~/talk

How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models

links:none on file
slides:none on file

Self-hosted GLM deployment cuts inference spend roughly 10x vs Anthropic/OpenAI and claims about 90% cost reduction with a blended token cost of 30¢ per million tokens, serving 2,000 engineers on 56 B200 GPUs. Results show ~60–70% faster lead time and 60% coding-time gains, plus an 88% performance uplift from tuning the same hardware (quantization VFP4/FP4, tensor-parallel tweaks, caching >90%). Targets: sub-3s time to first token and ~200 TPS interactive throughput; production needs OpenAI-compatible endpoints, sandboxed human-in-loop workflows, and strict telemetry for EU compliance.

AI-written summary · about

  1. 0:00Intro
  2. 0:36Who Revolut is
  3. 1:18What AI did to 2,000 engineers' lead time
  4. 2:45Why they started self-hosting
  5. 3:33Coding isn't the bottleneck
  6. 4:24What it takes to self-host GLM
  7. 5:46Tuning: same GPUs, 88% faster
  8. 7:06The parameters that matter
  9. 8:14Production setup
  10. 9:23For the CIO: compliance and safety
~/search
type to search 6,518 conferences, 143,918 speakers and 192,133 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
← →
on a talk: 5 seconds back / forward
0–9
on a talk: jump to 0%–90% of the recording
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.