Has the return on Amazon's massive AI capital expenditure been underestimated by the market?

Amazon's AI capital expenditures are becoming the core of valuation divergence: the market is concerned about the pace of investment being too fast and returns too slow, but signals released by AWS suggest that cloud migration, AI inference, and agent-driven workflows may be reshaping its growth curve.

According to The Information's Sunday report, Amazon is currently trading at a lower forward P/E than Walmart, despite its revenue growth being projected at about three times that of Walmart. Elevated capital expenditures and uncertainty around AI investment returns are cited as the main reasons for the discount from investors.

The focal point of this debate is whether Amazon’s approximately $220 billion capital investment can be absorbed by future demand. In the latest interview with a16z, AWS CEO Matt Garman stated, AI-related demand is enormous, capital expenditure will not slow down in the short term, and AWS continues to reserve scarce GPU capacity for startups.

For the market, the problem is not whether Amazon has AI opportunities, but whether these opportunities can be converted into sufficient revenue, profits, and cash returns. The AWS interview explained where the demand originates, but cannot yet prove that this round of investments has paid off.

According to Koyfin data cited by The Information, Amazon’s current forward P/E ratio is about 29, compared to Walmart’s 37. Walmart’s revenue growth this year is expected to be about 5.6%, only about one-third of Amazon’s estimated rate, but Walmart’s premium valuation has persisted for about 18 months.

This inversion reflects the market’s main concern: Amazon is significantly ramping up capital spending, primarily to expand AI data centers, betting on AI infrastructure demand alongside Alphabet and Meta. However, the return cycle for such investments remains unclear. If AI demand is released more slowly than anticipated, high capex could continue to suppress profitability.

However, segmental valuation analysis also points out that Amazon's business structure may not be fully reflected in its stock price. Amazon’s market value is about $2.8 trillion, but its asset mix includes AWS, online retail, advertising, and media businesses. Data shows Amazon’s online retail sales grew 13% year-on-year in the first half of this year, advertising generates about $80 billion annually with a growth rate of 20% to 25%, and AWS’s stand-alone value is conservatively estimated at not less than $1 trillion.

In other words, the market's discount on Amazon is not only pricing in the retail business, but also factoring in the uncertainty of AI-related capital expenditures, the complex conglomerate structure, and lack of visible returns. A lower valuation does not automatically imply undervaluation, but it provides a perspective to reassess AWS's growth logic.

In the a16z interview, Matt Garman stated that AWS revenue is between $160 billion and $170 billion, a 37% year-on-year increase. He believes AWS is still in the early stages of a long-term opportunity due to a large amount of workload still residing on-premise, continuously rising enterprise computing demand, and new momentum from AI.

The first layer of growth comes from enterprise cloud migration. Matt Garman noted that both cloud migration and AI demand are simultaneously driving AWS's growth. For large enterprises, AI is not just a single model call, but is tied to data, storage, security, permissions, and production systems. This means AWS is not just providing GPU capacity, but also the complete infrastructure necessary for enterprises to put AI into production.

The second layer of growth comes from startups. Matt Garman said that AWS estimates about 30% to 40% of its revenue now comes from companies that were once startups. AI-native startups differ from early internet companies, as they may need large-scale computing from day one, have higher model training costs, and larger fundraising and valuations. Thus, AWS chooses to reserve part of its scarce GPU and AI accelerator capacity for startups, instead of selling all resources to large frontier labs.

He also stated that AWS meets about 60% of those related requests in some way, sometimes delivering in different regions, at different times, or under different configurations. In coming years, AWS will purchase another 2 million Nvidia GPUs. This demonstrates that AWS is trying to maintain a broad customer ecosystem through capacity allocation, rather than only betting on a few large customers.

The third layer of growth comes from agent-driven workflows. Matt Garman explained that more and more teams want cloud environments not only suitable for people, but also for agents. Centered around Agent Core, Bedrock, and AWS Context, AWS is optimizing the ability for agents to make calls across data lakes, databases, and services. Agents are sensitive to tail latency and generate short-cycle, rapidly created and released computing needs.

This could change how cloud infrastructure is used. In the past, cloud resources mainly served human developers and enterprise IT teams. In the future, agents might automatically write code, call databases, manage infrastructure, and execute tasks with sandboxed and temporary permissions. For AWS, this implies new needs for inference, storage, security, permission management, and compute.

In addition, custom chips are also seen by AWS as key to underpinning this logic. Matt Garman said Graviton offers approximately 20% improvement in both cost and performance, Trainium has become the basic engine for most Bedrock traffic, and Trainium's capacity is essentially sold out until the end of next year. If these chips can improve unit computing costs, they will help enhance the economics of AI infrastructure investment.

For investors, the critical issue remains returns, not just the narrative around demand. AWS management has described multiple growth sources: enterprise cloud migration, AI inference, startup training demand, agent-driven workflows, and custom chips. But these opportunities ultimately need to translate into higher revenue, better profits, and adequate capital returns.

In the interview, Matt Garman emphasized that AWS's capacity is not highly concentrated with just one or two customers; usually, a single customer accounts for a single-digit percentage or even lower. He also stated that in core computing, storage, and inference, enterprise customers have already seen positive returns. These statements help alleviate market concerns about an AI bubble or customer concentration risk.

However, from a shareholder perspective, this still does not equate to Amazon's own investment returns having been proven. About $220 billion in capital expenditure requires consistently high utilization rates, stable customer demand, and controllable costs. If agent applications remain experimental, or enterprise adoption is slower than expected, capital investment could outpace revenue recognition, putting pressure on profitability.

Therefore, the real metrics the market needs to verify are whether AWS’s revenue growth can continue, whether services like Bedrock, Trainium, and Agent Core can expand paid usage, whether enterprises can shift from agent-involved workflows to more autonomous production processes, and whether this new demand can cover data center, chip, power, and supply chain investments.

Amazon's AI returns may be underestimated by the market, but the market's hesitation is not without reason. The AWS interview provided sources of demand and technical roadmaps, while The Information's valuation analysis suggests current multiples deserve to be reconsidered. Whether the discount can ultimately be fixed depends on Amazon’s ability to turn its AI infrastructure investments into visible and sustainable financial returns.