Analysis: how cheap, open V4.1 squeezes rivals and the compute market
The author argues DeepSeek's strong, cheap, no-strings V4.1 line starves domestic rivals like K3 of revenue, ties DeepInfer to its own models, and raises the value of owned compute.
> Am I being a schizo to think that this could be a Chinese government ploy Yes. Meds, now. But there does exist a cynical read of DeepSeek's moves. Because V4.1 line is very strong, cheap and has zero strings attached, they destroy the case for hosting anything like K3 ($3/$15) by Western partners, so starve their domestic rivals of revenue. Moreover DeepInfer is useless for anything not shaped just like V4.1 (so, it'll be DeepSeek's models and finetunes). They also make compute ownership more valuable by having created mutual knowledge about its profitability even at "shockingly low" prices. Frontier labs and hyperscalers already understood how valuable compute is, so they can't go harder on securing allocation, but neoclouds will now ask premium. Flash, in effect, requires a full NVL72 GB200 rack to hit these margins. V4.1 Pro and Ultra (hypothetical for now) will make the issue even more pronounced. When you need a whole cluster to run an 8T DeepSeek model efficiently, *but* it prints you like $40/GPU/hour, this removes a lot of colocated GPUs from easy access. And DeepSeek had little hope to tap into this pool, so they lose nothing, gain mindshare and keep building their new power base around Ascend, which will also become increasingly Whale-shaped due to co-design. In the end, DeepSeek will have defined the entire upstream and downstream AI ecosystem in China, harvesting a tiny slice of the profit in the process. Heavenly Demon Sage 梁圣, ruthless in his benevolence…
This post quotes another post: x.com/blinklebloop/status/2112889401495691643