The AI economy is two business models pretending to be one.
In 1845, investors cumulatively committed close to half Britain's entire GDP to railway construction. Most of the railway companies went bankrupt. The railways transformed civilisation.
In 1720, the South Sea Company reached a valuation twice the size of the British economy. It produced little of lasting commercial value, collapsed, and took the banking system to the brink.
Everyone asks which template fits AI. The question is wrong. The answer is yes. Both. Running simultaneously, on the same infrastructure, inside the same companies. And the inability to tell them apart is the actual risk.
A tale of two economies
The AI industry contains two fundamentally different businesses.
The first is building railways. Enterprise AI - where Anthropic has drawn level with OpenAI in enterprise revenue and is trending ahead [1], where customers pay based on value delivered, where demand is organic and expanding, where unit economics improve with every generation of model efficiency. A managing director at an investment bank will pay a thousand times more for AI that analyses a portfolio in minutes than a student pays to generate essay outlines. That asymmetry is the enterprise economy. It is real, it is growing, and it would exist without a single dollar of subsidy.
The second is less certain. Consumer AI - where 900 million weekly users access frontier intelligence, but only around 60m pay anything at all [2]. The other 94% use the free tier. The $20-a-month price point is an industry-wide convention adopted by every major provider not because it reflects cost, but because it manages demand. Nick Turley, the person who runs ChatGPT, was explicit: "Having an unlimited plan is like having an unlimited electricity plan. It just doesn't make sense" [3]. Power users consume orders of magnitude more compute than casual users at the same price. The internal economics are, in his words, "mind-boggling."
Same models. Same data centres. Same chips. Completely different economics. And a $20 flat rate that creates what you might call the fog of freemium - blending the enterprise user generating thousands of dollars of value per query with the casual user generating low-value slop, and reporting them as the same metric. The subsidised consumer tier is hiding the fact that enterprise AI may already have viable unit economics - obscured by a pricing model that makes both businesses look identical.
Every AI company sits somewhere on the spectrum between these two economies. The question is not whether AI is a bubble. It is which AI you are exposed to.
The railways case
The evidence for productive transformation is substantial and deserves its full weight.
Real infrastructure is being built - data centres, custom silicon, power plants. Physical assets that will outlive any individual company's financial difficulties, just as bankrupt railway companies' track was consolidated by survivors who built profitable networks on someone else's losses. If OpenAI failed tomorrow, its capacity would not disappear. It would be acquired.
Real efficiency gains are compounding. Inference costs are declining at roughly an order of magnitude every 12 to 18 months through custom silicon, model distillation, and optimisation [4]. DeepSeek proved frontier-grade reasoning can be delivered at dramatically lower cost. The cost curve is genuinely falling.
Enterprise demand is organic and expanding. Companies deploy AI not because it is fashionable but because it measurably reduces cost or increases output in specific, verifiable workflows. Anthropic's revenue trajectory - $14 billion annualised in February, with some estimates suggesting it surpassed $19 billion by early March [5] - is driven overwhelmingly by enterprise customers paying prices that reflect value, not subsidy.
And the hyperscalers are funding the buildout primarily from operating cash flow, not debt. This is not the leveraged railway mania of the 1840s. It is patient capital deployment by the most profitable companies in history.
Consumer AI also creates real value. Google was free. Facebook was free. The most valuable companies in history were built on free products monetised through adjacent means. Free does not mean fraud.
The uncomfortable parallels
But something else is running alongside the productive investment.
Nvidia invests in OpenAI. OpenAI buys Nvidia chips. CoreWeave borrows against Nvidia GPUs to serve OpenAI workloads [6]. Capital flows in loops. This can be rational ecosystem building - Intel invested in PC OEMs in the 1990s, Cisco financed its customers, and both reflected genuine supply-chain economics. But it can also inflate the appearance of demand beyond what independent customers would generate. The distinction matters enormously, and the industry has not yet been tested in conditions that would reveal which interpretation is correct.
Close to 30% of the S&P 500 and 20% of the MSCI World index sits in five companies - the greatest concentration on record [7]. When the South Sea Company fell, it took down its banker because they were entangled. The cross-dependencies in the AI complex - equity stakes, supply agreements, revenue loops between Nvidia, OpenAI, Microsoft, CoreWeave, and Oracle - are real and growing.
OpenAI projects $14 billion in losses for 2026 [8]. David Sacks and Jason Calacanis projected a cumulative $500 billion industry loss before breakeven in the mid-2030s [9]. Those projections assumed stable energy costs. Then escalating tensions between the United States, Israel, and Iran led to the effective closure of the Strait of Hormuz. Brent crude surged from approximately $73 to a peak above $125. Goldman Sachs revised its PCE inflation forecast to 2.9% and deferred the first expected rate cut from June to September [10]. The direct impact on data-centre electricity is modest - most run on domestic power. The indirect impact - inflation, deferred rate cuts, elevated cost of capital, compressed investor patience - is not.
The question is not whether consumer AI creates value. It does. The question is whether the capital deployed against it can be recovered before the pricing transition reveals the true demand curve.
The test
One event resolves both readings simultaneously.
When value-based pricing replaces the flat-rate subsidy - and Turley has told you it must - the industry discovers what intelligence is actually worth. Enterprise segments are already demonstrating durable pricing power with expanding ARPU. Consumer segments have never been tested at anything close to real cost.
Turley's comments suggest pricing changes are imminent. Enterprise contracts typically renew annually. The first earnings reports reflecting value-based pricing will likely arrive in Q3 or Q4 2026. Enterprise holds and expands? Railways confirmed. Consumer survives real pricing? Not South Sea after all. Consumer demand collapses at cost? The capital structure reprices.
The data required to validate the business model is the same data that could collapse the growth narrative. That is the trillion-dollar tension at the heart of the industry.
The private capital committed - OpenAI has raised $110 billion in its most recent round alone, Anthropic approximately $64 billion in total [11] - will demand liquidity within 18 to 24 months. Even the hyperscalers are feeling pressure: Amazon's free cash flow is projected to go negative in 2026, Meta's to decline significantly [12]. They play with house money, but their shareholders are counting it.
Whether unit economics arrive before or after the IPO window determines who absorbs the losses.
The precedent already playing out
Figma grew revenue 41%. Net dollar retention: 136%. The stock is down 84% from its post-IPO high. Last week Google launched a free AI design tool and Figma dropped another 12% in two days [13]. Cisco had real earnings in 1999 and still fell 86%.
Revenue does not prevent repricing when the competitive structure shifts beneath it.
Infrastructure concentrates like railway routes. Intelligence commoditises like rolling stock. The companies building routes will consolidate and profit. The companies selling rolling stock will compete on price until the margins disappear.
Railways or South Sea?
When the subsidy recedes, we discover who was creating value and who was just using free compute.
Notes
[1] Axios, "Anthropic turns the tables on OpenAI in critical revenue category," March 18, 2026. Enterprise revenue split shifted from 60/40 OpenAI to approximately 50/50, trending toward Anthropic. Not yet overtaken but closing rapidly. WSJ reported OpenAI considering strategic shift toward enterprise focus.
[2] ChatGPT: 900 million weekly active users and 50-60 million paying subscribers (OpenAI announcement, February 2026; confirmed Reddit/TechCrunch). Consumer tier priced at $20/month across all major providers (ChatGPT Plus, Claude Pro, Gemini Advanced).
[3] Turley, N., Head of ChatGPT, BG2 Pod, March 2026.
[4] Inference cost decline approximately tenfold every 12 to 18 months. Sources: custom silicon roadmaps (Google TPU, Amazon Trainium, Microsoft Maia), model distillation research, DeepSeek demonstration of frontier reasoning at dramatically lower cost (January 2025).
[5] Anthropic: $14 billion ARR February 2026. Some estimates (Sacra, Bloomberg) suggest this surpassed $19 billion by early March - a figure that implies extraordinary monthly growth and should be treated as indicative rather than confirmed. TIME Magazine ("Most Disruptive Company," 2026).
[6] Nvidia strategic investments in AI companies. CoreWeave debt facilities collateralised against Nvidia GPUs. Historical parallels: Intel invested in PC OEMs throughout the 1990s; Cisco provided customer financing during the networking buildout. South Sea Company lending practices documented in Larry Neal, "The Rise of Financial Capitalism" (Cambridge University Press).
[7] S&P 500: 30% in five companies. MSCI World: 20% in five companies. Source: Financial Times, Bank of America Global Research.
[8] OpenAI projected losses: $14 billion for 2026. Source: The Information, reporting on internal projections.
[9] Sacks, D. and Calacanis, J., All-In Podcast, March 13, 2026. J-curve projection of approximately $500 billion cumulative losses before mid-2030s breakeven. Their analytical estimate.
[10] Brent crude from approximately $73 to peak above $125. Goldman Sachs revised PCE inflation to 2.9%, deferred first rate cut from June to September. Source: Goldman Sachs Economics Research, March 12, 2026.
[11] OpenAI: $110 billion round announced February 27, 2026 (most recent); prior $40 billion SoftBank-led round. Source: CNBC. Anthropic: approximately $64 billion total raised including $30 billion Series G. Sources: TechCrunch, Crunchbase.
[12] Amazon free cash flow projected negative in 2026 on approximately $200 billion capex. Meta free cash flow forecast to decline significantly. Sources: Wall Street consensus estimates as of March 2026, based on company capital expenditure guidance.
[13] Figma: IPO'd 2025 at $33, rallied 333% to $142.92, now approximately 84% below high. Revenue growth 41%, NDR 136%. Google Stitch: 12% Figma decline in two days, March 19, 2026. Sources: CNBC, Forbes. Cisco: 86% peak-to-trough decline, 2000-2002.