A change in AI strategy
The build-vs-buy debate is over. Buy won decisively, in a single year. Now the real strategic question has shifted: not how to build AI, but what to build on top of it, and how to make sure your data is the thing nobody else can replicate.
Something changed quietly in enterprise AI strategy over the past eighteen months and most post mortems on it are still being written. The companies that spent 2022 and 2023 building proprietary AI models are now, many of them, integrating foundation models instead. The companies that started integrating early are outpacing them. And the data on which side of that bet is winning has become too clear to argue with.
I want to be precise about what shifted, because the shift is not what most people think. This is not a story about foundation models being good enough. It is a story about what competitive advantage in AI actually consists of and why most companies spent two years pursuing the wrong one.
The flip that happened in twelve months
Menlo Ventures’ third annual State of Generative AI in the Enterprise report documented something striking: the enterprise build-vs-buy ratio for AI completely reversed between 2024 and 2025. In 2024, the split was roughly even 47% built internally, 53% purchased. By 2025, it had flipped to 24% build, 76% buy. In a single year. Enterprise generative AI spending tripled over the same period, reaching $37 billion. More money flowing in, less of it going to custom model development. The market had voted.
Sources: Menlo Ventures State of Generative AI in the Enterprise 2025; S&P Global Market Intelligence 2025; a16z enterprise AI survey October 2025
The 42% figure from S&P Global deserves particular attention. Nearly half of the companies that built custom AI initiatives in 2025 scrapped most of them. Not pivoted. Not iterated. Scrapped. The economics of custom AI development ran directly into the economics of foundation model improvement and foundation models won almost every time.
What I watched happen and what it cost
I had a close view of what this shift looks like from the inside. A company I was watching not one I had backed, which is part of what made it instructive spent 18 months and approximately $3 million building a proprietary AI model for their domain. The logic was defensible at the time: proprietary models mean proprietary moats. The team was technical. The data was real. The thesis made sense in 2022.
By the time they finished, GPT-4 handled the core use case better than their model and the API cost for their expected volume was somewhere around $20 a month. The 18 months they spent building were 18 months they did not spend on distribution, on customer development, on building the data flywheel that would have been genuinely defensible. The model they built was a commodity the moment it shipped. The proprietary customer data they would have accumulated during those 18 months of active deployment was not.
I think about that company often. Not because the decision was obviously wrong when they made it. But because the lesson it contains about where AI moats actually live is one I now apply to every AI company I evaluate.
The decision that company made is not unusual. It was the consensus decision in 2022 and 2023. “Our AI advantage comes from our proprietary model” was in dozens of pitch decks I saw during that period. What most of those decks got wrong was the premise: that the model is where the defensibility lives. In almost every case, it does not. The model is the commodity. The data is the moat. The workflow is the switching cost. The integration speed is the competitive gap.
Why building the model was the wrong bet for most companies
The three forces that drove the build-to-buy reversal are worth naming specifically, because understanding them changes how you think about AI strategy going forward.
OpenAI, Anthropic, Google, and Meta have each invested billions in model capability. A company spending $3 million on a custom model is not competing with those teams it is racing against them while they run on a different track entirely. Every six months the foundation models improved enough to obsolete another category of custom model justification. By 2025, the capability threshold at which a custom model outperforms a well-prompted foundation model had moved so far up the capability curve that it was relevant only for a narrow set of specialized applications.
The upfront build cost for custom AI typically $60,000 to $250,000 for initial development was what most companies budgeted. What they did not budget for was the ongoing maintenance, the retraining cycles as the underlying task evolved, the engineering time to handle model drift, and the opportunity cost of the engineering talent that could have been building product instead. S&P Global found that 42% of companies that built custom AI initiatives scrapped most of them in 2025 not because the models failed on day one, but because the total cost of ownership was unsustainable relative to API alternatives.
a16z’s survey found that AI-native companies are outpacing competitors at 2–3× speed and reaching $100 million ARR significantly faster than prior software generations. The common variable is not model quality it is deployment velocity. The companies that integrated foundation models and iterated on their workflows shipped faster, learned faster, and accumulated proprietary data faster than the companies that spent 12–18 months in model development before getting a single customer interaction into production.
The new strategic question and where the moat actually lives
Once you accept that the model is not the moat, the strategic question changes entirely. It is no longer “which AI should we build?” It is “what do we build on top of AI that nobody else can replicate?”
The answer, in almost every durable case I have seen, is one of three things: proprietary data, workflow integration depth, or network effects that compound through AI usage. Sometimes all three together.
Proprietary data is the most durable moat because it cannot be bought at any price. If your product generates data that only exists because your customers use it behavioral data, transaction data, domain specific interaction data and that data makes your AI outputs meaningfully better than a generic model, you have something a competitor cannot replicate by throwing money at API costs. The data compounds the longer you run. The model is a commodity. The data is not.
Workflow depth is the second most durable moat because it creates switching costs that are genuinely expensive. When your AI system is embedded in the way a company runs its operations not as a standalone tool but as the connective tissue between steps in a process replacing it requires rebuilding the process, not just swapping a vendor. One enterprise technology leader in the a16z survey noted that after tuning prompts to a specific model across multiple integrated workflows, switching providers had become “a task that can take a lot of engineering time.” That friction is not a bug in the integration strategy. It is the point.
Microsoft invested $13 billion in OpenAI and then, in 2026, began quietly replacing OpenAI models across its M365 stack with its own MAI family. The same pattern that played out in cloud computing build vs. rent is now playing out in AI models. Even the companies at the frontier of model capability are making model-agnostic architectural decisions because they have learned that locking strategy to any single model is a structural vulnerability.
The lesson for every company below that level: if Microsoft is building model-agnostic infrastructure, you should almost certainly be doing the same. The model you choose today will not be the model you want in 18 months. Build your AI workflows to redirect traffic without touching application code. Own the data layer. Stay agnostic about the model layer.
What this means for founders, enterprises, and investors
For founders
If you are building an AI product and your competitive advantage is the model, you are building on sand. The question to ask is: what data does my product generate that gets more valuable the more customers use it? What workflow am I replacing so completely that switching back is more expensive than the product itself? Those are the questions that produce durable businesses. The model is the starting point, not the destination.
For enterprise leaders
The 76% buy consensus is directionally right but operationally incomplete. Buying AI without owning the data it generates is renting intelligence from someone else’s infrastructure. The strategic priority is not which AI vendor to choose it is building the data infrastructure that makes your AI outputs proprietary regardless of which model you run underneath. Own the data layer. Stay agnostic about the model layer.
For investors
The AI companies worth backing in 2026 are not the ones with the most sophisticated models. They are the ones accumulating data assets that compound with usage, building workflow integrations that create genuine switching costs, and deploying fast enough to build a data lead that a later entrant cannot close by paying for better API access. The model is a commodity. The moat is downstream of it.
For the companies still building
Custom model development is still the right answer in a narrow set of cases: when AI is your core product, when you have genuinely unique proprietary data that produces meaningfully better model outputs, or when data sovereignty requirements make third-party APIs legally untenable. Outside those cases, the engineering hours are almost certainly worth more building product and accumulating data than training models that will be obsolete before the training run completes.
The new AI strategy in one sentence
For most companies, the correct AI strategy in 2026 is: integrate the best available foundation model for your use case today, build model-agnostic infrastructure so you can switch tomorrow, and spend every engineering hour you are not spending on model development on building the data flywheel that will be your actual competitive advantage in three years.
The company I watched burn 18 months and $3 million had all the ingredients for a durable business. They had a real domain, real customers, and real data. What they did not have was 18 months of customer interaction data accumulated while running a production AI product. They spent those months building a model. Their competitors spent those months building a moat.
That is the change in AI strategy. The question was never who could build the best model. It was always who could accumulate the data and the workflow depth that made their AI outputs irreplaceable. The answer to that question does not require building a model at all. It requires shipping faster than everyone else, learning from every customer interaction, and never letting anyone else own the data your product generates.