Data sovereignty puts a strain on the best AI pipelines

Data sovereignty is becoming an architectural issue as companies focus artificial intelligence inference on their most sensitive information. Pipelines assembled from separate tools struggle to align embeds, permissions, and source data, leaving many valuable data sets beyond the reach of AI.

The shift from model training to inference is transforming the performance of AI infrastructure and has fueled demand for Vast Data Inc., which raised about $1 billion in April 2026 at a $30 billion valuation. Platforms designed for scalability and speed now also need to handle user identification, access control and regulated data, according to Vaughn Stewart (pictured), vice president of systems engineering at Vast.

“There are very different philosophical concepts in the market, aren’t there? One is the tailored model… a best-in-breed of all parts of a drug pipeline,” Stewart said. “And what we’re seeing is just too much of there being data that may be out of sync between the update and the vector embeds, or maybe out of sync between user access controls on the source data and what’s provided by an inference query.”

Stewart spoke with John Furrier of theCUBE Research in the AI ​​Data Pipeline Forum during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed data sovereignty, integrated data platforms for AI factories and how data regulated by confidential computing can be opened up to boundary models. (*Disclosure below.)

Why data sovereignty puts a strain on the best-of-breed pipeline

AI developers like Neocloud operators have little patience for underperforming infrastructure. They will delete and replace software that no longer works, marking a sharp break from the storage industry’s era of lock-in, Stewart noted. Vast’s disaggregated, shared-everything architecture unifies the roles of storage systems and data processing tools, he explained.

“Nobody has ever called a Hadoop data lake or a Kafka event bus a storage platform, but they stored petabytes or tens of petabytes of data. So we’re able to tightly tie the capabilities of these two systems together to deliver performance,” Stewart said.

When asked about Vast’s competition, Stewart referred not to a rival company, but to an “old way of thinking” – building an agent pipeline from interchangeable best-of-breed components. The trade-off is that data must be transferred from system to system and access controls are lost in the process, he noted.

“I think this construct of ‘I can build and move this huge volume of data and all these connections through a pipeline and make it happen in near real time’ is a fallacy when we start talking about the need to serve parallel requests and parallel functions at scale,” Stewart said.

According to Stewart, about half of the world’s data still resides in corporate data centers or colocation facilities because it is sensitive and occurs in regulated industries. Much of this cannot legally be integrated into a public cloud frontier model, making data sovereignty a hard constraint. Vast’s answer is DataEnclave, a confidential AI runtime environment launching in September 2026 with partners including Nvidia Corp. and Cohere Inc. was announced. It runs models in secured virtual machines, requires remote cryptographic confirmation before key sharing, and logs all accesses to a tamper-proof database, he explained.

“If there is no attestation, there is no access to the keys to the model and/or data,” Stewart said. “We believe this will allow us to introduce frontier models into these highly secure data sets today, bridging the last mile of unlocking AI tools, technologies and capabilities.”

Here is the full video interview, part of SiliconANGLE and theCUBE’s coverage of the AI ​​Data Pipeline Forum:

(*Disclosure: ScaleFlux sponsored this segment of theCUBE. Neither ScaleFlux nor other sponsors have editorial control over the content of theCUBE or SiliconANGLE.)

Photo: SiliconANGLE