The Moat Is the Data, Not the Model

Start with the single most consequential shift in AI economics for a board: the model itself is commoditizing. Gartner has gone so far as to call foundation models a "strategic commodity" — meaning differentiation based on raw model performance alone is unlikely to last.

Three forces drive this:

• The cost of intelligence is collapsing. The inference cost to reach GPT-3.5-level quality fell from about $20 per million tokens (late 2022) to about $0.07 (late 2024) — a roughly 280x reduction in ~18 months (Stanford HAI AI Index 2025). Treat the exact figure as time-stamped and re-verify it before each board cycle — but the direction is the durable point.
• Quality is converging. Leading models share similar architectures and training methods; open-weight models have nearly closed the quality gap. The frontier still moves, but the gap between "the best" and "good enough for the job" keeps shrinking.
• Anyone can deploy one. Cloud APIs, fine-tuning frameworks, and open weights let any enterprise use a capable model without building one.

The strategic translation: the intelligence you are buying gets dramatically cheaper and more interchangeable every year. That is wonderful news for adoption — and a warning for anyone who thinks owning or betting on one model is the strategy.

If the model is the commodity, where does durable advantage go? It migrates up the stack — away from raw intelligence and toward how that intelligence is applied, integrated, owned, and governed. As Marc Andreessen's firm a16z frames it: "The moat is not the model — it's what you build around it."

Four places durable advantage actually lives:

Notice what's not on this list: the model. The model is the raw engine everyone can rent. The moat is the proprietary asset the engine is pointed at.

Of the four, one stands apart: proprietary data is the one thing competitors cannot buy, borrow, or replicate. And uniquely, it can compound — through what's called a data flywheel.

The flywheel works like this: when AI is architected so that every interaction generates feedback data — what the user accepted, corrected, or rejected; what outcome followed — that data feeds back into improving the product, which attracts more usage, which generates more data. The advantage doesn't just persist; it widens over time. A rival starting later faces a gap that grows every day you operate.

This is why data is the only AI moat that compounds. The model gets cheaper for everyone. Your customer relationships grow linearly. But a well-architected data flywheel grows the gap non-linearly — and it is the asset a competitor with a bigger checkbook still cannot purchase.

"The one thing competitors cannot buy, borrow, or replicate is your data." As models commoditize, proprietary data becomes the differentiator — and the board's strategic asset to protect.

The board-level reframe: turn your unique operational knowledge into a defensible, revenue-generating, board-reportable strategic asset. That changes data from an IT cost center into a P&L-relevant moat the board is accountable for stewarding.

Here is the trap. Leaders hear "data is the moat" and assume they already have one — we have decades of data! But raw data is not a moat. Data is only a moat if it is governed and usable.

Proprietary data becomes an asset only when it is discoverable, accessible under policy, and auditable. If it lives in disconnected silos — no unified catalog, no access controls, no lineage — it is not a dormant moat. It is an active liability: a compliance exposure, a security surface, and a blocker to the very AI you want to deploy.

The phrase to remember, because it makes the abstract concrete:

"AI agents cannot query what they cannot find."

An agent pointed at your business is only as good as the data it can actually reach under policy. Roughly 80% of the real AI work is data, governance, and workflow integration (MIT Sloan) — not the model. McKinsey calls the precondition "data productization": treating data products as managed, owned assets, which is what unlocks scaling agents at all.

If the model is a commodity whose price keeps falling and whose quality keeps converging, the strategic posture follows directly: don't lock the company into one vendor's model or price card. Architect for model portability — the ability to swap the underlying model with minimal disruption — and assume multi-model is the norm, not the exception.

The evidence backs this up. Multi-model is already standard enterprise practice (a16z reports ~37% of enterprises use 5+ models, differentiating by use case), and roughly three-quarters of enterprise use cases are bought, not built. A telling nuance leaders miss: the household-name consumer brand is not the enterprise-spend leader. Per Menlo Ventures' 2025 survey, enterprise LLM API spend was led by Anthropic (40%) over OpenAI (27%) and Google (21%) — a reversal from 2023. Treat any single market-share snapshot as volatile; the durable lesson is the rankings move, so keep your optionality.

What portability buys you:

• Pricing leverage — as intelligence gets cheaper, you capture the savings instead of being locked into yesterday's rate.
• Resilience — if a vendor fails, is breached, changes terms, or deprecates a model, you switch rather than scramble.
• Best-fit per use case — different models lead at coding, reasoning, or multimodal work; portability lets you route each task to the right engine.

The inverse mistake is equally common on the buying side: blanket per-seat licensing on day one — reports show 30-40% of company-wide Copilot/Gemini seats sit unused within 90 days. Start with high-value cohorts and expand on evidence.

Bring this to the boardroom as a small set of durable questions that separate spend on commodities from investment in moats. They outlast any model ranking or price.

On every use case:

• Is this capability available to all competitors off the shelf (a commodity — buy it, move on), or does it compound a unique asset we own (a differentiator — invest to own it)?
• If this is a moat, which asset does it compound — proprietary data, workflow, customer relationship, or vertical depth?

On data:

• Is our data clean, governed, catalogued, and accessible enough to actually feed this use case? Could an agent find it?
• Are we treating our proprietary data as a board-reportable strategic asset — governed, protected, activated — or as an IT cost we ignore until it leaks?
• Does this design create a flywheel — does every interaction generate feedback data that compounds our advantage?

On vendors and architecture:

• Can we swap the underlying model without re-platforming? Are we architected for portability?
• Which critical capabilities are really someone else's model — and what happens if that vendor fails, is breached, or changes terms?
• Are we about to license seats for the whole workforce on day one, or start with high-value cohorts and expand on evidence?

The through-line: spend like a buyer on the commodity layer; invest like an owner on the moat layer.