The podcast opens with Matt, Amazon EC2's first General Manager in 2006, discussing AWS's monumental growth to $169-170 billion in revenue, growing at 37%. He emphasizes that despite this scale, the cloud opportunity is still in its early stages, with most workloads remaining on-prem.
Startups have always been the "lifeblood" of AWS, providing innovation and becoming the "enterprises of tomorrow." Matt notes that 30-40% of AWS revenue comes from companies that were once startups on their platform. The nature of startups has evolved; they now start with significantly larger funding (e.g., $1 billion valuations from day one) and often require substantial compute for model training. Despite this, their core needs for scalable architecture, security, and well-defined IAM setups remain constant, which AWS addresses.
A significant shift is the increasing demand for cloud services that cater to agents, not just people. AWS is optimizing existing services like S3 with context layers for agents and building new ones like "agent core" and Bedrock. Matt highlights that agentic workflows perform better on AWS due to its long-standing focus on low latency and high throughput. AWS is also simplifying the onboarding process for new accounts, removing initial complexities like VPC and IAM setup, to make it easier for agents to deploy. Challenges for AWS include adapting to agents' transient resource needs (e.g., databases created and destroyed quickly vs. five-nines durability) and building entirely new components like compute sandboxes and fine-grained agent permissions. Micro-VMs like Firecracker, developed by AWS, are proving ideal for agent sandboxing due to their rapid spin-up and strong security boundaries.
Regarding GPU capacity, Matt acknowledges the "massive" demand. AWS is investing heavily, with a CapEx of $220 billion for 2026, including plans to buy 2 million NVIDIA GPUs over the next couple of years. He explains that resource allocation is a constant challenge, constrained by factors like power, data centers, memory, chips, and even construction labor. While top frontier labs are significant customers, AWS intentionally reserves capacity for startups, saying "yes" to approximately 60% of requests, even if it means slight delays or alternative configurations. This strategy supports a healthy ecosystem and diverse future enterprises.
AWS’s unparalleled view of global demand influences its long-term planning, now extending 20 years for power and multiple years for hardware components. Matt draws an analogy to "The Goal," stating there's never one constraint, but rather a constantly shifting "latest constraint" in the supply chain.
Addressing public debate around data centers, Matt agrees the industry needs to be more vocal about the benefits they bring, such as renewable energy use, minimal water consumption, high-paying jobs, and significant tax contributions that reduce local citizens' tax burdens.
AWS's custom chip strategy, starting with Nitro and leading to Graviton and Trainium, is a key differentiator. Graviton, their Arm-based processor, is a "runaway hit," offering 20% better performance at 20% lower cost, with Graviton ships now outnumbering other types in their fleet. Trainium, initially for training, has proven to be an "excellent inference chip" as well, driving much of the inference on Bedrock due to its cost-performance and absolute performance.
For enterprise adoption of AI, CEOs are asking about safe agent deployment. Enterprises are moving beyond simple agent replication of existing workflows, starting to envision greenfield solutions. The key challenge is building trust in autonomous agents, requiring robust guardrails, permissions, and security. AWS is developing services and using its internal FTE (Full-Time Equivalent) teams to help customers with evals, testing, and data labeling for agent deployment.
Matt emphasizes AWS's commitment to customer data privacy, stating that Bedrock guarantees data never leaves a customer's VPC. AWS supports both proprietary and open-source models, enabling customers to fine-tune open-weight models with their proprietary data using SageMaker. Internally, AWS extensively uses AI for security (e.g., Continuum service), software development (leading to a "turbo boost" in product deployment pace), and across all business units, including HR and finance, through Amazon Q. This internal adoption highlights significant gains in the speed of software and product development, as well as enabling business teams to innovate more rapidly.