The Open Source AI Spending Paradox: Enterprise Budgets Are Growing, But Open Source Is Losing

**TL;DR:** Enterprise AI budgets are exploding (up 55% in two years, with 65% of enterprises planning further increases). But the share going to open-source models is shrinking, not growing — down from 19% to 11% of spend. The models themselves are cheaper than ever and nearly as capable as proprietary ones. So why are enterprises doubling down on closed? Because what they're actually buying isn't the model. It's the governed infrastructure that makes AI safe, auditable, and production-ready. The platforms that provide that layer, not the model providers, are the natural winners.

---

Here's a statistic that should make open-source advocates deeply uncomfortable.

According to the latest a16z CIO survey of 100 Global 2000 technology executives, enterprise spending on open-source AI models dropped from 19% to 11% of total AI budgets in the past year. Closed models moved in the opposite direction: 81% to 89%. The average enterprise now spends $7 million annually on LLMs, up from $4.5 million two years ago. Sixty-five percent expect to increase that further in 2026.

Total budgets are growing. The open-source slice is shrinking. That's the paradox.

And it gets stranger when you look at the model economics. Convly's June 2026 pricing analysis of 29 models found the median open-weight model costs about $0.15 per million blended tokens. The median proprietary model? $6.00. That's a 39x gap. The five cheapest models in the market are all open-weight. The five most expensive are all proprietary. On raw token economics, open source should be cleaning up.

It isn't. Why?

## What are enterprises actually paying for?

I think the answer is uncomfortable for both sides of the open-versus-closed debate, but it's obvious once you look at where enterprise AI budgets actually go.
The largest single category, 30 to 40%, goes to software and SaaS AI tools: ChatGPT Enterprise, Claude Enterprise, Microsoft 365 Copilot, GitHub Copilot. These are per-seat licensed products with SLAs, admin consoles, audit logs, and someone to call when things break.

The next 20 to 25% goes to cloud infrastructure. Then 15 to 20% on talent. Governance, security, compliance, and monitoring now consume 8 to 12% of AI budgets, up from roughly 3 to 5% in 2024. It's the fastest-growing line item.

Notice what's missing? Nobody has a budget line for "model weights."

Open-source models are free to download. But deploying them safely in an enterprise context — with access controls, audit trails, compliance documentation, model version pinning, fallback routing, cost monitoring, and a support escalation path — costs a fortune in engineering time. That's why Mozilla's July 2026 State of Open Source AI report found something striking: 79% of developers use open models, compared to 71% using closed ones. But only 51% of open-model teams reach production, versus 63% for closed. The gap isn't capability. It's operational tooling.

Mozilla put it bluntly: "Open ships easy. Open deploys hard."

## The model is the commodity. The harness is the product.

This is the real structural shift that the spending numbers reveal. The model layer is pricing toward zero: GPT-4-class inference dropped from $20 to $0.40 per million tokens in 36 months. Open weights are at or near parity on coding and general knowledge. The capability gap between open and closed models now sits at roughly 3.3 percentage points on average, and on the workloads most businesses actually run, it's negligible.

If the model were the product, enterprises would be routing everything through the cheapest open-weight option that passed their eval set. They're not.

They're paying for the harness: the layer above the model that handles governance, observability, access control, compliance, model routing, and cost management. And right now, the best-integrated versions of that harness ship attached to proprietary models. Anthropic, OpenAI, and Microsoft aren't winning because their models are better in ways that matter for most workloads. They're winning because their enterprise contracts bundle the model with the operational layer enterprises can't function without.

This is deeply familiar if you've watched any other enterprise software market mature. Nobody pays for the database engine anymore. They pay for the managed service wrapped around it. AWS didn't win because it had better servers.

## Where no-code platforms fit in all this

I've been building on no-code platforms for years, and this paradox is the strongest argument I've seen for why platforms like **Stacker** and **Bubble** are positioned to capture real value from the AI wave. In a way that pure model providers probably can't.

Here's the thing: when you build on a governed no-code platform, model selection becomes a configuration choice, not an architectural decision. A well-designed platform abstracts the model behind a governance layer that handles authentication, rate limiting, logging, and compliance. Doesn't matter whether the actual inference runs against Claude, GPT, or a self-hosted Llama variant. You swap the model without rewriting the application.

That's the enterprise dream, isn't it? Use the cheap open model for 90% of your volume, route the hard 10% to the frontier closed model, and have one audit trail for all of it. But building that routing layer yourself is a multi-month engineering project. Getting it from a platform? That's a feature toggle.

The platforms that win won't be the ones with the best models. They'll be the ones that make model selection boring.

## So what should builders actually do?

If you're building AI-powered applications right now, here's what I'd focus on.

**Pick platforms that treat models as interchangeable.** If your stack requires you to commit to a specific provider's API at the architecture level, you're building technical debt that will cost you when pricing shifts. And it will shift. Look at what happened when Anthropic killed flat-rate enterprise pricing and moved everyone to consumption.

**Governance is not a nice-to-have.** The EU AI Act carries fines up to €35 million or 7% of global turnover. Even if you're not in Europe, your enterprise customers probably are. The platforms that give you audit trails, access controls, and model versioning out of the box are saving you from building compliance infrastructure from scratch.

**Budget for the harness, not the model.** The raw inference cost is the smallest line item in any serious AI deployment. The people, the integrations, the monitoring, the compliance. That's where the money goes. Platforms that collapse those costs into a predictable subscription are worth far more than their sticker price suggests.

## The takeaway

The open-source AI spending paradox isn't really a paradox at all. It's the market telling us something we should have already known: the value in enterprise software has never been in the raw compute. It's in the layer that makes that compute safe, predictable, and auditable enough for someone to sign off on.

Open-source models are winning the token-volume war. OpenRouter's top five models by monthly volume in June 2026 were all open-weight. Developers love them. But the budget war, the one that determines which companies capture enterprise revenue, is being fought one layer up. And right now, the governed infrastructure layer ships with the closed models.

The no-code platforms that build that governance layer as their core product, while keeping the model selection flexible, are walking straight into the gap the market is creating. That's the bet I'd make.