Case Study Deskcasestudydesk.com
AI Web ResearchSourced

Case Study: Parallel Web Systems gets 3x throughput and 3x cost savings on Baseten

Parallel Web Systems Case StudySourced & dated by Case Study Desk
Key facts · TL;DR
Company
Parallel Web Systems
Industry
AI Web Research
Challenge
Agentic pipelines needed throughput-optimized inference under spiky traffic
Headline result
3x throughput and 3x cost savings vs. closed-source

Key results

3x
Higher throughput
50%
Lower latency
KV cache-aware routing + speculative decoding
3x
Cost savings vs. closed-source
~100x
Traffic scaled on Baseten
within a few months

The challenge

Parallel powers web research for AI-native teams, running structured agent pipelines where a single job can involve many model calls and tens of thousands of multi-turn requests per hour with bursty traffic. Two weeks before its Search API launch it was serving thousands of multi-hop searches on closed-source models, whose costs would be unsustainable at its projected long-term scale.

The solution

Parallel launched on Baseten Model APIs with pay-per-token pricing to absorb unpredictable demand, then moved to a dedicated deployment of an open-source model. Baseten's model performance team applied speculative decoding and KV cache-aware routing to optimize latency and throughput.

Our workloads can be very spiky. We need to be able to manage the peaks, but we also don't want to pay for the valleys. Pay-per-token with Model APIs enabled us to easily make the transition to open-source until our scale merited dedicated deployments optimized for throughput and flexible autoscaling.

ML
Matt Lee
Engineering Lead, Parallel Web Systems

The results, in context

Baseten's runtime cut latency by 50% and increased throughput 3x through KV cache-aware routing and speculative decoding, delivering 3x cost savings versus the equivalent closed-source model. Within a few months of the initial deployment, Parallel scaled traffic on Baseten roughly 100x.

Products used

Baseten Model APIsBaseten Dedicated deploymentsBaseten Speculative decodingBaseten KV cache-aware routing