An executive from Amazon.com NASDAQ: AMZN took the stage at Cisco’s AI Summit to discuss what separates early-stage AI experimentation from successful production deployments, how AWS is thinking about “AI-first cloud” infrastructure, and why security, scaling, and sovereignty concerns are shaping enterprise adoption.

From proofs of concept to production: defining success up front

The AWS speaker said one of the biggest gaps between companies running AI proofs of concept and those deploying AI in production is the lack of well-defined success criteria at the outset. Many organizations launched “hundreds of different experiments” as a learning journey but did not set clear goals, making it difficult to determine which initiatives should move into production.

To illustrate, he described two conversations at the JPMorgan Health Conference about “ambient listening” systems that automatically generate physician notes and route information to electronic health records and insurers. In one case, an administrator said the rollout improved doctors’ work-life balance but did not save money. In another, a different administrator viewed the same type of benefits as a major win because it reduced expected attrition among clinicians. The contrast, he said, showed that “having that metric” and understanding the intended business value is critical—cost savings may not be the right yardstick for every use case.

He added that the quality of metrics varies by function. Customer service deployments often come with clearer measurements, and coding-related rollouts also tend to have defined metrics. For broader workforce productivity use cases, he characterized metrics as “fuzzy,” with many companies lacking a strong measurement framework.

Security and scaling challenges rise with agentic workflows

Beyond measurement, he emphasized that security concerns are a major blocker, particularly as organizations move toward agentic workflows. He said companies worry about:

• Agents taking actions they are not supposed to
• “Sprawl” of agents across the enterprise
• Agent identity and permissions
• How to scale from a small proof of concept to global deployment

On scaling, he pointed to a common pattern where a team tests on a single GPU-backed instance and then faces operational questions when attempting to roll out broadly—especially when that initial hardware is underutilized. In his view, organizations are still learning how to “productize a proof of concept” from security, operations, and scaling perspectives.

What AWS means by an “AI-first cloud”

Asked about the idea of an “AI-first cloud,” the AWS executive said the company expects inference to become a “critical part of every single application in the world.” Rather than separate categories of AI and non-AI software, he expects all applications to incorporate inference capabilities.

That shift, he said, requires AI to be integrated into AWS infrastructure components, including how agents interact with storage, virtual private cloud (VPC) networking, security settings, and permissions. He also highlighted the need to manage production trade-offs such as testing and selecting models over time.

He described AWS’s Bedrock offering as a platform intended to provide “a ton of model choice,” broad capabilities, and security consistent with how customers operate inside VPC environments. He also stressed the role of partners, saying AWS does not expect to build every component customers need and that a broad ecosystem operating on top of AWS is important to customer success.

Silicon economics, Trainium, and capacity planning

The conversation also touched on AWS’s custom silicon efforts, including Trainium. When asked whether AWS’s margin profile could improve if inference becomes embedded in all applications, the speaker said he did not know and suggested it was “for the bankers to guess.” However, he did say that using AWS silicon can improve economics versus relying only on NVIDIA GPUs, while still viewing NVIDIA products as important to the market.

He said the company aims to offer better price-performance and customer choice, noting that high gross margins in the ecosystem can create room for alternatives with lower margin and strong performance. He framed AWS’s historical approach as passing savings to customers, suggesting the more likely outcome is maintaining margins while lowering prices, which he said can free customer budget for new workloads and strengthen AWS’s growth flywheel.

On chip development, he said AWS’s cycle time is roughly 18–24 months, sometimes compressing, while also being constrained by semiconductor process generation timelines and the practical challenge of manufacturing at scale. He estimated it can take nine to 12 months from the first chip to reaching scale in large AI systems.

In discussing capacity planning, he said AWS benefits from demand continuing for older chips. As an example, he said AWS is “completely sold out of and have never retired an A100 server,” aside from failures. He also cited technical reasons older generations can remain relevant, including workloads that require higher precision than newer AI-optimized approaches, and noted that smaller models can still run well on older systems. He added that some large labs are willing to make multi-year capacity commitments, which can mitigate risk for AWS.

He also gave a snapshot of data center expansion, stating that AWS added “just south of 4 GW” of new data center capacity over the past year. He described planning across multiple time horizons, with data centers amortized over 20–30 years, servers typically used for five to six years, and networking gear lasting roughly six to 10 years.

Coding gains, sovereign cloud, and advice for CIOs and CISOs

On AI-assisted coding, the speaker said AWS is seeing “massive acceleration” when teams start projects from the beginning with AI-driven coding in mind. He described some teams where the mandate is to write no code directly and instead prompt models or systems, arguing this can build context and documentation that supports testing and creates a flywheel for development. He estimated results of “10x improvement or more,” and in some cases “100x improvement,” in those greenfield-like settings.

He said the biggest gap remains complex, integrated legacy systems, including large distributed codebases. While some improvement is visible, he said AWS has not yet achieved the same magnitude of speedups, though he suggested progress could come within “the next six to nine months” through new approaches to help systems understand legacy code.

On sovereign AI and geopolitics, he said many European conversations begin with variations of “we trust you, I don’t know if I can trust your country,” and that companies ask what would happen if the U.S. government “decides to turn me off.” He said AWS views that as very unlikely but acknowledged it is a concern. He noted AWS recently launched the EU Sovereign Cloud, describing it as a fully separate EU-incorporated subsidiary with an independent governing board, with data—including metadata and account logins—kept within the EU region. He also said AWS tested disconnecting the region from the AWS backbone and co-designed the model with Germany’s BSI and other EU countries.

In closing advice to CIOs and CISOs, he used an analogy of crossing a canyon: without guardrails people move slowly, but with handrails and walls they can run. He said enterprise teams hesitate because they fear agents could cause outages or security issues, citing an internal example where an agent “was just about to go delete some infrastructure” because it believed that was the fastest path. His takeaway was that organizations need “building blocks” and guardrails—safe ways for teams to move quickly in production environments—to accelerate adoption.

About Amazon.com NASDAQ: AMZN

Amazon.com, Inc is a diversified technology and retail company best known for its e-commerce marketplace and broad portfolio of consumer and enterprise services. Founded by Jeff Bezos in 1994 and headquartered in Seattle, Washington, the company launched as an online bookseller and expanded into a global retail platform that sells products directly to consumers and provides a marketplace for third-party sellers. Over time Amazon has grown beyond retail into areas including cloud computing, digital media, devices and logistics.

Key businesses and offerings include Amazon’s online marketplace and fulfillment services, the Amazon Prime membership program (which bundles expedited shipping with streaming and other benefits), Amazon Web Services (AWS) which supplies on-demand cloud computing and storage to businesses and public-sector customers, and a range of content and advertising services such as Prime Video and Amazon Advertising.

Featured Stories

This instant news alert was generated by narrative science technology and financial data from MarketBeat in order to provide readers with the fastest reporting and unbiased coverage. Please send any questions or comments about this story to contact@marketbeat.com.

Before you consider Amazon.com, you’ll want to hear this.

MarketBeat keeps track of Wall Street’s top-rated and best performing research analysts and the stocks they recommend to their clients on a daily basis. MarketBeat has identified the five stocks that top analysts are quietly whispering to their clients to buy now before the broader market catches on… and Amazon.com wasn’t on the list.

While Amazon.com currently has a Moderate Buy rating among analysts, top-rated analysts believe these five stocks are better buys.

View The Five Stocks Here

Reduce the Risk

Market downturns give many investors pause, and for good reason. Wondering how to offset this risk? Click the link to learn more about using beta to protect your portfolio.

Get This Free Report