01 The platforms
| Platform | GPU families | Where it wins | Landing zone |
|---|---|---|---|
| Alibaba Cloud | ecs.gn8is (L20) · gn7i (A10) · ebmgn8v (H20) | The natural home for any workload where Asia-Pacific or mainland latency matters commercially, and competitive on price against the American platforms for comparable capacity. | Foundations |
| AWS | p5 / p5e (H100, H200) · p4d (A100) · g6e (L40S) · Inferentia2 | Best option when the rest of the platform is already on AWS and the data gravity is real. Managed container services for the workload itself, with the platform's own metered AI service ke... | Foundations |
| AWS China | g5 · p4d (Beijing / Ningxia, subject to allocation) | For companies that already run AWS globally and need a mainland presence that their existing tooling can mostly address. The partition boundary is real — plan for duplicated IaC, not a co... | Foundations |
| Tencent Cloud | GN10Xp (V100) · GT4 (A100) · PNV4 (A10) | A pragmatic second mainland option, particularly where Alibaba Cloud quota is tight or where an existing WeChat or game-adjacent integration already lives here. | Foundations |
| Huawei Cloud | Ascend 910B (ModelArts) · Pi / P-series NVIDIA where available | For organisations that need a non-NVIDIA path, or whose mainland customers require a domestic stack end to end. Expect to spend time on MindSpore/CANN compatibility rather than on the def... | Foundations |
| Baidu AI Cloud | Kunlun P800 · A800 / A100 where allocation allows | Considered when ERNIE is specifically required, when Kunlun capacity is easier to get than NVIDIA, or when a mainland partner has standardised on it. | Foundations |
02 How to choose
In order of how much it actually matters.
- Where your data already is. Moving inference to the data is almost always cheaper than moving data to the inference, and egress charges are how clouds make the second option expensive.
- Where your users are. A model answering in 800ms plus 250ms of transatlantic latency is a noticeably worse product than the same model answering locally.
- What jurisdiction requires. Sometimes this is the first consideration rather than the third, and when it is, it removes most of the list immediately.
- What capacity you can actually get. Published instance types and available instance types are different sets. We check before promising.
- Price. Genuinely last. The difference between providers is real but smaller than the difference between 25% and 80% GPU utilisation on any one of them.
MULTI-CLOUD, HONESTLY
Usually a bad idea. Occasionally the only idea.
Running the same stack on two providers doubles the operational surface for a benefit most companies never collect. We recommend against it by default.
Where it earns its cost: a jurisdictional split, where an EU deployment and a mainland China deployment are separate by design rather than by redundancy. That is not multi-cloud for resilience — it is two single-cloud deployments that happen to share a config repository, which is a much better thing to operate.
