AI news right now is like that movie “Groundhog Day” where Bill Murray keeps experiencing the same things over and over again. Every morning we’re served up some version of the same story: A lab just released a new closed model, and it topped some stellar benchmark.

I mean, sure, some of these closed models crack decades-old math problems or pull off some other impressive feat. But frontier labs are mostly selling to a handful of AI-native companies, not to the Fortune 500. Anthropic’s own leaked IPO prospectus notes that roughly 25% of its 2025 revenue came from just two customers, one of them being the AI coding company Cursor. 

So what exactly can closed models do for an enterprise like yours?

Not much, from where I sit. 

That’s why, when leaders like you ask me where enterprise AI adoption stands, I say we haven’t even reached the first quarter of the game yet.

Closed models might monopolize news cycles right now, but they’re a gamble many risk and regulatory teams will never take because they require you to hand over your most sensitive data, accept AI answers you can’t inspect, and build an AI roadmap on infrastructure you don’t own (infrastructure that could suddenly, unexpectedly become obsolete when the API deprecates, as many frustrated leaders have discovered firsthand). Even those who have made the leap won’t keep throwing money at closed AI models nobody can inspect or explain.

Open-source AI models, on the other hand, are quickly picking up steam and giving you what closed APIs never can:

• Weights you can fine-tune on your own data.
• Deployment inside your own environment.
• Freedom from vendors who can arbitrarily change the model, price, or terms of service your business depends on.

Chinese AI labs initially dominated the open-source space, but boards tend to tense up when you tell them you want to deploy a powerful Chinese AI model; right or wrong, some people are still concerned about malware, spying, and data walking out the door. And perception drives budget approvals. So the AI market’s begun shifting more toward Western models.

Nvidia, for example, released its Nemotron 3 models last December and is spending $6 billion to build a powerful U.S. alternative to Chinese AI. Nvidia-backed Reflection just unveiled its first model, which is built to take on Chinese open models directly. And I expect more hyperscalers and chipmakers to quickly follow suit. 

This is great news to me and everyone here at 4Minds because we believe Western open-source models will drive the enterprise AI adoption that benchmark headlines keep promising, yet still haven’t delivered.

But open weights will matter most if you can put them to work on your terms.

So before you sign with any vendor (including us), make sure these four key capabilities are built-in:

1. An air-gapped, fully on-premises deployment. Your most regulated, most secretive datasets will have to stay on-prem, and many will need to be air-gapped. So cross off your shortlist any inference or fine-tuning players that will require you to use an API or SaaS.
2. Use case-based pricing. Use case-based pricing is predictable and manageable. Plus, vendors that charge per use case work harder to keep you happy because they want you to have enough budget for a second, third, or 100th use case. Token-based billing, on the other hand, turns every month into a guessing game about spend, which is why Uber burned through its entire 2026 AI budget in just four months at the beginning of the year.
3. A fast path to a custom, private model. You want AI vendors that can give you speed, custom AI and a private deployment. All three. In contrast, LoRA takes some time and lacks power; API deployments move faster but aren’t custom or private; and most vendors simply can’t take you on prem.
4. An always-open exit path. You should be able to move the model, weights and work to different hardware anytime you want, without a rewrite. No exceptions. No excuses.

You’re not done yet, though.

Before you begin implementing that new model, be sure to take care of two important housekeeping issues I’ve seen stall or kill too many AI projects:

1. Fix the basics. No AI model can succeed if the organization still hasn’t developed strategies for data cleanup, data governance and other critical data initiatives. As I like to say, no plan is still a plan. It’s just a really bad one.
2. Align executive incentives to your AI plan. For 20 years I’ve watched a pattern play out time and again that will absolutely derail any AI plan, even those that might be well thought-out in every other way: an executive signs a comp plan built only around short-term KPIs, does everything he or she can to max it out, and then departs after 2-4 years. Organizations are essentially paying execs a lot of money to have no vested interest in the company’s long-term success because they’re only incentivized to focus on the short term. And the layers upon layers of technical debt that accumulate during their tenures from that narrow focus become a huge, expensive obstacle for your AI plan to overcome before it ever has a chance of succeeding.

The Bottom Line

Western open-source models are about to start driving the fast enterprise AI adoption that benchmark headlines keep promising and that we’ve all been waiting for.

But not all open-source models are the same. So evaluate them against real workloads. Then build a shortlist of the ones that guarantee you’ll always own the weights, your data and the outcomes. 

That’s the fastest, safest path to widespread enterprise AI.

Find out how 4Minds deploys contextual AI you control: fine-tuned on your institutional knowledge, priced per use case and running wherever you need it, including fully air-gapped on your own infrastructure. Visit us at www.4minds.ai.