Inside Capital One’s shift to a ‘serverless-first’ operating model

Capital One Financial Corp.’s five-year mission to transform its information technology operations to a “serverless-first” model is paying off in improved developer productivity, lower operating costs and a shift away from infrastructure focus toward customer-facing outcomes.

The move has been less a technology migration than a redefinition of how applications are built, operated and governed at enterprise scale, executives say. Reducing the need to manage infrastructure has required the financial services giant to shed some control, but the speed at which it can deploy new applications has more than made up for the tradeoffs.

Serverless computing is a cloud execution model in which the cloud provider automatically handles capacity provisioning, scaling and maintenance, charging only for the resources consumed during code execution. Serverless computing’s promise to cut cloud computing costs, improve scalability and reduce the need for developers to manage infrastructure has led to rapid adoption across enterprises. The global serverless computing market is expected  to grow more than 15% annually from $21.9 billion in 2025 to $44.7 billion by 2029, according to MarketsandMarkets Research Private Ltd.

Five-year mission

Capital One has been on a five-year mission to make serverless the default model for new development, with Amazon Web Services’ Lambda as its platform of choice. At the core of the strategy is a philosophical shift, “to see our engineers focus more on delivering value to our customers than managing infrastructure,” said Brian McNamara (pictured below), distinguished engineer in Capital One’s Cloud Operations and Resiliency Engineering group.

While serverless computing’s value is often defined by lower cloud consumption costs, Capital One has found that engineering efficiency is the more significant variable.

“People often index on cloud costs and say Lambda can be expensive to run at scale,” McNamara said. “The compute is much less expensive relative to engineering time.”

That has translated into measurable productivity gains. Teams that adopt serverless services “have significantly lower run-the-engine costs,” McNamara said. “Developers aren’t patching servers. They’re solving problems for customers.”

Capital One estimates engineering teams save roughly 30% of the time tending infrastructure by all but eliminating tasks like rebuilding operating system images. “You can focus on application-specific things,” McNamara said.

Mind shift

The transition required changes not just in tooling but in mindset. “It’s not that all of your operational responsibilities go away,” McNamara said. “It’s just that they shift from infrastructure management to operating your application.”

Though serverless is a goal, Capital One acknowledges that it isn’t suitable for all workloads. Its “serverless-first, but not serverless-only” strategy recognizes that very large workloads can often be run more economically on provisioned servers.

The same goes for applications requiring deep operating system control. “We’re governed by the cloud provider’s OS tuning,” McNamara said. “We don’t have access to a shell,” which can be used to customize or change system behavior at runtime or boot.

Similarly, ultra-low-latency or burst-heavy workloads may be poor candidates for serverless because of delays in activating idle functions. “If you need single-digit millisecond latency, Lambda is not going to be your choice,” he said.

Event-driven catalyst

Where serverless excels is in the event-driven use cases that are increasingly central to modern application design. Events are actions that are trigger state changes such as a click or an order.

The flexibility of serverless in defining events is one of the factors that has made it a dominant target for new development, McNamara said. “It’s remarkable what constitutes an event,” he said. “It could be an HTTP request or uploading an object to an S3 bucket. You don’t need to stand up an admin server to kick off tasks.”

The architecture also simplifies resilience by making it possible to spread workloads across multiple availability zones. “Because Lambda is a stateless service, it’s easy to spin up wherever you need,” McNamara said.

Application economics

Serverless computing limits user flexibility by design. “There’s really only one knob to turn: the memory you have allocated to a function,” McNamara said. Those restrictions are manageable when Lambda is combined with alternatives that offer more control like AWS Fargate, a serverless engine for software containers.

The company has also turned up his focus on instrumentation and observability, since serverless environments don’t allow direct access to infrastructure. Tooling to permit fine-grained visibility has matured significantly, McNamara said.

AWS CloudWatch provides access to logs, metrics and custom metrics and the OpenTelemetry standard supports data integration with third-party platforms.

Early concerns that serverless wouldn’t scale have been largely disproven in practice. “I’ve seen Lambda scale to tens of thousands of concurrent executions,” McNamara said.  “It’s a matter of being aware of what its scaling properties are.”

Nearly five years into the initiative, Capital One’s experience suggests that serverless is not a universal replacement for traditional infrastructure, but a foundational layer for modern, event-driven applications.

Photos: Wikimedia Commons, Capital One