Jun Zhou

➡️ We just dropped our whitepaper: The Monetization Operating Model. If you’re building in a world where pricing is product, billing is infrastructure, and customers expect real-time everything, I promise this was written with you in mind. Quick taste: – Don’t get fooled: Usage-based pricing is an operating model in pricing model’s clothing – Billing is now system design – Every pricing decision has become a product decision This model is based on what we’ve seen work across some of the fastest-moving teams in AI, infra, and PLG. Here's the link if you want to dig in (you should): https://bit.ly/4mfovHi

141

17 Comments

You can now train SOTA models without any storage! We completely revamped the Hub’s backend to enable streaming at scale. We streamed TBs of data to 100s of H100s to train SOTA VLMs and saw serious speed-ups. How? We rebuilt the backend for massive concurrency. ⚡️ Persistent Data Files Cache: Only the first worker resolves the data. No more traffic jam of requests. 🏎️ Parquet Prefetching: We fetch data while the GPU is busy, eliminating I/O bottlenecks. Andrés Marafioti, Quentin Lhoest and team are already using this to train our next-gen models with nanoVLM. Streaming directly from the Hub is now as fast as our cluster's local SSDs, but without the 3-hour wait to download and prep the data. We wrote a blog with more details, but the main takeaway is: pip install --upgrade datasets huggingface_hub And then you can stream datasets from the hub super fast! Have fun!

365

21 Comments

Meta’s Llama gets a home — and a strategy shift Meta just tapped Groq and Cerebras Systems (both private!) to host its open-source LLMs — but the real story might be Meta’s first real move into cloud AI. • The news: Meta partners with Groq + Cerebras to host Llama models, bypassing AWS, GCP, and Azure. • The strategy: Start with AI inference, add storage next? Meta might be quietly building an AWS competitor. • The play: Offer Llama hosting at-cost (or less) to win dev mindshare + undercut OpenAI. • Why it matters: Open-source + cheap inference = existential threat to OpenAI’s pricing model. • Trend watch: AI compute costs could race to zero — and Meta just joined the race.

31

1 Comment

I love the way Ozgun puts it: Ubicloud is an open-source alternative to AWS, at a fraction of the cost How do they dare take on these cloud giants? Özgun shares the backgrounds of the founding team (Daniel and Umur) and what it really takes to go head to head with big tech   More videos from e2vc Summit!

64

1 Comment

There’s been a lot of discussions on the best way to build agents:  OpenAI's general take is that increased model capabilities simplify the SDK  Others (incl. Anthropic agent pattern guide + dexhorthy's 12 factor guide) generally outline a more constrained approach. We’ve been asked by a bunch of people on where we stand. High-level we agree with both: ✅ We agree with OpenAI that a lot of existing agent approaches are a bit cumbersome/overengineered to build agents. This is why we actually moved from a declarative graph-based approach in LlamaIndex (our old QueryPipelines) to workflows as an event-driven pattern. Also as models get better inevitably you will trust general agent reasoning more. Not everyone wants or needs to build agents from scratch. ✅ At the same time there’s a reason these agent design patterns exist. They are best practices for a good swath of use cases where you need to encode a specific business process. I personally really like Dex's 12 factor guide - it’s a collection of best practices that ensure you actually deeply understand the agent you’re building. As a framework, LlamaIndex workflows supports the entire spectrum, from constrained/from-scratch to general reasoning. The really nice thing about it is it doesn’t feel that different from just writing normal Python/Typescript, so you don’t have to learn a completely new design pattern, which makes it way more seamless to transition from regular programming. Huge shoutout to Laurie Voss for the blog Blog: https://lnkd.in/dVtv7Mgq For those getting started check out our workflows guide: https://lnkd.in/gRsUHuV8 12 factor guide coming in soon! 

254

14 Comments

The real reckoning is the death knell of the 'non-venturing' VC. Most VC had turned into the 'first check' for ramen salary + first AWS check beyond the $10k credits. After that, opportunistic 'second check' on traction. If the gap between customer need and solution narrows significantly, thanks to GenAI, then that 'first check' isn't required. Entrepreneurs can keep tinkering, talking to customers, evolving that first solution (not MVP) and bank revenue once product/customer fit happens. If the 'first check' isn't required, the ability to write the 'second check' never appears. Thus, VCs end up having to take actual chances on _venture_ - i.e. funding crazy ideas that won't have revenue potential for years. If there is revenue potential, they won't be able to get in the startups early enough for it to matter. Banks will 'enter the chat' as they say in a big way since they are the best providers of capital when the revenue trajectory is clear and business de-risked. VCs that only invest in 'de-risked' startups weren't true VCs anyway. They will not be missed.

35

11 Comments

Martin Eberhard played a key role early at Tesla and even brought in Elon Musk. Then he was pushed out of the company for pennies. Founders get most of the attention in Silicon Valley, but the ecosystem is built on the people who take a chance and join early stage ventures. Unfortunately, those same people are often the easiest to cut out. Founders can dilute early employees, change vesting terms, and restructure equity with little effort and zero consequences (you can literally do it in Carta in 5 minutes). Daniel Kottke was a close friend of Steve Jobs and helped build the Apple I and II computers. He got no equity (Steve Wozniak eventually gave him a few of his own shares). Brian Acton co-founded WhatsApp and stayed through the Facebook acquisition. When he left Facebook over privacy concerns, he had to walk away from $850 million in unvested stock. There’s more awareness now around how vesting works, what happens during a liquidity event, and how to think about re-ups and dilution. That’s good. But the truth is, if someone really wants to screw you, they can. It still comes down to one thing: trust the people you work with. Or don’t work with them at all.

84

8 Comments

NEXTGEN INFRA STARTUPS MUST PLAY OFFENSE Historically, software infrastructure / middleware startups have played DEFENSE well. They have built deep, defensible technology moats over a period of time. The founders have been deeply technical, have sold to other developers, and have been content to not have a strong public-facing brand. However, this playbook (and archetype) must change. Infra founders and startups must play OFFENSE and become more dynamic and front footed. Why? First, foundation models are adding more and more infastructural components and nibbling away at what was traditionally considered infrastructure. No company can now comfortably sit and own a part of the stack. It will inevitably erode. Second, end customers are deciding or opining on what infrastructure to use. Guess what? Customers of infra products are no longer just career engineers but ordinary people (including VCs!) Infrastructure now has a PLG motion. Having a customer-facing brand is no longer optional. Finally, it’s possible (or even probable) that your company might need to fundamentally change its business model, and not be an infrastructure company any more. And in this crucible moment, you need to have the muscle and the ability to seize the opportunity and go for it. Consider OpenAI launching ChatGPT or Vercel launching v0. They could never have managed it with a legacy infrastructure mindset. Check out Groq’s homepage - it looks more like a vibecoding startup than a chip company. To summarize, infrastructure startups need to evolve. They must start thinking about the end customers, become much more public-facing with their brand, and develop the muscle to build customer-facing (not just developer facing) products. The archetype of infrastructure founders is changing. They cannot just be technical. That’s tablestakes. They must build a deep understanding of their customers’ customers, be superb at product development, marketing, storytelling and brand building, and be ready to move up the stack at any time. In other words, they must be able to play OFFENSE. That’s what greatness means in infra companies today.

146

8 Comments

REALTOR TO DEV: Learned Scratch as a kid and years later 19-yr old Chef 🧑‍🍳 Gallmeyer https://lnkd.in/d7NT6zuN is both a top realtor and now a digital product maker. But what's his secret recipe? 1. Chef Gallmeyer's deep in domain knowledge. He not only earned one of the top scores in Island County on his realtor exams, but he's forever been an apprentice to his realtor uncle. 2. Chef Gallmeyer knows how to speak machine. He's a computational thinker who works incrementally towards solutions and recognizes that engineering challenges can't get solved all at once. 3. Chef Gallmeyer is disciplined. When you're a lot younger (than me), your ability to both maximize and waste time is at an all-time high. Brandon uses his time most effectively the way that startup founders work. --- Realtor to Developer 🎂 Episode: https://lnkd.in/d7NT6zuN All 50 Cozy AI 🍰 Episodes: https://lnkd.in/g6upvbGX How To Speak Machine (2025) 📘 MIT Press: https://lnkd.in/gC9DGPb8

50

1 Comment

The big labs are expansive in their product ambition, especially since foundation models have largely improved in lockstep - in order to compete with them you have to do things they won’t which are: - building a very rich software ecosystem around a primitive - orchestration across multiple models - going insanely deep on product and growth for a narrow vertical domain

58

3 Comments

This section on how dbt Labs transitioned from a purely PLG motion to layering on enterprise is a fascinating one. Two instructive passages that I have bolded (below from First Round Review's path to PMF series featuring dbt Labs) TLDR: GTM comprises your ICP, channel, and message. When you transition from PLG / bottom up to Enterprise / top down motion, naturally your ICP and channel also changes, but your messaging / proposition needs to change too, e.g., the enterprise may be multiple personas and are buying assurance as much as solution. --- "Handy found product-market fit organically for dbt as an open-source tool mostly used by data practitioners and developers. But a few years into running a commercial business, he realized he had to build a growth curve all over again with C-suite data leaders. “Even though we had an unbelievable amount of market pull, as we initially commercialized, it wasn’t easy for us to transform this open-source command line tool into a product that enterprises would pay a million dollars for,” says Handy. “When you have enough product-market fit, sometimes it allows you to get away with not being super tight on product marketing or sales motions. So around 2022, we went from this gigantic acceleration curve and overnight we realized, we have to sit at the adults’ table and figure things out real fast,” says Handy. After the PLG flame started to fizzle, Handy turned his attention to layering on a sales-led motion for the cloud platform. “We had to focus our efforts on telling cohesive stories to senior data leaders. We had to have a very clear, explainable answer to the question, ‘Why should I use the commercial product and not the open-source product? And it had to be digestible by someone with a C in their title,” he says. Handy’s answer: dbt Cloud can handle complex data for companies of every size. “The longer people used dbt, the more complex their code became,” he says. “It was a problem for the most sophisticated dbt users, who were often at the largest companies. So there was a real opportunity for us to step in and solve that for them with dbt Cloud.” To tell that story to enterprise customers, Handy relied on data, naturally. “At a user conference we presented a chart that showed the number of dbt projects that had a certain number of models in them — over 100, over 1,000, et cetera,” he says. “We watched that number climb and we knew as users ourselves, ‘Oh my God, trying to work in a dbt project with 5,000 models in it is challenging.’ So we started with that quantitative data point and asked folks in our community about their experiences with these very large, complex dbt projects, and validated that this was a pain in the ass without a cloud platform.”

23

4 Comments

Requesty announced their $3M Seed round to build the gateway layer for the multi-model AI era. Every team working with LLMs in AI today wrestles with shifting APIs, model failures, ballooning costs, and governance gaps. Requesty fixes this in one line of code — routing across providers, standardizing access, and giving enterprises the control plane they’ve been missing. 25k developers have already taken note. We are bullish to be building with Thibault, Daniel and team; alongside 20VC, Tiny Supercomputer Investment Company, Michele A. and more.

58

4 Comments

Confluent for Startups is proud to support AI Accelerator cohort member Flexprice as they launch publicly on Product Hunt today! 🚀 Flexprice is a developer-first billing platform that helps companies roll out usage-based, subscription, or credit-based pricing models without the overhead of building billing systems in-house. It's an open-source platform built to give product and engineering teams more control over how they monetize their AI products. By leveraging the Confluent Data Streaming Platform, Flexprice powers mission-critical infrastructure for real-time customer usage tracking, live credit balance monitoring, and reliable webhook delivery.  You can check them out today at https://flexprice.io ✅ Check out their Product Hunt launch here and share your feedback: https://lnkd.in/gg-dE8nX

63

7 Comments