Recommended Articles
Tech & AI

The Build-vs-Buy Decision Is the Most Consequential AI Choice Founders Make in 2026

By Angeled · October 10, 2026 · Tech & AI
The Build-vs-Buy Decision Is the Most Consequential AI Choice Founders Make in 2026
The Build-vs-Buy Decision Is the Most Consequential AI Choice Founders Make in 2026

This week crystallized the open-versus-closed debate into dollars and deployments. Nous Research closed a $90 million Series B at a $1.5 billion valuation to turn its open-source Hermes agent — downloaded more than 22.7 million times since February — into an enterprise business, with Robot Ventures, Nvidia, Microsoft's M12, Samsung, and Union Square Ventures participating. Nvidia's State of AI in Telecommunications report, released October 6, found 89% of surveyed telecom companies acknowledge the importance of open-source models and software in their AI strategies. And Scality launched an "AI Inference Factory" on October 8: a supported, open-code stack for running inference on enterprise-owned infrastructure.

Meanwhile, closed models are getting cheaper at speed. Anthropic cut Haiku prices by as much as 90% for requests under 100,000 tokens — input from $1.00 to $0.10 per million tokens — landing exactly on OpenAI's GPT-6 Luna pricing, with real-world workload costs falling roughly 75%.

Founders in late 2026 face a genuinely hard question: do you build on open-weight models you can host yourself, or buy API access to closed frontier models? The answer is not ideology. It's arithmetic, risk, and strategy. Here's the framework.

When open wins: control, data, and sovereignty

The telecom story is the template. SoftBank built a Large Telecom Model on Nvidia's open Nemotron framework; Indosat adapted open models to Indonesian languages and culture; AT&T's chief data and AI officer called open models "essential" for a flexible, governed approach. AT&T's OTel 2.0 model was trained on 400 billion telecom-specific tokens. The common thread: proprietary network and customer data that can never leave the building, domain workloads that need fine-tuning on private data, and regulatory or sovereignty constraints.

Open architectures let companies run models on infrastructure they choose, modify systems internally, and set their own data policies. Nous's bet is that this desire — control without permanent lock-in to one provider — is worth $1.5 billion. Management expects annualized revenue to exceed $100 million by year-end, up from $36 million in mid-September. The enterprise version lets companies deploy customized agents while keeping data private.

For founders, the open case is strongest when: the workload touches sensitive customer or proprietary data; the task is narrow enough to fine-tune (support triage, document extraction, domain classification); you need cost predictability at high volume; or your customers — governments, banks, telcos, hospitals — demand data residency. Scality's launch is proof the market wants supported open stacks, not DIY plumbing: validated open-weight models, an inference-serving layer, and metering without assembling it yourself.

When closed wins: capability, speed, and the price war

Closed APIs win on time-to-market and frontier capability. The Haiku price cuts changed the math materially: high-volume classification, routing, and agentic workflows that were expensive on closed models now cost roughly a quarter of what they did. When both Anthropic and OpenAI are racing prices down 90%, the cost advantage of self-hosted open models narrows to the very highest volumes — and the engineering cost of running your own stack doesn't fall at all.

Buy the API when: you need frontier reasoning, tool use, or coding capability that open models can't yet match; you're pre-product-market fit and engineering hours are your scarcest resource; workloads are bursty and unpredictable; or your differentiation is in the application layer, not the model. A good rule of thumb for 2026: if the model is more than 20% of your burn, revisit the decision quarterly, because prices are moving that fast.

The trap: confusing openness with independence

Two subtleties founders miss. First, "open" is not "free." Self-hosting means GPU sourcing, serving infrastructure, eval pipelines, safety review, and maintenance — the exact demands Tenable highlighted when it built an enterprise-grade safety vetting process for open-source AI agents in its CyberAgents Exchange this week. Nous's $90 million raise exists precisely because converting open-source popularity into enterprise revenue requires security, support, and governance.

Second, "closed" is not "safe from lock-in." Anthropic's IPO prospectus showed 47% of its revenue flowing through Amazon and Google — concentration risk runs both directions. Smart founders abstract the model layer: build against a thin interface, keep prompts and evals portable, and treat the model as a replaceable component whether it's open or closed.

The founder's decision framework

Start closed, measure honestly, switch at scale. Ship with APIs to find product-market fit fast. Instrument per-token cost, latency, and quality from day one. When a stable, high-volume workload emerges — and only then — price out open-weight alternatives on your own or supported infra.

Keep data gravity in mind. The workload that should never have gone to an API in the first place is the one touching customer secrets or regulated data. That's your open-model beachhead.

Watch the price wars. A 90% price cut resets every build-vs-buy spreadsheet in the industry. Recompute quarterly, not annually.

The Nous round, the telecom survey, and the Haiku cuts are three readings of the same gauge: enterprises want control AND economics, and they'll pay whichever vendor delivers both. Founders who architect for portability win either way.

Angeled.com connects founders with investors.

Read Next