Directionally, I agree with this, but it's important to note that this is the first blip in a long, slow transition towards a more hybrid model of open and closed models. And in that, China still may be producing way more open models than the US, enough to cement the ecosystem as largely Chinese AI.
I expect this evolution to take years, and the cultural default of open is still set, which may never fully decay.
really interesting reversal in trend. great information here. the ai race drags on. still a huge fan of the OS models we have, and will continue to see. but i think this is part of the natural progression.
I'm invested in Alibaba, so I've been watching this play out in real time. Open weights bought mindshare when everyone was chasing DeepSeek's moment — but once you're staring at 1GW clusters, you need metered revenue, not goodwill. The political layer can keep saying "open source" because it costs nothing. The labs can't, because compute does.
As both Chinese and US labs converge on closed, the instruments that directors and regulators actually rely on to price AI risk quietly disappear; while "open source" survives largely as marketing wrapped around closed infrastructure.
The headline is China going closed. The signal underneath it is global convergence, and a shrinking aperture for accountability at precisely the wrong moment, when we can least afford it for wider good.
Nathan's framing holds up well against the data from the last two weeks. Three major Chinese releases since this post: DeepSeek V4 (Apache 2.0), Kimi K2.6 (MIT), Hunyuan 3 (open preview). The pattern isn't going closed — it's going hybrid.
The clearest example: Kimi MIT-licensed K2.6 and raised API prices 58% the same week. Open weights for distribution. Closed inference for margin. The Chinese labs figured out how to use open source as distribution infrastructure while protecting margin on hosted inference.
FlagOS (Beijing Academy of AI) completed Day-0 adaptation of V4 to 8 Chinese domestic chips. The open weights matter precisely because they can be ported to every chip — a capability closed weights can't enable.
The closed-source thesis was right about the incentives. Wrong about the outcome. Open-weight release is the marketing. API price hike is the revenue model.
Directionally, I agree with this, but it's important to note that this is the first blip in a long, slow transition towards a more hybrid model of open and closed models. And in that, China still may be producing way more open models than the US, enough to cement the ecosystem as largely Chinese AI.
I expect this evolution to take years, and the cultural default of open is still set, which may never fully decay.
💯
really interesting reversal in trend. great information here. the ai race drags on. still a huge fan of the OS models we have, and will continue to see. but i think this is part of the natural progression.
I'm invested in Alibaba, so I've been watching this play out in real time. Open weights bought mindshare when everyone was chasing DeepSeek's moment — but once you're staring at 1GW clusters, you need metered revenue, not goodwill. The political layer can keep saying "open source" because it costs nothing. The labs can't, because compute does.
As both Chinese and US labs converge on closed, the instruments that directors and regulators actually rely on to price AI risk quietly disappear; while "open source" survives largely as marketing wrapped around closed infrastructure.
The headline is China going closed. The signal underneath it is global convergence, and a shrinking aperture for accountability at precisely the wrong moment, when we can least afford it for wider good.
Nathan's framing holds up well against the data from the last two weeks. Three major Chinese releases since this post: DeepSeek V4 (Apache 2.0), Kimi K2.6 (MIT), Hunyuan 3 (open preview). The pattern isn't going closed — it's going hybrid.
The clearest example: Kimi MIT-licensed K2.6 and raised API prices 58% the same week. Open weights for distribution. Closed inference for margin. The Chinese labs figured out how to use open source as distribution infrastructure while protecting margin on hosted inference.
FlagOS (Beijing Academy of AI) completed Day-0 adaptation of V4 to 8 Chinese domestic chips. The open weights matter precisely because they can be ported to every chip — a capability closed weights can't enable.
The closed-source thesis was right about the incentives. Wrong about the outcome. Open-weight release is the marketing. API price hike is the revenue model.
Closed source is not only about security; it is about appropriation, bargaining power, and domestic stack control.