I have shipped to all three hyperscalers in 2026. The conventional wisdom (AWS for everything, Azure for enterprise, GCP for ML) is wrong now. Pick by what you actually do.
Compute cost is mostly identical
Source: Cloud provider public price lists, January 2026
Within 5 percent across all three at every tier[1][2][3]. If anyone tells you their workload is on AWS because it is cheaper, they have not done the maths.
AI services where they diverge
| Spec | AWS | Azure | GCP |
|---|---|---|---|
| Foundation models | Bedrock (Claude, Llama, Titan) | Azure OpenAI (GPT-4o, o1) | Vertex AI (Gemini, Claude, Llama) |
| Notable native model | Titan | Phi-4 | Gemini 2.5 |
| OpenAI exclusive? | No | Yes | No |
| Per-token cost vs OpenAI direct | Same | +10% (mostly) | Slightly cheaper |
| Vector DB | OpenSearch, Bedrock KB | Azure AI Search | Vertex AI Vector Search |
Azure has the OpenAI exclusive. If your stack depends on GPT-4o or o1 specifically, Azure is the safest enterprise route. AWS Bedrock has Claude (also available on GCP). GCP has Gemini exclusive plus Claude.
When I pick AWS
Anything where I need S3 specifically (the ecosystem of tools assuming S3 is enormous). Anything where the team already knows AWS deeply. Anything regulated (the FedRAMP / HIPAA / FINMA story is most mature here).
When I pick Azure
Anything where the customer is already deeply on Microsoft 365. Anything where the team needs to invoice through an enterprise agreement. Anything where I want OpenAI-direct without a separate vendor contract.
When I pick GCP
Anything where I want managed Postgres without surprises (Cloud SQL is the cleanest). Anything where I want Gemini specifically. Anything ML-heavy (Vertex AI is the most coherent of the three).
When I pick none of them
Solo projects and small SaaS up to mid scale: Vercel + Supabase + DigitalOcean. The hyperscaler bill stops making sense until you are at meaningful traffic.
What about cost optimisation
Reserved instances save roughly 35 to 60 percent depending on commitment. Spot/preemptible instances save 70 to 90 percent if your workload tolerates interruption. None of this is news. The real cost optimisation is not picking a hyperscaler in the first place if you do not need one.
About the data
A note on what the numbers in this post represent so you can read them with the right confidence:
- "My own bench" rows are personal measurements on my own hardware. They are honest about my setup and reproducible there, but they should not be treated as universal benchmark scores.
- Benchmark numbers attributed to public sources (Geekbench Browser, DXOMARK, NotebookCheck, FIA timing) are illustrative — the trend is what matters, not the third decimal place. Cross-check against the source for anything you would act on financially.
- Client outcomes and ROI percentages in business-focused posts are anonymised composites drawn from my own consulting work. Real numbers, real direction, sanitised so individual clients are not identifiable.
- Foldable crease-depth and similar engineering measurements are estimates pulled from teardown reports and reviewer claims; manufacturers do not publish these directly.
- Forecasts and "what I bet" lines are exactly that — opinions, not predictions with a track record yet.
If you spot a number that contradicts a source you trust, tell me — I would rather correct it than be the chart that was off by 6 percent and pretended otherwise.