At COLLIDE 2025, Aerospike founder and CTO Srini Srinivasan argued that infrastructure cost decides whether real-time AI pays off. He explained the company's hybrid memory architecture, how faster lookups improve fraud and recommendation decisions, and why agents will multiply database load.
Key takeaways
- Cost is the differentiator. Srinivasan said many systems scale, but the question is what scale costs.
- Index in memory, data on SSD. Aerospike's hybrid memory design can shrink clusters from about a thousand nodes to about 50, he said.
- Faster lookups mean better decisions. More data and inference steps within a 50 ms window can yield a three to four times better decision, by his estimate.
- Agents multiply load. Users dispatching several agents at once, each needing real-time data, raise scale by another order of magnitude.
- Not for every workload. Srinivasan said relational databases remain the right choice when SQL features matter more than extreme throughput.
Agentic AI will multiply the load on enterprise databases by another order of magnitude, Aerospike founder and CTO Srini Srinivasan told attendees at COLLIDE Data + AI Conference 2025 in Atlanta. In a keynote sponsored by Aerospike, Srinivasan explained how the company's database architecture supports predictive, generative and agentic AI in real time, and why he believes infrastructure cost decides whether AI pays off at scale.
What is the "dragon in the data center"?
Srinivasan opened with the story behind the dragon on Aerospike's t-shirts. More than a decade ago, he said, the CTO of AppLovin, an early customer handling very high internet workloads, described how his previous database behaved as traffic spiked. It used every available resource and then demanded more, growing from tens to hundreds to thousands of machines. The CTO called it the dragon in his data center.
After the switch to Aerospike, Srinivasan said, resource use came down and stabilized as load grew. The broader point, he said, is that online applications now generate enormous loads because nearly everyone carries a phone. "You can't really have a high performance personalized application without having a high performance database," he said.
How does a hybrid memory architecture reduce AI infrastructure costs?
"Virtually everybody can do things at high scale," Srinivasan said. "But what does it cost?" He called cost the key differentiator.
Aerospike's patented hybrid memory architecture keeps the index in DRAM and stores the data itself on SSDs, which he said are about 10 times cheaper per bit. A traditional database reads data into a buffer pool and performs only as well as that pool is large, so high-performance deployments end up holding data and indexes in memory across very large clusters.
With SSDs, Srinivasan said, a single node can hold 50 to 100 terabytes, compared with about a terabyte in memory. That can shrink a cluster from roughly a thousand nodes to about 50. Each node then has to handle many times more transactions, which he said Aerospike achieves through heavy optimization of its in-memory transaction paths in C, including multithreading and processor pinning. He said the result is sub-millisecond reads and writes on the same system and a total cost of ownership four to five times lower.
He also described a shared-nothing cluster design with no single point of failure, data rebalancing while the system stays online, and active-active replication between distant clusters, because mission-critical systems "need to be always on."
How does database speed improve predictive AI decisions?
Srinivasan said models are trained offline, but their decisions happen in real time. At PayPal, he said, a fraud score has to be generated within about 200 milliseconds, with the full transaction completed in under half a second. If no score is ready in time, the transaction goes through and the business takes on risk.
Within a fixed service-level window, such as 50 milliseconds for a recommendation, a typical database might fetch only enough data to run a rules-based step and part of an intermediate risk step. A faster database can fetch hundreds of items in a fraction of the time, leaving room for rules-based, intermediate risk and deep contextual steps, each with fresh feature-store lookups. Srinivasan estimated that the extra data and processing can produce a decision three to four times better.
He said Fidelity uses Aerospike alongside other databases, such as Snowflake, to cache query results and ingest data in real time. He cautioned that he didn't want listeners to "think Aerospike is a solution for everything."
Where does a real-time database fit in generative and agentic AI?
For generative AI, Srinivasan said, Aerospike's role is feeding real-time, relevant data into retrieval-augmented generation pipelines, including conversation context that may stretch over days for hundreds of millions of users. He described Myntra, an Indian fashion e-commerce company, combining a shopper's prompts across sessions with its recommendation engine to suggest items, such as a dress for a special occasion.
Agentic AI raises scale again, he said. A traveler might dispatch two agents with different instructions to find tickets and keep only one result, and agents have to work with real-time data because schedules and availability change. Agents also create their own working data, the way a person jots notes on paper. Srinivasan said these "micro databases" are a good fit for Aerospike. He acknowledged he had no production agentic use case to share yet and said the company is still working with customers to measure the ROI of agents.
When shouldn't you use Aerospike?
Asked in Q&A, Srinivasan said Aerospike does not natively support SQL, though it does support strictly serializable ACID transactions across geographically distributed clusters. When relational systems such as Oracle, MySQL or Postgres meet an application's needs, or when Redis works at the required scale, teams should keep using them, he said. Aerospike fits workloads that need high throughput, low latency, scale and consistency together.
He also outlined three deployment options: self-managed, a managed service that Aerospike runs in a customer's own cloud account on AWS, Azure or Google Cloud, and Aerospike Cloud, a fully hosted service that he said was in preview on AWS.
See sessions like this at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta for keynotes, panels and technical sessions on AI infrastructure. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. More information is on the COLLIDE event page.
Frequently asked questions
What is a hybrid memory architecture?
As Srinivasan described it, a design that keeps the database index in DRAM and stores the data on SSDs, which he said are about 10 times cheaper per bit than memory.
How fast does a fraud score need to be generated?
Srinivasan said PayPal needs a fraud score within about 200 milliseconds and the full transaction in under half a second.
How does agentic AI affect database workloads?
Srinivasan said users dispatch multiple agents at once, agents need real-time data, and they create their own working data, all of which multiplies load.
When should you not use Aerospike?
Srinivasan said Aerospike lacks native SQL, so relational systems such as Oracle, MySQL or Postgres remain better when their features meet an application's needs.
Watch: Driving ROI with High-Performance Data Infrastructure for AI, presented by Aerospike on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
