The GPU is what you sell. SynapseDC runs everything it depends on.
SynapseDC runs the power, cooling and liquid racks around your GPUs and joins every GPU to its job, its customer and what it costs, on one estate model. A CDU alarm reads as GPU time lost; a failed GPU reads as revenue at risk.
40–60 kW liquid racks on CDU loops. A cooling fault stops being a ticket and becomes lost GPU time within minutes.
Coupling
A training job spans hundreds of GPUs. One failed GPU or one flapping rail stops or slows all of them.
Cost
Power is the largest running cost, and it is billed whether the GPUs are earning or idle.
Blind spots
The GPU team sees a slow job. The facility team sees a CDU alarm. Nobody sees that it is the same event.
One chain
Fromtheutilityfeedtotheinvoice.
What a GPU depends on and what it earns, on one estate model. One event reads the same way on both sides of the building.
What the GPU depends on
Utility feedtariff · contract
UPS · buswayA and B sides
CDU loopflow · supply temp
Liquid rack40–60 kW
What the GPU earns
GPU node8 GPUs · NVLink
GPUclocks · throttle · Xid
Jobstep time · checkpoint
Customerreserved · on-demand
RevenueGPU-hours · credits
One chainCDU flow at 62 %144 GPUs throttlejob 24 % slowerincident on the cooling fault, not the job
Use cases
FivesituationseveryGPUcloudknows.
Each one names the modules that answer it.
GPU cloudUC-G1NeoCloud
One failed GPU stops a 152-GPU training job
An Xid error on one GPU stalls the whole allocation. The GPU team sees the error; nobody sees the job, the customer or the cost.
The job with the biggest blast radius is on the Command Center and GPU Fleet: the customer, GPU time lost per hour, minutes since the last checkpoint and work to redo, with a Critical incident opened on the node.
Per-GPU telemetry joined to the job and the customer
152GPUs stopped by one failure$625/hGPU time lost52 minsince the last checkpoint
GPU cloudUC-G2NeoCloud
A job is 24 % slower and the cause is in the plant room
CDU flow drops, GPUs throttle, the job slows. The facility team sees the CDU; the platform team sees a slow job.
Thermal slowdown is traced from the job to the GPUs to the rack to the CDU loop. The slowest job and its cause sit together, with the hardware-slowdown flags and the step-time share lost.
24 %step time lost62 %CDU flow at the loop144GPUs slowed
GPU cloudUC-G3NeoCloud
Where should the next 256-GPU job go?
Placement is a guess between free GPUs, rack headroom, cooling state and power already paid for.
Pods are ranked for the job: free healthy GPUs, rack power headroom, cooling state (CDU for liquid), contracted headroom, kW already paid under the block or floor, and the marginal power cost per GPU-hour.
8–256GPUs per placement query$0.19power per sold GPU-hourper podfit and headroom
GPU cloudUC-G4NeoCloud
GPUs are allocated, drawing power and doing nothing
Idle allocations are billed and powered, but invisible between the scheduler and the facility.
Idle jobs are listed with their GPUs, the power they draw and what that costs per day, next to the inference pods that could be capped in the peak tariff window.
6 hidle allocation found in the demoper tenantenergy and cost4finding types against contracts
CommercialUC-C4Colocation · NeoCloud
We pay for power we don't use
Take-or-pay floors, demand ratchets and idle colocation blocks are found when the invoice arrives.
The bill is priced against the tariff and the contract: energy, demand charge on the floor or peak, idle kW, and findings when a block is nearly used, a new peak sets the ratchet, or draw rises with no change in progress.
Does SynapseDC cover liquid cooling as well as the GPUs?
Yes. CDU loops, liquid racks, the power path to the rack and the power bill are in the same estate model as the GPUs, the jobs and the customers, so a cooling fault is read as the GPU time it costs.
Do you read our customers' jobs or models?
No. SynapseDC reads GPU telemetry and the scheduler's job metadata (who, which GPUs, start, state). It never reads a job's data, model or code.
Which accelerators are covered?
NVIDIA B200, H200, H100, A100, L40S and T4, and Intel Gaudi 3 and Data Center GPU Max 1550, with vendor-aware telemetry labels. Other accelerators are to be discussed.
Do we need a scheduler integration?
It helps. GPU telemetry alone shows the GPUs; the scheduler's job metadata (Slurm or Kubernetes) is what joins a GPU to its job and its customer.
BringoneGPUcluster.
We map it onto a demo estate shaped like yours: its power path, its cooling loop, its jobs and its customers. Then we agree what a one-pod pilot should prove.
