The burgeoning world of AI agents has put the value of enterprise data even more into focus. Agents are only as good as their last meal. Feed them stale data, and their given task becomes, as Aerospike founder and CTO Srini Srinivasan puts it, "nonsense, basically." After years of the GPU getting all the attention, now comes the hard part.

In Srinivasan’s view, the tricky element is in fact the unglamorous layer beneath the inference stack, the databases that have to capture state, serve it back, and do it fast enough that an expensive GPU never has to wait.

“You can't slow down the GPU much, it's expensive,” he said, “You need the CPUs to capture the state and run into the database just fast and then restore it so there is minimal disruption to the CPU-based processing.”

The Mountain View-based real-time NoSQL database company, which counts the likes of PayPal, Airtel, and Wayfair as clients, has spent years betting that the answer to slowing down hardware and performing fast database runs lies in a hybrid memory architecture.

Aerospike pairs solid-state drives (SSDs) with dynamic random access memory (DRAM) to deliver in-memory performance at a fraction of the cost. As agentic workloads threaten to multiply data demands by an order of magnitude, that bet is looking increasingly well-placed.

“You’re having your cake, and your cost is low,” Srinivasan said. “That is actually the main reason [for] being successful. And we continue to invest in all kinds of new storage structures. And no storage technology that is not widely adopted in the cloud, I don't think, is going to be super successful.”

The proof of that cloud-first discipline arrived, perhaps uncomfortably, with the collapse of Intel's Optane persistent memory or Crow Pass, a storage technology that never quite won the hearts of any major cloud vendor, before being unceremoniously shelved.

A cautionary tale, then, invoked by Srinivasan with the lesson learned at Aerospike being a simple one: build for where the workloads actually run. But that principle is being tested at a pace the Aerospike CTO compares to nothing less than the dot-com era.

“It's like dot-com all over again in many ways,” he said. “It's even bigger in some ways, because it is increasing productivity ... and I think the AI technology that we have right now can do better in many of those situations than what [human] programmers can do.”

To help build to meet the massing workloads while also ensuring enterprise compliance, the firm recently unveiled native Dynamic Data Masking (DDM) for its flagship NoSQL Database, a move the firm believes helps to simplify the protection of personally identifiable information (PII) to meet regulatory requirements.

The need for such an offering, Srinivasan argued, is urgent. For example, rising pressure from regulatory bodies, including the Reserve Bank of India, the European Central Bank, and the Federal Reserve, means that for Aerospike's largest customers, a compliance slip caused by an AI agent could go from an internal headache to front-page news.

“These are materially affecting hundreds of millions of users,” he said. “Their regulators are very active in making sure that there are no violations, and any violations will be the front-page news. And this is not something anybody wants.”

The compliance push around agents is only one part of Aerospike's recent product offensive.

Last month, the company also unveiled a Developer SDK – or as it called it, an “AI-native developer experience,” that combines a visual workspace called Voyager, an embedded model context protocol (MCP) server compatible with Claude Code, Cursor, and Codex. It’s designed to get developers from the first cluster to the first query in as little as five minutes.

With agents taking on even more of the coding work, and companies spending tens of thousands of tokens, the Aerospike team contends the infrastructure underneath needs to be something both humans and AI can reason about quickly and ship confidently.

“Essentially, we just have to be very cognizant of all that,” Srinivasan said. “Our goal as an infrastructure provider is to provide all of these features that can help our customers implement these things seamlessly, but also with minimal loss and performance.”