Doubleword

Newcastle AI community! In 1 weeks time, Seb Ringrose will be speaking at AI in Production 2026 on one of the biggest challenges emerging in production AI systems: what happens when async agents start running for hours - or days - without supervision. The talk, “Async Agents in Production: Failure Modes and Fixes”, will cover the operational realities of long-running agentic systems, including: → error compounding across multi-step workflows → context drift and recovery strategies → controlling runaway token spend → infrastructure patterns for reliable async execution → and lessons from running high-throughput inference systems in production As more teams move from copilots to autonomous workflows, reliability and cost control are quickly becoming the defining engineering problems. If you’re around in #Newcastle next week, come along and say hello! 👋 Tickets + event info: https://lnkd.in/ePWQdYxR   Jumping Rivers Ltd Amieroh A. Emily Wales Gift Kenneth Colin Gillespie Tim Brock

Newcastle AI community! In 1 weeks time, Seb Ringrose will be speaking at AI in Production 2026 on one of the biggest challenges emerging in production AI systems: what happens when async agents start running for hours - or days - without supervision. The talk, “Async Agents in Production: Failure Modes and Fixes”, will cover the operational realities of long-running agentic systems, including: → error compounding across multi-step workflows → context drift and recovery strategies → controlling runaway token spend → infrastructure patterns for reliable async execution → and lessons from running high-throughput inference systems in production As more teams move from copilots to autonomous workflows, reliability and cost control are quickly becoming the defining engineering problems. If you’re around in #Newcastle next week, come along and say hello! 👋 Tickets + event info: https://lnkd.in/ePWQdYxR   Jumping Rivers Ltd Amieroh A. Emily Wales Gift Kenneth Colin Gillespie Tim Brock

Doubleword AI is now officially supported provider. https://www.doubleword.ai/

The token explosion is becoming very real. Median token usage per request: ~2,000 → ~5,200 in a year and it’s only trending in one direction. As models improve, teams are becoming far more willing to: → add more context → run deeper reasoning loops → chain more tools together → keep agents alive longer Which means the bottleneck increasingly becomes cost at scale. This shift is exactly why Doubleword focuses so heavily on throughput optimisation and low cost-per-token inference for high-volume AI workloads. Source: Datadog

Memory-bound GPU kernels are capped by bandwidth between memory and compute. But LLM weights are losslessly compressible — so you can beat the cap by moving fewer bytes. We built a proof of principle fused tANS-decode + vecadd kernel that runs at ~1TB/s effective, faster than the achievable memory bandwidth on an RTX 4090. More here: https://lnkd.in/eQv-wmBr

One of the clearest signs that async inference matters is workloads that simply wouldn’t be economically viable in realtime. OpenMed used Doubleword to process and annotate 119K medical images using frontier vision-language models. → 94% lower cost → 119K images annotated → cross-validation pipelines → higher benchmark robustness This is the kind of workload that changes completely once inference becomes cheap enough to scale aggressively. Link to the full project in the comments.

🚨 London! Tomorrow, join Meryem Arik as she takes the stage at LLMday! Tickets still available with the discount code LLM30.

QCon Software Development Conferences Boston! On June 3rd, Doubleword CEO Meryem Arik will be delivering the closing keynote at QCon Boston 2026. Looking forward to chatting to teams building AI systems in production - especially as inference costs, long-running agents and autonomous workflows become core engineering challenges rather than future problems. If you’re attending and want to chat about inference infrastructure, async systems, or reducing token spend at scale, come say hi! 👋 More info + tickets: https://lnkd.in/efrtRQi7 #QConBoston

This week Doubleword partnered with SGLang, Ai2 and Crusoe to host a happy hour during #MLSys2026. It was a great turnout in Bellevue chatting to builders working on open models, inference infra, and agent systems. Thanks to everyone who joined us on the night - and shoutout to Jamie Dborin for sharing the work the Inference Lab at Doubleword have been working on recently! Want to check out the Doubleword inference platform? Try today at: https://app.doubleword.ai Or want access to the slides from the evening? Comment SLIDES below.

Most teams think moving to open-source models is the optimisation. For high-throughput, non-realtime workloads, that’s only step one. This chart is the same model - Kimi K2.6 - across multiple providers. The cost difference comes from how efficiently the provider serves the model at scale: batching, scheduling, utilisation, routing, queueing, and inference system design. When you’re running large async workloads - evals, synthetic data generation, background agents - those infrastructure efficiencies compound extremely quickly. Doubleword comes in ~25% below the industry average. Same model. Same intelligence. Lower token cost.