Claude Desktop on AWS, GCP & Microsoft Foundry: Per-Cloud Cost Differences for Enterprise Coding

What Anthropic Shipped on June 23

Anthropic announced full-feature Claude Desktop access — Chat, Claude Cowork, and Claude Code integration — for organizations consuming Claude through AWS, Google Cloud, and Microsoft Foundry. Inference stays in the customer's cloud environment. Conversation history stores locally on the device. IT teams get IAM Identity Center, Workforce Identity Federation, Microsoft Entra ID, and Okta sign-in support. Policy templates export to Intune, GPO, and Jamf. Microsoft 365 connectors run through Entra apps with GCC High and DoD endpoint support.

For enterprise coding teams, the operational story is clean — but the cost picture varies meaningfully across the three clouds. Choosing where to consume Claude is no longer just a vendor preference; it's a budget decision worth several percentage points on a multi-million-dollar AI spend.

Per-Token Pricing Across Clouds

The same Sonnet 4.6 call carries different effective costs depending on the cloud. Three drivers create the spread:

Base token markup. AWS Bedrock, Vertex AI, and Microsoft Foundry each negotiate their own wholesale rate from Anthropic and apply their own retail markup. Historically the spread has been 0-5% — small per call, meaningful per quarter.

Egress pricing. Claude Desktop pulls inference from the cloud and stores history locally. That means data egresses your VPC for every conversation. AWS charges roughly $0.09/GB out, GCP $0.12/GB out, Azure $0.087/GB out. For teams running 50-100GB of Claude traffic per month, that's $5-$12/month per developer, materially.

Committed-spend discounts. All three clouds offer enterprise discount programs (EDPs) that pull Claude consumption into broader commitments. AWS's Private Pricing Agreements typically deliver 10-25% off list. GCP's Committed Use Discounts go to 30-40% on three-year terms. Microsoft Foundry leans on Azure Consumption Commitments for similar ranges.

A Realistic Cost Model

Picture a 200-developer engineering org running Claude Code across all desks, at $80/developer/month in raw token spend (typical for active users on Sonnet 4.6 with occasional Opus 4.8). Annual all-in across the three clouds, before EDPs:

The all-in spread is roughly 2.4% — small but real. After applying typical enterprise discounts, the numbers compress further; the cloud you already commit to almost always wins on price.

Identity Integration Costs the Same Until It Doesn't

All three clouds support the same identity stack on paper — Entra ID, Okta, IAM Identity Center, Workforce Identity Federation. The cost difference shows up in deployment friction:

Microsoft Foundry + Entra: If you're already an M365 shop, this is a same-day rollout. M365 connectors flow through your existing Entra app registrations. Practical setup cost: ~16 engineering hours.

AWS Bedrock + IAM Identity Center: Clean for AWS-native shops, more friction if you SSO through a third party. Typical setup cost: ~32 engineering hours.

Google Cloud + Workforce Identity Federation: The most flexible but the most configuration-heavy. Workforce Identity Federation works across all major IdPs but requires explicit attribute mapping. Typical setup cost: ~48 engineering hours.

At a $150/hour blended rate, the deployment delta runs $2,400-$7,200 — one-time, but it does shift the first-year break-even.

Data Residency: A Cost Lever, Not Just Compliance

Microsoft Foundry's GCC High and DoD endpoint support is the only path to Claude on classified or regulated US government workloads. Vertex's region pinning is the cleanest for EU sovereignty requirements. AWS Bedrock's regional availability is broadest but doesn't yet cover every isolated region.

Treating residency as a cost variable is useful because the alternative is paying twice — once for an unrestricted cloud and again for a bespoke compliance program. Picking the cloud that natively meets your residency profile usually saves more than the 2-3% pricing spread.

A Three-Question Decision Path

Most enterprise teams converge on the right cloud by answering three questions in order:

• Where does your committed spend live? If you already have a $5M+ Azure commitment, Foundry leverages that capacity at marginal cost
• What residency boundary do you need? US Federal → Foundry. EU sovereignty → Vertex. Multi-region with broad SOC2 → Bedrock
• What identity stack do you already run? M365/Entra → Foundry. Okta-only → AWS or GCP, both support it cleanly

In nearly every case, the cloud that wins on those three questions also wins on price after enterprise discounts apply. The pricing spread isn't the decision driver — it's the tiebreaker when commitments and compliance don't dictate the answer.