AI, Moats, and the Question of a Bubble

AI is the new gold rush -- fortunes for a few, illusions for many. The winners will be those with moats deep enough to survive the flood.

Introduction

Every gold rush produces fortunes, but not every prospector strikes it rich. The same is true of artificial intelligence (AI). Thousands of startups are chasing opportunities, but enduring profits will flow only to firms protected by barriers to entry -- durable “moats” that keep rivals at bay.

AI may also follow another historical pattern: the bubble. Railroads, electricity, the internet, and clean tech all saw valuations surge before painful corrections. The right question for AI is not simply whether a bubble exists, but who survives it.

Michael Porter’s Five Forces remains a useful starting point. Industry profits are shaped not just by management skill but by structural forces: the threat of entrants, buyer power, supplier power, substitutes, and rivalry. Moats work by reshaping these forces.

Crucially, AI is not a single industry but a stack of interdependent layers -- from chips to cloud, electricity, models, middleware, applications, and distribution. Each layer faces different pressures, and each is defended (or not) by different moats.

Let’s begin with the moats that protect firms from competitive encroachment. 

The Moats

Across industries, a dozen or so recurring moats explain why some firms thrive while others fade. Each reshapes the competitive forces that determine profitability.

1. Network Effects -- Each new user makes the platform more valuable -- Facebook’s social network, Visa’s payments system, Airbnb’s marketplace.
2. Brand Power -- Trust, reputation, and recognition sustain loyalty and pricing power -- Apple in consumer electronics, Nike in sportswear, Coca-Cola in beverages.
3. Switching Costs -- Customers hesitate to move if the process is costly or disruptive -- SAP and Oracle enterprise software, Adobe Creative Cloud, Bloomberg terminals in finance.
4. Economies of Scale -- Large volumes drive down unit costs and increase bargaining leverage -- Walmart in retail, Amazon in logistics, Toyota in car manufacturing.
5. Capital Intensity -- Massive upfront investment requirements deter new entrants -- Boeing and Airbus in aircraft, ExxonMobil in oil and gas, TSMC in semiconductor fabrication factories.
6. Regulatory Barriers -- Licenses, patents, and compliance rules create protective walls -- Pfizer and Merck with pharmaceutical patents, JPMorgan in banking, utilities protected by government franchises.
7. Data and Learning Effects -- Products improve with usage and feedback loops -- Google Search and Maps, Netflix’s recommendation engine, Amazon’s retail algorithms.
8. Exclusive Resources -- Control over scarce assets or intellectual property -- Disney with Star Wars and Marvel, De Beers in diamonds, Apple’s exclusive supply contracts for iPhone components.
9. Culture and Organisation -- Distinct capabilities and ways of working that rivals cannot easily copy -- Toyota’s lean production system, Netflix’s experimentation culture, Southwest Airlines’ service model.
10. Human Capital — Concentrations of scarce skills or visionary leadership. -- Pixar’s creative teams, Goldman Sachs’ finance talent, Apple under Steve Jobs.
11. Vertical Integration -- Owning multiple layers of the value chain to capture synergies and block rivals -- Apple designing chips, software, and devices; Tesla making batteries, cars, and charging infrastructure; Amazon combining e-commerce, logistics, and cloud.
12. Ecosystem Lock-In -- Developer tools, APIs, or compatibility standards that make switching prohibitively costly -- Microsoft Windows and Office suite, Apple’s App Store, Sony’s PlayStation ecosystem.

Now, let's consider how these moats protect each layer of the AI value chain.

The AI Value Chain

Chips and Hardware

• Players: NVIDIA, AMD, Intel, TSMC, ASML.
• Moats:
• Capital intensity: $20B+ fabrication plants (fabs) deter entrants.
• Scale: mass production lowers costs.
• Exclusive resources: NVIDIA’s CUDA software ecosystem is a lock-in.
• Strength: Extremely strong. Hardware is the deepest moat in AI.

Cloud and Compute

• Players: AWS, Microsoft Azure, Google Cloud, Oracle.
• Moats:
• Scale: hyperscale data centers.
• Switching costs: enterprise workloads hard to migrate.
• Regulatory barriers: compliance certifications restrict rivals.
• Weakness: Compute is still a commodity; customers can multi-home.
• Strength: Moderately strong.

Power and Energy Infrastructure

• Players: utilities, independent power producers, grid operators, and energy majors.
• Moats:
• Capital intensity: multi-billion-dollar generation and transmission assets deter entry.
• Regulatory barriers: long permitting cycles and monopoly grid rights limit competition.
• Exclusive resources: access to land, cooling water, and interconnection points.
• Vertical integration: direct power-purchase agreements (PPAs) and on-site generation by Microsoft, Amazon, and Google tighten control over energy security.
• Switching costs: data centres cannot easily relocate once grid-tied.
• Strength: Moderately strong. As AI demand reshapes electricity markets, reliable power becomes a profit centre rather than a mere utility commodity.
• Weakness: Political and regulatory risk; rising community opposition to transmission; exposure to fuel and carbon costs. If the cloud and compute boom falters, overbuilt capacity could leave the power sector exposed -- though for now, supply remains well behind demand, and the greater near-term issue is scarcity, not surplus. 

Foundation Model Builders

• Players: OpenAI, Anthropic, DeepMind, Meta, Cohere, Mistral.
• Moats:
• Capital intensity: training frontier models costs $100 - $500 million.
• Brand: "ChatGPT" and "Claude" enjoy consumer trust.
• Data & learning: reinforcement from millions of human interactions.
• Human capital: concentrations of elite researchers.
• Weaknesses: Open-source challengers, low-cost competitors in China, cross-trainable talent, falling compute costs.
• Strength: Transitional. Strong today, but likely to erode.

Middleware and Deployment

• Players: Hugging Face, LangChain, Weights & Biases.
• Moats: Network effects from developer communities; some switching costs.
• Weaknesses: Features easily absorbed into clouds.
• Strength: Weak to moderate.

Applications

• Players: Perplexity, Jasper, GitHub Copilot, Notion AI.
• Moats: Brand in niches, workflow integration.
• Weaknesses: Low entry barriers, constant substitutes.
• Strength: Weak.

Distribution and Platforms

• Players: Microsoft, Google, Apple.
• Moats:
• Exclusive resources: Office, iOS, and Google Search are gateways.
• Switching costs: enterprises and consumers deeply embedded.
• Brand: trust and default status.
• Weakness: Antitrust scrutiny could bite.
• Strength: Very strong.

The Human Capital Question

Isn’t AI just math, and isn’t math abundant?

At the frontier of research -- scaling laws, alignment, novel architectures -- talent remains scarce. DeepMind’s AlphaGo moment depended on rare expertise. But beyond that, the moat is thinner: thousands of mathematicians and engineers can cross-train.

Human capital in AI is thus valuable but not structural. Unlike TSMC’s fabrication know-how, which takes decades to replicate, AI talent can diffuse quickly. Concentrated labs matter now, but the moat is leaky.

Ranking the Moats

• Strongest: chips (NVIDIA, TSMC), and distribution platforms (Microsoft, Google, Apple).
• Moderately strong: Power providers (regulated utilities, vertically integrated energy suppliers), and cloud/compute providers (AWS, Azure, Google Cloud Platform).
• Strong at the outset, but likely to weaken over time: foundation model builders (OpenAI, Anthropic).
• Weakest: Middleware (Hugging Face, LangChain) and applications (Jasper, Perplexity).

Profits will accrue where moats are deepest: capital-intensive infrastructure and distribution choke points.

Looking through the window of Porter's Five Forces:

• Threat of entrants: High at apps; close to zero at chips.
• Buyer power: Strong against middleware; weak against NVIDIA.
• Supplier power: Graphics Processing Unit (GPU) suppliers wield enormous leverage.
• Threat of substitutes: Apps are flooded with substitutes; hardware is insulated.
• Rivalry: Brutal at the app layer; oligopolistic at chips and distribution.

In time, energy provision may rival chips as the most defensible profit pool. Owning reliable power is capital-intensive, heavily regulated, and geographically constrained -- precisely the ingredients of a durable moat. The firms that control energy inputs to computation will quietly set the pace of the AI revolution.

Commoditisation: The Profit Gravity

Why are moats so critical? Because commoditisation drives profits down.

When customers see products as interchangeable, price wars follow. Margins fall toward the cost of capital. This is why airline seats, generic Android phones, and basic cloud compute all produce slim returns.

AI applications risk the same fate. Anyone can wrap an API in a user interface. Unless firms build brand loyalty or deep workflow integration, they will face relentless churn. Commoditisation is profit gravity; moats are the only way to resist.

The Bubble Question

All this leads back to the obvious question: are we in an AI bubble?

NVIDIA’s price to earnings (P/E) ratio hovers around 50; OpenAI is valued at \$300 billion on perhaps \$12 billion in revenue, despite heavy losses. Investors are paying for tomorrow's profits, not today's.

The danger is that expectations outpace reality. If every AI firm is assumed to maintain dominance, then yes -- valuations are bubbly. Applications and middleware are most exposed; model builders are vulnerable too.

But hardware and distribution look different. NVIDIA’s moat is real: capital intensity, CUDA lock-in, and supply chain dominance. Microsoft and Google’s moats are real: control of platforms billions already use. These firms may justify their lofty valuations.

The right answer, then, is both. There are bubble-like dynamics in much of the sector. These dynamics may continue for some time before any reckoning. But AI itself is not hype. It is a general-purpose technology as transformative as electricity or the internet. The challenge is separating moat-holders from mirage-chasers.

Conclusion: The Shape of the Gold Rush

AI is the 21st-century gold rush. But history shows miners rarely profit most; it is the suppliers of picks and shovels, and the owners of land and railroads, who win.

• NVIDIA and TSMC control the “picks and shovels.” Their moats are the deepest.
• Microsoft, Google, and Apple own the “railroads” of distribution, capturing a disproportionate share.
• Electricity firms power those railroads -- and set their limits -- though they, too, may be exposed if demand falters. For now, supply still trails the boom, which for the moment protects AI-linked energy investments.
• Cloud providers have scale but face intense rivalry.
• Model builders shine today but risk commoditisation and competition.
• Middleware and applications will churn; only brand and integration will preserve value.

Bubbles are almost inevitable. They fund infrastructure, flush out weak players, and leave behind enduring giants. AI’s bubble will be no different: it will mint fortunes, but not evenly. The lasting winners will be those who own the picks and shovels -- and the railroads of distribution.