Amazon Is Buying Billions in AI Chips It Can't Turn On—Here's Why

Amazon's artificial intelligence ambitions are running into a constraint that money alone cannot quickly solve: the physical infrastructure needed to run powerful computing systems. The challenge is no longer simply securing advanced chips, but ensuring that data centres have enough electricity, cooling, and operational capacity to put them to work.

The distinction matters as Amazon commits billions of dollars to NVIDIA hardware. Equipment delivered ahead of infrastructure readiness can help secure scarce supply, but it can also leave expensive assets waiting to generate returns.

Why AI Infrastructure Can Outrun Its Power Supply

Business Insider reported on 21 October that Amazon Web Services (AWS) expected to receive around $5 billion (£3.75 billion) in NVIDIA GB300 server racks through an Early Rack Receipt process, with internal emails projecting another $3 billion (£2.25 billion) in monthly deliveries for the foreseeable future.

The report said some equipment was being held at intermediary warehouses because certain AWS facilities lacked the power needed to operate it. That does not establish that all the incoming systems are idle, or that the company has no deployment plan.

An Amazon spokesperson defended staging finished hardware close to deployment, describing it as 'best business practice', and said the company had established asset-tracking systems. The technical demands help explain why installation is more complicated than unpacking servers. NVIDIA's GB300 NVL72 specifications describe a liquid-cooled rack containing 72 Blackwell Ultra GPUs and 36 Grace CPUs.

NVIDIA's technical documentation says a full rack can require up to 142 kilowatts of power. Facilities must also support cooling, networking, and integration before the hardware can run workloads reliably. Expanding these systems therefore requires coordination between chip deliveries, construction schedules, and electricity infrastructure.

The Commercial Trade-Off Behind Early Hardware Purchases

Ordering equipment ahead of deployment can be a rational response to intense competition for AI hardware. Waiting until a data centre is fully ready could mean missing delivery windows or delaying capacity customers are already seeking. But early procurement shifts the timing of the investment risk. Capital is committed before every rack can necessarily contribute to revenue, while the company continues to pay for construction, power connections, and other infrastructure.

Amazon is not pursuing this expansion without evidence of demand. In an official announcement on 26 August 2026, AWS and NVIDIA said they planned to deploy two million additional NVIDIA GPUs across AWS's global infrastructure in 2027 and 2028. That commitment indicates the scale of the companies' growth expectations, but it is a future deployment plan, not evidence that all the capacity is already installed or operating.

What the Spending Means for Amazon's Cash Flow

The timing gap matters because infrastructure investment can weigh on cash flow before the resulting capacity earns money. Amazon's second-quarter results reported a free-cash-flow outflow of $7.6 billion (£5.7 billion) over the trailing 12 months, primarily driven by increased spending on property and equipment.

Amazon's capital expenditure forecast has also been reported at $220 billion (£165 billion) for 2026, reflecting the scale of its infrastructure build-out. Chief executive Andy Jassy has described AI as a 'once-in-a-lifetime opportunity', arguing in his 2025 shareholder letter that the company expects substantial investment to support future growth.

The key question is not whether AI demand exists, but how quickly Amazon can turn spending into usable capacity. If electricity and construction delays persist, the company could face a longer wait before some investments contribute to sales, even if demand remains strong.

Amazon's advantage will depend on coordinating the entire system: securing hardware, bringing power online, completing facilities, and matching new capacity with customer demand. In the AI infrastructure race, buying the chips is only one part of the job.