GPT-6 Astra 与 Cerebras 算力传闻梳理
有人梳理 OpenAI 在 Cerebras 与 Nvidia 上跑 GPT-6 Astra 的算力取舍，认为 300 tok/s 的「ultrafast」只是营销话术。
The $CBRS / OpenAI / $NVDA / ‘Ultrafast' commotion on here is getting out of hand. Here’s what we know, or might reasonably assume: - Up until GPT-5.6 Sol Ultrafast launched on Cerebras hardware in limited preview in August, OpenAI were using 100% of their Cerebras capacity internally and they love it (and have publicly said as much) - GPT-5.6 Sol is widely taken to be around 2 to 2.5T parameters - We don’t know the parameter count for GPT-6 Astra, but it’s obviously a larger model, potentially of the order of 5T, as numerous pundits are suggesting - As such, just to store the weights on the minimum CS-3 footprint, Astra might require twice as many systems as GPT-5.6 Sol - Cerebras’ inference capacity is maxed out and remains extremely constrained while they continue to ramp - To immediately switch from serving GPT-5.6 Sol to GPT-6 Astra might double the required footprint, which for obvious reasons would be uneconomical given the trade-offs, especially if there was a better option right around the corner… - Assuming therefore that GPT-6.1 Sol is a similar size to 5.6 Sol, it would make far greater sense to make that switch - OpenAI have said 6.1 Sol ultrafast is 'coming soon', so we’ll see - In the meantime, 6 Astra ‘ultrafast' is running at 300 tok/s on Nvidia, which numerous pundits are dramatically taking to mean that Cerebras has been displaced. The problem is that 300 tok/s isn’t even ultrafast by any reasonable up-to-date standard. I think Cerebras would be pretty embarrassed to serve models that slow. So it feels like a marketing language slip-up by OpenAI, having already reserved ‘fast mode’ for an even slower service - In any case, Cerebras is gearing up to start serving the significantly more capable CS-4 at scale which, in combination with disaggregation, will be much better suited for serving very large models to high numbers of users at 1000+ tok/s. The indications are that Cerebras switches into this gear from Q4 2026 onwards.