AI made building easy. It didn't make ownership cheap.

A dangerous assumption sits inside most AI strategies right now: that building is the easy choice, because the tools have gotten so good. It used to be a reasonable assumption — if you built something, you owned it, you controlled it, no vendor risk, no dependency, no lock-in. But the economics of AI have quietly inverted that logic, and most mid-market companies and their PE backers haven't caught up yet.

Here is what actually happened: AI tools made the act of creating software dramatically cheaper. What they didn't change — and what keeps getting more expensive — is the cost of owning what you built.

01The cost nobody is counting

When a team builds an AI capability today, the build cost is visible. The ownership cost is not.

McKinsey's data puts it plainly: model costs represent roughly 15% of the total cost of an AI system. The other 85% goes to integration, operations, maintenance, security, compliance, and governance. That 85% doesn't show up in the initial project proposal. It shows up eighteen months later, when you're wondering why the thing you built costs more to run than it did to create.

This is the trap. The question isn't whether you can build. You can. The question is whether you should own the whole stack when owning the whole stack is exactly the wrong place to spend your competitive energy.

02Where the market actually went

To understand the strategic reality, you need to understand what happened at the infrastructure layer over the last two years.

Sequoia's research on capital intensity in AI tells a stark story. The leading AI providers have collectively deployed over $229 billion in infrastructure capital, running clusters of 300,000 or more GPUs. Four major alliances — Microsoft-OpenAI, Google-Anthropic, Amazon-Anthropic, and Meta — now control the frontier model layer. The compute requirements, the research talent, and the capital needed to compete at this level are simply not available to 99% of companies. Building your own foundation model is not a strategic option for mid-market firms. It is a distraction.

What this consolidation means for you is actually clarifying: the question of whether to build at the model layer is already answered. The real question is which layers above that model layer represent genuine competitive differentiation for your business — and which ones do not.

03The layer that actually matters

JPMorgan figured this out faster than most. Rather than attempting to build or fine-tune foundation models, JPMorgan built a model-agnostic control plane — a platform layer their teams use to access AI capabilities across the organization. The result: 200,000 users onboarded in eight months. The proprietary investment wasn't in the model. It was in the orchestration, the governance, the data integration, and the workflow layer around the model. That's where their competitive advantage lives.

BCG made a similar calculation. They built a proprietary capability called GENE — their internal AI platform and integration layer — but partnered with OpenAI at the frontier model layer rather than attempting to compete with it. They then certified all 33,500 of their employees on AI, embedding that capability advantage into the human layer of their business. In February 2026, BCG and OpenAI announced a multiyear expansion of their relationship through OpenAI's Frontier Alliance. CEO Christoph Schweizer described BCG's role as helping organizations redesign functions and workflows, upskill people, and create changes in P&L performance — rather than simply increasing model or token usage.

04Partnership isn't what you think it is

There is a mental model problem that holds many mid-market companies back from smart partnerships. They hear "partner with an AI vendor" and think of outsourcing, dependency, loss of control.

That's the wrong frame. The right frame is this: you are accessing a rapidly evolving capability at a cost basis you could never replicate internally, while retaining full ownership of the layer where your business actually differentiates. Partnership in the current AI market is how you stay current without burning your engineering capacity on infrastructure that is not your core business.

The companies in the Powered by Claude ecosystem — over 50 organizations now shipping production AI products built on Claude — understand this. Block, Apollo, Arctic Wolf, Augment Code, Attention, Base44: these are not companies that outsourced their AI strategy. They partnered at the model layer and built differentiated products on top. Gainsight offers a useful mid-market illustration: they partnered for AI infrastructure — the underlying model and compute — while building proprietary AI workflows specific to customer success management. The partnership gave them access to frontier capabilities. The proprietary build gave them differentiation. Neither alone would have been sufficient.

05The lock-in problem just got solved

The most legitimate concern about AI partnerships has always been lock-in. What happens when the vendor relationship changes? What happens when you're dependent on a capability you don't control?

That concern is now substantially addressed by an open standard called the Model Context Protocol, originally introduced by Anthropic in late 2024 and now adopted across the major AI providers including ChatGPT and Claude. MCP functions as what its documentation describes as a "USB-C port for AI" — a standardized interface between AI models and the data sources, tools, and systems they need to work with.

What MCP means practically is that the integration layer you build today is not locked to any single model provider. If you build your orchestration and integration logic on MCP, you can switch underlying models without rebuilding everything. The portability that used to require significant engineering investment is now available as a design pattern. Partnership risk just got meaningfully lower.

06What the data says about how companies are actually deciding

If you're wondering whether your instinct to buy more and build less is supported by how sophisticated organizations are moving, the answer is yes — decisively.

BCG surveyed 270 organizations across 15 sectors about how they developed and deployed generative AI. About half of the organizations experimenting with GenAI were developing solutions entirely in-house. But companies combining their own teams with external vendors reported greater satisfaction, clearer deployment roadmaps, and better ROI tracking than organizations going entirely alone. Among organizations already using vendors, 84% worked with two or more partners, while 62% combined external vendors with internal capabilities.

Menlo Ventures tracked enterprise AI sourcing behavior and found that organizations shifted from 47% purchased AI capabilities in 2024 to 76% purchased in 2025. That shift happened in twelve months. It reflects organizations learning that the cost of internal ownership is higher than anticipated, and that the quality of external partners has improved faster than internal teams can match.

The portfolio approach — different sourcing strategies for different layers — is not a hedge or a compromise. It is the mature operating model. Vista Equity Partners formalized this logic into what they call an Agentic AI Factory approach across their portfolio companies, creating shared frameworks, shared evaluation criteria, and shared implementation patterns rather than leaving each portfolio company to figure out AI strategy independently. The PE model for AI is increasingly portfolio-level rather than company-level, and the firms that recognize this early are compounding their advantage.

07The framework you actually need

Before deciding what to build, buy, or partner on, there are six questions worth sitting with honestly: Does this capability create genuine strategic differentiation, or is it table stakes every competitor will have? How quickly is the external market evolving in this area — fast enough that anything you build today might be obsolete in eighteen months? Do you have a realistic right to win, given your current talent, data, and domain expertise? How reversible is the decision if you get it wrong? What is the true operating burden — not just the build cost, but the ongoing cost of running, maintaining, securing, and evolving this over three years? And most importantly: where does your proprietary advantage actually reside in this capability area?

The answers should drive you toward one of four positions.

Most companies, when they work through this honestly, find that they have overinvested in the Build quadrant and underinvested in the Partner and Rent quadrants. The result is too much engineering overhead on capabilities that don't differentiate, and too little investment in the orchestration and workflow layers where real competitive advantage is built.

08The honest strategic conversation

The companies winning in AI right now are not the ones who built the most. They are the ones who made the clearest decisions about what to own and why.

JPMorgan didn't win because they built a model. They won because they built the right layer around the model. BCG didn't win because they competed at the frontier. They won because they built proprietary glue and embedded capability into their people. The pattern across every compelling AI success story in enterprise right now is the same: disciplined clarity about which layer matters, followed by concentration of proprietary investment at exactly that layer.

If you are a mid-market CEO or PE principal and you haven't yet had the honest conversation about layer-level AI strategy — not just "what should we build" but "where does our proprietary advantage actually live in an AI-enabled world" — that conversation is overdue. The market is moving fast enough that the cost of delayed clarity is no longer abstract. It is compounding every quarter.