Amazon CEO Jassy: AWS compute orders booked through 2028, enterprise AI adoption not yet widespread

Article body EM_StockImg_Start

EM_StockImg_End

On October 23 local time, Amazon CEO Andy Jassy said in an interview with the US venture capital firm a16z that even with roughly $220 billion in capital expenditure in 2026, Amazon will still be unable to meet a large amount of customer demand. At present, AWS's main computing capacity for 2026 has already sold out, most of its 2027 capacity has been booked, and a considerable portion of its 2028 capacity has also been committed to by customers.

It is reported that Amazon will release its third-quarter 2026 results on October 29 local time and hold an earnings call at 5 p.m. US Eastern Time that day.

$220 billion in capital expenditure

still insufficient to meet customer demand

Speaking about AI infrastructure investment, Jassy said the current growth rate of the artificial intelligence business far exceeds the technology waves Amazon has experienced before. In the first three years after the rise of generative AI, AWS's AI business had already surpassed $25 billion in annualized revenue, whereas three years after AWS was founded, its overall business had annualized revenue of only about $60 million.

Amazon's second-quarter earnings report showed that AWS revenue reached $42.2 billion, up 36.7% year on year, with an annualized revenue scale of about $169 billion. Jassy said that even if the company invests capital expenditure according to the $220 billion plan this year, it will still be unable to meet a large amount of customer demand.

On how to decide the scale of infrastructure investment, Jassy explained that AWS capital expenditure is mainly divided into two parts: data center construction and the procurement of servers and network equipment. Data centers usually require investment about two years in advance, and the assets have a useful life of about 30 years; servers and network equipment can be purchased a few months before being put into use according to customer demand.

He disclosed that investment in servers and network equipment generally breaks even in about three years, the equipment has a useful life of about five to six years, and most related transactions currently have terms of about five years. Jassy said that because data centers take a long time from construction to being put into use, rapid expansion in the early stage will put pressure on free cash flow, and then as facilities are put into operation and generate revenue, the related investment will be gradually recovered.

Regarding future supply, he said AWS will still reserve some capacity for on-demand services such as Bedrock and EC2, but its main booked capacity for 2026 has sold out, 2027 is basically sold out, and a considerable portion of 2028 has also been booked. He expects tight supply of some infrastructure components to persist for several years.

Inference costs still need to fall

Speaking about the structure of AI demand, Jassy described the current market as "dumbbell-shaped": at one end are foundation model labs that need large amounts of computing power, as well as a small number of fast-growing AI applications; at the other end are enterprises using AI to reduce costs and improve production efficiency, including scenarios such as automated customer service, fraud detection and business process automation.

By contrast, the existing production applications of a large number of enterprises have not yet widely used AI inference capabilities. Jassy expects that in the next five to seven years, a considerable portion or even most enterprise applications will gradually introduce AI inference. He said that promoting the adoption of enterprise AI applications requires solving issues such as talent, technology and organizational capabilities, among which inference cost is particularly important. Although inference costs have continued to decline over the past few years, in his view current costs are still too high and need to be substantially reduced further.

On self-developed chips, Jassy said Amazon acquired Annapurna Labs in 2015, initially to improve server computing efficiency, and subsequently developed the Graviton CPU and the Trainium AI accelerator.

According to him, among AWS's top 1,000 EC2 customers, 99% have widely used Graviton, whose price-performance ratio is about 40% higher than other x86 processors; Trainium has a price-performance advantage of about 30% to 40% over comparable processors. Jassy said that reducing inference costs helps enterprises move more existing applications to AI architectures, and is also an important reason why Amazon continues to invest in self-developed chips.

Responding to investment in OpenAI and Anthropic

On why Amazon invested in Anthropic and OpenAI, Jassy said the company is optimistic about the two companies' model technology, application products and long-term development prospects, and that both sides have extensive business cooperation with AWS. Among them, Anthropic has long used Trainium to develop models, and the two sides cooperate deeply on chips and model operation.

Regarding OpenAI, Jassy said that previously, due to commercial arrangements, some AWS customers could not use OpenAI models through AWS. As cooperation between the two sides deepened, AWS Bedrock has integrated OpenAI models, and the two sides have also carried out joint development around the stateful runtime environment required by agents. He said both OpenAI and Anthropic have made multi-year, multi-gigawatt-scale commitments to use Trainium.

Speaking about AI's impact on corporate organizations, Jassy said Amazon's Bedrock team once rebuilt an inference engine with six engineers in 76 days, whereas similar projects in the past might have required about 40 people working for a year. He also gave an example that a consumer business team once completed a runnable prototype of an agent product with four or five people in a few days. Jassy believes that as AI tools become more capable, the number of people needed to complete existing work may decrease in the future, team sizes will shrink further, and employees will also have stronger cross-functional capabilities.

However, he also stressed that AI cannot replace people's critical thinking, judgment and creativity. For enterprises, the changes brought by AI are not just about improving work efficiency, but also require rethinking existing products, software architectures and customer experience. Jassy said Amazon is redesigning businesses including Alexa, retail and Prime Video to adapt to the technological changes brought by generative AI.

(Source: Shanghai Securities News) Mid-article embedded advertisement