On October 20, DeepSeek open-sourced DeepInfer, the inference framework it says powers its models in production. It runs DeepSeek’s V4.1-Flash model with many of the optimizations behind the company’s own API, and the code is licensed under Apache 2.0, which allows commercial self-hosting.
The detail that made headlines is the hardware list. DeepInfer supports Nvidia’s SM100 data-center GPUs and Huawei’s Ascend 950 chips from day one. A Chinese lab has put a Chinese chip on the same footing as Nvidia’s, in public, in code anyone can download.
What DeepInfer Is, and Is Not
DeepSeek has long published model weights that developers can run with third-party software such as vLLM. Running a model efficiently is a separate problem. The software that schedules requests, spreads work across chips, and generates responses can swing speed, throughput, and cost.
DeepInfer hands developers DeepSeek’s own answers to those problems. According to The New Stack’s report, DeepSeek said last year that releasing it was not yet practical.
The release has limits. The Next Web and MarkTechPost report that it supports only V4.1-Flash for now, with other models promised for the same repository. It is built mainly on vLLM, with ideas from SGLang and Nvidia’s TensorRT-LLM, and vLLM has said it will borrow ideas back.
Several pieces are missing. The release leaves out the KV cache system, the split between prompt processing and answer generation, and load balancing across machines, which DeepSeek says are tightly tied to its internal infrastructure. MarkTechPost describes it as offline inference only, not a complete serving service.
DeepSeek’s benchmarks used four Nvidia SM100 GPUs for prompt processing and 64 for output generation, so this is engineering for large clusters, not a laptop download.
The performance and cost figures are DeepSeek’s own. Per Bloomberg, as relayed by The Next Web, no outside party has independently tested DeepInfer yet. One more detail matters for the chip story: MarkTechPost reports that the deterministic, batch-invariant inference mode is available on Nvidia only.
Why Chips Matter
Chips decide what an AI lab can build, how fast it can serve customers, and what each answer costs. Four forces make hardware the center of the race.
- Training and serving are different jobs. Training a frontier model takes enormous, tightly connected clusters over weeks or months. Serving it to millions of users takes a different mix: memory, bandwidth, and low latency. A chip can be adequate for one and weak at the other.
- Software is part of the chip. Nvidia’s CUDA platform, estimated by the New York Times at about four million developers, is a large part of why Nvidia is hard to replace. Any rival chip needs compilers, kernels, and libraries before it is usable at scale.
- Supply is political. Export controls decide who can buy what. As of July, the Commerce Department’s export chief called H200 shipments to China “trivial,” and Beijing reportedly told firms to use domestic chips, mainly Huawei’s Ascend, for inference. A loophole that let Chinese-owned subsidiaries abroad buy Blackwell chips was closed by guidance dated May 31, per TechTimes.
- Cost per answer is the business model. Cheaper serving is how a lab undercuts rivals, which is the pricing war we covered in our report on the DHS order against DeepSeek.
How Close Is Huawei, Really?
The headline claim is “equal billing.” The evidence is more mixed. It shows real progress in software and in some workloads, and gaps that the glossy framing skips.
- The software push. On September 30, DeepSeek open-sourced a version of TileLang, a programming language it positions as a simpler alternative to CUDA, along with compute and communication libraries for Ascend. Six modules mirror DeepSeek’s earlier tools for Nvidia chips. Huawei gave “full support,” and the two built a supernode from 128 Ascend 950 chips. The “simpler” claim is DeepSeek’s own, with no benchmarks.
- The training claim is narrower than it sounds. A Huawei-led report dated July 22 documents post-training of DeepSeek’s V4 family on Ascend systems, at 34.22% model FLOPs utilization. It covers post-training only. Tsinghua professor Liu Zhiyuan told MIT Technology Review that DeepSeek appears to have adapted only part of V4’s training for Chinese chips, and that the model may have been trained mainly on Nvidia hardware.
- Inference is where Chinese chips look best. Multiple sources told MIT Technology Review that Chinese processors remain better suited to inference than training. Earlier DeepSeek testing reportedly showed about 60% of an Nvidia H100’s inference performance.
- Supply is a constraint. The New York Times reports Huawei cannot meet demand for its AI chips and plans to restrict sales outside China. Bloomberg reports DeepSeek plans to deploy at least 160,000 Huawei accelerators in an Inner Mongolia data center. A July report says Huawei targeted up to 750,000 Ascend units in 2026.
- The software gap shows up in the code. DeepInfer’s deterministic mode is Nvidia-only. Parity on the support list is not parity in features.
Read together, the picture is a credible second platform for serving models, with a long road to full parity. That is still strategically significant. A lab that can serve on Ascend is less exposed to export controls.
How AI Labs Expose or Hide Their Chips
Labs differ sharply in how much they reveal. Some publish deals and engineering detail. Some say almost nothing. Others disclose in code what they refuse to say in policy. Here is what the public record shows.
- DeepSeek: shows the software, guards the history. It publishes weights, kernels, TileLang, and now its engine, and in a Stanford transparency index it was among the few companies that disclosed hardware for its final training run. But which chips did V4’s pre-training remains disputed. DeepSeek cited H800 GPUs and Ascend 910C chips when denying an allegation by a senior Trump administration official that V4 was trained on smuggled Blackwell chips. Nvidia called the allegation “far-fetched.” DeepSeek has not commented on the Huawei-chip training claim, per Tom’s Hardware.
- Anthropic: names its platforms, not its split. Anthropic has said publicly that it runs Claude on AWS Trainium, Google TPUs, and Nvidia GPUs. It says it uses over one million Trainium2 chips, committed up to 5 gigawatts with Amazon, and announced up to 1 million TPUs from Google plus 3.5 gigawatts more through Broadcom starting in 2027. What it does not publish is how work is divided. Its Amazon announcement describes a “diversified hardware strategy” with no percentages.
- OpenAI: announces deals, withholds model-level detail. OpenAI has announced a 10-gigawatt custom accelerator program with Broadcom, with rollout starting in the second half of 2026, alongside agreements with Nvidia, AMD, and Cerebras. An estimate in Tom’s Hardware puts combined commitments near 26 gigawatts. But in the Stanford index’s review of OpenAI’s o3, hardware type and quantity, compute provider, and training compute were all marked “not disclosed.”
- Google: owns its silicon, and shows it. Epoch AI estimates Google holds roughly a quarter of the world’s AI compute, mostly its own TPUs, and that about half of its machine-learning compute goes to Google Cloud. The company is also a chip supplier to a rival lab, Anthropic.
- Meta and xAI: scale without detail. Epoch estimates Meta owns about 2.3 million H100-equivalents, though much of it serves recommendation systems, and about 550,000 for xAI’s Colossus 1 and 2 facilities. In the Stanford index, xAI scored 14 out of 100 and Meta 31.
- Other Chinese labs: domestic chips as a selling point. Meituan says its 1.6-trillion-parameter LongCat-2.0 was trained on Chinese-made chips, a claim made by the company itself and reported by Fortune. Where US labs treat hardware as a competitive secret, some Chinese labs treat domestic hardware as a message.
Much of what outsiders know comes from third parties. Epoch AI notes that OpenAI disclosed its own data center power, while Anthropic’s 1.4 gigawatts at the end of 2025 came from an internal OpenAI memo reported by Bloomberg, and the chip-by-chip totals for most labs are Epoch’s own estimates. It also finds that OpenAI and Anthropic mostly rent their compute.
Why the secrecy? The Stanford index authors point to incentives such as ongoing copyright litigation and the treatment of compute spending as a competitive advantage, and they note the wide variation between companies suggests organizational choice matters.
For Chinese labs, the incentives pull both ways: domestic chips are a point of pride, while any use of restricted Nvidia hardware invites legal and political scrutiny. That last point is our reading of the record, not a finding from any source.
What It Means for Companies
- Treat hardware lineage as vendor diligence. Ask AI suppliers where their models run and are trained, and what they will put in writing. Supply, export-control, and sanctions exposure follow the chips.
- Do not mistake open code for easy self-hosting. DeepInfer targets large clusters. DeepSeek’s own benchmark used 4 plus 64 GPUs, and key serving components are missing. Count the engineering and hardware cost, not just the license fee. Our AI ROI Measurement Playbook covers counting the full cost.
- Watch the federal angle. DeepInfer arrived one day after DHS ordered civilian agencies to remove DeepSeek products, including third-party products built using DeepSeek implementations. An open engine makes it easier for other companies to build on DeepSeek’s stack. Suppliers selling to federal agencies should know what sits inside their products.
- Keep your stack portable. Hardware and model supply can change quickly. Prompts, evaluations, and integrations that can move between providers are cheap insurance. Our AI Compliance Readiness Playbook covers building the inventory and vendor file.
Bottom Line
DeepInfer is a software release with a geopolitical edge. It does not prove that Huawei has caught Nvidia. It does show that a leading Chinese lab is building, publishing, and shipping a serving stack that treats a domestic chip as a first-class option.
The larger lesson is about opacity. The industry that asks the world to trust its models is mostly quiet about the silicon underneath them, and the exceptions are telling. Companies buying AI should not need leaked memos and third-party estimates to learn what their vendors run on. Until labs say more, buyers should ask.
Sources: AI Pulse · Big Picture · workplaceai.ai. DeepInfer: The New Stack newsletter report (October 2026); The Next Web (October 20, 2026); MarkTechPost (October 20, 2026); Quartz (October 20, 2026). Ascend tools: The Next Web (October 1, 2026). Huawei post-training report and expert views: Implicator, citing MIT Technology Review and Tom’s Hardware (July 2026). Export controls and Nvidia’s China share: TechTimes (July 15, 2026). Lab compute: Anthropic (April 20, 2026), Silicon Republic (April 7, 2026), Tom’s Hardware (OpenAI and Broadcom), Epoch AI (compute ownership estimates), Anthropic engineering postmortem (September 2025). Transparency: Stanford Foundation Model Transparency Index 2025 (December 2025). Meituan: Fortune (July 26, 2026). Performance, cost, and training claims by DeepSeek, Huawei, and Meituan are the companies’ own and have not been independently verified. Compute estimates from Epoch AI are estimates.