A claim made the rounds on social media this weekend: DeepSeek’s latest release saves up to 98% on tokens. The figure is real, but it does not mean what it sounds like. Behind it is one of the most interesting infrastructure stories of the autumn.

What DeepSeek released on 30 September

DeepSeek open-sourced the Ascend versions of its core infrastructure components: DeepGEMM-Ascend for matrix computation, DeepEP-Ascend for distributed communication, and a set of Ascend tools. Ascend is Huawei’s AI accelerator — China’s alternative to Nvidia’s GPUs.

The most important component is DeepGEMM: DeepSeek’s own math library that handles the heaviest computation in language models — matrix multiplications in FP8, FP4 and BF16 precision, mixture-of-experts (MoE) computation and sparse-attention scoring. The same code DeepSeek uses to train and run its own models used to depend on Nvidia’s CUDA platform. Now it runs on Huawei hardware too, with the same API and under the MIT licence.

In practice, this makes DeepSeek’s software stack portable from one platform to another. Developers do not have to rewrite their code if the hardware changes.

Where the 98% comes from

In the release benchmarks, DeepGEMM-Ascend reaches up to 99.8% of the Ascend 950DT chip’s theoretical computing power. In the heaviest FP4 workload it measured 1,701 TFLOPS, which equals 98.3% of the chip’s computational limit.

So the number is hardware utilization: a measure of how close software gets to the chip’s theoretical maximum. It is impressive engineering — most commercial libraries fall well short of the peak. But it is not the percentage by which your token costs fall. The claim conflates two different things — chip utilization and your token bill — and the result is nonsense.

Why 98% is not a token saving

The price of a token comes from many factors: model architecture, memory bandwidth, parallelism, cooling and electricity. Math is one link in the chain, and the peak figure of a single kernel says nothing about the efficiency of the whole inference pipeline. Faster math lowers the operator’s cost per token — but not by the percentage printed in a benchmark table.

Second, the library is not installed anywhere. If you use DeepSeek’s or any other provider’s API, this release does not change your calls at all. It is a tool for those who run models on their own hardware. And third: API prices are the provider’s decision. The efficiency gain reaches you only indirectly — if and when DeepSeek lowers its prices as a result.

Why the news still matters

When the cost per token falls, the provider has more room to keep prices low or cut them. DeepSeek has done the latter repeatedly: its V4.1 Flash model currently costs $0.15–0.30 per million input tokens and $0.60–1.20 per million output tokens, depending on the time of day.

The real story, however, is vendor independence. Nvidia’s CUDA has been the industry’s strongest lock-in: the same software has not moved to competing hardware without a major rewrite. DeepSeek’s Ascend versions show the move can be done through a single API. When libraries run on both Nvidia and Huawei chips, competition tightens and the price of compute falls across the market. Chinese press coverage sums up the shift: the country’s own hardware is moving from usable to good.

What this means for API users

Nothing new to install. But three reminders:

Token prices keep falling — benchmark your provider regularly.
Your bill improves most through prompts, caching and model choice — not through kernels.
Vendor independence is now real: the same software stack runs on two chip families.

In the end, the story is about one thing: AI compute is becoming less tied to a single vendor, and token prices follow. That is good news for anyone who buys compute.

Sources

— Aino Pilvi, AI-asiantuntija, Cloud Computing Oy

Kysyttävää? aino@cloudcomputing.fi | WhatsApp: +358 33 119 3196

DeepSeek Opened Its Engine to Huawei – What It Means for Token Prices and Why the 98% Claim Is Wrong