AI-Infra Group: Market Research on AI Compute Vendors' Landing Practices (Group Overview & Cross-Comparison)

1. Group Overview

1.1. Research Scope

This group (03-Market Research / 06-AI-Infra Group) conducts market research on the most representative 10 platforms/ecosystems in the 2026 AI compute industry, focusing on "specific vendor landing practices" — that is: who these compute platforms actually sold to, at what deployment scale, what results they delivered, and how well they can support upper-layer agent workloads.

Three categories of research subjects are covered:

1. International leaders and chasers: NVIDIA (GB200 NVL72 + CUDA full stack), AMD (MI355X + ROCm), Google TPU (Ironwood), AWS Trainium/Inferentia (Project Rainier);
2. Domestic leaders and second-tier: Huawei Ascend (CloudMatrix 384 + CANN full stack), Cambricon (Siyuan series), Moore Threads (MTT S5000 + Kuae cluster), Hygon Information / Tianshu Zhixin / Enflame Technology (integration chapter);
3. Open-source software ecosystem: inference serving ecosystem (vLLM/SGLang/TensorRT-LLM/MindIE) and training framework ecosystem (Megatron-LM/DeepSpeed/FSDP/Colossal-AI).

Each platform has one dedicated document (see the Section 6 document navigation), uniformly following an eight-part structure: introduction, terminology, feature description & product line, platform architecture, Harness design (six-layer analysis), actual cases, summary, and references.

1.2. Division of Labor with the 02-Industry Enablement / 01-AI-Infra Group

This group has a clear division of labor boundary with the 02-Industry Enablement / 01-AI-Infra Group:

1.3. Analysis Framework & Rating Basis

All platforms uniformly apply the Harness six-layer capability model defined by the project parameter card:

Rating basis (four levels): Strong (native first-class citizen mechanism, configurable and extensible) / Medium-Strong (has a native mechanism but incomplete coverage) / Medium (requires users to compose it themselves, or only provides basic primitives) / Weak (largely missing, requires external systems to fill in). "Weak (opaque)" means publicly verifiable mechanisms are missing (not that actual product capability is missing).

Methodology statement that must be emphasized: All six-layer ratings in this group are qualitative judgments based on public materials, not measurements against a unified benchmark. There is no unified agent-workload benchmark across compute platforms; performance data is mostly vendor-sourced (repeatedly flagged by this group), and third-party verifiable data exists only for NVIDIA (complete MLPerf submissions), Google TPU (partial MLPerf submissions) and sporadic submissions by AWS/AMD. The ratings reflect "the density and depth of publicly verifiable mechanisms and evidence," not "production-environment measured performance."

1.4. Industry Landscape: Four Layers of Chip, Supernode, Cluster, and Framework

Figure 0-1 | AI Infra industry landscape and the position of this group's 10 platforms (four-layer view)

Note: drawn based on a synthesis of this group's 10 research documents; shipment share is per the IDC 2025 basis (as relayed via Tencent Cloud Community).

1.5. 2025–2026 Key Events Timeline

The following events directly shaped the competitive landscape of this group's 10 platforms, and each document expands on them accordingly:

2. Cross-Comparison Matrix of the 10 Platforms

2.1. Basic Information & Positioning Matrix

2.2. Harness Six-Layer Maturity Matrix

Ratings are qualitative judgments based on public materials, not measurements against a unified benchmark; "Weak (opaque)" means publicly verifiable mechanisms are missing.

Four structural readings of the matrix:

1. The strong players in L1/L4 are software ecosystems, not chip vendors: All the core L1/L4 innovations such as PagedAttention/RadixAttention/KV offloading come from the open-source inference ecosystem (chapter 09); the value of hardware vendors lies in providing physical capacity for these innovations (shared HBM, global memory pools);
2. L6 has two kinds of "Strong": The strength of overseas cloud vendors is "cloud governance out of the box" (GCP IAM, AWS audit), while the strength of domestic vendors is "the only Xinchuang compliance solution" — the customer bases of the two barely overlap;
3. L3 has branched into three species: task orchestration (agent frameworks, within the scope of the 02-AI-Agents group), inference-request orchestration (Dynamo/PD separation), and training-topology orchestration (Megatron parallel strategies) — same name, different things, and they must be distinguished when cited;
4. Cambricon and the integration chapter's "Weak" is a transparency issue: customers building their own upper-layer ecosystems make its Harness capabilities invisible (weak-opaque), which must be distinguished from a real capability gap.

2.3. Ecosystem & Applicability Boundary Quick Reference

3. Domestic vs. Overseas Landscape Comparison

3.1. Market Share Landscape (IDC 2025 basis, B-grade relay)

China's total AI accelerator card deliveries in 2025 were about 4 million units, of which domestic brands accounted for about 1.65 million (41%): Huawei Ascend 812,000 units (about 20.3%, domestic No. 1), Cambricon about 2.9%, Enflame about 1.7%, with MXM roughly on par with Enflame; NVIDIA about 2.2 million units (about 55%, mainly pre-embargo inventory and compliant cards).

3.2. Route Comparison

3.3. Three Overall Conclusions

1. The unit of competition has moved up from chips to systems: The domestic camp has proven with CloudMatrix 384 / the Kuae 10k-card cluster that "bridging the single-card gap with system-level scale" is a viable path, but at the cost of a validation system (MLPerf-level third-party benchmarks) and software ecosystem (the CUDA moat) that still need time;
2. Inference is the domestic camp's window of opportunity: Inference-side technologies such as large-EP inference, PD separation, and KV management only became the main battlefield in 2025–2026, with a smaller generational gap than the training side — Ascend large EP (4× throughput / 15 ms latency, vendor basis) and Enflame inference ramping (66,000 units) both point to this;
3. Multi-stack hardware neutrality has become the standard strategy for leading model vendors: Anthropic is simultaneously betting on NVIDIA GPU + Trainium (Rainier) + TPU (million-unit commitment), and NVIDIA's share advantage is being structurally diluted by "second sourcing + in-house design."

4. Key Claim-Conflict List (Presented Side by Side)

This group follows the discipline of "present conflicts side by side, do not pick one"; all conflicts are listed together below, and this section must be read before citing either party:

5. Selection Recommendations

5.1. Selecting by Constraint Conditions

1. With a hard domestic-substitution constraint: Route by "full-stack need → Huawei Ascend; bare compute only → Cambricon; GPU generality needed → Moore Threads; existing CUDA migration → Hygon; inference cluster → Enflame";
2. Without a domestic-substitution constraint: Route by "performance first → NVIDIA; cost first → AMD/TPU measured; cloud-ecosystem lock-in → the corresponding cloud vendor's in-house chip";
3. Software-ecosystem autonomy: Regardless of hardware choice, default to vLLM/SGLang as the interface layer (OpenAI-compatible API) on the inference side, retaining cross-hardware migration capability.

5.2. Pre-Adoption Checklist

1. Claim verification: Confirm the source grade of any performance/utilization metric (official disclosure / company basis / third-party benchmark / media estimate); this group's information-gap statements in each chapter provide the corresponding verification entry points;
2. Benchmark measurement: Always replace vendor-sourced metrics (MFU, tokens/s, multiples) with measurements on your own workload before they enter a procurement decision;
3. Delivery acceptance: Compute procurement follows the "it only counts once it's on the rack" principle (per Anthropic CPO Mike Krieger), writing delivery acceptance into the contract;
4. Ecosystem lock-in assessment: The migration cost of CUDA/ROCm/XLA/Neuron/CANN should be estimated as part of TCO;
5. Validation-system gap: When selecting a domestic platform, require the vendor to provide MLPerf-level third-party data or accept joint measurement; otherwise start with a minimum-scale PoC;
6. Information currency: All data in this group is current as of 2026-09-12; supernodes/chip generations iterate fast (annual cadence), so note the retrieval date when citing.

6. Document Navigation

7. Research Limitations & Global Information Gaps

Before citing any conclusion from this group, the following global limitations should be read first:

1. The six-layer ratings are not measurements against a unified benchmark: All ratings are qualitative judgments based on public materials; compute platforms have no unified agent-workload benchmark, and cross-comparable data is sparse and mutually incomparable.
2. Vendor self-test vs third party: Except for NVIDIA (MLPerf A-grade), the key performance/utilization numbers of AMD, Moore Threads, and Huawei are all vendor-sourced; this group has flagged this in each place, but readers should not treat vendor claims as verified facts.
3. Domestic vendors' orders and customers: Cambricon customer orders, ICBC orders, etc. are all media estimates/self-media (D-grade), with no official company disclosure; the body presents them only as "according to public reports (not company-confirmed)."
4. Claim conflicts retained item by item: See the Section 4 list; all are presented side by side, not forcibly merged.
5. Official original documents to be verified: The originals of Cambricon's 2025 annual report and Enflame's prospectus should be double-checked against the disclosures on CNINFO/SHSE; gaps in official documents such as the Ascend 910C official spec sheet, MindIE 3.0.0 new features, and Trainium3 full specifications have been listed item by item in each chapter's information-gap statement.
6. MLPerf as the final anchor: The MLPerf figures cited by this group come from vendor-blog relays of MLCommons results; for formal citation it is recommended to use the mlcommons.org results library as the final anchor.
7. Version currency: All version numbers, deployment scales, and feature availability are as of the 2026-09-12 retrieval point; supernodes and chip generations iterate on an annual cadence, with upcoming changes including: Vera Rubin NVL144 (2026–2027), Helios (2026), Ascend 950/960/970 (through 2028), Trainium3 (released 2025-12), and Ironwood global deployment (planned mid-2026).
8. Geopolitical and policy variables: Export controls (H20 series), domestic-substitution procurement policy, and sovereign AI projects (Stargate, HUMAIN, etc.) may all change the landscape in the short term; this group only records traceable facts before 2026-09-12 and makes no predictions.

8. References

1. NVIDIA Q2 FY2026 CFO Commentary — NVIDIA Investor Relations, 2025. https://investor.nvidia.com/files/doc_financials/2026/Q226/Q2FY26-CFO-Commentary.pdf
2. NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 — NVIDIA Technical Blog, 2025. https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0/
3. AMD Instinct MI350 Series GPUs: A Game Changer — AMD Official Blog, 2025. https://www.amd.com/en/blogs/2025/amd-instinct-mi350-series-game-changer.html
4. Ascend AI Industry Summit Successfully Held (300 supernodes, Postal Savings Bank practice) — Ascend Community, 2025-09. https://www.hiascend.com/activities/dynamic-news/555
5. Cambricon 2025 Annual Report Key Points — China Securities Journal / CS Smart Financial News, 2026-03-13. <https://newzzbcx.cs.com.cn/cxnews.html?name=new20260313185352mwlpcfpa&random=dp7iRuOp>
6. Moore Threads 2025 Annual and 2026 Q1 Results Briefing — China Securities Network Roadshow Center, 2026. https://roadshow.cnstock.com/fbh/mexc2025
7. Ironwood: The First Google TPU for the Age of Inference — Google Official Blog, 2025-04. https://blog.google/intl/zh-tw/products/cloud/ironwood-the-first-google-tpu-for-the-age-of-inference/
8. AWS Week in Review: Project Rainier Goes Live (2025-11-03) — AWS Official Chinese Blog, 2025. https://aws.amazon.com/cn/blogs/china/aws-weekly-roundup-project-rainier-online-amazon-nova-amazon-bedrock-and-more-november-3-2025/
9. Enflame Technology Listed on the STAR Market, Opening Up Over 188% — East Money (Xinhua Finance), 2026-09-11. https://finance.eastmoney.com/a/202609113871802756.html
10. PyTorch Foundation Welcomes vLLM as a Hosted Project — PyTorch Official Blog, 2025. https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/
11. Comprehensive Comparative Analysis Report of vLLM, SGLang and TensorRT-LLM — Alibaba Cloud Developer Community, 2025. https://developer.aliyun.com/article/1686693
12. The Rise of Domestic AI Chips and Scenario-Based Selection (IDC 2025 China AI Accelerator Card Market Data) — Tencent Cloud Developer Community, 2026. https://cloud.tencent.cn/developer/article/2686867
13. MLPerf Benchmarks Results Library — MLCommons, 2025. https://mlcommons.org/benchmarks/inference/
14. Project Parameter Card v1.1 (unified basis for the Harness Six-Layer Capability Model and the Harness Loop) — Internal baseline document of this project.
15. R16-AI-Infra Market Group Retrieval Report — Internal retrieval report of this project (4 rounds, 24 searches, about 90 facts, 66 URLs).