At The AI Enterprise Conference 2026, Everpure VP of AI Kaycee Lai argued that the fastest path to production AI is running it where sensitive data already lives. He explained what AI-ready data requires, why unstructured data is the biggest blocker, and who he thinks will win the AI race.
Key takeaways
- Bring AI to the data. Running AI where sensitive data already lives avoids new governance work and data transfer costs, Lai said.
- AI-ready data has five traits: trusted, governed, traceable, properly vectorized and searchable across the data estate.
- Unstructured data is the main blocker. Lai said 80% of most enterprises' data is unstructured, and each file type needs different handling.
- Fragmented pipelines stall production. When every application builds its own data access and search, IT and infosec often block it.
- Start with one measurable outcome. Lai advised a use case with an isolated data source, governance applied early, and a plan for inference at scale.
Many enterprises have tried AI in the cloud, but their most sensitive data, including PII, PHI, production data and large volumes of unstructured files, still sits on premises. At The AI Enterprise Conference 2026 in New York City, Kaycee Lai, VP of AI at Everpure, formerly known as Pure Storage, argued that the faster route to production is to stop moving that data and "bring AI to the data."
Lai, previously founder and CEO of data fabric company Promethium, said that over the past seven years he has seen many customers stuck experimenting with AI because governance, cost and security made moving data to the cloud too hard.
Why bring AI to the data instead of moving data to the cloud?
If AI runs where the data already lives, Lai said, there are no new governance complications and no ingress and egress costs to manage. Data stored on an enterprise storage system has typically already passed IT governance and security review.
He also described a broader shift in how companies share data with software. For years, every application, from SAP to Salesforce, kept its own copy of a company's data. Lai said customers increasingly want the data to stay in place, with applications given access but never owning or storing it. For chief data officers, he said, applying GDPR discovery and governance policies centrally is "a heck of a lot easier" than doing it in every application.
What does AI-ready data actually mean?
Faster GPUs, networks and storage don't fix the core problem, Lai said: the data may not be ready for AI. He listed what AI-ready requires:
- Trust: removing duplicates and corrupted files
- Governance: knowing who can access what, and who shouldn't
- Traceability: being able to find and track data
- Vectorization: embedding data properly for both accurate search results and performance
- Searchability: making data findable across the whole data estate
Unstructured data makes this harder, he said, because what works for PDFs doesn't work for PowerPoint files or videos. Lai said 80% of most enterprises' data is unstructured, and customers tell Everpure that getting it AI-ready is their number one challenge. He pointed attendees to a recent IDC survey that he said confirmed it.
Most organizations take a do-it-yourself approach. "If you talk to five people how you're making data AI-ready, you might get 25 different responses," Lai said. He put the typical setup at two to four months, before enterprise data access rules and firewalls are added.
What slows AI applications on the way to production?
Teams building AI applications focus on capabilities and user experience and take the infrastructure work for granted, Lai said. Every application still has to work out which data it can access and how to do so securely, how to schedule GPUs, which models to apply to which data types, how to run a RAG pipeline and how to expose semantic search.
Idle GPUs are a particular cost problem, he said, because GPUs are the most expensive part of the infrastructure stack in acquisition, power, space and cooling. When storage can't keep up and GPUs sit idle, "that's not a good thing to tell your CFO."
And when each application solves these problems its own way, IT and infosec teams often block them from production, he said.
How does Everpure approach multimodal AI on premises?
Lai described multimodal AI as applying different models to different data types, and even to different stages of the data pipeline, rather than relying on one frontier model. Using one model for everything, he said, is inaccurate at worst and needlessly expensive at best.
His pitch was a single automated layer that every application can use. Everpure's product, which Lai called Data Stream, is packaged as a containerized application co-developed with NVIDIA and integrated with the NVIDIA AI Enterprise software suite and its NIM models, he said. It runs on Kubernetes and, according to Lai, can connect to storage from other vendors, so data doesn't need to be migrated first. One early customer described it to him as "kind of like AI in a box."
Lai also described a key-value cache feature that moves data out of GPU memory into a persistent cache layer to extend the context window. He said it delivers up to 20 times the performance and reduces repeated token costs when inference scales to many users.
Who wins the enterprise AI race?
"They're not the ones who buy the most GPUs," Lai said, nor the ones who consume the most tokens. The winners, in his view, are the organizations that operationalize their data, activate it with AI and do so without giving up control of it.
His advice was to avoid boiling the ocean. Pick a business outcome, choose a measurable use case with an isolated data source, apply governance rules, and plan for scale in inference as well as storage. Many organizations, he said, hit a wall the moment they try to go live with tens of thousands of users.
Learn from enterprise AI leaders in 2027
Sessions on getting AI into production with sensitive enterprise data are a regular part of The AI Enterprise Conference, which returns to Pier Sixty in New York City on Tuesday, August 31, 2027. Registration is coming soon at the conference page. Data Science Connect also runs 24 free live webinars on its 2027 schedule.
Frequently asked questions
What does AI-ready data mean?
As Lai described it, data that is deduplicated and trusted, governed, traceable, properly vectorized for search and performance, and findable across the data estate.
Why is unstructured data hard to prepare for AI?
What works for one file type, such as PDFs, doesn't work for others like presentations or videos, so preparation has to be repeated many different ways, Lai said.
What is multimodal AI in a data pipeline?
Lai described it as applying different AI models to different data types and pipeline stages instead of relying on a single frontier model.
Why do enterprises keep sensitive AI data on premises?
Lai said moving sensitive data to the cloud raises governance, cost and security problems, while data on enterprise storage has usually already passed IT and security review.
Watch: Turning AI Ambition into Business Impact on YouTube
Event: The AI Enterprise Conference · More coverage: The AI Enterprise Conference insights
