Something quietly changed in AI infrastructure: the same GPT model is now available from three different cloud vendors, at three different price points, with three different SLAs, data residency guarantees, and committed-spend discount structures. The model is identical. The price is not.
That's a FinOps lever. Most teams haven't picked it up yet.
In 2025, if you wanted GPT-4o, you called OpenAI directly. In 2026, you have options. AWS Bedrock now offers GPT-5.4 and GPT-5.5 natively in US East regions — same model, billed through your AWS account, applied against your existing AWS committed-spend discounts (Amazon Bedrock Pricing). Azure OpenAI Service has offered GPT models for years under enterprise agreements, with data zone and regional deployment options across 27+ regions (Azure OpenAI Pricing).
The practical result: a FinOps team negotiating cloud spend for the first time has genuine leverage over AI inference costs — not just through model substitution, but through provider arbitrage on identical models.
Here's the live rate card as of June 2026:
| Model | OpenAI Direct (input / output per 1M) | AWS Bedrock (input / output per 1M) | List Premium |
|---|---|---|---|
| GPT-5.5 | $5.00 / $30.00 | $5.50 / $33.00 | +10% |
| GPT-5.4 | $2.50 / $15.00 | $2.75 / $16.50 | +10% |
| GPT-5.4 (Batch/Flex) | $1.25 / $7.50 | $1.375 / $8.25 | +10% |
Sources: OpenAI API Pricing, AWS Bedrock Pricing. At face value, Bedrock costs more. But that's before enterprise committed-spend discounts enter the picture. Large AWS customers routinely receive 15–30% discounts on compute spend through EDPs (Enterprise Discount Programs) — and Bedrock inference qualifies for those discounts. A team spending $500K/month on AWS infrastructure that routes AI inference through Bedrock can see net effective rates below the direct OpenAI price, after discount stacking.
The same logic applies on Azure, where PTU (Provisioned Throughput Units) and enterprise agreements create a separate cost structure entirely — often with fixed monthly commitments that provide cost predictability alongside the discount.
The reason this lever goes unpulled is an org chart problem, not a technical one. In most companies:
This is compounded by the fact that most organizations don't know their effective per-model cost across providers. They see a monthly OpenAI invoice and a monthly AWS bill, but no unified view of what GPT-5.4 actually costs them end-to-end, inclusive of the discount structures they've negotiated. FinOps teams that expertly manage EC2 pricing tiers routinely struggle to get this visibility into AI inference.
The operational challenge with multi-cloud AI procurement isn't the decision — it's the execution. Manually managing which traffic goes to Bedrock vs. direct OpenAI vs. Azure requires application-level changes, credential management across three systems, and per-team implementation effort that doesn't scale.
The architecture that makes this practical is a routing proxy that:
This is what least-cost routing for identical models looks like. It's meaningfully different from model substitution (routing GPT-5.4 traffic to Gemini or Claude) — it's routing the same model call to the cheapest provider currently serving that model. The output is identical. The bill is not.
Worked example. Company has:
If that traffic routes through Bedrock at $2.75/M input (vs. $2.50/M direct), the list-price premium is 10%, or $100K/year gross premium. But with the 20% EDP discount applied, effective Bedrock pricing comes to ~$2.20/M input — $30K cheaper than direct for the input component alone. Output token savings compound further. Total annual delta against direct-API pricing: well above six figures on a $1M spend base.
The same analysis works in reverse: if your cloud discount structure is thin and your AI spend is small, direct API stays cheaper. The routing layer makes the calculation continuous and automatic — it doesn't require a quarterly procurement review to capture the savings.
Cost arbitrage is one reason to route across providers. Compliance is another — and sometimes it's the forcing function.
AWS Bedrock's GPT models are currently available in US East regions only. Azure's GPT deployments span 27+ regions including EU data zones, with GDPR-compliant data residency options and FedRAMP High authorization for government workloads (Azure OpenAI Service). Google Vertex AI offers its own Gemini models — including Gemini 3.5 Flash at $1.50/M input for workloads where a non-OpenAI model fits — across a global network with strong EU data residency controls (Vertex AI Pricing).
A healthcare company processing patient data through an AI workflow may have no choice but to use Azure's HIPAA BAA — but can route all non-regulated traffic to whichever endpoint is cheapest that day. When compliance constrains some workloads and cost optimization governs others, the routing layer becomes the enforcement point — not a manual engineer decision on each deployment.
The window to capture this is not permanent. As more companies figure out multi-cloud AI routing, cloud vendors will rationalize their discount structures and the arbitrage will compress. The teams that instrument their spend now — and build routing infrastructure before they need it — will have a durable cost advantage over teams that negotiate once a year with OpenAI's account team.
Three concrete moves:
Trimio is the LLM API gateway built to make multi-cloud AI routing operational — live cost comparison across providers, automatic least-cost routing for identical models, and Finance Alpha attribution that gives your FinOps team the per-model, per-endpoint visibility they need to capture procurement savings. See how it works.