Build, buy, or partner

Where deep-tech AI startups should own the stack, where they should rent it, and where ownership has quietly stopped being an option.

1. What actually changed

Two years ago the build-versus-buy debate for AI was a debate about capability: could you get a good enough model without training one? That question is closed for all but a handful of frontier labs. The 2026 debate is about commitment — which parts of the stack a company can realistically own, maintain, document, and evolve over a multi-year horizon.

Three forces did the work. First, price. Stanford's AI Index reported that querying a model at GPT-3.5-equivalent accuracy fell from about $20 per million tokens in November 2022 to roughly $0.07 by October 2024 — a ~280× reduction. Later 2026 write-ups extend the trend to ~1,000× over three years, though the multiplier depends on which benchmark and which two dates you pick.

Second, the open-weight floor. The open-vs-closed performance gap narrowed from 8% to 1.7% on some benchmarks within a single year. For a startup, the open floor is optionality. It is what makes "swap the model underneath" a real plan rather than a slide.

Third, the failure record. MIT's Project NANDA found that about 95% of enterprise generative-AI pilots produced no measurable P&L impact. We cite it because it is the most-referenced datum in every 2026 build-buy-partner discussion. The failures were concentrated in integration and organisational learning, not model quality or regulation.

2. Three postures, honestly costed

The vocabulary is loose in practice, so fix it. Build means permanent internal ownership of a capability and its lifecycle. Buy means acquiring the capability as a product, licence, or company. Partner is a collaborative engineering relationship in which external expertise contributes to architecture while the client retains ownership of the outcome.

3. The allocation map for a deep-tech stack

A single company-level answer is the wrong resolution. The decision is per layer, and for science-based ventures the layer that carries the defensibility is rarely the model — it is the instrument, the assay, the simulation, the physical process, or the proprietary measurement loop that produces data nobody else has.

4. The economics an investment committee should model

Margin. The marginal cost of serving an AI query is not near zero. Cost of goods scales with usage, which is why practitioner analyses place AI gross margins around 50–60%, compared to 75–85% for traditional SaaS. For a build decision this matters twice: you carry both the inference bill and the amortised cost of the team maintaining what you built.

Deflation cuts both ways. If inference cost per million tokens halves, a product priced unchanged moves to higher margin. But the same deflation erodes any moat built on model access, and there is a credible argument that current token prices reflect a subsidy-driven buyer's market.

The deep-tech time tax. BCG found that deep tech investments take 25–40% more time between funding stages, carry greater failure risk at each stage, and that among funds above $1B in assets 42% of investments are multi-round.

5. "Buy" now points at you: the licence-and-hire era

For a startup, the most consequential form of "buy" in this market is one where the startup is the asset. The structure — variously called the reverse acqui-hire or quasi-merger — pairs a non-exclusive technology licence with the hiring of founders and core researchers, leaving the legal entity intact. The mechanical driver is that traditional acquisitions above roughly $120M trigger mandatory Hart-Scott-Rodino pre-merger notification; a licence-and-hire package generates no such filing and can close in weeks.

Three practical consequences for founders and their boards. One: team composition and research output carry exit value independent of revenue. Two: standard single-trigger acceleration does not fire on a licence-and-hire; acceleration clauses should be redrafted to cover talent-and-licence transactions. Three: for anyone whose business depends on being seen as neutral infrastructure, a strategic investor's stake can be a customer-retention risk.

6. Partnering, written so it does not rot

Partnering is the option most often omitted from the analysis and the one that most often fits: a differentiated capability that needs architectural expertise, delivery speed, or an investment level an internal team cannot sustain. The failure modes are contractual, not technical. The clean form is a scoped engagement rather than an open-ended retainer.

The distinction worth writing into the statement of work is simple: a partner building with you leaves the organisation stronger after the engagement; a partner building for you leaves a system and a dependency.

7. The regulatory clock as a cost line

If you build a general-purpose model, you take on provider obligations. If you buy one, you inherit a supplier's compliance posture and your own deployer duties. For EU-facing companies the calendar is concrete. GPAI model obligations have applied since 2 August 2025. The Digital Omnibus on AI defers high-risk duties but does not dismantle them.

Penalty tiers: €35M or 7% of turnover for prohibited practices, €15M or 3% for high-risk and transparency breaches, €7.5M or 1.5% for supplying misleading information.

8. A decision protocol you can run in an afternoon

Score each candidate capability — not the company — against six tests. Rate where low internal capability with high complexity and high criticality points to partnering.

1. Substitution test. If a competitor can buy the same capability tomorrow, building it is overhead. Buy.
2. Compounding test. Does using it generate proprietary data or process knowledge that makes it better next quarter? If yes, build.
3. Deflation test. Will the cost or difficulty of this layer fall by an order of magnitude within your build window? If plausibly yes, do not build it.
4. Milestone test. Does the build consume time you owe to the next scientific or regulatory milestone? This test overrides most cost analysis.
5. Obligation test. Can you staff the lifecycle — drift, retraining, evals, documentation — for three years? If not, ownership is theatre.
6. Reversibility test. How expensive is it to undo? Bias toward the cheaply reversible option when evidence is close.

Ask any AI company for its model-routing and caching strategy with the same rigour as customer acquisition cost — in 2026 it is the larger swing factor. And ask what happens to the product if its primary model supplier changes price, terms, or availability next quarter.