Build, Buy, or Partner: How to Actually Invest in AI Capability

Most advice on “how businesses should invest in AI” is written for someone deciding whether to buy a stock. It talks about diversification, market size, and getting in before the window closes. None of that transfers to the actual decision an engineering-led company faces, which isn’t a portfolio allocation question at all. It’s a capability question: build it, buy it, or bring in someone to build it with you — and each of those three paths has a genuinely different cost structure, not just a different price tag.

We’re an engineering firm, not a financial advisor, so this is the only version of “invest in AI” we’re actually positioned to write well: what the decision looks like from inside the budget, not from outside the ticker.

Skip the market-size number

It’s tempting to open a piece like this with a big number — global AI spending is enormous and getting more enormous — but the ones that circulate are usually stale the moment they’re quoted. A widely-repeated “$300 billion by 2026” figure traces back to an IDC forecast from 2022; IDC’s own current numbers already put worldwide AI infrastructure spending alone at roughly $497 billion for 2026, and Gartner’s current estimate for total worldwide AI spending this year is roughly $2.6 trillion.

IDC, AI infrastructure spending, 2026 forecast raised to $497 billion — idc.com. Gartner, worldwide AI spending forecast to grow 47% in 2026 — gartner.com. The two count different things: IDC measures infrastructure hardware, Gartner the whole stack.

The number moving that fast should tell you something on its own: the market size was never really the useful part of the question. What matters is not how big AI spending is in aggregate, but how much of it is landing anywhere near its promised value — and on that specific point, the honest data is much less flattering than the spending totals suggest.

The gap between potential and realized

In 2023, McKinsey estimated that generative AI could add on the order of $2.6 to $4.4 trillion a year in value across the functions it studied. That figure gets cited constantly, almost always without its most important qualifier: it was always a theoretical potential, not a forecast of what companies would actually capture.

McKinsey Global Institute, “The economic potential of generative AI: The next productivity frontier,” June 2023 — the $2.6–4.4 trillion range is explicitly framed as potential annual value across analyzed use cases, not realized spend. mckinsey.com

How much of that potential has actually been captured is the harder question, and the survey most often quoted on it — which reports that almost no enterprise generative-AI pilot has moved a profit-and-loss number — is disputed on its method. The case for reading it as a direction rather than a rate is in Most AI Failures Are Missing a Gate, Not a Model, and it does not need re-arguing here. For an investment decision the exact rate changes nothing anyway: a theoretical upside is not a forecast of your own return, and no spending total tells you which side of that line your budget lands on.

Three paths, honestly compared

Build. You own the whole thing, including the parts that are tedious and unglamorous — evaluation harnesses, data pipelines, the on-call rotation for a system that can now fail in new ways. Time to first value is the slowest of the three paths. The payoff is a real, defensible capability that a competitor cannot buy off a price list, which matters enormously if the capability touches something core to what makes your product different and matters much less if it doesn’t.

Buy. Fastest path to something working. The ongoing cost is real but predictable — a subscription line, not a headcount line — which is genuinely attractive for anything that isn’t core to your differentiation. The tradeoff is that your competitor can buy the exact same tool. There is no moat in a capability every company in your market has equal access to.

Partner. An embedded senior team building the thing with you, inside your stack, with the explicit goal of leaving you able to run it yourselves. Slower than buying, faster than building alone, and — done properly — the capability transfers to your own team over the engagement rather than staying locked inside a vendor relationship you have to keep renewing forever. This is close to describing our own model, which is exactly why it is named here as a real third option rather than left as a marketing footnote — declaring the interest is the honest move, hedging about it is not.

Cost mechanics people skip past

“How much will this cost” depends enormously on which of three different technical approaches you’re actually funding, and they are not interchangeable line items:

• Prompting against a general model. Cheapest to start, fastest to iterate, and the right choice for a large share of use cases. The cost is almost entirely inference — pay per call, scale with usage — and the ceiling is whatever a well-prompted general model can do without seeing your proprietary data.
• Retrieval over your own data. More engineering upfront — a pipeline, an index, a permission boundary around what the system is allowed to retrieve — in exchange for answers grounded in information a general model never saw. The ongoing cost is mostly the pipeline and its maintenance, not the model calls themselves.
• Fine-tuning. The most expensive and slowest of the three, and the one most often reached for before it’s justified. It buys behavior a prompt genuinely can’t — a consistent house style, a narrow specialized skill — but it is a real, recurring cost every time the underlying model updates, and it is frequently the wrong tool for a problem retrieval would have solved for a tenth of the price.

The decision, stated plainly

Buy the capability that every competitor can also buy. Build the capability that is genuinely core to what makes you different, and be honest that it will be slow. Partner for the capability that sits in between — important enough to want in-house eventually, urgent enough that you can’t wait for your own team to ramp on it from zero. Most companies do not need a bigger AI budget. They need to know which of these three questions they’re actually answering before they spend it.