Inference platform map — three segments, six vendors
custom silicon
Groq (LPU)
voice agents, real-time translation
5-10x faster decode than GPU
Cerebras (WSE)
wafer-scale engine
very large context decode
SambaNova (RDU)
reconfigurable dataflow
enterprise regulated
pay premium per-token
win on latency SLAs
GPU platforms
Fireworks — latency
FireAttention, $4B valn
$1/hr GPU raise May 2026
Together — breadth
200+ models, 50-70% under Replicate
"AI Native Cloud"
Baseten — enterprise
$5B valn Jan 2026, SOC2+HIPAA
Truss, per-min billing
Modal — Python DX
$1.1B valn, per-sec billing
infrastructure-as-Python
Anyscale — Ray-native
RayTurbo inference engine
managed Ray clusters
API-first marketplaces
Replicate
pay-per-prediction
multimodal breadth default
DeepInfra
aggressive per-token pricing
hobbyist and prototyping
OpenRouter
unified API over 100+ models
single credit for many providers
Fal
image / video focus
serverless per-call
time-to-first-call optimized
cheapest for bursty and small-scale
rule: per-minute beats per-token above ~30% sustained utilization of one GPU