What Counts as a Real Moat in the AI Era?

The Old Playbook Is Breaking Down

As artificial intelligence rapidly commoditizes software development, product design, and process automation, a fundamental question is reshaping how investors and founders think about competitive advantage: what actually constitutes a defensible moat in 2025 and beyond? The uncomfortable answer, according to a growing chorus of venture capitalists and strategists, is that most of the answers people give are wrong.

The prevailing wisdom — that complex integrations, proprietary algorithms, or deep product embeddedness will protect businesses — rests on a single fragile assumption: that intelligence won't become 100x faster, 100x more capable, and 100x cheaper than it is today. But every trend line in AI suggests exactly that leap is imminent. OpenAI, Anthropic, Google DeepMind, and Meta are all racing to deliver it.

Key Takeaways:

• AI is collapsing the cost of 'doing hard things,' making execution-based moats increasingly fragile
• Physical assets, regulatory licenses, proprietary data flywheels, and network effects are emerging as the most durable defenses
• Static datasets will be replaced by synthetic data; only continuously accumulating private data retains strategic value
• Companies like TSMC, ASML, and major energy infrastructure players exemplify 'hard to get' moats
• The critical distinction for investors is now 'hard to do' vs. 'hard to get'
• Capital intensity and regulatory complexity are becoming features, not bugs

The Critical Distinction: 'Hard to Do' vs. 'Hard to Get'

Defensibility has historically come from two sources: things that are hard to do and things that are hard to get. AI is systematically destroying the value of the former while dramatically amplifying the value of the latter.

Consider what used to be 'hard to do.' Building enterprise software required teams of dozens of engineers working for years. Maintaining complex integrations across customer systems created switching costs that could lock in contracts for a decade. Embedding a product so deeply into a client's workflow that ripping it out would take 12 months of migration effort — that was a moat.

Now consider the AI-powered reality. Tools like GitHub Copilot, Cursor, Devin, and a growing ecosystem of AI coding agents can replicate software functionality in hours or days, not months. Claude and GPT-4 can analyze codebases, generate integration layers, and automate migration plans. The 'difficulty' of execution is compressing at an exponential rate.

But owning 10 million active users? Holding a government-issued broadcasting license or banking charter? Operating a semiconductor fabrication plant? Having $1 billion in deployable capital? These are 'hard to get.' AI compresses the time it takes to do things, but it cannot compress the time it takes for things to happen. That distinction is now the single most important filter for evaluating any business or investment.

Five Moats That Actually Survive the AI Revolution

Strategists and investors are converging on a short list of defensive advantages that hold up even when intelligence becomes essentially free. Here are the 5 that pass the 'hard to get' test.

1. Continuously Accumulating Private Data

Not all data qualifies. A static dataset that exists merely because it was expensive to collect will eventually be replaced by synthetic data or circumvented entirely by foundation models trained on broader corpora. The datasets that matter are those that grow richer with every user interaction, every transaction, every sensor reading — creating a compounding flywheel that widens the gap over time.

Think of Tesla's autonomous driving data, which improves with every mile driven by its fleet of millions of vehicles. Or consider Bloomberg's financial data terminal, which aggregates real-time market information that no AI model can simply hallucinate into existence. The key characteristic is that the data must be proprietary, continuously refreshed, and deeply integrated into the product experience.

Companies sitting on static archives — even large ones — should be nervous. OpenAI's models are already demonstrating the ability to generate plausible synthetic replacements for many types of training data. The moat isn't the data you have; it's the data you're generating right now, faster than anyone else.

2. Regulatory Licenses and Government Relationships

In an era where anyone can build the software, permission to operate becomes the binding constraint. Banking charters, pharmaceutical approvals, spectrum licenses, defense clearances, energy permits — these take years to obtain and involve navigating bureaucratic processes that AI cannot accelerate.

JPMorgan Chase doesn't just have good software; it has a banking charter that took over a century of regulatory compliance to maintain. SpaceX's competitive advantage isn't just rocket engineering — it's FAA launch licenses and NASA contracts that took years of relationship-building. In healthcare, companies like Tempus hold FDA clearances that create multi-year barriers to entry regardless of how good a competitor's AI model might be.

Regulatory moats are actually strengthening in the AI era, as governments worldwide — from the EU's AI Act to emerging frameworks in the US, UK, and Japan — layer on new compliance requirements that favor incumbents with existing regulatory infrastructure.

3. Physical Infrastructure and Atoms-Based Assets

The most obvious 'hard to get' moat is physical. TSMC operates semiconductor fabs that cost $20 billion or more to build and take 3-5 years to bring online. ASML is the sole manufacturer of extreme ultraviolet (EUV) lithography machines, each costing roughly $380 million. No amount of AI can conjure these facilities into existence overnight.

Energy infrastructure is another prime example. As AI data centers consume ever-growing amounts of electricity — Microsoft, Google, and Amazon collectively committed over $80 billion to data center spending in 2024 alone — companies that own power generation, transmission, and grid infrastructure are sitting on increasingly valuable hard assets.

• TSMC: 90%+ market share in advanced chip manufacturing
• ASML: Sole supplier of EUV lithography, $380M per machine
• Equinix: Operates 260+ data centers across 72 metros globally
• NextEra Energy: Largest generator of renewable energy in the world

These are businesses where the 'moat' is literally made of concrete, silicon, and steel. AI makes the software layer on top more efficient, but it cannot replace the physical substrate.

4. Network Effects at Scale

Network effects have always been powerful, but in the AI era they become even more decisive. When AI can replicate any product's features, the network itself becomes the only thing that can't be copied.

Consider the dynamics at play. A new AI startup can build a better messaging app than WhatsApp in a weekend. But it cannot replicate WhatsApp's 2 billion users. It can build a more sophisticated marketplace than Amazon, but it cannot recreate the flywheel of 300 million active customer accounts and 2 million third-party sellers.

The strongest network effects in AI specifically belong to platforms like Hugging Face (with its model-sharing community of over 500,000 developers), GitHub (100 million+ developers whose collaboration data feeds Copilot), and app ecosystems like Apple's App Store and Google Play. These networks generate data, attract participants, and create lock-in that compounds over time.

5. Deployable Capital at Scale

Perhaps the most underappreciated moat is simply having enormous amounts of capital ready to deploy. In a world where AI makes execution cheaper, the ability to fund massive infrastructure projects, acquire companies, and subsidize market entry becomes disproportionately powerful.

Microsoft's $13 billion investment in OpenAI wasn't just a bet on technology — it was a capital moat. Few companies on earth could write that check. SoftBank's Vision Fund, Nvidia's $500 billion+ market cap enabling strategic investments, and sovereign wealth funds like Saudi Arabia's PIF (which committed $100 billion to AI initiatives) all represent capital-based defensibility that no amount of clever AI engineering can overcome.

• Massive upfront capital requirements deter new entrants
• Capital enables patient, long-term infrastructure buildouts
• Acquisition capability lets incumbents absorb emerging threats
• Balance sheet strength supports subsidized pricing to win markets

Why This Framework Matters Now

The timing of this strategic shift is critical. We are entering what some analysts call the 'intelligence deflation' era — a period where the cost of cognitive work drops precipitously, similar to how cloud computing deflated the cost of server infrastructure in the 2010s.

When AWS launched in 2006, it didn't destroy all competitive advantages — it destroyed advantages based on owning servers. Companies that had built moats around proprietary hardware infrastructure suddenly found those moats evaporating. The winners of the cloud era were those who owned the things that couldn't be virtualized: customer relationships, brand trust, regulatory positions, and network effects.

The AI era is following an identical pattern, but at a larger scale. Intelligence itself is being commoditized. The companies and investors who recognize this shift early — and reallocate toward 'hard to get' assets — will capture the strongest returns of the next decade.

Looking Ahead: The Compounding Advantage of Hard Assets

As AI drives the marginal cost of intelligence toward zero, businesses built on physical assets, regulatory permissions, network effects, and compounding data flywheels are entering what may be the strongest compounding cycle in history. They can now deploy near-free intelligence to optimize operations, reduce costs, and accelerate growth — all while sitting behind barriers that AI itself cannot breach.

For founders, the implication is clear: build on top of something that can't be replicated by a prompt. For investors, the filter is equally straightforward — ask whether a company's advantage comes from doing something hard, or from having something rare. In 2025, only the latter counts.

The AI revolution won't eliminate competitive advantages. It will simply reveal which ones were real all along.