Pretrain Runway
84 Days
Completion: ~2.8 Months
Effective Pretrain Velocity
18.4M tok/s
14.88 ZFLOPs / Day
Datacenter MW Load
168.8 MW
Headroom: 31.2 MW (84.4%)
API Serving Capacity
62.5k QPS
Video: 0 Gen/min (Killed)
Cluster FLOP/s Capacity Allocation
Pretrain
RL
Video
Inference
Frontier Model Delivery Roadmap (Milestones)
Day 0 to Day 180
Workload Partition Telemetry Ledger
MFU Model: 44.5% Effective
| Workload Domain | Allocated GPUs | Effective FP8 PFLOPS | Estimated MW Load | Domain Output Metric |
|---|---|---|---|---|
| ■ Foundation Pretrain ('Bel') | 137,500 | 137,672 | 92.8 MW | 18.4M Tokens/sec |
| ■ Post-Training Reasoning RL | 50,000 | 50,063 | 33.8 MW | 650k Rollouts/hr |
| ■ Multimodal Video Gen (Sora) | 0 | 0 | 0.0 MW | 0 Streams/sec |
| ■ Production API Serving | 62,500 | 62,578 | 42.2 MW | 62,500 Queries/sec |