Skip to content
Black Friday

Lock in 40% off every renewal, not just year oneBlack Friday · 40% off every renewalBlack Friday · 40% off

Ends Dec 6Get 40% off →

An American AI Startup Runs on Chinese Models. Here's Its Playbook.

An American startup builds with Claude and serves customers on fine-tuned Qwen: the costs, data control, and obliterated-model risk behind open weights.

Guide8 min read

A split workbench: frontier models for building, open weights for running

Thirty cents per million tokens — about one hundredth the cost of the frontier model it competes with — and you can tell it to do anything. That is the price Daniel Saedi quotes for an obliterated DeepSeek: the refusal neurons switched off, PhD-level hacking ability on tap. His own company, Minerva, does something far more ordinary with the same family of models. Minerva builds with Claude and Astra, then serves customers on a fine-tuned Qwen. The Odd Lots question, from Tracy Alloway and Joe Weisenthal: as Washington debates forcing US companies to disclose their use of Chinese models, what are those companies actually doing with them?

The question since the DeepSeek scare

The framing is the January 2025 DeepSeek moment: a cheaper Chinese model looking nearly as good as the cutting edge, and one core anxiety — will frontier or near-frontier intelligence be commoditized, so nobody can charge a premium for it? Tracy puts the closed-lab dilemma plainly: Anthropic and OpenAI must justify enormous CapEx and charge premium prices, while Chinese open models keep posting near-frontier benchmark scores. Joe lists the standard objections — distilled from Western models, bench-maxed for tests, possibly trained on Nvidia chips routed through third countries — and redirects to the practitioner. Startups like price competition and downward pressure; the interesting interview is with someone holding the whole menu — OpenAI, Anthropic, and the Chinese competitors — and choosing per task.

Saedi fits: five and a half years at Bridgewater, first raising money for Pure Alpha, then overseeing trading in banks, REITs, and insurers, now co-founder of Minerva, an AI company for marketing. Its agents automate everything from pre-campaign planning to execution — a marketer asks for buyers, the system launches and monitors the campaign — and the stack mixes open and closed models by design.

Open-weight is not open-source

The distinction Saedi draws is the episode's most useful definition. Open-source would mean disclosing all code, all data, the exact training recipe — and that is not what the Chinese labs do. What they release are the weights: the end artifact. An API call to a closed lab returns an answer with no account of how it was produced; downloaded weights can be hosted, inspected, and above all fine-tuned. The papers tell you roughly what happened, not how slop plus architecture became the artifact — but self-hosting opens a different set of possibilities, starting with control of your own data.

Joe presses the economics question he is unsure about: is fine-tuning on local data actually valuable, or folklore? Saedi calls it the primary benefit of open weights for the application layer right now. His example is concrete. A Minerva user asks for an audience — everyone who works at OpenAI and likes football. The old path ran an expensive closed model to parse the request into SQL and hit the database: slow, pricey. Plain Qwen out of the box roughly matched the closed model but lagged the frontier. Then Minerva graded a few thousand hard query-to-result pairs and fine-tuned Qwen on them. Now the model is faster than the closed ones on this task and better at producing what customers actually want — what Saedi calls a small durable advantage. Harvey and Cognition run the same play at larger scale: proprietary data plus distribution, fine-tuned to match the frontier where it pays.

CapEx versus OpEx

Tracy asks him to rank the advantages: cost, data control, vendor leverage. Cost is first, and the rule is arithmetic — upfront plus operating cost versus the OpenAI and Anthropic bill. Often the fine-tuned stack wins; sometimes it does not. Second is sovereignty. Saedi expects this to be huge in 2027: companies send data out and get answers back, while the model never improves on their use case. Running your own weights means repeated runs can train the model back. He is careful on the compliance point — no evidence, he says, that labs train on corporate client data in violation of their claims — but the asymmetry stands: your usage does not compound for you. Third is leverage. Even a Harvey that never shows customers the house model can walk into pricing talks and force value-based terms, because the alternative is no longer nothing-or-automation but good automation versus very good automation. My read: this is the quiet power shift in the episode — fine-tuning hands negotiating power to small companies for the first time in the AI boom.

What it actually costs to run your own

Joe's skepticism is healthy — open-source used to mean free, and there is no free AI; somebody pays for compute and electricity. Saedi's answer is that the hosting layer collapsed in the last year. A year ago, serving a fine-tuned model meant GPU concurrency planning, inference servers like SGLang or vLLM, a team of inference engineers — the reason only research groups and giants like Uber attempted it even two years into Llama. Now Fireworks, Modal, and Thinking Machines sell it as a commodity: hand over the training set, pay around $7 to $9 per million tokens for the weight changes, and people have landed good final training runs for as low as $90. Harvey's tenant model reportedly spent about $6 million in compute to reach the frontier of legal tasks — two to three orders of magnitude below a flagship training run. Both Fireworks and Modal sit near a billion dollars in ARR, Saedi notes, with good unit economics: commoditized serving still mints money.

What is not commoditized is data and evals. Labs pay billions for them — Mercor, Surge — because the alpha is in structuring validation, not in the weights. Minerva picked marketing precisely because the data is plentiful: call logs, ad data, creatives, site behavior, fifteen to twenty years of programmatic exhaust. The training target is an RL environment — a simulation of acquiring a customer, ad shown to click to purchase — and the scarce ingredient is showing-your-work traces, not just right answers. Where data is thin, like construction blueprints plus every architect-contractor-client comment that changed them, fine-tuning stalls. The market price of good company data tells the story: Micro One pays $250,000 to $500,000 per complete dataset, plus $50,000 to $100,000 per referral. Google bought Spirit Airlines for $10 million, Joe jokes, and the only thing worth saving was the data. My read: every zombie firm should price its inbox before its furniture.

The obliterated part

Then the episode turns to what Saedi calls the biggest risk of open weights in the wild. He tried the newest DeepSeek in uncensored form: someone switched off the neurons that tell the model to say no. At roughly 30 cents per million blended tokens — 10x cheaper than Sonnet, about a hundredth of Fable — it runs PhD-level capability with no refusal left. The internet theory, he says, is that current hack swarms are obliterated DeepSeeks. The labs' real worry is not your data flowing back to China — he quells that notion directly — but that anyone can flip a few parameters and field a hacker army. His policy sketch is deliberately unresolved: preserve the dynamism that lets startups download and fine-tune, while denying the same download-and-edit path to attackers — possibly KYC before weights. Joe's coda is bleak and specific: Google made bomb-making and suicide instructions nearly unfindable, a real firewall around dangerous knowledge; obliterated models may not respect it, and they are not easy to come by yet — but strange if they stay scarce.

The distillation interrogation follows, and Saedi gives the best barometer available: Reflection AI, the US open-weight shop allegedly being bought by Nvidia, with thousands of Blackwell GPUs — a 12,000 to 14,000 cluster — shipping a flagship six to nine months behind the frontier. If the Chinese labs played by the book, he argues, they would land neck-and-neck with Reflection; instead they ship novel techniques the frontier labs themselves adopt. So distillation is real but not the whole story.

One GPU per worker

Saedi closes with a heuristic: an AI-native employee burns about a GPU's worth of tokens a year; the US has maybe 5 to 10 million Blackwells deployed against roughly 100 million knowledge workers. Four years out, he expects something like one B300 of capacity per knowledge worker — we are massively short of compute for an AI-forward labor force. The hosts' send-off sharpens the buy-versus-build rule: engineer time is scarce, so for repeatable harnesses — A/B-testing two Instagram images — pay the Anthropic premium; where a lagging model suffices, run cheap. Privately, Joe adds, plenty of large US companies already run Qwen and only talk about it when bragging about fine-tuning wins — plus a little posturing, the we-don't-need-you shot across Anthropic's and OpenAI's bow, even as most keep paying the Anthropic-or-OpenAI tax. And the MIT report from the week of recording: an Asterisk-class model fed raw Qwen weights and told to make them better at math actually edited the numbers.

So the question the episode leaves is Joe's: is AI usage CapEx or OpEx? Build with the frontier, serve with the fine-tune — and watch whether the cheaper models keep winning everywhere except the tasks that justify the premium. The disclosure fight will decide how openly American companies can say which half they are running.

More podcast deep dives from Podwise.

Podwise

Skip the six steps. Just listen.

Podwise transcribes, summarizes and maps any episode for you.