> 200 EFLOPS of compute & generates 5m tokens/s I hate how we're doing "tokens" as if they're equal. 5M tokens of what? Flash? Anyway, 5M/s is not impressive (200 EFLOPS is, very much so). I guess they mean 5M per… ah, no, this is fine. That's about 12 SuperPODs I guess.
Users praise the compute power of Huawei Cloud's new Agentic AI Platform as explosively impressive despite minor token caveats.
interestingly, 100K cards with 5m/kilocards = 500M tok/s, 1.8T tok/hr, ≈1.3 quadrillion tokens/month. This is where Google was, in Oct 2025, «processed across all surfaces». Token factory is pretty big.
> 200 EFLOPS of compute & generates 5m tokens/s I hate how we're doing "tokens" as if they're equal. 5M tokens of what? Flash? Anyway, 5M/s is not impressive (200 EFLOPS is, very much so). I guess they mean 5M per… ah, no, this is fine. That's about 12 SuperPODs I guess.
@teortaxesTex tokens不等价，但算力确实牛
@teortaxesTex 这个算力牛炸了 token 速度有点迷