Introduction
The debate around enterprise AI has long centered on geography—whether a model was built in the United States, China, or elsewhere. But after a year of conversations with business leaders who run open models locally, a different pattern has emerged. Executives are not losing sleep over a model’s country of origin. Instead, they are asking practical, urgent questions: Where does our data go? Who can access it? And how long is it stored?
What Happened
Over the past year, coding agents, internal assistants, and security tools have pushed enormous volumes of proprietary information into AI systems. Source code, customer records, and internal documentation now move into models as part of everyday workflows. At the same time, security incidents such as the PHP2Shell campaign exposed millions of websites to risk, sharpening concerns about data exposure.
Industry leaders are taking notice. NVIDIA CEO Jensen Huang highlighted the issue during his GTC Berlin keynote, and Anthropic’s deputy chief information security officer, Jason Clinton, has warned that the threat environment will likely worsen before it improves. Their message is consistent: organizations need to know exactly what happens to their information when AI is used.
The provider landscape is not binary. Some low-cost services explicitly trade lower prices for the right to train on customer data. Others advertise zero retention, meaning prompts and responses are processed but not stored or reused. Between those extremes sits a broad middle ground of vendors that promise not to train on your data yet may keep it for days. None of these approaches are automatically wrong, but leaders must understand where their providers fall on that spectrum—especially as coding agents become near-universal among developers and can act without a human explicitly choosing a vendor.
Why This Matters
When engineers send sensitive code or customer details to a model, that information is valuable not just to the company but to anyone seeking to train a competing model. As token demand surges, more providers are handling heavier traffic, and not all have built adequate safeguards. An agent operating on a developer’s behalf could end up with a provider on the less protective end of the spectrum without anyone making a deliberate choice.
Open models add another dimension. While they have sometimes been portrayed as less secure, many enterprises are discovering the opposite. Running open models locally means data never leaves the organization’s environment. Alternatively, companies can use a trusted provider in a region approved by compliance teams. In either case, the business retains control over where data resides.
Usage data underscores the shift. Roughly 10 million developers worldwide use Ollama each month to run open models, processing trillions of tokens daily. Menlo Ventures research shows open-source models are claiming a growing share of enterprise usage, and about 85 percent of Fortune 500 companies have Ollama installed on desktops. Cost is a major driver: one fintech customer cut token spending by 30 percent within two weeks after routing most engineering work to open models while reserving frontier models for complex tasks. Another consumer AI firm reported that an open model matched a frontier system in human evaluation at roughly one-tenth the cost.
Security teams also benefit. Open models allow defenders to analyze malicious code, probe systems, and respond to incidents without waiting for access to gated programs controlled by model labs. Anthropic’s S-1 filing notes that cyber and safety risks could slow the release of its most capable models, which means organizations relying on a single closed provider could find themselves without critical tools when they need them most.
Key Takeaways
- Treat data retention as a standard vendor risk: ask the question directly, get the answer in writing, and make zero retention the default.
- Know the full chain: where the model runs and every party that touches the data.
- Open models can be run locally, keeping sensitive information inside your own infrastructure.
- Cost savings are significant; some enterprises have reduced token expenses by 30 percent or more by shifting routine work to open models.
- Security teams should maintain a second option if frontier tools are gated or restricted by the labs that build them.
- Corporate AI policies should catch up to current practice, including rules for non-U.S. models running on approved domestic infrastructure.
Conclusion
AI is becoming embedded in nearly every enterprise workflow. The organizations that succeed will not necessarily be those that selected the “right” model. They will be the ones that stayed in control of their data—by asking hard questions, enforcing clear retention policies, and running models on infrastructure they trust.

Discussion
Join the conversation
Thoughtful reactions, questions, and follow-up ideas help shape the next story.