Sovereign AI Compute: Pretraining on an NPU at 170% of an H100, 75% of a B200

TL;DR: NPUs are sold as inference-only chips. VKORE, a hardware-independent acceleration layer, turns them into training hardware too. In our measurements, pretraining throughput on an NPU reached 170% of a single Nvidia H100 and 75% of a single B200 per device, while inference ran 3.8x faster. The point: you can build and improve models on sovereign silicon, without depending on Nvidia GPUs.

Why "sovereign compute" suddenly matters

Access to top Nvidia GPUs is no longer a given in large markets. In China, domestic firms have been steered away from buying Nvidia accelerators, and local silicon from Huawei, Cambricon, Alibaba and Baidu is rising to fill the gap. The same pressure shows up elsewhere as supply, price, and policy risk.

That creates a hard question for any operator building on non-Nvidia NPUs: inference is fine, but what about training? Until now the honest answer was "you still need GPUs for that." VKORE is our attempt to remove that dependency.

What VKORE is

VKORE is a software acceleration layer for NPUs. If CUDA is what makes GPUs programmable for deep learning, VKORE plays that role for NPUs. The chip does not change. Only the software layer does. It is designed to be hardware-independent so a data center can mix GPUs and NPUs and still train and serve from one stack.

The pretraining result

The headline is that pretraining now runs on an NPU, not just inference.

• Pretraining throughput per device reached 170% of a single H100.
• The same workload reached about 75% of a single B200 per device.
• Numerical behavior matched standard training, so this is real training, not an approximation.

In plain terms: a sovereign NPU that most of the industry treats as inference-only outran an H100 on pretraining throughput in our setup, and came within range of a B200, the current flagship.

Inference was already strong

Before training, VKORE improved serving on the same NPU hardware:

• Single-user response: 3.8x faster.
• Server throughput: 4.1x higher.
• Under heavy concurrency, per-token latency stayed stable with no failed requests.

So the same chip serves faster by day and can train by night.

Why this changes data center economics

AI data center competition has moved from "buy more GPUs" to "get more intelligence from the same power and the same silicon." If an NPU fleet can both serve and pretrain, operators avoid a separate GPU build-out just for training. For teams on sovereign silicon, it means the full loop, build, serve, and improve, can run without Nvidia in the critical path.

Pair this with a decision engine that judges in a single forward pass, and the night-shift loop becomes concrete: serve inference by day, score the day's traffic, and pretrain or fine-tune on the same NPUs by night.

FAQ

Does NPU pretraining lose accuracy?
No. Numerical behavior matched standard training in our measurements.

Is 170% a typo? An NPU beating an H100?
It is per-device pretraining throughput in our configuration, with VKORE. The H100 is the reference at 100%. Against a B200, the NPU reached about 75%.

Is VKORE locked to one vendor's NPU?
No. It is a hardware-independent layer meant for mixed GPU and NPU data centers.

Can I use it for inference only?
Yes. Inference acceleration (3.8x single-user, 4.1x throughput) works independently of the training path.

Further reading

• Edge AI in Practice: Running large models on-device without a GPU
• Edge AI in Practice: Decision models that judge in one forward pass