BREAKING
NVIDIA Bets on 'Intelligence per Dollar'
Why Post-Training Is Continuous
1
Pre-train fluency
↓
2
RL post-train skills
↓
3
Rollout = inference
↓
4
Repeat continuously
0
PFLOPS
FP4 compute
0
GB
HBM4
0
TB/s
NVLink 6
Rubin vs Blackwell Gains
Inference
5
Training
3.5
NVLink BW
2
0
B
parameters
0
B
active params
0
%
SWE-bench
Verdict Waits on Q4 2026
AI NEWS BLITZ
NVIDIA is pitching its Vera Rubin platform around cost-per-token for AI agents.