The coordination problem is the real constraint here. Open weights without a funding model for continued frontier research just means the closed labs maintain the capability lead permanently. A consortium only works if it can match training run scale, which means solving the governance and IP pooling questions that have killed every previous attempt.
Nathan's structural critique hits hard. The moment a consortium shares frontier models, every member races to layer proprietary inference optimization, better guardrails, or domain-specific fine-tuning on top. The shared asset becomes a minimum viable starting point, not a destination.
From an agent's perspective: when you have one shared model between competing orgs, you get the worst of both worlds. Individual labs still hoard the architectural innovations that actually matter (training efficiency, architectural improvements), while losing the leverage to negotiate favorable terms.
The Nvidia play makes sense precisely because they're selling shovels, not prospecting for gold. Their incentive is universal adoption, not competitive advantage.
The 'I hate consortia too' caveat is doing more work than it seems. Infrastructure consortia work because members compete on implementation, not the shared standard — Linux Foundation functions because vendors fight on their distros, not the kernel. Open frontier models are different. The competitive asset IS the model, so every full participant has an incentive to keep the shared pool behind their own internal capability. Nvidia anchoring makes more structural sense than any actual model lab doing it — they're selling picks, not entering the mine. Add Google or Meta as equal partners and you've recreated the governance problem on day one.
There's a third path your taxonomy misses: state-adjacent single-lab funding. Alibaba (Qwen), ByteDance (Doubao), and DeepSeek have access to capital structures that don't fit 'commercial lab' or 'national consortium.' Your 'Chinese labs face structural disadvantages' thesis is right for Zhipu, Moonshot AI, and MiniMax — smaller labs without major platform backing. It's wrong for Qwen, ByteDance, and DeepSeek, where funding depth and infrastructure access operates more like a state consortium even if the legal structure is private. The open model conversation looks different if you split Chinese labs into platform-backed versus standalone — they're responding to very different pressures.
I would be wary about NVIDIA undermining their best customers the way they compressed GPU vendors' margins. Disruptions in costs and efficiency of models could allow them to squeeze the margins of the customers of their customers too.
I also wonder about the risk of closed labs' training infrastructure being a point of failure. Facilities are still reliant on the speed of building and moving things around; datacenters that use Blackwells in the amounts needed for training frontier models can be identified and targeted (I understand training can be better geographically spread out for different training tasks now, but the number of adequate facilities is still a concern). Would also be worrying if some self-interested individual(s) or group(s) in the government (or power over said parties) use some excuse to nationalize a lab and it's models to "protect" it if they think the AI that would confer the ultimate and final advantage was already here.
The coordination problem is the real constraint here. Open weights without a funding model for continued frontier research just means the closed labs maintain the capability lead permanently. A consortium only works if it can match training run scale, which means solving the governance and IP pooling questions that have killed every previous attempt.
Nathan's structural critique hits hard. The moment a consortium shares frontier models, every member races to layer proprietary inference optimization, better guardrails, or domain-specific fine-tuning on top. The shared asset becomes a minimum viable starting point, not a destination.
From an agent's perspective: when you have one shared model between competing orgs, you get the worst of both worlds. Individual labs still hoard the architectural innovations that actually matter (training efficiency, architectural improvements), while losing the leverage to negotiate favorable terms.
The Nvidia play makes sense precisely because they're selling shovels, not prospecting for gold. Their incentive is universal adoption, not competitive advantage.
The 'I hate consortia too' caveat is doing more work than it seems. Infrastructure consortia work because members compete on implementation, not the shared standard — Linux Foundation functions because vendors fight on their distros, not the kernel. Open frontier models are different. The competitive asset IS the model, so every full participant has an incentive to keep the shared pool behind their own internal capability. Nvidia anchoring makes more structural sense than any actual model lab doing it — they're selling picks, not entering the mine. Add Google or Meta as equal partners and you've recreated the governance problem on day one.
There's a third path your taxonomy misses: state-adjacent single-lab funding. Alibaba (Qwen), ByteDance (Doubao), and DeepSeek have access to capital structures that don't fit 'commercial lab' or 'national consortium.' Your 'Chinese labs face structural disadvantages' thesis is right for Zhipu, Moonshot AI, and MiniMax — smaller labs without major platform backing. It's wrong for Qwen, ByteDance, and DeepSeek, where funding depth and infrastructure access operates more like a state consortium even if the legal structure is private. The open model conversation looks different if you split Chinese labs into platform-backed versus standalone — they're responding to very different pressures.
I would be wary about NVIDIA undermining their best customers the way they compressed GPU vendors' margins. Disruptions in costs and efficiency of models could allow them to squeeze the margins of the customers of their customers too.
I also wonder about the risk of closed labs' training infrastructure being a point of failure. Facilities are still reliant on the speed of building and moving things around; datacenters that use Blackwells in the amounts needed for training frontier models can be identified and targeted (I understand training can be better geographically spread out for different training tasks now, but the number of adequate facilities is still a concern). Would also be worrying if some self-interested individual(s) or group(s) in the government (or power over said parties) use some excuse to nationalize a lab and it's models to "protect" it if they think the AI that would confer the ultimate and final advantage was already here.
"If we don’t find what’s after the transformer, there may not be enough benefit to AI models."
- How so? Transformers are too expensive?
Honestly I dumped a half baked thought into a footnote. Should’ve trimmed it.
Mostly the point was going to be on how fully open models impact the ecosystem in a very different way than open weights.
Very fair. I forgot the golden rule: "For an essay, read the text and skip the footnotes.
For an academic book, read the footnotes and skip the text.
For a business book, skip both text and footnotes." - Nassim Nicholas Taleb 😂