Amazon's AI capital spending is becoming the central point of disagreement over its valuation: the market worries that the company is investing too quickly and seeing returns too slowly, but demand signals from AWS suggest that cloud migration, AI inference, and agentic workflows may be reshaping its growth trajectory.

Amazon is currently trading at a lower forward price-to-earnings ratio than Walmart, even though its revenue growth is expected to be roughly three times faster, according to a Sunday report. High capital spending and uncertainty over the returns on AI investment are the main reasons investors are applying a discount.

The focus of this debate is whether Amazon's roughly $220 billion in capital investment can be absorbed by future demand. In a recent a16z interview, AWS CEO Matt Garman said AI-related demand is enormous, capital spending will not slow down in the near term, and AWS is still reserving scarce GPU capacity for startups.

For the market, the question is not whether Amazon has AI opportunities, but whether those opportunities can translate into sufficient revenue, profit, and cash returns. The AWS interview explained where demand is coming from, but it has not yet proven that this round of investment has paid off.

The Mystery of Amazon's Valuation Discount

Citing Koyfin data, Amazon currently trades at about 29 times forward earnings, while Walmart trades at about 37 times. Walmart's revenue is expected to grow about 5.6% this year, only about one-third of Amazon's estimated growth rate, yet its valuation premium has persisted for about 18 months.

This inversion reflects the market's main concern: Amazon is sharply increasing capital spending, mainly to expand AI data centers, joining Alphabet and Meta in betting on AI infrastructure demand, but the payback period for such investment remains unclear. If AI demand materializes more slowly than expected, heavy capital spending could continue to weigh on earnings performance.

However, sum-of-the-parts analysis also suggests that Amazon's business structure may not be fully priced in. Amazon has a market value of about $2.8 trillion, but its asset portfolio includes AWS, online retail, advertising, and content businesses. Materials show that Amazon's online retail sales grew 13% year over year in the first half of this year, its advertising business generates about $80 billion in annual revenue with growth of 20% to 25%, and AWS's standalone value is conservatively estimated at no less than $1 trillion.

In other words, the market's discount on Amazon is not merely assessing the retail business, but pricing in uncertainty over AI capital spending, the conglomerate structure, and insufficient visibility on returns. A low valuation does not automatically mean undervaluation, but it provides an entry point for re-examining AWS's growth logic.

AWS's Growth Logic Is Undergoing Three Major Shifts

In the a16z interview, Matt Garman said AWS revenue is about $160 billion to $170 billion, up 37% year over year. He believes AWS is still in the early stages of a long-term opportunity, because a large number of workloads remain on-premises, enterprises' daily computing needs continue to increase, and AI is bringing additional tailwinds.

The first layer of growth comes from enterprises moving to the cloud. Matt Garman said cloud migration and AI demand are driving AWS growth simultaneously. For large enterprises, AI is not a single model call, but a workload tied to data, storage, security, permissions, and production systems. This makes AWS not just a provider of GPU capacity, but also a provider of the full infrastructure enterprises need to put AI into production.

The second layer of growth comes from startups. Matt Garman said AWS estimates that about 30% to 40% of its revenue currently comes from companies that were once startups. AI-native startups differ from early internet companies in that they may need large-scale computing power from day one, training models is costly, and their funding and valuation sizes are larger. AWS therefore chooses to reserve some scarce GPU and AI accelerator capacity for startups, rather than selling all resources to large frontier labs.

He also said AWS has met about 60% of relevant requests in some way, sometimes delivering in different regions, at different times, or under different configurations. AWS will also buy 2 million Nvidia GPUs in the coming years. These statements show that AWS is trying to maintain its customer ecosystem through capacity allocation, rather than betting only on a few large customers.

The third layer of growth comes from agentic workflows. Matt Garman said more and more teams want cloud environments to be suitable not only for human use, but also for agents. AWS is optimizing agents' ability to call across data lakes, databases, and services around offerings such as Agent Core, Bedrock, and AWS Context. Agents care about tail latency and also generate short-duration computing demand that can be created and released quickly.

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 may automatically write code, call databases, manage infrastructure, and execute tasks under sandboxes and temporary permissions. For AWS, this means new demand for inference, storage, security, permission management, and computing.

In addition, custom chips are seen by AWS as key to supporting this logic. Matt Garman said Graviton offers about 20% improvement in both cost and performance, Trainium has become the foundation for most Bedrock traffic, and Trainium capacity is roughly sold out through the end of next year. If these chips can improve the cost per unit of computing power, they will help improve the economics of AI infrastructure investment.

What Really Needs to Be Verified Are the Key Metrics

For investors, the key question remains returns, not the demand narrative itself. AWS management described multiple sources of growth: enterprise cloud migration, AI inference, startup training demand, agentic workflows, and custom chips. But these opportunities must ultimately show up as higher revenue, better profit, and sufficient returns on capital.

In the interview, Matt Garman emphasized that AWS capacity is not highly concentrated in one or two customers, with any single customer usually accounting for a single-digit percentage or even less. He also said enterprise customers are already seeing positive returns in core computing, storage, and inference. These comments help ease market concerns about an AI bubble and customer concentration.

But from a shareholder perspective, this still does not mean Amazon's own investment returns have been proven. Roughly $220 billion in capital spending requires sustained high utilization, stable customer demand, and controllable costs to support it. If agent applications remain in the experimental stage, or enterprise deployment is slower than expected, capital investment may precede revenue recognition and weigh on earnings.

Therefore, the metrics the market really needs to verify are whether AWS revenue growth can continue, whether services such as Bedrock, Trainium, and Agent Core can expand paid usage, whether enterprises can move from agent processes with human involvement to more autonomous production workflows, and whether this new demand can cover investment in data centers, chips, power, and supply chains.

Amazon's AI returns may be undervalued by the market, but the market's hesitation is not without reason. The AWS interview provided the sources of demand and the technology path, while the valuation analysis showed that current multiples leave room for re-examination. Whether the discount can ultimately be repaired depends on whether Amazon can turn AI infrastructure investment into visible, sustainable financial returns.