A startup says it found hidden memory behavior in NVIDIA GPUs and is building a security layer around it

Chicago startup Newtonian Standard says it has identified a memory behavior in NVIDIA GPUs that could serve as the foundation for a new kind of hardware-level security.

J.P. O’Donnell, the company’s founder and a former Okta engineer, says he spent 18 months characterizing the behavior across NVIDIA Turing, Lovelace and Blackwell GPU architectures. The company is a member of NVIDIA Inception, the chipmaker’s startup program.

The claim centers on Base Address Register 1 (BAR1), the PCI Express window that lets a host computer access GPU memory. O’Donnell says that on the NVIDIA systems he has tested, part of that window returns nothing to the host while the GPU continues computing. CUDA kernels keep running against the same memory the host can no longer inspect.

“There’s more unaddressable space here than is needed for the standard use case,” O’Donnell said. “So there’s this big gap of space that you could do computation in. But nobody’s tried to do it yet.”

On May 1, Newtonian Standard says it ran Qwen2.5-1.5B-Instruct, a small large language model, in the unaddressable region on both Blackwell and Lovelace hardware in bfloat16 precision. The company reports that host-side probes of the BAR1 aperture returned no plaintext from the model or its output. Independent academic replication has not yet been published.

O’Donnell is also developing what he calls a Liveness Encryption Key, which reads involuntary neuromuscular tremor from a mouse or touchscreen to provide continuous biometric authentication rather than one-time checks.

The three-person company plans to license its ZERO runtime and APIs to security companies, cloud providers and hardware vendors. The company says it has filed more than 77 provisional patent applications, with intellectual property managed by Marshall Gerstein & Borun LLP. A live demo, Aura, is available on the company’s website.