AWS VP Swami Sivasubramanian on agentic AI going into production — Bedrock request volume in Q1 exceeded all prior years combined

Swami Sivasubramanian has spent 20 years at AWS, starting as an intern whose project became DynamoDB, then building out the database and analytics businesses before pivoting fully to agentic AI. His current read is that the platform inflection is real and accelerating.

Bedrock demand

AWS Bedrock request volume in Q1 2025 exceeded all prior years combined. Sivasubramanian attributes this to enterprises moving from proof-of-concept to production, where actual business outcomes are visible. He says there are no signs of the growth rate slowing.

“In Q1 alone, the amount of workload — request rate — that we are seeing is greater than all of the previous years combined... frontier teams see 10x to 20x productivity improvement, and when you see their bill on tokens, it's something like around $2,000 to $3,000 a month.”

New launches

AWS launched two products at the time of this conversation. Quick (the Quake Autonomous Agent) is designed to eliminate what Sivasubramanian calls "walled gardens" — the context gaps between tools like Outlook and Slack that force workers to spend an hour catching up after any absence. Quick gathers cross-tool context with built-in governance and security. GoDaddy is using it to eliminate roughly 15,000 hours of manual work.

Continuum is a continuously-running security service. The design logic is that security should work like antibodies rather than as a discrete periodic process — always on, always scanning. The motivation is a genuine organizational tension: AI is accelerating code production dramatically, but faster shipping without faster security review creates compounding risk.

AWS also launched a release agent and a pen testing agent. Sivasubramanian's argument is that code velocity gains become worthless if deployment stays at the old pace — you end up with technical debt accumulating faster than it can be cleared.

The no-single-model bet

When AWS built Bedrock, the belief that no single model would dominate was not a popular position. Sivasubramanian says it now looks obvious. Bedrock offers access to models from Anthropic, OpenAI, and others, with model routing and disaggregated inference as core infrastructure capabilities.

ROI discipline

Within Amazon, Sivasubramanian says the organization has moved past "get everyone familiar" into active spend monitoring. Per-user, per-model, and per-organization cost breakdowns are now integrated into AWS Cost Explorer. The data is surfaced to every VP.

The number he cites to justify the spend is striking: the teams he calls "frontier teams" — those seeing 10x to 20x productivity gains rather than 50% — are spending roughly $2,000 to $3,000 per month on tokens to get there. Southwest Airlines and Delta are among the external enterprises following a similar trajectory, with Southwest building crew planning agents on AWS's Agent Core platform.

The frame he leaves with is "write it correctly, ship it fast, keep it modern" — and his position is that all three legs of that now require agents, not humans, to scale.