Amazon is buying $3B of AI chips a month but cannot turn them on
AWS is reportedly stacking billions in Nvidia GB300 server racks in warehouses because its data centers lack the power to run them.
Business · Source: The Circuitry
What happened
Amazon Web Services is aggressively stockpiling Nvidia AI hardware before it can actually use it. Internal emails reviewed by Business Insider reveal that AWS launched a new Early Rack Receipt program. Under this system, the cloud giant expected to take delivery of roughly five billion dollars in Nvidia GB300 server racks at facilities across the United States. Following that initial massive delivery, one internal email projected AWS would take in about three billion dollars more in GB300 racks every single month for the foreseeable future.
The problem is that Amazon cannot actually plug all these machines in. People familiar with the effort reportedly told Business Insider that while some AWS sites have the physical floor space for more racks, they simply do not have enough available power to turn them on. Because of this severe electricity shortage at the data centers, finished and highly expensive Nvidia equipment is currently being staged and stored in intermediary warehouses.
This is a major shift in how Amazon handles its infrastructure. Previously, AWS only took ownership of a finished rack when it was delivered directly to an active data center. An internal email noted that at least one AWS team did not have a clear forecasted path for getting these incoming racks into data centers. An Amazon spokesperson defended the practice to Business Insider. They stated that staging finished hardware near deployment sites is a best business practice so it can be installed the moment capacity opens up. The spokesperson also disputed the claim that the company lacked a clear path for deploying the racks.
Key facts
- $5 billion — Value of Nvidia GB300 server racks AWS expected to receive early under the new program.
- $3 billion — Projected monthly spend on additional GB300 racks for the foreseeable future.
Why it matters
The fundamental bottleneck in the artificial intelligence industry has officially flipped. For the past two years, the entire market was constrained by chip supply. Everyone was fighting to get their hands on Nvidia hardware. Now, the hardware is flowing, but the physical infrastructure cannot support it. The new speed limit for AI is electricity. When a company as massive as Amazon is forced to park billions of dollars of depreciating compute assets in warehouses, it shows that grid upgrades and power permits are failing to keep pace with technological ambition.
This creates a massive second order effect for the entire AI ecosystem. Server racks sitting idle in warehouses generate zero revenue while rapidly losing their cutting edge value. The capital expenditure required to buy three billion dollars of hardware a month is staggering. If cloud providers are absorbing the costs of these logistical nightmares and idle periods, those inefficiencies will eventually trickle down. The cost of training and running models will reflect the reality of these expensive parking lots.
For builders
Compute costs will remain artificially high
Warehousing idle chips is a massive capital drain. Cloud providers will inevitably pass the costs of these logistical bottlenecks down to the developers and founders paying for compute. You should plan your runway assuming compute prices will not drop anytime soon.
Infrastructure pivots entirely to power
The immediate bottleneck is no longer silicon supply. Startups that can solve grid upgrades, power management, or energy efficiency will capture massive budgets from hyperscalers. If you are building in the infrastructure space, pivot your messaging toward power availability.
Deployment timelines face hard physical limits
You cannot software engineer your way out of a power shortage. Founders relying on cloud providers rolling out next generation chips must factor physical grid delays into their product roadmaps. The hardware exists, but access will be gated by electricity.
My take
I find it absolutely wild that Amazon is paying billions of dollars for the most advanced chips on earth just to park them in a warehouse. The AI race is completely bottlenecked by the power grid right now. Hardware sitting idle is hardware losing value every single day. If you want to build the next massive infrastructure unicorn, stop looking at software layers and start looking at electricity.
Original reporting: The Circuitry. This is my rewrite and opinion.