What DeepSeek Isn't Doing
Notes from DeepSeek CEO Liang Wenfeng's leaked investor call.
The most surprising thing in the leaked transcript of Liang Wenfeng’s May 20 investor call is how little DeepSeek seems to want users at all.
This is in stark contrast with US frontier labs who are knife-fighting over developer, consumer, and enterprise mindshare through a torrent of product launches.
Liang seems only interested in the pursuit of AGI. Users are “sesame seeds, not watermelons”, he said on the call. Don’t stop to pick up small wins on the road to a large one.
There are two ways to build a frontier AI company. Build a chatbot, fight for users, sell subscriptions and ads. Or, alternatively, chase AGI, and the chatbot is a by-product. Liang clearly sits in the latter camp. He says DeepSeek’s consumer app and the enterprise API aren’t the point, and only needs a handful of people to run the API. With the hype around DeepSeek last spring, fighting ByteDance for consumer users was a real business he could have a shot at winning. He passed, because he thinks what’s coming makes today’s consumer market look tiny. He’ll pick up the sesame seeds in passing, but just won’t stop for them.
The strongest evidence for this is how DeepSeek prices. They chose to set their API price to a ten-month hardware payback, roughly 6x on compute, and stop right there. Liang acknowledges that demand in that band is inelastic, so doubling the price would nearly double revenue. He knows this and won’t do it.
Two reasons. One is team - they cheered when he cut prices, and that’s part of why they work there. But mostly because at just 6x margin, nobody else can profitably run DeepSeek’s own models cheaper than DeepSeek can. Fat margins would invite everyone to undercut him using his own weights. Thin margins are the moat.
A note on the source: DeepSeek has not confirmed the document, and it is a speech-to-text transcript that flags its own error rate on names and figures. Everything below is what the transcript says he said.
On Compute:
Compute-capped: Frontier models activate ~800B params; Chinese labs are at tens of B. To train at 800B he’d need ~50,000 GB300s or 200,000 Huawei 950s just on training alone. Even spending the whole ¥50B raise he couldn’t afford it, or afford to serve it.
He believes that DeepSeek is one to two years behind on one-twentieth the compute. The goal is to hold the compute ratio and compress the time gap to three or six months.
He argues that export controls caused the substitution they were meant to prevent. In a normal market where he could buy Nvidia freely, substitution with domestic chips would be a harder call.
Nvidia’s software moat is loosening. Nvidia’s CUDA software platform is breaking for three reasons: AI can now write the ecosystem code cheaply; TileLang makes kernel rewriting fast; and CUDA is architecturally welded to gaming cards, which stopped making sense once compute cards outgrew gaming.
The 16,000 Huawei Ascends everyone reported as a big deal: he says it equals ~4,000 B-series chips, isn’t enough for a next-gen model, and the real reason for buying is to help Huawei get its ecosystem right. Big Chinese internet firms got an order of magnitude more chips than DeepSeek.
On Research
The staircase: language model → chain of thought → agents → continual learning → self-iteration → embodiment (physical ai).
Continual learning is the named next bottleneck and nobody in the world has a working method - we are all still searching.
He wants continual learning before general intelligence explicitly because it’s the lazy path. The next model’s first customer is DeepSeek itself: “first it has to be useful to us.”
(Editor’s Note: the next model that trains itself is a challenge the top minds in AI are circling around. Anthropic hired Andrej Karpathy in May to build a team using Claude to accelerate its own pre-training. Jeff Dean and three Google colleagues left this week to start Discovery Loop, automating large-scale ML experiments.)
Open-ended research’s main edge is researcher attention, rather than compute. (slight contradiction) He calls it akin to drawing lots.
He won’t do video generation or world models: commercially good, off the intelligence main line. He found the post-Sora pile-in odd, like it had become mandatory to jump on the bandwagon
Anthropic’s lead over OpenAI is transitory; the code-agent advantage isn’t large. Half his company thinks OpenAI is better.
They have no org chart, no KPIs, no written vision. He admits this breaks with scale and he’s already building departments.
Too many Chinese labs building base models; the US has three. It will converge to maybe two big and two small.