A third option is emerging in the fight over AI and your data
The New Stack Alex Wilhelm ● Covered by 3 sources
A new tool from Vast Data tries to let companies use proprietary AI without handing over their data. It’s meant to cut the trust problem in both directions.
Based on reporting by The New Stack, Alex Wilhelm — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
For months, the enterprise AI fight looked like a bad two-choice quiz. Use the most capable proprietary model and risk leaking company data, or stick with open-weight models and accept that you may not be getting the frontier. Vast Data thinks there’s a third path, and it is betting on secure compute to make it real.
The company’s new product, DataEnclave, is built to let AI labs and large enterprises run proprietary models in secure environments without moving sensitive information back and forth. That matters because the trust problem cuts both ways. Companies want to use an outside model without teaching the model maker how to eat their lunch. Model makers, meanwhile, do not want to hand over their weights so anyone can run the model on their own GPUs.
DataEnclave uses Nvidia’s Confidential Computing technology. Vast Data’s core business, AI OS, sits underneath a company’s AI applications, so this is very much an infrastructure play rather than a new model release. Jeff Denworth, Vast Data’s co-founder, told The New Stack that AI agents are creating a different set of requirements at the data layer, which is exactly the kind of sentence that tends to show up when a real market pain point is finally getting productized.
The timing matters. Nvidia started rolling out Confidential Computing in a serious way in 2024, but Denworth argues the market also needed demand, not just hardware support. Until late 2025, AI usage was still modest compared with today’s token totals. Then agentic coding tools took off, corporate demand surged, and the early-2026 pricing crisis followed. After that came the summer’s secure-AI debate. Performance drove demand, demand drove usage, and usage exposed the next layer of problems.
Whether DataEnclave actually solves enough of those problems is still the open question. It is in early access now, with general availability still ahead. But the bigger story is that enterprise AI is moving past the cartoon version of the debate. This is no longer just open versus closed, or fast versus safe. It’s about whether the industry can stop forcing companies to choose which piece of their IP they want to surrender first.
My take — AI-written commentary, not fact-checked reporting
The cleanest part of this story is also the least glamorous: everyone wants the model, nobody wants the leakage. That’s why confidential computing keeps showing up in these debates; it’s the corporate version of asking for a receipt. The awkward truth is that the AI industry spent a year pretending trust was a feature it could add later, and now it’s being billed like rent.
Read more about this at: The New Stack