AI Data Center Power Capacity: Why AWS Has GB300 Racks Waiting in Warehouses

What was reported

On October 20, Business Insider reported, from internal August emails, that AWS had set up a new Early Rack Receipt process for an expected wave of NVIDIA GB300 server racks at US facilities. One email projected about $5 billion of GB300 racks in August and roughly $3 billion more each month "for the foreseeable future". Some AWS sites had room for more racks but not enough available power to turn them on, so finished racks were staged at intermediary warehouses until the facilities were ready. At least one AWS supply chain team did not have "a clear forecasted path" for getting the racks into data centers, one email said.

Amazon usually does not pay for racks until they are installed in a data center, according to the internal documents, which leaves suppliers carrying equipment in transit. Under the new process it takes ownership earlier, so cash goes out sooner and revenue waits for power. An Amazon spokesperson said the company routinely stages finished hardware near deployment sites so it can be installed once power capacity is available, called that "best business practice", and disputed that the company lacks a forecasted path for getting racks into data centers.

The context is Amazon's roughly $220 billion of capital expenditure this year. AWS revenue grew 37 percent to $42.2 billion in the second quarter, and free cash flow for the twelve months to June 30 was an outflow of $7.6 billion. Amazon reports its third quarter on October 29.

Why it matters for an AI data center operator

Few operators will receive $3 billion of racks a month, but the mechanism is the same at any scale. A GPU rack arrives on the vendor's schedule. The power to run it arrives on the utility's schedule, after the transformer, switchgear and PDU circuits are in and tested. When the first runs ahead of the second, the operator owns a rack that cannot be energized, ties up cash and produces nothing.

The phrase to keep is a clear forecasted path. The question is how many racks can be switched on each month, on which circuits, and which racks go first when the power arrives. That forecast of energized capacity needs data from several teams: the utility's energization date, the commissioning schedule, the rack space and cooling free in each hall, and the inventory of racks received, staged and installed. In many organizations each lives in its own spreadsheet, and they drift apart. AWS's infrastructure leader Prasad Kalyanaraman told staff in September that the forecast "is changing on a daily basis".

There is a second gap after power-on. An energized rack is unproductive until it has passed burn-in, joined a pool and received its first job. Business Insider notes that the term dark GPUs covers both cases: hardware waiting for power, and hardware plugged in and underused.

What the operator sees in Sensaka

Sensaka for AI Data Centers keeps the physical estate, the accelerator inventory, the power readings and the tickets in one record. The AI Cockpit takes the questions in plain language: how many accelerators are online, faulty or idle right now; which racks or zones are near their rated power; which assets come off warranty soon. Each is matched locally to a preset question and answered as a board or a 30-second data video. No external language model is called.

GPU pool operations keeps a lifecycle state for every accelerator from intake to retirement, with its node, rack, room and site, warranty dates and owner team. A rack received but not energized sits in the same inventory as the racks in production, and a server joins a GPU pool only after bare-metal provisioning and the burn-in gate. The CMDB holds facilities, rooms, racks and rack units next to power equipment, and logs every change with its previous and new value, source and time, so the delivery-to-operations handover is a recorded step. The names of the states, received, staged, racked, energized or in production, are settled during implementation, when configuration item types and required attributes are defined.

On the facility side, Sensaka monitors power distribution and PDU circuits, shows rack power density as a heat map with limit alerts, and analyses rack space, power and cooling capacity together. When a hall is energized, the work to rack, cable, provision and burn in each unit runs as tickets, synced with ServiceNow or Jira where those are the system of record. Once the racks draw power, the same data feeds PUE and WUE per zone and the reporting an EU site needs under the Energy Efficiency Directive.

Sensaka does not forecast when a utility will energize a substation. Its part is the operator's count: owned, installed, energized and busy racks, the gaps between them, and the tickets that close them. It works with NVIDIA, AMD and other accelerators.

Takeaway

The gate on new AI capacity has moved from GPU supply to the energization date, and the hardware now arrives before the power. An operator who can say, for each rack, whether it is received, installed, energized or running, and when the next tranche of power is due, has the forecasted path the AWS email was missing. That is an inventory and ticket discipline, easier when the inventory is read from the hardware and shared by the facility, compute and finance teams.

Frequently asked questions