A year ago, mentioning DeepSeek or Qwen in an enterprise planning meeting got you blank stares. Today those two names alone represent roughly 15% of the global AI market, up from barely 1% at the start of 2025. That's not a rounding error — it's the fastest share shift the AI infrastructure market has seen. The question enterprises are asking has changed. It's no longer "are these models good enough?" It's "how do we deploy them without creating a liability?"

The number that gets everyone's attention

The pitch is usually cost. Teams that route routine work to cheaper open models and escalate only the hardest tasks to premium ones report cutting inference bills by more than 80% compared to a closed-model-only stack. That's the kind of number that gets a CFO's attention in a budget meeting. But it comes with a catch that rarely makes it into the pitch deck: those savings assume the deployment is architected correctly. Bolt an open model onto the wrong infrastructure, and the "50% cheaper" promise evaporates into egress fees, redundant compute, and compliance rework. 

Why the hesitation is real, not just optics

The skepticism isn't manufactured. In early 2025, Florida's state CFO banned DeepSeek across the Department of Financial Services, calling it a national security risk, and several federal agencies followed with device-level restrictions of their own. Security researchers gave the concern some teeth: independent audits found weak encryption, potential SQL injection flaws, and undisclosed data transmissions in DeepSeek's R1 model.

But it's worth separating what was actually banned from what wasn't. A banking attorney put the underlying tension well, noting that institutions face both a substantive concern about data reaching the Chinese government and a separate political climate that's simply hostile to anything China-linked right now. Those are two different problems with two different solutions — and conflating them is exactly how enterprises end up avoiding a genuinely capable model for the wrong reasons. Nearly every ban to date has targeted the consumer-facing DeepSeek app running on Chinese-controlled servers, not the open-weight model file deployed inside a company's own environment. That distinction is the whole ballgame.

What "deployed safely" actually means

The regulated industries that have figured this out aren't avoiding open models, they're just not touching the hosted version. The pattern finance, healthcare, and government teams have converged on is self-hosting MIT- or Apache-licensed open models so that no prompt data ever leaves their own network, something that's structurally impossible with a closed API. Once the model runs inside infrastructure the enterprise actually controls, most of the headline risk (data transmission to a foreign server, opaque logging, unclear retention) simply doesn't apply anymore. What's left is a more ordinary vendor-risk conversation: model provenance, patch cadence, access controls, the same due diligence any serious CTO already runs on other software.

This is also where the "self-hosted" conversation tends to get too binary. Standing up your own servers solves the "whose cloud is this" problem, but it doesn't solve the "what happens if this one environment is compromised" problem, a single point of failure is still a single point of failure, whether it's someone else's server or your own. Aphanarc's approach starts from that gap: rather than processing and storing sensitive data in one centralized place, it atomizes compute, model execution, and data handling so no single point ever holds the full picture. What's left after that is a more ordinary vendor-risk conversation: model provenance, patch cadence, access controls. This is the same due diligence any serious CTO already runs on other software.

The calculus has changed for CTOs

This isn't a fringe move anymore. Coinbase's Brian Armstrong and Lindy's Flo Crivello have both spoken publicly about using Chinese-origin models to cut costs, and Cursor built its Composer 2 model on top of Moonshot AI's Kimi. At the same time, U.S. lawmakers have opened an inquiry into the growing use of Chinese AI models inside American companies. This is a reminder that the scrutiny is real and the deployment decisions enterprises make now will be the ones they have to defend later.

The AI spend conversation used to start and end with "which model." It now has to include "hosted where, controlled by whom, and governed under what policy.” That's what actually determines whether the cost savings are real or just theoretical.

**About Aphanarc \ Aphanarc is AI infrastructure built for sovereign AI. Enterprises and privacy-focused users run powerful models and agentic workflows without exposing proprietary data to centralized systems. Aphanarc atomizes compute, model execution and data handling, so sensitive information is never fully processed, stored or reconstructed in one place. In development since 2024, it delivers private, high-performance AI without centralized-cloud premiums.