Uploaded October 2026 | Updated November 2026, 4 days ago
If your token spend keeps going up and productivity isn't keeping pace, start here. Eugene Cheah lays out four steps for cutting AI inference costs at scale: measure, right-size, buy capacity well and check your processes. Along the way he covers splitting API keys per team, agentic vs. process workloads, open models that match frontier models on specific tasks, classifiers, fine-tuning, the speed-vs-cost tradeoff and "slop grenades."
Recorded live at AI Engineer NYC 2026.
🎟️ Join us at the next one: AIE CODE Summit → https://aie.codes
Speaker
Eugene Cheah — X: https://x.com/picocreator
Featherless AI — featherless.ai
Timestamps
0:00 Intro: Eugene Cheah, Featherless AI
1:48 What "scale" looks like: 2T tokens/day
2:13 Disclaimers: who this is for
3:13 The 4 moves
3:23 Move 1: Measure (stop sharing one API key)
5:12 Agentic vs. process workloads
9:27 Move 2: Right-size your models
12:42 Classifiers: a new class of cheap models
14:37 Plan with the smart model, build with the cheap one
15:47 Nobody switches models to save money
19:27 Give teams model labels, not model names
25:27 Bootstrapping your own fine-tuned model
27:42 Shadow traffic to keep improving
28:57 Move 3: Buying GPU capacity
30:07 2x the speed = 3x the cost
32:22 Bursty traffic & dedicated capacity
34:32 Move 4: Check your processes
37:07 Tokens are not work done
42:07 Don't throw slop grenades
44:27 Q&A: Fine-tuning & data sets
48:02 Q&A: Golden traces
50:04 Q&A: Escaping harness lock-in
52:37 Q&A: Caching & auto-routing
If your token spend keeps going up and productivity isn't keeping pace, start here. Eugene Cheah lays out four steps for cutting AI inference costs at scale: measure, right-size, buy capacity well and check your processes. Along the way he covers splitting API keys per team, agentic vs. process workloads, open models that match frontier models on specific tasks, classifiers, fine-tuning, the speed-vs-cost tradeoff and "slop grenades."
Recorded live at AI Engineer NYC 2026.
🎟️ Join us at the next one: AIE CODE Summit → https://aie.codes
Speaker
Eugene Cheah — X: https://x.com/picocreator
Featherless AI — featherless.ai
Timestamps
0:00 Intro: Eugene Cheah, Featherless AI
1:48 What "scale" looks like: 2T tokens/day
2:13 Disclaimers: who this is for
3:13 The 4 moves
3:23 Move 1: Measure (stop sharing one API key)
5:12 Agentic vs. process workloads
9:27 Move 2: Right-size your models
12:42 Classifiers: a new class of cheap models
14:37 Plan with the smart model, build with the cheap one
15:47 Nobody switches models to save money
19:27 Give teams model labels, not model names
25:27 Bootstrapping your own fine-tuned model
27:42 Shadow traffic to keep improving
28:57 Move 3: Buying GPU capacity
30:07 2x the speed = 3x the cost
32:22 Bursty traffic & dedicated capacity
34:32 Move 4: Check your processes
37:07 Tokens are not work done
42:07 Don't throw slop grenades
44:27 Q&A: Fine-tuning & data sets
48:02 Q&A: Golden traces
50:04 Q&A: Escaping harness lock-in
52:37 Q&A: Caching & auto-routing










