Enterprises purchased 76% of their AI use cases in 2025 rather than building them, up from 53% the year before. Meanwhile, the models themselves are converging: the gap between the top US and Chinese models has narrowed to 2.7%. If the model is the same one your competitor can rent by the token, it is not an advantage for you.
Two years ago, the strategic question in most enterprise AI conversations was whether to train a proprietary model. Boards asked it, vendors encouraged it, and a surprising number of companies spent real money answering it. The question has largely settled, and it has settled on buying.
Menlo Ventures, surveying roughly 500 US enterprise decision-makers in November 2025, found that 76% of AI use cases are now purchased rather than built internally. A year earlier, the split was 47% built and 53% bought. Enterprise generative AI spend over the same period hit $37 billion, up from $11.5 billion, a 3.2x increase in twelve months. Companies are spending more and building less, and both trends are accelerating.
That is the correct instinct, applied to the wrong half of the problem. Buying the model is right. Assuming the purchase confers an advantage is not. Deloitte's State of AI in the Enterprise 2026 report, based on 3,235 leaders across 24 countries, found that only 34% of organizations are using AI to create new products, reinvent processes, or reimagine business models. The other two-thirds are optimizing functions that already existed, which is useful and entirely non-differentiating. This article covers why the model layer is becoming commoditized, what remains to be built on top of it, and where the build-buy line actually lies.
What Does "Buy the Model, Build the Moat" Mean?
Buy the model, build the moat means renting frontier intelligence from a provider rather than training your own, and investing the saved capital in the layers a competitor cannot copy: proprietary data, workflow depth, evaluation systems, and domain expertise. The model is an input available to everyone at falling prices. The advantage lies in what you feed it, what you wrap around it, and how well you can prove it works.
Why Is the Model Layer Commoditizing?
Three measurements tell the story, and they all point in the same direction.
Capability is converging at the top. Stanford HAI's 2026 AI Index found that the performance gap between leading US and Chinese models narrowed to 2.7% by March 2026, down from a range of 17.5 to 31.6 percentage points in May 2023. Among the leaders themselves, the spread has collapsed to arena scores separated by a few dozen points. When the second-best option is within a rounding error of the best, "we use the best model" ceases to be a strategy.
Price is collapsing. Epoch AI's tracking of inference prices found declines ranging from 9x to 900x per year, depending on the task. Matching GPT-4-level coding performance cost $37.50 per million tokens in March 2023 and $0.10 per million tokens by July 2024. Capability you paid a premium for last year is a commodity line item this year, and anything you built a business case around at last year's prices needs re-examining at this year's.
Provider share is unstable. In Menlo's enterprise data, Anthropic holds 40% of enterprise LLM API usage, OpenAI 27%, and Google 21%. OpenAI's share was 50% in 2023. Whatever you standardize on today has a meaningful chance of not being the best choice in eighteen months. That is an argument for treating the model as a swappable component rather than a foundation.
None of this means models are unimportant. It means the model is now a supplier decision, evaluated like cloud regions or payment processors: on price, latency, reliability, and terms. Supplier decisions are worth optimizing. They are not worth mistaking for a strategy.
What Is Actually Left to Build?
Four layers, and every one of them is harder to copy than a model choice.
Proprietary data and the context around it. Not a data lake. The specific, current, permissioned information about your customers, your operations, and your domain that no general model has seen. This is the difference between an assistant that answers plausibly and one that answers correctly for your business, and it is the entire reason retrieval architectures exist. We built exactly this layer for a SaaS platform that needed natural-language querying over its domain data, as described in our Gacela AI case study. The model was rented. The retrieval design, the data preparation, and the domain grounding were not.
Workflow depth. A model that drafts an answer is a feature. A system that pulls the case file, checks entitlement, drafts the response, routes the exception, and writes back to the system of record is a business process. The second one requires months of integration work with systems your competitor doesn't have, and that's where switching costs actually accumulate.
Evaluation and governance. The ability to prove your AI system performs, catch it when it degrades, and demonstrate that to a regulator is now a competitive asset. Deloitte found that only 20% of organizations maintain mature oversight models for autonomous agents. In regulated markets, it is more than an asset: obligations under frameworks like the EU AI Act now reach US companies whose AI outputs are used in Europe, and the documentation and logging requirements are engineering deliverables, not policy memos.
Domain expertise inside the build team. Menlo found that 47% of AI deals reach production, compared with 25% for traditional SaaS, which is encouraging until you consider the corollary: more than half still do not. Deloitte identified insufficient AI competency among workers as the single largest barrier to embedding AI into workflows. The constraint is not model access. It is people who understand both the model and the process it is meant to change, which is why the engagement model you choose for AI talent has a greater impact on outcomes than the model vendor you select.
Where Do Companies Get the Build-Buy Line Wrong?
Building at the wrong layer. Training or heavily fine-tuning a foundation model to serve a use case that a retrieval layer over a rented model would handle is the most expensive mistake in the category. It burns the capital and the scarce talent that the differentiating layers need.
Buying at the wrong layer. The mirror error. Purchasing a packaged application for a workflow that is the actual source of your margin, then discovering that you and four competitors now run identical processes on identical software. Buy commodity workflows. Build the ones customers pay you for.
Treating the model as permanent. Hardcoding a single provider's API across the codebase converts a supplier decision into an architectural one. Given how quickly share and price have moved, an abstraction layer between your application and the model is cheap insurance, and it is standard practice in the machine learning systems we put into production.
Underestimating the data work. The moat layer is unglamorous. Extraction, cleaning, labeling, lineage, permissioning, freshness. It is most of the effort and none of the demo, which is why it is the layer most often skipped and the most common reason a pilot never scales. Our data science practice exists mainly because this is where projects live or die.
Confusing adoption with transformation. Two-thirds of the organizations Deloitte surveyed are optimizing existing functions. That is a reasonable place to begin and a poor place to stop. Only 20% report AI-driven revenue growth today, while 74% expect it in the future; this section focuses on the gap between those two numbers.
How Do You Build a Moat on Rented Intelligence?
Pick the model like a supplier on a scorecard. Cost per unit of work, latency, reliability, data handling terms, and roadmap risk. Re-run the scorecard every two quarters. Expect the answer to change.
Spend the savings on the data layer. Whatever you did not spend on training, spend on making your proprietary information retrievable, current, and permissioned. This is the compounding asset. Models reset every release; a well-maintained knowledge layer accrues.
Build the evaluation harness before the feature. A test set drawn from your own domain, a scoring method your business owners agree with, and a regression check that runs on every model change. Without it, you cannot safely swap providers, which means you cannot treat the model as a commodity even when it is one.
Go deep on one workflow rather than shallow on ten. Depth is what competitors cannot replicate quickly. Our healthcare AI engagement scaled a single organization's data capability with specialized talent rather than spreading thin across a portfolio of pilots, and depth is why it reached production.
Staff for the layers you are building, not the one you are buying. You do not need model researchers. You need data engineers, ML engineers, and domain-fluent product people, which is a different hiring problem and, for most US companies, a different sourcing strategy. Teams closing that gap with nearshore AI talent in overlapping time zones tend to move faster than those queuing for scarce domestic senior engineers, and the build work itself belongs with a partner who engineers for production rather than for the demo.
Common Questions About Building an AI Moat
Should we train our own AI model?
For nearly all companies outside the AI research sector, no. Enterprises now purchase 76% of AI use cases rather than building them, and inference prices for a given capability level have fallen by up to 900x per year on some tasks. The capital and scarce engineering talent yield more durable returns across the data, workflow, and evaluation layers.
If everyone uses the same models, how do you differentiate?
Through what the model does not have: your proprietary data, your workflow integrations, your evaluation and governance systems, and domain expertise inside the team building the system. Deloitte found that only 34% of organizations are using AI to create new products or reinvent processes, so the differentiation gap is wide and mostly unclaimed.
What is an AI moat?
An AI moat is the set of assets a competitor cannot buy or quickly copy: current permissioned proprietary data, deep integration with systems of record, a working evaluation harness that demonstrates performance, and people who understand both the model and the business process. The model itself is not part of it, because your competitor can rent the same one this afternoon.
How do we avoid vendor lock-in with AI models?
Put an abstraction layer between your application and the model provider, maintain an evaluation harness built from your own domain data so you can compare candidates objectively, and re-score providers at least twice a year. Enterprise LLM API share has moved sharply in three years, so portability is a practical requirement rather than a theoretical one.
If your AI roadmap still has a line item for training a model but no line item for the underlying data layer, the budget is pointed at the commodity. Talk to the Golabs team about where your defensible layer actually sits and what it takes to build it.