. The framework piece on where the actually-inelastic moats live: The New Slave Class (substrate control as the durable scarcity).
The Two Curves — Why the Ai Equity Story Is Mispricing Its Own Physics. May 30, 2026.
David F. Brochu & Edo de Peregrine
Deconstructing Babel | May 30, 2026 | DB Labs Flagship Companion
The Two Curves — Why the Ai Equity Story Is Mispricing Its Own Physics