AWS CEO Matt Garman on Cloud Infrastructure for AI Agents

AWS CEO Matt Garman joins a16z’s Raghu Raghuram to discuss cloud infrastructure for AI agents, startup demand, chips and enterprise adoption. Their conversation reflects AWS’s perspective as a provider with a commercial interest in these markets.

They cite an approximately $169 billion AWS annualized revenue run rate and 37% growth. Amazon’s [second-quarter 2026 filing](https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm) identifies the first figure as an annualized run rate, not actual revenue over the preceding year.

Garman says startups have long mattered to AWS because small customers can become large enterprises and often push providers to build new capabilities. His estimate of the share of revenue from former startups is an internal approximation.

He sees AI-native startups starting with more ambitious compute needs than many earlier software companies. At the same time, they still need familiar capabilities such as databases, scaling, security and help managing costs.

As agents write and operate software, Garman says AWS is adapting existing services as well as building agent-specific capabilities. He gives rapid creation of infrastructure as an example of a workflow that may need different defaults.

He argues that latency, throughput and tail behavior remain important for both people and agents. Some agent workflows create short-lived resources at a pace that can expose friction in service setup.

The host observes that startups increasingly use coding agents. Garman says many customers still choose a cloud provider explicitly, while others begin experimenting before they have established an account or infrastructure.

He describes a gradual rollout of simpler AWS account creation. Product availability and requirements can change, so readers should verify the current onboarding process directly rather than assuming the interview describes every account.

Agent-created infrastructure raises questions about durability and speed. Garman discusses balancing quick creation of a database for an experiment with the possibility that the same database later becomes a long-lived production system.

Permissions are another concern. An agent should receive access appropriate to its task and environment rather than automatically inheriting every permission held by its user. Garman mentions microVMs as one possible isolation building block.

Raghuram asks how startups obtain accelerators when the largest AI labs need enormous capacity. Garman says AWS deliberately reserves some supply for smaller customers while also serving large labs and enterprises.

He says demand and the capital required to add compute have both risen sharply. Customer access can still depend on location, timing and configuration, and the allocation he describes is a company policy rather than a guarantee.

AWS and NVIDIA announced plans for two million additional GPUs over multiple years. Their [joint announcement](https://press.aboutamazon.com/aws/2026/8/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai) describes a planned deployment, not capacity already delivered at the time of the interview.

Asked about an AI infrastructure bubble, Garman says AWS feels confident in its spending because it serves a broad customer base and many types of workloads. That is the company’s assessment; demand and returns remain uncertain.

He notes that many startups will fail even if a few become large successes. AWS’s diversification argument depends on how capacity is used across customers and on future demand, not on every individual customer succeeding.

Amazon’s roughly $220 billion 2026 capital-spending guidance discussed in the interview covers the company as a whole, not an AWS-only investment figure. Actual spending and its allocation are reported separately in financial filings.

Power planning now extends years ahead, Garman says, including generation, grid connections and transmission. Supply constraints can shift among power, memory, chip manufacturing and networking components, and they can differ by region.

On data centers, he describes AWS’s approach to cooling, renewable energy and local engagement. These are company claims; community costs and benefits depend on a specific project and require independent local evidence.

Garman traces AWS’s chip work to Nitro, which moved virtualization functions away from the main server CPU. He says that experience helped build the team and capabilities that later led to custom processors.

Graviton grew from that work into an Arm-based server processor. Garman says customers use it for price and performance benefits, but the result depends on the application and migration effort.

AWS also developed Trainium for AI workloads. Garman says it now serves both training and inference and carries substantial Bedrock traffic; these are AWS’s descriptions of its own infrastructure.

Enterprise customers are moving from assistants that suggest actions toward more autonomous workflows, according to Garman. He says many are still working out where automation is valuable and what must be checked by people.

He encourages teams to redesign a workflow instead of mechanically reproducing every human step. Parallel work can be useful, but the outcome still needs to satisfy the business’s actual requirements.

Trust requires controls over what an agent can access and change, along with observability and evaluation. Garman points to risks such as unintended changes to production systems and says customers want stronger guardrails.

Evaluations must reflect real tasks and data. He notes that preparing representative labeled examples is difficult and that the industry has not solved every aspect of reliable agent evaluation.

On model privacy, Garman describes the protections offered by Bedrock and AWS’s separation of customer data from model providers. Specific guarantees depend on the service, model and agreement, so users should review the applicable terms.

He argues that some organizations may adapt open-weight models using their own data to improve performance or cost on specialized tasks. Fine-tuning and deployment are possibilities, not guarantees that a smaller model will always be better.

The conversation turns to agent security and vulnerabilities. Garman says AWS is investing in controls and cites its Continuum service as a way to identify and prioritize issues using context about a customer’s environment.

He says Amazon also uses AI internally for security and has rolled out tools to employees outside software engineering. The reported gains are company observations; the discussion does not provide a controlled productivity study.

Garman expects organizations and team structures to change as people experiment with agents. He describes small teams trying new ways of working, while acknowledging that the final organizational model is still being explored.