Which Cloud Provider to Choose? AWS vs Alibaba Cloud vs Tencent vs Azure vs GCP

Which cloud provider should you choose? There is no default answer, because "best" depends on where your users are, what kind of workload you run, how big the bill is, and whether the account has to survive a corporate security audit. The most common mistake is comparing sticker price and region count — those two decide almost nothing about what your bill and failure rate look like three months in.

This compares AWS, Alibaba Cloud International, Tencent Cloud International, Microsoft Azure and Google Cloud: region coverage and latency to Greater China and Southeast Asia first, then billing models and compliance, then a verdict for four workload families — media, gaming, data and AI, and traditional enterprise systems — and finally a seven-question checklist. All pricing is expressed as discount ranges; treat official live quotes as the source of truth.

Short answer: no best cloud, only the matching one

If you want the three-sentence version:

• Global multi-region and the widest service catalogue → AWS. Most regions, longest product list, deepest third-party ecosystem; the trade-offs are a complex console and a long bill.
• China-to-global expansion with a Chinese-speaking team → Alibaba Cloud International. Chinese-first console and docs, solid interconnect to China, and a good spread across Europe and the Middle East.
• Media, live streaming and multiplayer gaming → Tencent Cloud International. Its transcoding, live and VOD product set is the most complete of the five.
• Enterprise security audit plus an existing Microsoft stack → Azure. Entra ID, Microsoft 365 and Azure DevOps line up, and the certification list is the easiest one to defend to a security team.
• Big data, machine learning and Kubernetes → GCP. The most mature BigQuery / GKE / Vertex AI ecosystem on a stable global network.

These five are a starting point, not a conclusion. The same provider can swing by more than 20% in cost-effectiveness depending on region and commitment term, so the real answer has to land on your specific workload. Every provider has deep pages on instance families, storage and discount mechanics, all linked from the cloud provider overview — follow the one or two you shortlist.

The comparison table

Here is the whole picture. The first columns decide the technical choice; the last two decide whether it can actually ship.

Two caveats when reading this: the discount ranges are typical public-channel magnitudes, not quotes — real pricing depends on instance family, region, term and volume, so treat official live quotes and written channel quotes as the source of truth; and region counts change every year, so the figures above are the public-record view, best confirmed against each provider's own region list.

Region coverage: who has more nodes

The number of regions has no value by itself. The number you can actually use does. It shows up in four places: whether residency rules are satisfiable, whether multi-region disaster recovery is realistic, whether users can connect nearby, and whether you have an alternative region when one goes bad.

• Multi-market or globally synced workloads: AWS and GCP have the densest footprints; a new market usually has available capacity within weeks
• Hard residency requirements (data must sit in the EU, Southeast Asia or a specific country): filter on region existence first — it can halve the candidates immediately
• Disaster recovery: the number of availability zones per region decides whether you can use a multi-AZ design without extra cost; few AZs means a zone fault hits both nodes
• Emerging markets: Africa, the Middle East and Latin America are usually opened by AWS, GCP or Azure first, with Alibaba and Tencent following later

There is a second layer people skip: the primary region decides the cost structure of everything after it. It decides which jurisdiction runs your compliance review, whether on-call has to work overnight, where database backups legally sit, and what cross-region traffic costs once you expand.

• Choose the primary region for residency requirements first, not for physical proximity to the office
• Keep the primary region inside your team's working hours — overnight incident response stretches mean time to recovery badly
• Discounts are often tiered by region, and commitment pricing can be materially better in one region than another, so price them separately
• Egress unit price in the primary region belongs in the cost model: the same instance can cost more than double in traffic across different regions

Latency: real ranges to Greater China and Southeast Asia

For Chinese teams, latency usually matters more than node count. Below are typical RTT ranges from southern and eastern China to each provider's best Asia-Pacific node. They vary by carrier, hour and plan — measure your own.

Notice the last row inverts: for Southeast Asian users, Singapore is usually faster than Hong Kong. By the same logic Tokyo wins for Japan. "Pick Hong Kong" is a rule of thumb that holds for mainland China, Taiwan and Hong Kong — and stops there.

• Where all three have a node, latency differences are often smaller than carrier differences — China Telecom, Unicom and Mobile can differ by 2×
• Peak hours (20:00–23:00) are the truth; loss on shared lines can rise from under 1% to 3–5%
• Public clouds rarely peer directly, so cross-cloud calls pay both round-trip latency and egress fees
• How to measure: launch a minimal instance in each candidate region and run ping and traceroute from your real user network for 24 hours

There is also a layer of latency people ignore: a static asset served from a CDN edge and a dynamic request that goes to origin are completely different latencies. The first is typically 10–30ms and stable for most users; the second is governed by regional RTT, queueing and connection reuse, and varies a lot. So split the "site is slow" problem in two: if images and scripts are slow, fixing the CDN beats changing region; if the API itself is slow, only a closer region helps.

A rough but useful rule: every additional origin domain in your first paint costs one more cross-border round trip. Merging domains, trimming third-party scripts and self-hosting fonts and icons usually pays off faster than a region migration, and without any downtime.

Pricing models: five different philosophies

You cannot compare the five on headline unit price, because their discounts hang off different commitments.

Any cross-provider comparison has to include these four items, or the "cheaper" number is fiction:

• Egress traffic: the least predictable line on the invoice; a low CDN hit rate or unoptimised images multiplies it
• Availability-zone surcharges: some network and managed services are priced per specific zone
• Licences and images: Windows, Oracle, SQL Server and commercial images can bill hourly on top
• Minimum spend and prepayment terms: cheap plans often require a one-year term or a large upfront payment, and shortening it removes most of the discount

The most common way commitment discounts lose money is not overpaying — it is over-committing. Reserved instances and committed-use discounts bill against what you promise to use: if actual usage only reaches part of the commitment, you still pay the full commitment, and the unused portion is generally not refunded.

1. Run on-demand for two full billing months first and get a stable usage curve
2. Commit only the low baseline of that curve and leave peaks on on-demand or spot
3. Start with one year, not three, and confirm usage and architecture before extending
4. Scope savings plans to the instances you know are long-running; leave elastic capacity uncommitted

Compliance and certification

The moment your company has to pass ISO 27001, a PRC classified-protection review, GDPR obligations or a customer security questionnaire, the certification list decides which providers are even eligible. There is no shortcut here.

• International baseline: ISO 27001, ISO 27701 for privacy, and SOC 1/2/3 reports — all five have them in major regions
• PRC regulation: if you need MLPS 2.0 level 3 or a CSP licence, Alibaba Cloud and Tencent Cloud have direct mainland entities; a purely overseas region does not satisfy it
• Industry addenda: HIPAA BAA, PCI DSS, C5, IRAP and FedRAMP are not satisfied by "we are certified" — confirm the specific region you plan to use is in scope
• Residency: verify that data really stays in the selected region, and ask where control-plane metadata lives

Tooling differs too: AWS Artifact plus Organizations SCPs for centralised governance, Azure Entra ID reusing your existing conditional-access policies and audit logs, and GCP Assured Workloads as the least fiddly path for compliant data and AI workloads.

Workload 1: media, VOD and live streaming

In media workloads the cost is not compute — it is transcoding and egress. So the order is: transcoding capability and billing granularity first, CDN coverage and unit price second, servers last.

Two things get missed in media budgets: transcode template count, which decides whether you encode every rung once or pay repeatedly; and whether recording and restream are billed separately, which shows up clearly at peak concurrency.

Workload 2: gaming and real-time multiplayer

Multiplayer games are far more latency-sensitive than websites. A 1ms jitter spike is nothing on a page and a disconnect in a mobile shooter. So the metrics that matter are jitter, packet loss and node density inside the matchmaking radius — not average latency.

• Managed sessions and matchmaking: AWS Game Servers, Azure PlayFab and Google Cloud Game Servers ship session allocation, player data and leaderboards, saving you a custom matchmaking service
• Asia-Pacific density: Tencent Cloud International has deeper experience in multiplayer sessions and mini-program backends
• Protocol and orchestration: large-room sync is highly sensitive to your frame-sync versus state-sync choice; the orchestration layer often matters more than the hardware
• Anti-cheat and identity: managed anti-cheat and account security are separately billed — price them in early

Practical advice: stand up a minimal environment on two providers and load-test with real room counts and real lossy-network conditions, for a full night. Game traffic is spiky, and a daytime benchmark will mislead you.

Workload 3: data platforms and AI

This is where the five diverge most, and where the conclusion is clearest.

• GCP: the least operational work around BigQuery with query-based billing, a strong managed Kubernetes story, and a coherent Vertex AI toolchain
• AWS: the widest catalogue of managed model services, and the deepest data-engineering community built up over years
• Azure: ties into Power Platform and Fabric, and suits organisations whose BI already lives on the Microsoft stack
• Alibaba and Tencent: fast-moving model and AI platforms with strong Chinese documentation and domestic capacity — but confirm availability in your target regions before committing

When costing this, GPU unit price is rarely the deciding variable. Three others often dominate:

1. Per-token versus per-second billing: on long-context workloads the two can diverge a long way
2. GPU quota and spot availability: hot families are capacity-constrained in some regions, and a discount on capacity you cannot get is worth nothing
3. Egress and request counts: repeatedly reading training data from object storage makes storage read requests a real line item

Workload 4: traditional enterprise systems

If your stack is .NET, Windows Server, SQL Server or Active Directory, the answer is rarely in doubt: Azure has the lowest migration friction, because identity, operations and backup patterns can be reused wholesale.

• Microsoft-first stack (SQL Server, C#, domain controllers, IIS) → Azure; Entra ID maps onto existing conditional-access policies
• Java, Oracle, SAP estates → AWS or Alibaba Cloud, where third-party images and middleware certifications are broader and operational knowledge is more widely shared
• Domestic finance and government workloads → Alibaba Cloud or Tencent Cloud mainland entities, where the compliance chain is much shorter
• Cross-border commerce → Alibaba Cloud International or AWS, decided by where your buyers and fulfilment are

This category also hides a cost people forget: migration labour. The same module is written differently on every cloud, and manual rework can exceed three years of infrastructure savings. Put "migration hours × loaded hourly rate" into the TCO instead of comparing server unit prices alone.

And the trap specific to this audience: adding a second cloud "for backups" and ending up with permanent active-active. The sane version uses the second cloud only for interruption-tolerant work — backup archives, cold data, DR drills — not for online traffic. Cross-cloud replication fees plus two sets of operational knowledge usually cost more than the availability gain is worth.

Decision checklist: seven questions

Answer these in order and the field narrows to one or two providers on its own.

1. Which three countries or regions are your users in? → Fix the target regions and drop any provider without them.
2. What are your measured peak-hour RTT and packet loss? → If you have not measured, stop here; region choice is not reversible.
3. Which of the four workload families are you? → Media, gaming, data/AI and enterprise systems each have a clear leader.
4. Are there mandatory compliance requirements? → MLPS, HIPAA, FedRAMP or IRAP eliminate providers that lack the relevant regional accreditation.
5. What is the existing stack? → Staying usually costs less than switching, unless the gap is large.
6. What is the monthly bill size and commitment term? → This decides whether you qualify for good pricing and whether commitment discounts make sense.
7. Will a second market open within 12 months? → If yes, favour region density and service breadth, to keep the exit path open.

After these you should have a primary and a backup. The sensible next step is one week of on-demand instances on each, running real traffic, quantifying bill, latency and operational effort together. Only then negotiate, and get the discount basis, billing cycle and delivery time written down. We help open accounts and act as payment intermediary for all five providers, with discounts coming from official volume channels; specific pricing follows official live quotes and whatever is confirmed in writing before you order. Where the discount actually comes from, which variables move it, and when you should not use an agent at all: how AWS discounts work.