Audit inference stack before hardware lock-in

DeepSeek is planning to install at least 160,000 Huawei Ascend 950DT accelerators in an Inner Mongolia data center under construction. The planned accelerators are intended for inference, not model training. That is a serving decision: the accelerator sets tokens per second per user and cost per token before the model is chosen. DSpark improved aggregate throughput by 51% for DeepSeek-V4-Flash at an 80-token-per-second-per-user target and by 52% for DeepSeek-V4-Pro at a 35-token-per-second-per-user target. At the 80-token-per-second-per-user target, that gain means the same user load can be served with fewer accelerator-hours. The arithmetic, not the demo, decides whether a chip is worth buying.

Production teams often buy the model first and treat the accelerator as an afterthought. The reverse order is cheaper: you can change a model family faster than a rack, a driver stack, a scheduler, and a billing line.

Keep the model team's weights, prompt, and evaluation set separate from the platform team's accelerator, scheduler, SLO, and capacity plan. If the decisions merge, the vendor with the best benchmark can win the hardware before the serving math is done. The audit below forces the serving question before the procurement question.

The chip decision is a serving decision

Peak tokens per second is a lab number. Production SLOs are user-level: how many tokens per second can each user receive while the fleet stays within latency, error, and capacity limits. A vendor can show a high aggregate number and still miss your per-user target after context length, batching, and concurrency change. The audit asks for the per-user number at the SLO, not the headline number.

The DSpark release is available through public GitHub and Hugging Face pages under an MIT license. DeepSeek's DSpark evaluation and published checkpoints span other open model families, such as Alibaba's Qwen and Google's Gemma. Model-family support is an exit path: if the serving stack can run alternative families, you are less trapped by a model roadmap.

Price the five checks before signing

A vendor can win the demo and still lose the production audit. Each check exposes a different lock-in and a failure mode you can price before signing. Run the audit with the same traffic profile you expect in production: same context length, same concurrency, same error budget. A short-prompt benchmark is not comparable. A single-accelerator benchmark leaves the exit path untested.

• Model-family support. Confirm the serving stack can run alternative open-weight models, not only the demo model. If the answer is a roadmap, treat it as unproven.
• Tokens per second per user at SLO. Ask for the per-user token rate at your latency target, context length, and concurrency. Aggregate throughput does not equal user experience.
• Cost per token. Convert accelerator-hours, power, cooling, and support into cost per token at the SLO. A cheaper chip can become expensive if it needs more replicas to hold the target.
• Accelerator abstraction and exit path. Check whether the serving layer can move workloads across accelerators without rewriting the model, prompt, or client contract. If the scheduler is vendor-specific, the exit is a migration project.
• Production capacity lead time. Ask when production chips, racks, and support are available, not when samples ship. A cluster that cannot be delivered on schedule is not a serving option.

Capacity lead time is the quiet lock-in

The planned Huawei setup would rank among the largest known Huawei AI chip clusters, though its rollout timing hinges on Huawei's production capacity. The likely aim of the inference plan is reducing reliance on both Huawei and Nvidia. A large inference cluster can be planned and still not delivered when it is needed.

Procurement ends when the serving stack can prove the per-user token rate, the cost per token, and the path to another accelerator. A promise on any check means the hardware is not ready for real production traffic.