Alibaba RISC-V CPU runs 27B Qwen at 30 tps — Nvidia isn't the only bet

What this is

A hot post surfaced this week on Reddit's LocalLLaMA community: Alibaba's T-Head unit's RISC-V processor, the XuanTie C950, ran a 27-billion-parameter Qwen model at 30 tps (tokens per second — the number of text fragments generated each second) without any Nvidia GPU. RISC-V is a royalty-free open-source chip architecture (unlike ARM and x86, which require paid licensing). The 30 tps mark is a threshold worth naming: humans read at roughly 5–10 tps, which means this chip can sustain fluid conversation.

Our read: the real signal is not the 30 tps number itself, but that Alibaba has closed the loop on "in-house inference silicon + in-house foundation model" — precisely the chain US export controls are trying to choke off.

Industry view

The optimists see a domestic closed loop: open-source architecture + homegrown model + homegrown silicon — all three now assembled. Data-sensitive sectors — finance, government, healthcare — are the most obvious beneficiaries, since enterprises can deploy AI without betting the farm on Nvidia's supply chain.

Three pushbacks also deserve airtime. First, the numbers come from community testing and have no independent third-party verification. Second, the RISC-V software ecosystem (compilers, inference frameworks) is far less mature than CUDA (Nvidia's parallel-compute platform), so developer migration costs are steep. Third, this only solves the "inference" side; "training" still depends heavily on GPUs, and the core bottleneck has not gone away.

Impact on regular people

For enterprise IT: CIOs evaluating on-prem deployment now have a non-Nvidia option on the menu, and the leverage at the procurement negotiating table is starting to shift.

For working professionals: running large models on consumer-grade hardware may stop being a hobbyist toy and become a real cost-saver — though for now, developers will still mostly run on Nvidia cards.

For the consumer market: phones and home appliances could ship with built-in RISC-V AI silicon in the future, which would further compress the price of cloud AI calls.