Last week, I shared why @20vcFund invested $10m in Fireworks in 24h: Lessons from 40 reference calls, and why we believe they will be a >$100bn company 👇 First - I have never seen that level of advocacy across so many datapoints. 35 out of 40 were strong positives, and two things stood out: 1) *the customers do not talk like users of a commodity* - it matters, because it is the first pushback we tend to get here. Instead, they literally talk as if they found an unfair advantage, and do not want to give it back. So what about the specifics? - Our friends at Figma benchmarked it at 3x faster than Baseten in production, and now run their new AI "Assistant" on a small in-house model, on Fireworks, at >5x lower cost than a frontier model - The team at Uber ran its own benchmark, and said it plainly: Fireworks "crushes it" compared to anyone else. Specifically, they get to sub-100ms time to first token, and they are ~60% cheaper than Google - Engineer at HuggingFace we spoke to said that they were world leading on speed, latency and throughput - ElevenLabs scored them "at least a 9/10" and thought the same we did on the team: "they are the core PyTorch team, that is an unfair advantage" and "they are the best at fine tuning" - A large bank managed to get higher throughput than NVIDIA's own TensorRT, and watched Fireworks rewrite a kernel to support a brand-new model overnight I can go on for ages, but the theme is not "cheap tokens". It is instead: speed, reliability and customisation, that these teams cannot easily reproduce themselves. That is where @FireworksAI_HQ absolutely crushes it (and the competition at the same time). 2) *The part that's actually even more convincing*: the competition agrees. Here is what we discovered there: - a former head of marketing at one of their closest competitor told us any team looking at doing their best inference work should look at partnering with someone like Fireworks - an ex-engineer at another rival platform called them the leader in the space, and the company with most momentum. He specifically highlighted that they were converting developer adoption into a real, enterprise-scale business, as opposed to just running AI pilots - A former PM at the largest chip company in the world also admitted that they had a superior user experience and a better TCO (total cost of ownership) Those concrete examples (and there were many more) are why we see Fireworks as the un-disputed leader in the space. So while the pushback is real (open-source engines like vLLM and SGLang keep improving, and the hyperscalers would love to swallow this layer), there are two things we see: (a) across these calls, engineers still put Fireworks "several months ahead" of open source and rivals, so the gap is not closing as fast as the some would claims - if at all -, and (b) Fireworks keeps climbing up the stack, into fine-tuning, reinforcement learning and evals. I believe the latter is exactly where value moves as AI goes agentic. This leads me to the bigger idea I already shared: in AI, the model are the ones getting the limelights. But in the end, the layer that makes every open model fast, cheap and usable is where a lot of the durable value will sit. So to re-iterate: Invisible to users. Indispensable to developers. That is the bet. Thoughts @HarryStebbings @jaminball @GavinSBaker @EverettRandle @sahirazam @ericvishria?