--- title: "Build, buy, or partner | Avanti Technologies" description: "Where deep-tech AI startups should own the stack, where they should rent it, and where ownership has quietly stopped being an option — with decision protocols." url: https://www.avanti.ie/resources/build-buy-partner section: "Research Note · Strategy" published: 2026-07-28 modified: 2026-08-06 publisher: "Avanti Technologies" --- // Avanti Technologies · Research Note RN-2026-04 28 July 2026 # 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. // Executive summary - **The three-way question has collapsed into a layer question.** Rent commodity model and infrastructure layers, own the intelligence layer — domain logic, proprietary data, decision rules — and partner where execution expertise outruns internal capacity. - **Falling inference prices are an argument against building models, not against building products.** Capability at a fixed quality bar has deflated by roughly two orders of magnitude in three years. Anything priced on 2023 token economics is mispriced. - **Deep tech pays a time tax that changes the arithmetic.** BCG puts stage-to-stage maturation 25–40% slower than other tech ventures. A build decision consumes runway measured in scientific milestones. - **"Buy" now runs in both directions, and startups are usually the object.** Licence-and-hire structures moved billions into Big Tech without triggering merger review between March 2024 and early 2026. - **The regulatory clock is now a build-cost input.** GPAI model obligations have applied since 2 August 2025; transparency duties land on 2 August 2026. ~1,000× Reported fall in cost for a fixed GPT-4-class quality bar over roughly three years. 25–60% Gross margin band now attributed to AI-first software vs 75–85% for traditional SaaS. $120M Approximate HSR threshold that licence-and-hire structures are designed to sit beneath. ## 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. **Building commodity infrastructure yourself does not make it strategic.** The durable test is whether the capability makes you meaningfully better than competitors, or merely keeps you level with them. ## 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. Dimension | Build | Buy | Partner | Time to first production value | Quarters to years; one estimate puts 18–24 months and $1.5–2M+ per year | Days to weeks for a hosted product; months if it is an acquisition | Weeks; a fixed-scope, time-boxed proof of concept (e.g. four weeks) | What you actually own | Code, weights, data pipeline, and the whole lifecycle | A contract and a roadmap dependency you do not control | The outcome and the IP — if the engagement is written that way | Recurring obligation | Model drift, retraining cycles, eval maintenance, compliance documentation | Price and terms risk; vendor regret is among the most expensive problems | Knowledge-transfer discipline; a partner who builds for you leaves a dependency | Principal failure mode | Sunk cost in a commodity layer; first-time builders reported to fail over 60% | Undifferentiated product — if any competitor can buy the same tool, it is overhead | Open-ended retainers, black-box delivery, no structured handover | Best fit | The one or two capabilities that are your defensible advantage | Commodity capability, mature market, low switching cost | Differentiated but not permanent: low internal capability + high complexity | ## 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. Layer | Default posture | Reasoning | Compute & serving | Buy / rent | Utilisation break-even rarely favours owned GPUs pre-scale. Inference routing and caching matter more than ownership. | Foundation models | Buy, with a portability plan | Price falls ~10× a year at fixed quality; the open-weight floor is close. Build only if the model is the science. | Domain data & labelling loop | Build | The only asset that compounds and cannot be bought by a competitor at the same price. | Evaluation & safety harness | Build | Your eval set encodes domain knowledge that makes improvement possible — and it is what regulators will ask about. | Orchestration & agent logic | Build (thin) | Owning orchestration lets you swap the model when a supplier changes terms or availability. | Hard integrations (EHR, ERP, lab) | Partner | High complexity, low strategic content, brutal calendar risk. A scoped engagement compresses timeline. | Compliance & audit evidence | Partner, own the artefacts | Buy expertise; keep documentation and model cards in-house. They are due-diligence assets at the next round. | Go-to-market motion | Partner | Distribution via incumbents whose customers you serve is the pattern attributed to fastest-growing companies. | ## 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. // The implication A build decision in deep tech spends the scarcest resource — the 25–40% longer interval to the next milestone — on a layer that may deflate by an order of magnitude before it ships. Every quarter of build time on a commodity layer is a quarter not spent producing the proprietary data that makes the next round raisable. ## 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. Deal | Reported structure | Outcome for the shell | Microsoft / Inflection Mar 2024 | ~$650M total; ~$620M for model licence + ~$30M non-suit; ~70-person team hired including CEO | No HSR filing. Last valued near $4B; investors took modest return or were repaid $1.3B | Amazon / Adept Jun 2024 | ~$25M licensing fee; roughly two-thirds of team hired; no equity purchased | Drew regulatory questions; FTC opened probe following Microsoft–Inflection inquiry | Google / Character.AI Aug 2024 | $2.5–2.7B for non-exclusive model licence; founders returned to Google | Proceeds used to buy out investors; staff received cash and equity — most employee-favourable case | Meta / Scale AI 2025 | Large minority stake plus hiring CEO to lead internal lab; voting rights transferred back | Customer flight — Google terminated. A live warning about neutrality-dependent business models | 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. // Structure it this way - Fixed scope, fixed price, fixed end date - Named knowledge-transfer deliverables, not a promise - You own the code, the data, and the decision - Exit criteria defined before kickoff - A written model-portability requirement // Walk away from - A timeline too fast for the scope described - Black-box delivery with no handover plan - Vague answers on data ownership and security - No reference clients at your scale or vertical - A proposal that leads with tools before outcomes 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. Date | What applies | Who it binds | 2 Aug 2025 · in force | GPAI model obligations (Art. 51–56); AI Office operational | Anyone who builds a general-purpose model | 2 Aug 2026 · live | Article 50 transparency (disclose AI interaction, AI-generated content); watermarking grace period to 2 Dec 2026 | Almost every AI product touching EU users | 2 Dec 2026 | Watermarking grace period ends; Art. 5 prohibition on non-consensual intimate imagery and CSAM | Generative image, video, and audio providers | 2 Dec 2027 | High-risk obligations for Annex III systems — recruitment, credit, education, law enforcement | Applied-AI startups in those verticals | 2 Aug 2028 | High-risk obligations for AI embedded in Annex I regulated products — medical, machinery, vehicles | Most hardware-bearing deep tech | 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. - **Substitution test.** If a competitor can buy the same capability tomorrow, building it is overhead. Buy. - **Compounding test.** Does using it generate proprietary data or process knowledge that makes it better next quarter? If yes, build. - **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. - **Milestone test.** Does the build consume time you owe to the next scientific or regulatory milestone? This test overrides most cost analysis. - **Obligation test.** Can you staff the lifecycle — drift, retraining, evals, documentation — for three years? If not, ownership is theatre. - **Reversibility test.** How expensive is it to undo? Bias toward the cheaply reversible option when evidence is close. // Default posture by stage — deep-tech AI venture Pre-seed → seed Buy everything except the science. Prove the mechanism, not the platform. A short prototype sprint on existing tools answers architecture questions faster than months of vendor evaluation. Series A Build the data loop and the eval harness. Partner for hard integrations and first-customer delivery. Still buy models and compute. Series B → C Hybrid by default: own the layer your customers pay for, plus compliance artefacts and portability. Consider owned inference only when utilisation math and margin pressure both justify it. 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. ### Discuss this research with our team Whether you're evaluating build-versus-buy decisions, designing a deep-tech stack, or planning an investor roadshow, we can help map this framework to your context. [Book a discovery call →](https://calendly.com/kumar-avanti) ## // Keep reading [White Paper · ManufacturingOptimising fabrication for cost and deliveryHow AI and software engineering cut avoidable cost and shorten lead times for Irish and EU metal fabricators — with a reference architecture and pilot plan.](https://www.avanti.ie/resources/fabrication-cost-and-delivery)[Field Notes · Enterprise AITwo disciplines, one goal: lowering AI riskHow AI governance and AI security differ, what each one guards against, and how to combine them into a single layered defence for enterprise AI systems.](https://www.avanti.ie/resources/ai-governance-security)[Practical Guide · 2026GDPR & the new rules of dataA plain-English GDPR handbook for SMEs — and how the EU AI Act, Data Act and a wave of new regulation now sit on top of it.](https://www.avanti.ie/resources/gdpr) [All resources →](https://www.avanti.ie/resources) ## Frequently asked questions ### When should a company build technology rather than buy it? Build where the capability is genuinely differentiating and where owning it compounds — the parts customers pay for and competitors cannot replicate quickly. Buy or rent commodity layers where a vendor's economies of scale beat anything you can run yourself, and revisit the decision as costs and regulation move. ### What is a licence-and-hire or reverse acqui-hire? A structure that pairs a non-exclusive technology licence with the hiring of a company's founders and core researchers, leaving the legal entity intact. Because it is not an acquisition, it avoids the mandatory Hart-Scott-Rodino pre-merger notification that US deals above roughly $120M trigger, and can close in weeks. Microsoft-Inflection (March 2024) and Amazon-Adept (June 2024) are the reference examples; both drew regulatory attention. ### How long does it take to build an AI capability in-house? One commonly cited estimate is 18 to 24 months and $1.5-2M or more per year to reach first production value, against days to weeks for a hosted product and weeks for a fixed-scope partner engagement. First-time builders have been reported to fail over 60% of the time. That gap is why building is only defensible for the one or two capabilities that are genuinely your advantage. ### What is the difference between buying and partnering for technology? Buying acquires a capability as a product, licence or company — you own a contract and a roadmap dependency you do not control. Partnering is a collaborative engineering relationship where external expertise contributes to architecture while you retain ownership of the outcome and the IP, provided the engagement is written that way. Partnering fits capabilities that are differentiated but not permanent; its failure mode is open-ended retainers and black-box delivery with no structured handover.