Amazon faces a $220 billion capital spending hurdle for AI growth

The massive scale of AI demand

Amazon is currently facing a situation where it has very little trouble finding customers for its cloud and AI infrastructure. According to reports from ad-hoc-news.de, the company's AWS capacity is already sold out for 2026 and is largely committed for 2027. A significant portion of 2028 is also already booked.

Despite this high demand, CEO Andy Jassy noted that a substantial amount of demand will still go unmet even with the current buildout. The primary challenge for the company is not finding users, but rather managing the economics of the business. Amazon must ensure that the revenue from these services can keep pace with the massive amount of money being poured into the hardware.

The company is looking toward an October 29 earnings report to show how these investments are progressing. Investors are watching to see if the growth in the cloud unit can sustain high margins while the company continues to spend heavily on infrastructure. The goal is to turn these massive contracts into actual sales as capacity comes online.

The high cost of running AI

A major hurdle for the widespread adoption of AI is the cost of inference, which is the expense of running models that have already been trained. Andy Jassy argued that these costs need to fall significantly before many companies will adopt the technology at scale. To solve this, Amazon is leaning heavily on its in-house Trainium chips.

Amazon claims that the Trainium chips offer a 30–40% price-performance advantage over comparable processors in the market. The company has already disclosed future revenue commitments exceeding $225 billion from these chips, including agreements with major players like Anthropic and OpenAI. However, these figures represent contracted future revenue rather than cash that has been collected today.

The success of this strategy depends on how quickly Amazon can bring the promised computing capacity online. If the company can provide cheaper hardware and a friendly software environment, it creates a stronger incentive for developers to move their applications to Amazon's infrastructure. The company wants to make it easy for people to port their existing workflows onto its specific technology stack.

Infrastructure vs. Profitability

Amazon plans to spend about $220 billion on capital expenditures in 2026, a figure that was recently raised due to higher memory prices. This spending is necessary to build the capacity for the growing AI business, but it also puts a heavy burden on the company's free cash flow. Because of depreciation, these costs will also raise the company's expenses in future periods.

Data from Morningstar shows that depreciation climbed 67% in the prior quarter, and trailing free cash flow was slightly negative. This means that growth and investment cannot be judged by looking at them in isolation. Instead, the return on the capital that has already been deployed is the most important metric for the company's health.

The cloud unit remains the primary yardstick for success, with AWS revenue growing 37% in the second quarter of 2026. While the company faces risks like fuel costs, analysts are looking for upward revisions to operating income. The challenge is squaring the long-term outlook for AWS and advertising with the immediate weight of the current capital burden.

Technical progress and paper gains

Technical progress is broadening what Amazon's Trainium chips can do beyond just running models. New reports indicate that Trainium2 and Trainium3 can now natively run research training on PyTorch, which is a common software framework. This extends the use of the chips into the actual training of models, rather than just running them for end-users.

However, there are still limits to this technology compared to the more mature CUDA ecosystem. While the chips can handle some workloads, such as the Qwen3-32B model, they do not yet support every large training workload. For Amazon, the battle is as much about software compatibility as it is about the raw performance of the hardware itself.

The financial picture is also complicated by paper gains that can obscure the actual performance of the retail and cloud businesses. Of the $120.7 billion in pre-tax profit Amazon reported in the first half of 2026, more than $69 billion came from unrealized valuation gains on its stake in Anthropic. These are not operating earnings, meaning they do not reflect the day-to-day profitability of the company's core services.