Odd Lots · · 56 min
How an American AI Company Uses Chinese Open-Weight Models
With Daniel Saedi
ChorusBring the market's conversation to your AI. Learn more
Summary
Daniel Saedi, whose company Minerva builds AI tools for marketers, explains how American application startups use Chinese open-weight models such as Qwen and Kimi. He argues that startups compete by fine-tuning open weights on proprietary customer data, a route made cheap because hosted services like Fireworks and Modal have commoditized fine-tuning, though he sees their margins as healthy. He frames the economics around cost per completed task rather than token prices, and says companies increasingly mix frontier models from OpenAI and Anthropic during development with cheaper fine-tuned open models for repeatable work. He also discusses distillation's disputed role in Chinese model quality, the risk posed by abliterated uncensored models, the upgrade problem when new open-model versions arrive, and his view that benchmarks matter less than day-to-day performance, with compute demand still likely to outstrip supply.
Key takeaways
- Daniel Saedi on AI pricing & unit economics: For repeatable tasks, cost per completed task matters more than headline cost per million tokens.
- Daniel Saedi on Open-source AI models: Fine-tuning open-weight models on customers' proprietary data is the main advantage for application-layer AI companies.
- Daniel Saedi on AI pricing & unit economics: Fine-tuning economics have collapsed, with some fine-tuned models costing as little as $90 per final run.
- Daniel Saedi (bullish on Fireworks AI): Fireworks is making a lot of money at about $1B ARR, with likely strong unit economics despite commoditization.
- Daniel Saedi (bullish on Modal Labs): Modal is making a lot of money at about $1B ARR, with likely strong unit economics despite commoditization.
What was said, by company and theme
16 views from 1 speaker. Every quote is checked against the episode's transcript (how).
Fireworks AI
Across the shows: 2 speakers in other episodes, 2 bullish and 1 bearish calls, past 30 days
Where everyone stands on Fireworks AI
Daniel Saedi · MinervaBullish
Fireworks is making a lot of money at about $1B ARR, with likely strong unit economics despite commoditization.
Fireworks and Modal sell the infrastructure that makes fine-tuning and hosting open-weight models easy. Saedi says both are at about $1 billion in annual recurring revenue. He thinks their unit economics are probably quite good. The fact that fine-tuning has become commoditized has not driven their margins to zero, because the ease they offer is worth a lot to customers. Minerva uses Fireworks to host and fine-tune its models.
“fireworks and modal make quite a boatload of money actually in making it easy for folks. Both of them are sitting at a billion dollars in ARR, and I imagine their unit economics are actually quite good.”at 23:28
Modal Labs
Across the shows: 2 speakers in other episodes, 3 bullish and 0 bearish calls, past 30 days
Daniel Saedi · MinervaBullish
Modal is making a lot of money at about $1B ARR, with likely strong unit economics despite commoditization.
Modal, like Fireworks, sells the hosting and fine-tuning infrastructure that makes open-weight models easy to use. Saedi says both are at about $1 billion in annual recurring revenue and that their unit economics are probably quite good. Commoditizing the underlying fine-tuning has not driven their margins to zero, because the ease they provide is what customers pay for.
“fireworks and modal make quite a boatload of money actually in making it easy for folks. Both of them are sitting at a billion dollars in ARR, and I imagine their unit economics are actually quite good.”at 23:28
Open-source AI models
Where everyone stands on Open-source AI models
Daniel Saedi · Minerva
Fine-tuning open-weight models on customers' proprietary data is the main advantage for application-layer AI companies.
Startups lack the billions of dollars of training data the frontier labs buy, so they compete by fine-tuning open weights on the data their customers allow them to use. Minerva fine-tuned Qwen on a few thousand labeled examples of complex audience queries. The result was a model faster than the closed models, and better at the customers' tasks. Out of the box, Qwen was roughly as good as an OpenAI model but slower. Saedi says Harvey and Cognition are doing the same with Kimi, using their distinctive data to get better results at lower cost, which lets them charge a margin. He calls this the primary benefit of open-weight models for American application companies right now.
“I actually think that this is the primary benefit to the application layer right now is the ability to fine tune on top of those open”at 10:16
Daniel Saedi · Minerva
Open-weight models are often cheaper than paying OpenAI and Anthropic, though not always, so cost comes first.
Cost is the top priority when a business chooses open-weight models. The question is whether the upfront costs of hosting and fine-tuning plus the operating costs come in below paying OpenAI and Anthropic. In many cases the answer is yes, and in some cases it is no. Saedi says this comparison is the first thing a company has to calculate.
“I think the most important thing is cost. And what you have to do for as a business is essentially do the calculation of I have to pay some upfront costs and then some operating costs. And is that cheaper than paying OpenAI and Anthropic? And in many cases, it's true.”at 13:23
Daniel Saedi · Minerva
Fine-tunes do not carry over to new open-model versions, so upgrading adds ongoing cost and uncertainty.
When a new version of an open model comes out, such as a Kimi update, a fine-tune cannot simply be transferred to it. Upgrading to a new version is part of the ongoing operating cost of fine-tuning, and it is the biggest decision point for founders right now. The same dataset can be reused, and a reliable, expertly graded set of answers is transferable. But Saedi does not know whether retraining on the same data converges to the same optimal weights, which adds uncertainty to the decision.
“is part of the ongoing operational cost of fine tuning a model. It doesn't simply transfer over.”at 18:47
Daniel Saedi · Minerva
Fine-tuning data and evals are not yet commoditized and separate good fine-tunes from bad ones.
Fine-tuning tools have become commoditized, and the margins for hosts like Fireworks and Modal have not gone to zero. What is not commoditized is the data and the evaluations behind a fine-tune. That is why the labs pay data firms like Mercor and Surge billions. For Minerva, the key decision is whether it has enough good, validated data for the model to learn from, or whether it would only learn noise. Marketing has plentiful data, from call logs, ad data and website activity, which can be built into simulated environments for training.
“The things that aren't commodified yet and what actually separates a good fine tune from a bad fine tune is the data and the actual evals that are put in there.”at 24:00
Daniel Saedi · Minerva
Chinese open models' parity with closed frontier models will be judged by swapping them into real workflows, not benchmarks.
Saedi argues that as companies care more about sovereign data and about the tasks they run internally, benchmarks are becoming meaningless. Benchmark-maxing means a higher terminal-bench score for Kimi 3.2 over a closed model says little. What matters is whether a company can swap Kimi into an existing chatbot or coding workflow and see similar performance day to day. Once it replaces the daily driver for coding without the user getting frustrated, the open model has caught up. He notes that everyday users struggle to tell GPT-5.1 from GPT-5, so they will struggle more to tell GPT-5 from Kimi 3.1.
“The benchmarks are are starting to become meaningless as you guys said... if I can take Kimi 3.2 and start replacing it as my daily driver and how I'm coding, okay, now I know that things have caught up.”at 38:45
AI pricing & unit economics
Where everyone stands on AI pricing & unit economics
Daniel Saedi · Minerva
For repeatable tasks, cost per completed task matters more than headline cost per million tokens.
Once a product moves from building an AI workflow to running it repeatedly, the economics change. Companies then care about cost per completed task, which is different from the cost per million tokens that the media focuses on. The goal is to reach the minimum acceptable task completion rate at the lowest cost. Token prices are purely an input-output measure and don't capture whether the job gets done. Saedi says that for marketers' repeatable tasks, the gap between that minimum and the cost is where the margin comes from.
“you care about cost per completed task, which is a lot different than, you know, looking in the media where everyone cares about cost per million tokens. It's purely an input output thing.”at 7:05
Daniel Saedi · Minerva
Owning a homegrown open model lets smaller firms push frontier labs toward value-based pricing.
Even a company that does not sell its open model to customers can use it as leverage. If a company can tell Anthropic or OpenAI that it has a homegrown model, the labs have to engage in value-based pricing rather than charging for simple automation. Saedi says this returns leverage to smaller companies, which they lacked early in the AI boom. The frontier models are now very good automation, or potentially much better, so the labs can charge on the value they deliver.
“the fact that you can go to Anthropic and OpenAI and say, hey. I have this homegrown model. You have to now engage in value based pricing for me”at 14:45
Daniel Saedi · Minerva
Fine-tuning economics have collapsed, with some fine-tuned models costing as little as $90 per final run.
Fine-tuning used to need a team of inference engineers. Now it is commoditized, and hosted fine-tuning services such as Fireworks charge around $7 to $9 per million input and output tokens, where output means the changes to the model weights. Saedi cites people who have fine-tuned models that perform well for as little as $90 on the final training run. He contrasts this with Harvey's reported roughly $6 million in compute to reach the frontier on legal tasks. The gap is several orders of magnitude, so the economics have collapsed.
“there are people out there who have have fine tuned models that perform really well for as low as $90 on their final training run. So the economics have collapsed.”at 21:59
Daniel Saedi · Minerva
AI apps must cut variable per-token costs with fine-tuned or cheap open models to protect margins on fixed-price seats.
Harvey charges a fixed per-seat price but its costs rise with usage as the product improves, so it has a fixed revenue line and variable cost. Harvey therefore went to a legal fine-tuned model, Tenet, to get better leverage from the labs and reduce variable spend. Legora, which charges by usage, is also exploring fine-tuning. Saedi says his own company asks the same question: where the highest variable, non-usage-tied spend sits and how to replace it with a fine-tuned or cheap open model. For highly repetitive simple tasks they may just use a very cheap open-source model. The economics reduce to per-token cost paid to the labs times task count, which must exceed the fixed cost of training plus monthly running costs.
“Harvey has a per seat cost, which is fixed, and Legora charges usage... they have fixed revenue coming in... you want to reduce that variable cost as much as you can.”at 40:07
Enterprise AI adoption & AI software
Where everyone stands on Enterprise AI adoption & AI software
Daniel Saedi · Minerva
Companies use frontier models to build products, then switch to cheaper open models for repeatable tasks.
During development, Minerva uses frontier models such as Claude and the other frontier models because they save time and money, and because developers should not spend time on loops and mistakes. When a task becomes repeatable and the goal is a result that customers are happy with, the economics change. Then the company moves the task to a cheaper model, such as a fine-tuned open model, that meets the minimum acceptable completion rate at the lowest cost. Saedi describes this as mixing and matching models for each use case.
“we're still using the Frontier because you want to save as much time. You wanna save as much money. You honestly don't want your developers to be spending time dealing with with loops and with mistakes.”at 6:37
AI regulation & safety
Where everyone stands on AI regulation & safety
Daniel Saedi · Minerva
Abliterated open-weight frontier models are the biggest risk, since cheap uncensored copies can be run to hack at scale.
Someone turned off the neurons that make the uncensored DeepSeek model refuse requests, which is called obliteration. Saedi says the model is probably not as capable as Anthropic's Mythos from February, but it is close, and it costs about 30 cents per million blended tokens. That is roughly ten times cheaper than Sonnet, and about one-hundredth the cost of Fable. The risk is that a swarm of these models can be told to hack anything. The labs' real concern is not data leaking to China or model vulnerabilities, but that the model's safeguards can be removed. Saedi has no compelling policy yet. He floats requiring know-your-customer checks before weights can be downloaded, while preserving the ability of US startups to download and fine-tune.
“that's really the biggest risk, I think, personally, to having open weight frontier models out in the wild.”at 28:04
US–China AI & chip race
Where everyone stands on US–China AI & chip race
Daniel Saedi · Minerva
Distillation is real but not the main reason Chinese open models are strong; their research and techniques matter more.
Reflection AI is a US open-weight company backed by NVIDIA, with thousands of Blackwell GPUs and a cluster of roughly 12,000 to 14,000. Its flagship model was released about a month ago and sits about six to nine months behind the frontier. The Chinese open models were about three months behind the public frontier at the time, and Saedi argues they would be neck and neck with Reflection if they had played by the book. So distillation is not the pure reason for their position. Chinese labs have published novel techniques, such as sparse models and their inference framework, which the frontier labs also use. Distillation is real, but how much it drives quality is up for debate.
“And so what that basically tells you is that distillation is not the pure reason that the Chinese labs are where they are.”at 33:07
Frontier AI labs & model race
Where everyone stands on Frontier AI labs & model race
Daniel Saedi · Minerva
Frontier model differences are now too small for everyday users to notice, even between successive versions.
Saedi says that models have become night and day better than early ChatGPT, but they still make silly mistakes and produce slop. For everyday use, even he finds it hard to tell the difference between GPT-5.1 and GPT-5. If he cannot tell those apart, he questions how anyone could tell GPT-5 from Kimi 3.1 in ordinary use. The implication is that frontier improvements are hard to perceive in practice, which makes the case for cheaper open models that perform adequately on specific tasks.
“For everyday use, even for me, it's like it's really hard for me even to tell the difference between a 5.1 and a five.”at 38:01
AI infrastructure & data centers
Where everyone stands on AI infrastructure & data centers
Daniel Saedi · Minerva
The world will remain massively short of compute, with about one B300 per knowledge worker in four years.
Saedi uses a first-principles heuristic: an AI-native employee should consume roughly one GPU's worth of tokens per year. The US has roughly 5 to 10 million leading-edge Blackwell GPUs deployed and about 100 million knowledge workers. If every knowledge worker needs a GPU running an agent at all times, the economy is massively short of compute. He expects compute to net out at about one B300 per knowledge worker in four years, which he still sees as a shortfall. Short-term swings in the narrative are noise, since the loudest compute-short voices on Twitter are often the most short-compute people. He also cites the Musk and Altman and Brockman compute dispute as evidence that labs are still fundamentally short.
“Roughly, that's kind of the heuristic where I think things are gonna net out four years from now... we have kind of 1 B300 worth of computational capacity for every knowledge working employee in the economy.”at 45:36
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
- Fireworks AI2 speakers · 2 bullish, 1 bearish calls
- Modal Labs2 speakers · 3 bullish, 0 bearish calls
Topics
- AI inference hosting margins
- AI margins and model cost
- GPU compute supply and demand
- abliterated open-weight models
- cost per completed task
- distillation's role in Chinese model quality
- fine-tune durability across model versions
- fine-tuning costs
- fine-tuning data and evals
- fine-tuning open weights on proprietary data
- frontier model differentiation
- mixing frontier and open models
- open-weight cost versus closed APIs
- open-weight model performance
- value-based pricing leverage
More from Odd Lots
- This Is The Problem With AI's Most Bullish Chart · November 19, 2026
- How GTA VI Became the Most Expensive Video Game Ever Made · November 18, 2026
- The Impossible Trilemma That Hangs Over the Nvidia Crisis · November 16, 2026
- How Wall Street Is Bracing For the Post-Jensen Huang Nvidia Era · November 14, 2026
- Why a $2 Billion Law Firm Suddenly Collapsed · November 13, 2026
Ask your AI about this episode
Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.
Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus