We are witnessing a fundamental shift in the scale of digital infrastructure. The era of megawatt data centers is over; we have entered the age of the 1-Giga...
From Megawatts to Gigawatts
Data centers used to be sized around racks and rooms. Power budgets lived in the megawatt range, cooling was an afterthought bolted onto IT design, and capacity planning meant adding servers until the floor or the breaker panel said stop. That model no longer fits workloads that treat entire buildings as single training or inference systems. When the planning unit jumps to roughly one gigawatt, every decision—site selection, power contracts, cooling topology, network fabric, and operations—has to be made as if the facility itself is the computer.
At this scale, power is not a utility bill line item; it is the binding constraint. Availability, ramp rate, and redundancy of supply determine how much compute you can actually run, not how many accelerators you can buy. Cooling shifts from “enough CRAC units” to engineered heat rejection that can move continuous thermal load without throttling chips. Layout, cabling, and maintenance access must assume concurrent high density across large floor plates, not a few hot rows in an otherwise conventional hall.
Thinking Machines at Facility Scale
Thinking Machines, in this context, means systems that keep large models and agents in continuous, high-bandwidth contact with data and with each other. Training and serving both become long-lived, tightly coupled jobs. They punish latency spikes, partial failures, and uneven power or thermal conditions across the fleet. The useful unit of reliability is no longer a single server; it is a coherent slice of the facility that can finish a job without restarting from scratch.
Design for that world looks different from classic cloud tenancy. You prioritize fabric bandwidth and topology over raw node count, checkpoint strategy over ad hoc restarts, and fleet-level observability over per-box dashboards. Capacity is planned in coherent blocks that can be powered, cooled, and networked together—not in isolated VMs scattered wherever space happens to open up.
Rubin and the Accelerator Density Problem
Rubin stands for the next wave of dense accelerators: more compute and memory bandwidth per rack, and harder power and thermal envelopes per square meter. That density is what makes gigawatt campuses worth building, and what makes weak facility design fail first. An accelerator generation that outruns the building leaves expensive silicon underclocked, idle, or waiting on interconnect.
Practical response is to co-design silicon, rack, and room. Power delivery must reach the tray without excessive conversion loss. Liquid or hybrid cooling becomes baseline where air cannot carry the heat. The network must match the accelerator’s appetite for all-reduce and collective traffic, or utilization collapses even when every GPU is “up.” Buying the chip without the electrical, thermal, and fabric plan is how projects miss their effective capacity targets.
What to Build and Operate Differently
Teams entering the 1-gigawatt compute era should treat the following as first-order work, not follow-on polish:
- Lock power availability and ramp schedules before locking accelerator purchase volumes.
- Specify cooling and heat rejection for peak sustained load, not nameplate averages.
- Design the fabric and job scheduler for large, sticky workloads rather than many small tenancies.
- Instrument power, temperature, and interconnect together so operators can see facility-level bottlenecks, not only node health.
- Plan maintenance and failure domains as coherent compute blocks so one fault does not strand a multi-rack job.
The shift from megawatt halls to gigawatt campuses is not a branding change. It is a change in what “the machine” is: a power-limited, thermally constrained, network-coupled system that Thinking Machines–class workloads and Rubin-class density both demand. Get the facility right, and the accelerators deliver. Get it wrong, and the era’s headline capacity never shows up in useful work completed.