Protected Is Not the Same Word as Superior

I was in Paris. Eighth arrondissement.

Someone at the table said the model is Chinese. We cannot trust it.

What they were using instead cost four times more per inference call and was losing on their own internal benchmarks. But it had the right name. American. Familiar. Nobody got fired for choosing it.

That is not a technology decision. That is fear wearing a procurement policy.

The Numbers Nobody Wants to Read Out Loud

89% of enterprises now run at least one open source model in production. Those organizations report 25% higher ROI than the ones running exclusively on closed APIs.

That is not a small number. That is a structural advantage showing up in quarterly results while procurement teams are still having the same conversation that happened in that Paris conference room.

DeepSeek V4 matches or exceeds GPT-4o on 7 of 12 standard benchmarks. GLM-5.1 from Tsinghua matches Claude Opus on coding. Qwen 3.7 from Alibaba is competitive with GPT-5.5 on reasoning. Five independent model families from five different organizations across multiple countries hit frontier quality in the last eighteen months simultaneously.

That is not an anomaly. That is where intelligence is actually being built right now.

The performance gap that justified the closed model premium two years ago is gone for most enterprise use cases. What remains is not a capability argument. It is a habit argument dressed as one.

What Closed Actually Buys You

Convenience. The model is hosted, maintained, updated. No infrastructure burden. That is real value.

Brand certainty. Your board recognizes the name. Your legal team has reviewed the terms before. That social capital matters inside organizations where AI decisions need to survive internal review.

Frontier capability on the narrow tasks where the gap still exists. For the most demanding reasoning and multimodal work, the best closed models still lead. That lead is narrowing. On many tasks it has closed entirely.

What you are not buying is ownership. You do not control the weights. You cannot audit the training. You cannot fine-tune on your own data without going through their pipeline. You cannot deploy in a jurisdiction that requires data sovereignty. And you cannot negotiate the price when your inference volume scales. You are on their meter.

What Open Actually Gives You

You own the inference stack. Closed API costs scale linearly with usage. Open source inference on your own infrastructure scales at marginal cost. At millions of inference calls per month that difference is not theoretical.

You control the model version. No closed provider has ever told you before updating the model your production system depends on. Open weights do not change until you decide they do.

Your proprietary data stays inside your infrastructure when you fine-tune. Your competitive differentiation does not travel through someone else's pipeline.

And you solve the passport problem. An open model on your own infrastructure in your own jurisdiction answers to your laws. Not the CLOUD Act. Not a vendor's terms of service update at 2am on a Tuesday.

The Honest Part

Open is not free at the operation. Running your own inference stack requires engineering capacity, GPU resources, governance, and monitoring that a closed API handles invisibly. For organizations without that capacity the convenience premium is a legitimate call.

The frontier gap still exists at the edges. Intellectual honesty requires saying that.

The right architecture is not closed or open. It is both, deliberately, with a routing layer that sends each task to the model that serves it best regardless of which flag that model flies under.

What the Paris Room Got Wrong

The hierarchy the enterprise AI industry built was never based on evidence.

Closed models from American companies at the top. Everything else somewhere below. That hierarchy reflected marketing budgets and sales relationships. It did not reflect capability.

Being first is not the same as being best. Being protected is not the same as being superior.

The organizations winning on AI right now evaluate every model on merit. They route to the best available intelligence regardless of where it was built. They do not pay a premium for a name when what that name once monopolized is now running under an MIT license on their own hardware.

Ils font des affaires là où les preuves pointent.

They do business where the evidence points. Not where they are told.

Andrew Quillen is the founder of AndMaverick, a global Enterprise AI Orchestration consultancy. The MAO Framework evaluates, routes, and orchestrates across the full landscape of available intelligence regardless of origin. To continue the conversation, visit andmaverick.com.