Two years ago the question about domestic AI chips was whether they worked at all. In 2026 that question has an answer. The new questions are which workloads to run on domestic chips, how the economics pencil out, and whether the software ecosystem can keep up.
In our consulting work, domestic compute has moved from "not a candidate" to "evaluate it once". Imported high-end cards are hard to buy and hard to afford; domestic cards now hold their own on inference.
The 2026 numbers that matter
| Metric | Figure |
|---|---|
| Domestic AI chip shipments | 1.65 million units, above 40% of the market for the first time |
| Ascend shipments and share | 812,000 units, a 20.3% share, first place |
| Cambricon revenue, H1 2026 | RMB 5.996 billion, up 108% |
| Daily token calls in China | Above 500 trillion in June 2026 |
| DeepSeek's Ascend deployment | 160,000 Ascend 950DT in a 1GW data center in Ulanqab, inference only, no training, delivery beyond one year |
| GLM-5.3-Flash | 320B parameters with 18B active (MoE), 100,000-chip domestic cluster, performance on par with the top models, price about 1/40 of theirs |
The 950DT specs deserve a look: 144GB of HiZQ 2.0 HBM, 1 PFLOPS of FP8 compute and 2TB/s of interconnect. The 144GB matters most: it sets how large a model fits on one card.
The signal to watch: inference only, no training
DeepSeek's 160,000-chip cluster is explicitly for inference, not training. The logic behind that split applies directly to any local deployment.
Training and inference are different loads
| Dimension | Training | Inference |
|---|---|---|
| Compute profile | Hyperscale parallelism. Interconnect bandwidth is the bottleneck and thousands of cards must synchronize | One card or a small cluster. Memory capacity and throughput are the bottleneck |
| Software stack | Heavy CUDA dependence: distributed training frameworks, operator libraries, mixed-precision toolchains | Comparatively light. Mainstream inference frameworks and their domestic equivalents are already ported |
| Cost of failure | A run lasts weeks; a mid-run error is expensive and stability demands are extreme | A failed request is retried, so there is room for fault tolerance |
| Domestic card maturity | Still catching up, with the ecosystem the main obstacle | Usable and cost-effective |
99% of enterprise demand for local AI is inference. You are deploying an existing open model to serve internal work: knowledge base Q&A, document processing, customer service assistance, code completion, image recognition. That is exactly where domestic cards already compete.
GLM shows the ecosystem is turning
GLM-5.3-Flash runs on a 100,000-chip domestic cluster at 320B parameters with 18B active, matching the top international models at roughly 1/40 the price.
- Model builders now design for domestic hardware. MoE's sparse activation lowers demand on per-card compute and interconnect bandwidth, which suits domestic clusters.
- The price is the decisive business variable. An order-of-magnitude cut in API cost turns scenarios that were not worth considering into live projects: intelligent archiving across a document estate, real-time quality scoring on every support call.
A feasibility checklist before you buy
Start with four questions, not "which card".
1. Does this workload need to run on premises?
Three hard triggers justify it, and any one is enough:
- The data cannot leave. Medical images, financial customer records, confidential government documents, manufacturing process parameters: uploading them to a cloud API breaks policy or regulation.
- Volume beats the API on cost. If the monthly API bill tops RMB 15,000 for more than a year, owning starts to win. A reasoning server costs RMB 150,000-400,000 over three years.
- Offline operation or low latency. Factory floors, field equipment, isolated intranets, or anything needing a response under 100 ms.
If none applies, a cloud API is more rational: no operations staff, no depreciation, and upgrades arrive with the service.
2. How large is the model, and what hardware does it need?
| Workload | Model size | Hardware floor | Reference configuration |
|---|---|---|---|
| Individual or small-team knowledge base | 7B-14B quantized (INT4) | One card with 16-24GB | RTX 4090 or 5090 workstation, RMB 20,000-40,000 |
| Departmental document processing, support assistance (10-50 people) | 14B-32B quantized, or a small-activation MoE | One 48-96GB card, or two 24GB cards | Domestic inference card (Ascend 300I, Cambricon MLU370) or RTX 6000 Ada, RMB 80,000-250,000 |
| Multi-department concurrency (50-500 people) plus RAG | 32B-70B, or a MoE such as the GLM 320B/18B family | 4-8 card inference server, 96-144GB per card | Atlas 800I A2 or a Cambricon Siyuan cluster, RMB 500,000-1,500,000 |
| Fine-tuning or LoRA | Domain adaptation on a 7B-32B base | Roughly 3-5 times the memory of inference | Rent cloud capacity; a self-built fine-tuning cluster rarely pays |
Memory capacity matters more than TOPS or FLOPS. If the weights do not fit in memory, the model does not run at all; too little compute only makes it slower. The 950DT's 144GB is valuable because one card holds a 70B-class model at FP8 without splitting across cards, and splitting adds communication overhead.
3. Who runs the software stack and the operations?
This is the real bottleneck, and the top reason these projects fail:
- CUDA migration. If your team knows only CUDA, moving to CANN or Cambricon Neuware costs learning time, and operator coverage and debugging tools still lag for custom operators and unusual architectures.
- Porting is not download-and-run. Qwen, GLM, Llama and DeepSeek are ported to different degrees. Quantization scheme, framework version and driver version must all line up, and trial and error runs to weeks.
- Operations needs a dedicated owner. Concurrency scheduling, memory fragmentation, hot-swapping and monitoring are a different order of work from running a web server. Without one, projects reach "it works but nobody dares touch it" within six months.
Pick a combination the vendor has already certified, and package implementation plus a year of operations with an integrator. That is cheaper than hiring.
4. How much domestic-stack (xinchuang) compliance is required?
Government and state-owned enterprise purchases often require a domestic catalog, and the selection logic changes completely:
- Check that the product is in the catalog and holds the relevant certifications: commercial cryptography, Multi-Level Protection Scheme (MLPS), localization rate.
- It is not just the AI card. The whole machine (Kunpeng, Hygon or Phytium CPUs), the operating system (Kylin or UnionTech UOS), the database (Dameng, Renmin Jincang or openGauss) and middleware have to be sourced as a set, or acceptance fails.
- A domestic stack typically costs 1.5-2.5 times a commercial one and takes longer to implement.
Four buying mistakes
- Buying on one compute number. 1 PFLOPS of FP16 with 32GB of memory cannot run a 70B model; 144GB with low interconnect bandwidth parallelizes badly. Look at the match, not the headline.
- Ignoring lead times. DeepSeek's 160,000 units take over a year to deliver. Popular domestic cards run 3-12 months, longer in 2026 with compute this tight. Put that on the critical path, not a two-week delivery plan.
- Buying in bulk without a POC. Run 2-4 weeks on your own data, concurrency and models, and measure throughput, latency, stability and operational effort. Projects that skip this fail at a high rate.
- Underestimating power and cooling. An 8-card server draws 5-8kW, which office power and air conditioning cannot absorb. Build it into a server room after confirming supply and cooling, or colocate in an IDC. Past 20kW per rack, air cooling is finished.
ACCPC advises on and implements enterprise AI infrastructure: a free feasibility assessment, hardware selection and procurement across domestic and imported cards, server room power and cooling, IDC colocation through a licensed provider (dual mains, UPS, diesel backup and 24/7 staffing), integration, and a year of operations. We also wholesale servers in Guangzhou from Dell, HP, Lenovo, Supermicro, Inspur and Huawei. If the assessment says you do not need local deployment, we will say so and help you configure the cloud API route instead. Call 020-39029800.