Oct 8, 2026, 10:30 AM ETInterview, Fireside Chat

Building the Cloud for AI Agents | AWS CEO Matt Garman

AWS views startups as the "lifeblood" of its ecosystem, estimating that 30–40% of current AWS revenue derives from companies that were once startups on the platform.
AWS intentionally reserves GPU capacity for emerging startups rather than allocating 100% to frontier labs (e.g., Anthropic, OpenAI, Meta) to ensure the future enterprise base is nurtured.
Startups today differ significantly from the past; they often launch with billion-dollar valuations and $200M+ in funding, requiring larger capital and compute resources from day one compared to the iteration-heavy model of a decade ago.
AWS is prioritizing the "agentic" workflow over the "human-in-the-loop" model, optimizing infrastructure specifically for agents that require defined API interfaces, low tail latency (P999), and high throughput.
New entrants face a shift where they are often expected to build "greenfield" agentic solutions rather than merely replicating human workflows, requiring a fundamental rethink of problem-solving architecture.

AWS announced a commitment to purchase 2 million NVIDIA GPUs over the next few years to address massive demand.
AWS projected Capital Expenditure (CapEx) of $220 billion for 2026, a figure significantly higher than historical norms, driven by the AI compute build-out.
The primary supply chain constraints are shifting cyclically; as soon as one bottleneck (e.g., power) is resolved, the next (e.g., HBM memory, TSMC capacity, or networking components) becomes the limiting factor.
AWS is actively managing long-term power constraints by procuring its own renewable and nuclear energy projects, sometimes paying for grid infrastructure behind the meter or contributing to the wider grid.
AWS expects to fulfill approximately 60% of GPU requests eventually, though fulfillment timelines vary by region, configuration, and timing.
AWS maintains a diversified customer base with single-digit percentage concentrations per client, mitigating the risk of a potential bubble better than competitors who may rely on 30–60% of capacity for one or two customers.

AWS is building new "agentic" building blocks distinct from traditional user interfaces, including compute sandboxes, time-boxed agent permissions, and granular tool access controls.
AWS introduced a simplified account onboarding process allowing users to launch full AWS accounts in under 30 seconds via Gmail without initial VPC or IAM setup, though these can be hardened later as the organization scales.
AWS developed "AgentCore" and "Bedrock" services, alongside "AWS Context," a beta layer enabling agents to navigate complex data lakes across S3, Aurora, and other storage services.
Security concerns, including the Hugging Face attack and extension risks, are driving demand for "Continuum," a service that uses AI models to identify and prioritize vulnerabilities within customer environments.
Enterprises are currently holding back full autonomy for agents due to trust deficits; AWS is investing in "eval systems" and guardrails to prevent agents from making catastrophic production errors (e.g., deleting databases).
AWS has rolled out "Amazon Quick" to all employees internally, enabling line-of-business teams (HR, Finance) to build agents for tasks like tax compliance and resource management, previously blocked by software engineering backlogs.
Internal software development velocity has seen a "turbo boost" due to "frontier teams" that utilize agent-first workflows where agents write the code, significantly accelerating product deployment.

AWS has achieved a "runaway hit" status with Graviton chips, which offer 20% lower costs and 20% better performance; over 90% of the top 100 customers use Graviton in some capacity.
AWS has deployed three generations of Tranium chips; Tranium 3 is sold out through next year and serves as the primary inference engine for Bedrock and partners like Anthropic and OpenAI.
Despite the name, Tranium is now being used extensively for inference workloads due to its superior memory bandwidth and cost-performance ratio compared to competitors, often surpassing specialized inference chips.
AWS is ramping support for open weights models, with SageMaker emerging as a primary platform for enterprises performing post-training and fine-tuning on proprietary data to build custom models.
Bedrock enforces strict data privacy guarantees, ensuring enterprise data never leaves the customer's VPC or reaches the model provider, a feature cited as a primary reason for migration from other cloud providers.

AWS leadership believes the current AI investment is not a bubble, citing durable enterprise ROI across core compute, storage, and inference, similar to the internet infrastructure boom.
AWS is experimenting with organizational structures that prioritize smaller, agile pods (3–4 people) capable of rapid innovation and rotation, moving away from large, long-standing product teams.
Andy Jassy and leadership anticipate that the AWS market opportunity will remain large over the next decade, justifying continued aggressive CapEx spending despite supply chain challenges.
AWS plans to be more vocal about the community benefits of data centers, such as tax reductions for local communities (citing a specific example of $5,000/year tax reduction per resident in one county) and renewable energy initiatives.
AWS expects the industry to move toward "machine speed" security, leveraging AI models to detect and patch vulnerabilities faster than human teams can manage.