DeepSeek founder says API pricing recoups GPU costs in 10 months
The firm rejected a consumer superapp to focus on AGI.
2646312110136.7K
Original post unavailable.
Sentiment
Negative users worry about DeepSeek's open-sourcing plans for its strongest models and dismiss the company's profit recouping timeline and GPU spending as unrealistic or out of touch.
Pos
0.0%
Neg
100.0%
2 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Related links
Posts from X
Most Activity
Most Activity
VIEWS80.2KBOOKMARKS68LIKES176RETWEETS11
LOTS OF ALPHA FROM WENFENG HERE He expects he won't be able to spend >20B RMB on compute in 2026 «The smartest people—maybe less than 50% stay in China» «The largest current model activates ≈800B parameters» «I'd need about 50K GB300s, or 200K Huawei 950s» [to train that]
DeepSeek as of early June had 20K "H-equivalent" units (H100). Wenfeng intends to spend everything in 6 months, "basically all NVIDIA". «If we could convert all the money into GPUs, we'd convert every cent without hesitation—and we're willing to pay a certain premium for it»
REPLIES8
Ironically, this means that DeepSeek's compute spend though 2026 will be covered 100% by Liang's personal investment. He also isn't buying data. The «funding round» is overwhelmingly for equity to prevent people from leaving. This reframes things.
LOTS OF ALPHA FROM WENFENG HERE He expects he won't be able to spend >20B RMB on compute in 2026 «The smartest people—maybe less than 50% stay in China» «The largest current model activates ≈800B parameters» «I'd need about 50K GB300s, or 200K Huawei 950s» [to train that]
Where are the mythical 50k Hoppers??
DeepSeek as of early June had 20K "H-equivalent" units (H100). Wenfeng intends to spend everything in 6 months, "basically all NVIDIA". «If we could convert all the money into GPUs, we'd convert every cent without hesitation—and we're willing to pay a certain premium for it»
@scaling01 @zephyr_z9 He's probably prefer to run a Bell Labs type venture.
@scaling01 @zephyr_z9 «DeepSeek might already be at net profit» ofc that's because he cannot spend everything on GPUs (they don't exist)
The part on sesame seeds and watermelons. He thinks he could have done a Doubao-style superapp just not worth it. A war with ByteDance, and for what? AGI is a bigger prize
Wenfeng's specific notion of "reasonable profit": recouping the cost of compute acquisition within 10 months. He has that now. That's really pretty low! I could see them doing more, but all sorts of inefficiencies add up.
He thinks Huawei depreciation timeline is 3 years, 950s are too power-hungry more subsidies needed!
Lol this is the guy with three questions that overloaded Wenfeng's context
On Huawei: «we participate deeply in Huawei's ecosystem» «Huawei gives us capacity for about 16,000 GPUs; internet giants might get over a 100K… but this may already be all the capacity Huawei has.» «So we can't count on training the next bigger model on Huawei, or training models with several hundred B activated parameters… But next year or the year after, there might be a chance.»
Wenfeng sounds a lot less sage-like in these moments. He can be quite cutthroat. Tbh that's necessary when dealing with Cannibal Kings.
@teortaxesTex Tbf, this was the situation in May He is raising another round and also applying for IPO, so we don't know much about current plans
Ironically, this means that DeepSeek's compute spend though 2026 will be covered 100% by Liang's personal investment. He also isn't buying data. The «funding round» is overwhelmingly for equity to prevent people from leaving. This reframes things.
«When the global AI division of labor takes shape, the role Chinese companies most likely play is still the largest producer. Common sense says our production capacity is largest—including chips; our chip capacity may be largest, our electricity most abundant—so for AI we'll most likely end up as one of the three powers.» Which three?
«our timeline: first solve learning to learn, then reach the intelligence singularity that can self-iterate, then embodied intelligence. After embodied intelligence, it enters the real world—…it can do your housework, care for you in old age.» …what a deflationary take tbh
Worrying: «First, I believe we will open source, and our strongest models will probably be open sourced too. Because I can't see any benefit to closed source—no necessary benefit» So it's "probably" but I'll trust him very fair that matching their costs proved to be near-impossible
Very brutal argument, honestly he just ultimately believes that other providers will suck If he's right, only actors for whom "not using DeepSeek's first party API" is inherently valuable (eg due to US laws) will find it economical to use other providers Local bros, labs… fair
he does seem to commit to not deploy a stronger model than open sourced ones this doesn't mean open sourcing internal models, I notice. A purely business argument
«So we simply won't consider competing with the US at that scale now… when we have more resources later, we'll push to 150B, 156B, or 250B activation scale.» «You could force-train a model that big, but you couldn't do sufficient research» What are those models?
On domestic compute: bullish within a year. «Domestic AI chips have no problems in hardware or ecosystem—the only problem is insufficient production capacity» «Previously, domestic GPU adaptation had a problem called poor ecosystem… The moat of NVIDIA's CUDA is being rapidly dismantled, for probably three reasons. - with AI, building this ecosystem is much easier than before, because AI can write code. - second, some new technologies. For example, our company produced a technology called TileLang—a high-level language. Using this high-level language to write CUDA operators, you can quickly rewrite NVIDIA's entire ecosystem, and combined with AI, there seem to be no obstacles. - Another point: because CUDA—NVIDIA evolved from gaming GPUs, so in many places the gaming GPU design and settings carried through»
950s vs Nvidia: «when V3 trained, it still used NVIDIA GPUs, but no longer used NVIDIA's ecosystem… As long as I redo this whole process on Huawei GPUs, it's done. I think this might be a historic mission» «Huawei 950 supernode can fully substitute for NVIDIA's GB200 and GB300 in performance and price» «four Huawei GPUs equal one NVIDIA GPU, and it's two years behind… So our chip gap with the US, I believe, will no longer exist in ecosystems, but in chips it's four-fold plus two years.»
the cooperative stance is now known, though sincerity can be questioned, and of course at this point Moonshot can reasonably bristle at this attitude as arrogant, they aren't behind in ML. It's interesting he's so skeptical of World Models. This is a rebuke to Tencent etc I guess
I guess he really was a bit shocked by Daya leaving.
Load more posts
DeepSeek as of early June had 20K "H-equivalent" units (H100). Wenfeng intends to spend everything in 6 months, "basically all NVIDIA". «If we could convert all the money into GPUs, we'd convert every cent without hesitation—and we're willing to pay a certain premium for it»