Show Notes
Every month your inference bill climbs, and you tell yourself it is the cost of doing business. What if it is actually a tax on what you do not know?
Funding stage: Series A Plus. A $5M seed in 2025 plus a $20M Series A in April 2026 led by AMD Ventures and Airbus Ventures puts Featherless firmly past the early stages, roughly $25M raised total.
In this episode, Eugene Cheah, the founder of Featherless, makes a blunt case: the best model for most of what your startup does is open source, often runs for basically peanuts, and is frequently built in China. He has put real money behind that thesis, about $25M across a seed and a Series A led by AMD Ventures and Airbus Ventures, and a platform that holds tens of thousands of open models online at once through a single API key.
The throughline is freedom. Eugene's grandmother speaks seven languages and none of them are English or Chinese, which is roughly half the planet that the closed, English-and-Chinese-first future would leave behind. Open source, he argues, is not just free as in money. It is free as in freedom: when the model runs on your terms, nobody can ever take it away from you.
He walks through why the database wars of the past, Oracle and Microsoft and IBM, then MySQL and Postgres, are replaying in AI at ten times the speed, why “lazy” models are really just a mirror of us, and why the labs chasing superintelligence may be solving the wrong problem while businesses quietly beg for one thing: reliability.
Named Frameworks
Optimize After You Scale
Collect revenue first. There is no point building an AI solution for a market that does not exist.
- →Real traffic shows you where the money actually burns, instead of you guessing in an endless optimization loop.
- →Once customers are hitting it, the natural division lines between tasks appear on their own.
Hotspot Optimization
Find the single highest-token step in your agent flow, not the whole flow.
- →Swap that one step to a specialized smaller model. One customer moved button-identification from the largest Opus-class model to a 7B and 3B vision model and ran it at about 1/100 of the cost.
- →Repeat for the next hotspot. Optimization is surgical, not wholesale.
The Sonnet-Class Swap
Most frontier leaders cluster at a Sonnet class that is, for most people, indistinguishable in practice.
- →Open models like DeepSeek, GLM, Kimi, and Qwen can drop straight into that tier.
- →The surprise: customers on Opus-class models try a Gemma 3 1B, a model small enough to run on a laptop, and find it just works.
Trusted Provider as Security Layer
A model is text in, text out. Run the weights on a trusted provider and the request cannot leave that server.
- →Run it on a provider you trust and the Chinese government does not get your data, and neither does anyone else.
- →Censorship is real on hot-button topics, so do not let these models write your policy. For generating HTML for your video game, it is a non-issue.
- →Bifurcate: route sensitive work to a US frontier model you trust, route everything else to cheaper open models, and run them in tandem.
Reliability Over Superintelligence
The labs sell curing cancer and orbital physics. Businesses want a model that will not order ten pizzas to the wrong address.
- →The bar most founders actually need is 99.9% reliability on day-to-day tasks, reliable enough to hand it a credit card with a $100 limit and stop worrying.
- →The open source ecosystem is positioned to iterate toward that bar faster than the frontier labs are chasing it.
The Intern-and-Manager Loop
Looping is just emulating human agency. One model holds a goal it can validate, another keeps working until the goal is met.
- →That is literally a manager and an intern. You can get a lazy intern, so you keep looping. It is perfectly fine for business.
- →The harder, unsolved problem is a model that learns innately from your corrections, not through RAG, the '50 First Dates' model that wakes up, reads its notes, and resets every time the context window fills.
Founder Experiment
Run a reality-check audit on your own stack. List every distinct AI task in your product. Beside each one, mark whether it truly needs your most expensive frontier model or whether a Sonnet-class or smaller open model could plausibly do it. Take your top three highest-volume tasks, run the same prompts side by side through your current model and a cheaper open alternative for one week, and log both cost and quality. Migrate only the tasks where quality holds. You are not rebuilding anything. You are just checking the bill you assumed you had to pay.
Glossary
Inference
Running a trained model to get an output. The recurring cost in production.
Token
The unit models read and write. You pay per token, which is why volume tasks add up.
Serverless inference
Running models on demand without provisioning or maintaining the GPU yourself.
Open source / open weights
Models you can download, run, and keep, with no provider able to revoke access.
Fine-tuning
Adapting a base model to a narrow use case, often turning a small model into a specialist that beats a giant.
RWKV
An attention-free, linear architecture Eugene co-created, now under the Linux Foundation, designed to lower inference cost.
Frontier model
The largest, newest, most expensive closed models.
Opus-class vs Sonnet-class
Shorthand for top-tier versus the still-excellent, cheaper tier most workloads can live in.
Model routing
An emerging category that tries to send each request to the ideal model automatically. Nobody has perfected it yet.
RAG
Retrieval augmented generation, giving a model external notes instead of true built-in memory.
Agentic loop
A model running repeatedly toward a validated goal rather than answering a single prompt.
GPU orchestration / hot-swapping
Loading and releasing models in GPU memory on demand, which is how one platform keeps tens of thousands of models online at once.
Sovereign AI
A nation or enterprise controlling its own models and infrastructure rather than renting from a few foreign providers.
Dedicated capacity
Booking your own GPU capacity for very large workloads, sometimes for a custom fine-tune.
Q&A: What Founders Ask After This Episode
How do I lower my AI inference costs as a startup?
Find the highest-token step in your workflow and move it to a smaller or specialized open model. Founders routinely cut costs 10X or more without users noticing.
Should I optimize my AI stack early or late?
Late. Get customers and revenue first, then optimize the cost centers that real traffic reveals.
Are Chinese open source models safe to use in production?
The data risk is largely averted when you run the weights on a trusted provider, since the request cannot leave that server. The real caveat is censorship on political topics, which matters if the model sets policy and does not if it writes your app's code.
How do I pick from tens of thousands of open models?
There is no perfect tool yet. Most teams find their match through the platform or through Reddit, then validate by simply trying it. The fastest win is checking which tasks can drop to a smaller model at all.
What is Featherless.ai?
A serverless inference platform giving you one API key, instant access to tens of thousands of open source models, flat pricing, and unlimited tokens, with no infrastructure to manage.
Why are models lazy, and can they learn from mistakes?
They were trained on us, so they cut corners like we do. Today, fixing a mistake means generating thousands of corrected examples for retraining. The goal is a model that learns from a handful of corrections the way an intern does.
Five Founder Questions This Episode Answers
- →How do I stop my inference bill from quietly eating my margins?
- →When in my company's life should I optimize my AI stack, before or after I have paying customers?
- →Is it actually safe, and actually smart, to run Chinese open source models in production?
- →With tens of thousands of models out there, how do I find the right one for my specific use case?
- →Should I self-host on a tower of Mac Minis, or run everything serverless for $200 a month?
URLs Mentioned in the Episode
- Featherless.aihttps://www.featherless.ai
- Eugene Cheah on Xhttps://x.com/picocreator
- Tech Talk CTO (Substack)https://techtalkcto.substack.com
- RWKV (Linux Foundation)https://wiki.rwkv.com
- Ryan Estes on LinkedInhttps://www.linkedin.com/in/estesryan/
- AI for Foundershttps://aiforfounders.co
- Inbox Alchemyhttps://inboxalchemy.co/
- AI Native Studenthttps://ainativestudent.com/
