IN STEALTH / COMPUTE DEPLOYMENT REQUIREMENTS
Plan your AI
compute deployment.
LightningDC is building a way to match qualified power with suitable AI infrastructure. For compute buyers, the starting point is the work that needs to run and the conditions it requires. A useful brief connects hardware and service expectations to power, cooling, networking, and an achievable deployment sequence.
01 / THE WORKLOAD
What needs to run?
Describe whether you are planning training, fine-tuning, inference, or another compute-intensive workload. Specify the accelerator type, memory, quantity, and any hardware you already control. Explain how the job uses storage and how tightly its machines must work together.
For inference, distinguish interactive requests from work that can be scheduled. Define the latency, throughput, completion deadlines, and recovery behavior that matter to your customers.
02 / THE INFRASTRUCTURE
What must support it?
State the planned IT load and whether a quoted power requirement includes cooling and other facility loads. Share rack density, electrical distribution, and cooling requirements, including any liquid-cooling specifications.
Define network connectivity, bandwidth, latency, storage, and security expectations. Make the boundary clear between the site, the modular infrastructure, the installed hardware, and the service you expect to receive.
03 / OPERATING CONDITIONS
When and how must it work?
Share the desired start date, initial capacity, ramp schedule, and expected duration. Distinguish a firm deadline from a preference, and describe dependencies such as hardware delivery, software readiness, and customer commitments.
Specify availability requirements and what happens during an interruption. Include maintenance windows, notice, checkpointing, restart time, and who would operate each part of the deployment.
What should a compute buyer submit?
You do not need a finished engineering specification to start. A concise brief should cover:
- Workload: The application, job profile, and performance requirements.
- Hardware: Accelerator preferences, quantity, ownership, and compatibility constraints.
- Capacity: Initial megawatts or rack requirements, plus the expected ramp.
- Location and network: Geographic needs, connectivity, data handling, and latency limits.
- Timing and operations: Target date, duration, availability, and acceptable interruptions.
- Commercial scope: The capacity or service sought, operating responsibilities, and contracting expectations.
Mark requirements as essential or flexible. That distinction makes it easier to evaluate a potential site against the job it must do.
Practical questions
Can the same deployment serve training and inference?
The workload requirements need to decide that. Share the cluster topology, interconnect, storage, latency, and scheduling needs rather than treating a GPU count as the complete specification.
Does a flexible workload mean availability does not matter?
No. Define the service outcome and the interruption behavior you can accept. A completion deadline, a maximum outage duration, and a recovery requirement are more useful than an undefined promise of flexibility.
Does an inquiry reserve capacity?
No. LightningDC is in stealth and exploring deployment opportunities. An inquiry starts a discussion about fit, remaining requirements, and potential next steps. Any capacity, timing, and service commitments would need a separate agreement.