The first measured, cross-cloud VM performance comparison of Mainland China’s four clouds — 35 instances, 3 time slots, ~2,100 data points — finds that same-generation CPUs have converged to within 8%, while storage performance diverges by up to 6.3×. The Chinaready engineering team tested Alibaba Cloud, Tencent Cloud, AWS China, and Azure China across two instance tiers and three synchronized measurement windows in October 2026. Every number in this report is the median of 10 measured runs, with p95 and coefficient of variation disclosed.
This report follows the published methodology — the contract that defines what is measured, how, and under what disclosure rules.
Test matrix
| Cloud | Region | 4 vCPU/16 GB | 8 vCPU/32 GB | CPU generation |
|---|---|---|---|---|
| Alibaba Cloud | cn-hangzhou-k | ecs.g8i.xlarge | ecs.g8i.2xlarge | Sapphire Rapids (2023) |
| Tencent Cloud | ap-beijing-6 | S6.LARGE16 | S6.2XLARGE32 | Ice Lake (2021) |
| AWS China | cn-northwest-1 | m6i.xlarge | m6i.2xlarge | Ice Lake (2021) |
| Azure China | chinanorth3 | Standard_D4s_v5 | Standard_D8s_v5 | Ice Lake (2021) |
All VMs ran Ubuntu 22.04 LTS with identical kernel versions. Each slot created fresh instances, measured inside a synchronized window, then destroyed. Data disks (100 GB) were attached where available; the test suite ran sysbench (CPU, memory), fio (4k random + 1M sequential, read + write), and iperf3 + sockperf network pair tests.
Instance generation matters more than cloud choice. Slot 1 additionally tested AWS m5 (Cascade Lake, 2019) and found single-thread CPU at 1,059 events/s — 65% below the same cloud’s m6i (Ice Lake, 2021) at 2,986 events/s. The initial “AWS is slow” finding was entirely a generation artifact. All results below use the generation-aligned m6i unless explicitly noted.
CPU: convergence at the top
| Cloud | 4c single-thread (eps) | 4c all-thread (eps) | 8c all-thread (eps) | 4c→8c scaling |
|---|---|---|---|---|
| Azure China | 2,992 | 6,218 | 12,440 | 2.00× |
| AWS China (m6i) | 2,986 | 6,195 | 12,398 | 2.00× |
| Tencent Cloud | 2,834 | 5,913 | 11,831 | 2.00× |
| Alibaba Cloud | 2,763 | 5,705 | 11,412 | 2.00× |
(median of 10 runs per slot, 3 slots; cross-slot CV < 0.1% for all clouds)
Single-thread performance has converged. The spread between the fastest (Azure, 2,992) and slowest (Alibaba, 2,763) is 8.3% — a far cry from the 2.8× gap that appears when comparing across CPU generations. All four clouds deliver competitive single-thread performance when using the same silicon generation.
Multi-thread scaling is perfect everywhere. All four clouds achieve exactly 2.00× scaling from 4 to 8 vCPU — no CPU oversubscription detected.
Memory: Alibaba leads with DDR5
| Cloud | Bandwidth (MiB/s) | Relative |
|---|---|---|
| Alibaba Cloud | 28,036 | 100% |
| Azure China | 25,628 | 91% |
| AWS China (m6i) | 25,562 | 91% |
| Tencent Cloud | 24,232 | 86% |
(median of 10 runs, sysbench memory 1M blocks)
Alibaba Cloud’s g8i family uses DDR5 memory with Sapphire Rapids, delivering a 9-16% bandwidth advantage over the Ice Lake / DDR4 trio.
Storage: a 6.3× divergence
| Cloud | 4k random read (IOPS) | 4k p99 latency | 1M seq read (MB/s) | 8c 4k IOPS |
|---|---|---|---|---|
| Azure China | 19,534 | 0.35 ms | 398 | 38,900 |
| Tencent Cloud | 6,072 | 0.52 ms | 154 | 6,072 |
| Alibaba Cloud | 3,936 | 0.45 ms | 145 | 3,936 |
| AWS China (m6i) | 3,100 | 0.60 ms | 129 | 3,100 |
(median of 10 runs; fio libaio direct=1 iodepth=32; root disk, not attached data disk)
This is the largest gap in the entire benchmark — and it is architectural. Azure Premium SSD uses locally attached NVMe storage, while AWS EBS gp3 is network-attached with a 3,000 IOPS baseline. Alibaba’s ESSD and Tencent’s Cloud Premium sit between these extremes. The 6.3× spread between Azure and AWS on 4k random read is the single most consequential finding for any workload that is storage-bound.
At the 8-core tier, Azure’s IOPS nearly doubles to 38,900 (leveraging per-vCPU disk credits), while the other clouds show no scaling — their storage performance is capped by the disk type, not the instance size.
Stability: no evening-peak effect
| Cloud | Slot 1 (morning) | Slot 2 (evening peak) | Slot 3 (next morning) | Cross-slot CV |
|---|---|---|---|---|
| Alibaba Cloud | 2,765 | 2,760 | 2,763 | 0.1% |
| Tencent Cloud | 2,834 | 2,835 | 2,833 | 0.03% |
| AWS China (m6i) | — | 2,986 | 2,984 | 0.03% |
| Azure China | 2,993 | 2,992 | 2,991 | 0.03% |
(CPU single-thread, median of 10 runs per slot)
The evening peak (20:00-23:00 CST) has no measurable impact. Cross-slot coefficient of variation is below 0.1% for all clouds. This is consistent with dedicated (non-burstable) instances receiving guaranteed CPU allocation regardless of neighbor activity.
Price-performance
| Cloud | 4c16g list price | CPU score/¥ | 4k IOPS score/¥ | Region |
|---|---|---|---|---|
| Tencent Cloud | ¥0.99/h | 2,861 | 6,133 | Beijing |
| Alibaba Cloud | ¥1.05/h | 2,631 | 3,749 | Hangzhou |
| AWS China (m6i) | ~¥1.20/h | 2,488 | 2,583 | Ningxia |
| Azure China | ¥2.46/h | 1,216 | 7,941 | Beijing North 3 |
(list prices from official pages, 2026-10-27; CPU = single-thread events/s; IOPS = 4k random read)
Tencent Cloud’s Beijing pricing delivers the best CPU price-performance. Azure’s premium price is partially offset by its dominant storage performance — on a storage-weighted workload, Azure’s IOPS-per-yuan leads.
Field notes
- Tencent Cloud t8 stock: S6.2XLARGE32 was unavailable in ap-nanjing-1 across multiple attempts. Slot 2 and 3 used ap-beijing-6 where both tiers were available at ¥0.99 and ¥1.94/hour respectively.
- AWS m5 → m6i: Slot 1 tested m5 (Cascade Lake) as the initial choice; subsequent slots switched to m6i (Ice Lake) for generation-aligned comparison. The m5 data is preserved as a legacy baseline.
- AWS root disk: Slot 1 used a 20 GB root disk which was too small for fio’s 20 GB test file, causing test hangs. Subsequent slots used 40 GB.
- Network pair tests: iperf3 and sockperf tests were attempted but did not produce valid results in any slot. The iperf3 server started correctly on each VM, but client connections to peers failed. This is a known limitation of the current test configuration and will be addressed in the next wave.
- Instance consistency: Pair VMs (same cloud, same spec, same window) showed <1% variance across all metrics — evidence of consistent provisioning within each cloud.
Disclaimers
This program is independent research, funded by Chinaready, using list prices only, and accepts no vendor money. Chinaready is not affiliated with, sponsored by, or endorsed by any cloud provider named in this report; all vendor names and trademarks belong to their respective owners. Measurements are observations from specific, disclosed time windows and configurations — they are not performance guarantees, SLAs, or certifications, and they do not constitute procurement advice. The research program operates independently of Chinaready’s commercial services.
Dataset
This is a repeatable research program — future waves will extend coverage to additional VM families, regions, and time windows on an irregular cadence. Each wave’s processed dataset is published below as a downloadable JSON file.
- Version: v2026-10 (frozen 2026-10-28)
- Test windows: Oct 27 ~10:00, Oct 27 ~21:00, Oct 28 ~10:00 (all CST)
- VM instances: 35 total (12 + 12 + 11; AWS m5 baseline in Slot 1)
- Measurements: ~2,100 data points (35 instances × 10 rounds × 6 metrics)
Downloads
| Wave | Window | Description | File |
|---|---|---|---|
| Slot 1 | Oct 27 ~10:00 CST | Morning baseline (AWS m5 legacy) | slot1.json |
| Slot 2 | Oct 27 ~21:00 CST | Evening peak (AWS m6i) | slot2.json |
| Slot 3 | Oct 28 ~10:00 CST | Next-morning baseline (AWS m6i) | slot3.json |
| Combined | All three slots | Full dataset with all metrics | combined.json |