“These days, when entrepreneurs pitch at Andreessen Horowitz (a16z), a major Silicon Valley venture-capital firm, there’s a high chance their startups are running on Chinese models. “I’d say there’s an 80% chance they’re using a Chinese open-source model,” notes Martin Casado, a partner at a16z.”
If the AI bubble is going to burst, you’ve got to wonder how many of today’s AI stars like OpenAI will survive it. Are they already yesterday’s people, and the future is leaner, cheaper, and built on free open-source AI? If 80% of new American start-ups are choosing Chinese open-source, you can bet that figure rises to near 100% for the rest of the world.
Silicon Valley thought they were soon going to get an AI unicorn, another world-conquering Google or Meta. Maybe, one day. For now, it looks like Chinese Open-Source AI may be the model about to spread all over the world.
I hear the sound of a thousand tiny Unicorns, lightly clapping in support. Sounds like coconuts. Tiny coconuts.
That might change a little with more of OpenAI’s gpt-oss releases, which might be too new to exist in this data.
It’s not great, and its extremely censored compared to non-US models (yes, the irony), but it is sparse and cheap to run on H100s for how fast it is.
But yeah, it seems like Meta is imploding from tech bro overload. Google Gemini is good, but Google’s purposfully constricting their open source team to not compete with Gemini API. Anthropic is a censorship meme and will never release anything, and smaller US startups just don’t have the attention nor the funding to compete. Their hope seems to be to invent something proprietary and hypey that gets them bought up, not to actually build something functional.
And one unspoken observation among the ML crowd is that the Chinese firms are both training on the outputs of American models, and maybe sharing high quality, questionably sourced (Chinese govt?) datasets with each other under the table, given how many share the same quirks.
Their hope seems to be to invent something proprietary and hypey that gets them bought up, not to actually build something functional.
They all seem to be chasing the dream of being unicorns (for the unintiated reading this, monopolist giants like Google/Meta, not magical horses).
Do American VCs even bother with start-ups who want to be small/medium sized firms, and have a solid case for making a few hundred million dollars every year?
Do American VCs even bother with start-ups who want to be small/medium sized firms, and have a solid case for making a few hundred million dollars every year?
There are startups in small niches; I was part of one. But you don’t hear about them much.
Honestly, most ‘public’ startups know they have zero shot at being a unicorn (because giants will crush them like bugs if copycats don’t get them first) and chase getting acquired by Big Tech instead. That’s literally the business plan. An example would be some AI firm that comes up with a novel long context scheme (of which there have been several), maybe offering an API. They hype the snot out of it… But it’s not sustainable at all. They never open source it, they never work it into anything sustainable, the whole unspoken plan is literally for someone to just eat them and the tech to disappear into corporate oblivion.
In terms of advancing software, its extremely inefficient, which is one reason the Chinese are doing so well sharing stuff instead. Not that the Chinese firms don’t have their own issues.
In terms of advancing software, its extremely inefficient,
It amazes me how their BS on ‘innovation’ has infected broader culture and politics.
Look how little fundamental innovation there is in health, education and housing. All getting more expensive and out of reach.
OpenAI will have to do a lot better than their current open release to compete, frankly. It made a little blip and then Deepseek put out a new model that sent it off to obscurity.
You aren’t wrong, but Deepseek’s is a 600B model, so it’s kinda a different class. 20B/120B are much cheaper to run for agenic tasks. Beside, I think GLM 4.5 is where its at now :P
That being said, there’s a boatload of competition in the small model class anyway.
Yeah, there were already other models in that weight class that match or exceed GPT OSS. I just recall that it was Deepseek that took the headlines away because they happened to be the next ones to release a model afterward.
There were some in between too! Bytedance’s 36B (for instance) is incredible for its size.
