Your favorite frontier generative media team just shipped! @ComfyUI Great job, @yoland_yan @robinjhuang and Comfy team

.@ayaanmomin has never met a problem he didn’t want to investigate. He’s worked on antibiotic-resistant bacteria and surgical robots. When GPT-3 landed and everyone else was asking it to write poems, he was experimenting with an early version of agentic coding. That led to another question: if there were going to be many models, how would you know which one to use? So he co-founded Neutrino to route requests across them. Then the problem moved down another layer: how do you actually serve these models efficiently at scale? That took him to @FireworksAI_HQ, where he has done a little bit of everything over the past two years. Along the way, he developed the kind of familiarity with chips, infrastructure, and compute that means he can probably tell you more about AMD GPUs than you were planning to ask. Ayaan is now joining @chemistry as an investor. What ties all of this together is an instinct to keep going deeper. Ayaan doesn’t tend to stop at how something works; he wants to understand what makes it possible in the first place. That perspective feels particularly relevant right now. As AI becomes more interactive, multimodal, and embedded in the physical world, some of the most important questions are moving below the application layer: how compute is priced, which silicon is available, where the power comes from, and how all of those constraints shape what gets built next. His interests extend well beyond technology. He has strong opinions on Goya’s Black Paintings and disappears into national parks for days at a time despite being very afraid of bears (but not mountain lions). Welcome to Chemistry, Ayaan! @Mark_Goldberg_, @kshenster, @loubohan, and I can't wait to work with you!

One thing I omitted in talking about the disappearing middle ground for startups: founders must get the timing right in terms of momentum and theme... in other words, are you working on the 'right' problem at each stage? As the cycle time of each theme shrinks (more on this in a future post), predicting this precisely is becoming a hard moving target. @WSJVC covers this a bit this morning: wsj.com/pro/venture-capital/…

Pink elephants in the @WSJVC this morning. Achievement unlocked! wsj.com/pro/venture-capital/… h/t @mstrozier

I first met @gpuemi right after YC demo day – in the amazonian expanse of Plant Cafe Dogpatch. Over a pesto tofu scramble, I could see that Emi was clearly smart, talented, passionate, and driven – but it was clear his product - an inference optimization platform for the big inference companies - just wasn’t working. I could see the tension between belief and realism even as he pitched me on his original premise, and the breakout push that was just around the corner. My honest reaction was: I hope he works on something else. Turns out a breakout was around the corner. Emi was working on something else, deftly pivoting in April of this year to a new, more expansive and adjacent vision - making inference more accessible to the masses. In so doing, he leaned into the company’s competitive advantage - moving into an exciting and demand-laden, but yes highly competitive, market that many would have found way too daunting to tackle. Emi would need that passion and drive – which was blatantly obvious in that first meeting – in order to battle it out with much bigger incumbents in the wild west of a newly emerging market. One name change later, @wafer_ai was reborn as a real-time inference cloud – taking advantage of their agents that continuously tune their inference engine, custom kernels, and hardware to each company’s exact traffic, model and SLA – in the end delivering the best performance per dollar in the market (at 30 to 50% more tokens per dollar) and climbing to #1 on the Artificial Analysis leaderboard (artificialanalysis.ai/provid…). In the process, they became the first provider to make AMD GPUs competitive with NVIDIA’s – at a materially lower cost and shorter lead time. When Wafer posted their performance numbers, customers took notice; an engineer recently texted asking for eight figures worth of throughput “by the end of the week,” not a random ‘lead’ in Wafer’s funnel - but a soon-to-be live customer. Wafer already powers Vercel, Venice, Flip, Neon Health, Brilliant, and dozens of other fast-growing AI companies, servicing 2T+ tokens a month. And it’s not unusual for me to get a random text from a CEO or CTO sharing just how Wafer’s performance gains are speeding up the underlying performance of their products. So Emi and I kept meeting up and I'm thrilled that @chemistry is co-leading their $40M Series A alongside @MarathonMP, @ycombinator and many angels. Congrats @gpuemi, @gpusteve, and team. @kshenster, @Mark_Goldberg_, @loubohan and I are excited to be on team Wafer for the road ahead.

The venture market is increasingly bifurcating into two camps: massive consensus bets and contrarian companies as far from the heat as the dark side of the moon. There will be gems - and false gold - in both.