Google Vertex AI as a Claude Deployment Target

If your infrastructure is on Google Cloud or you operate in the U.S. public sector, Vertex AI is the first choice for Claude deployment. The platform carries full FedRAMP High certification, supports Claude across 35+ global regions with data-residency guarantees, and is one of only two routes where Anthropic guarantees that conversation data does not reach Anthropic systems. This chapter explains why Vertex stands out for GCP-first enterprises and what you need to verify with your account manager before committing.

Vertex AI is a unified, fully-managed AI platform. Claude sits alongside Google's Gemini, open models like Llama and Gemma, and partner offerings. From an enterprise perspective, the critical distinction is that Claude on Vertex uses a managed API with no infrastructure provisioning required. Your team writes applications that call Claude endpoints in your chosen region. Google handles scaling, availability, and compliance. The data residency and Anthropic data-guarantee questions that dominate procurement discussions around other routes are settled here by default.

One piece of context that affects how you read the rest of this chapter. On 22 April 2026, the same day this chapter was verified, Google announced Gemini Enterprise Agent Platform, which Google describes as "the evolution of Vertex AI." The announcement states that future Vertex AI services and roadmap evolutions will be delivered through the Agent Platform, and explicitly names "Anthropic's Claude Opus, Sonnet and Haiku" as supported third-party models in Model Garden. The technical substrate this chapter describes, meaning regions, the data-residency guarantee, FedRAMP High, VPC Service Controls, and Customer-Managed Encryption Keys, appears to continue to apply. Google has not said Vertex AI branding is going away, and the Vertex AI documentation URLs cited in this chapter all resolve as of this date. For procurement purposes, treat Vertex AI as the operational surface you deploy on today, treat Agent Platform as the roadmap wrapper that future Google features will reach you through, and treat the migration path between them as something to raise with your Google Cloud account manager before signing any multi-year commitment.

Claude model coverage on Vertex AI

Vertex AI carries ten Claude model versions as of 2026-04-22. Anthropic's latest releases land on Vertex on the same timescale as other cloud providers. The current lineup includes Claude Opus 4.7, Claude Sonnet 4.6, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, Claude Opus 4.1, Claude Haiku 4.5, Claude Opus 4, Claude Sonnet 4, and Claude 3.5 Haiku. Coverage is comprehensive and updates regularly. The key for your team is that the Claude documentation on Vertex states plainly that "the Anthropic Claude models use a managed API, there is no need to provision or manage infrastructure."

This differs fundamentally from routes where you own the infrastructure or license a model and run it yourself. Vertex AI provides a serverless endpoint. You configure region, model version, and authentication. Google handles the rest.

The breadth of versions available is material because your team may already use an older Claude variant in development or production and want consistency. Verify current coverage via Model Garden before locking in a version for a multi-year contract.

Regional availability and data residency

Vertex AI is available across 35+ regions globally as of 2026-04-22. The regional breakdown spans North America (nine regions including Iowa, Virginia, and Oregon), Europe (eleven regions including Belgium, Germany, and France), Asia Pacific (eleven regions including Tokyo, Singapore, and Mumbai), South America (two regions), Africa (one region), and the Middle East (three regions). This geographic footprint is deep and diversified compared to some competitors.

The essential data-residency statement in the Vertex AI regions reference reads as follows. "Google stores and processes your data only in the region you specify for all features of Vertex AI except for data labeling tasks and any feature in experimental or preview launch status." This guarantee is straightforward. When you configure a Vertex AI Claude endpoint in, say, eu-west-1 (Belgium), your inference requests and responses flow to that region and do not leave it during normal operation.

The caveat is that experimental and preview features do not carry the data-residency guarantee. If your team uses early-access features, verify their status with Google Cloud. For general availability features, the residency guarantee holds.

This regional model enables enterprises to meet GDPR, provincial data-residency mandates, and internal governance policies that restrict data flows across borders. If your compliance framework requires data to remain within a specific geography, Vertex's regional model is one of the mechanisms that makes that possible.

FedRAMP High and public-sector readiness

U.S. federal government agencies, defence contractors, and other organisations operating under FedRAMP mandates can deploy Claude via Vertex AI because the platform holds FedRAMP High certification. FedRAMP is the Federal Risk and Authorisation Management Program, a rigorous assessment baseline that requires compliance with NIST 800-53 standards and demonstrates that systems meet security, operational, and governance controls. FedRAMP High is the second-highest authorization level and is the standard required for agencies handling unclassified but sensitive information.

This certification is not a rubber stamp. It means Google Cloud's Vertex AI infrastructure, including Claude model serving, has been independently assessed against federal requirements. Agencies can use Vertex for AI integration without additional compliance work on top of their existing ATO (Authority to Operate) processes. The certification applies across Americas, EMEA, and Asia Pacific regions, though regional compliance footprints vary.

Public-sector procurement teams often treat FedRAMP as a hard requirement that eliminates entire categories of vendor solutions. Vertex removes that barrier for Anthropic. If you work with U.S. government clients or operate under federal contracts, this distinction is decisive.

The Anthropic guarantee on Vertex

Chapter 1 of this hub established that Anthropic's promise that conversation data does not reach Anthropic systems applies only to two deployment routes: Bedrock and Vertex AI. This is not a semantic distinction. It is the core reason many enterprises choose one of these two routes over others.

Anthropic's Cowork 3P overview states that "the data-residency, compliance, and 'no conversation data sent to Anthropic' statements throughout these pages apply only when inferenceProvider is vertex or bedrock. They do not apply when using Azure Foundry or a gateway." Read plainly, this means your prompts, responses, and file attachments do not reach Anthropic's infrastructure when you use Vertex. They remain between your users and the Vertex endpoint you configure. Conversation history is stored on the user's device, not Anthropic's backend. This is the compliance guarantee that regulated enterprises are often buying.

Bedrock and Vertex are the only two routes that currently carry this guarantee. If your compliance posture depends on the absolute assertion that no conversation data reaches Anthropic, these are the only options. Foundry is in progress for equivalent guarantees, and custom gateway routes may earn them over time, but as of 2026-04-22, neither has Anthropic's guarantee.

VPC Service Controls and encryption

Vertex AI supports both VPC Service Controls and customer-managed encryption keys (CMEK), two mechanisms that enterprises use to lock down access and encrypt data at rest. VPC Service Controls create a security perimeter around a set of Google Cloud resources, controlling egress from inside the perimeter and preventing unauthorised access from outside. CMEK allows you to manage encryption keys in your own key-management system rather than relying on Google's default key material.

Both capabilities are documented as available on Vertex AI generally. Claude-specific configuration guidance for these features is sparse in public documentation and appears to require direct discussion with Google Cloud's security team. This is not a blocker. It means that when you reach the detailed procurement phase, you ask your account manager how to configure VPC Service Controls and CMEK specifically for Claude inference endpoints. These are normal enterprise security questions and account managers expect them.

What the primary sources confirm

The Google Cloud and Anthropic documentation confirms five material facts that shape Vertex as a deployment choice. First, Claude is available on Vertex in ten model versions, from Haiku 4.5 to Opus 4.7. Second, deployment is serverless and managed, requiring no infrastructure provisioning. Third, regional availability spans 35+ locations globally with data-residency guarantees for non-experimental features. Fourth, FedRAMP High certification applies, enabling use in U.S. federal agencies and defence contractors. Fifth, Vertex is one of two routes where Anthropic guarantees conversation data does not reach Anthropic systems.

These five facts are the foundation. The sections that follow identify what the primary sources do not confirm, which is where your procurement work begins.

Material gaps and what to verify

Four critical information gaps exist in publicly available documentation. None is a showstopper, but each requires verification with your Google Cloud account manager.

First, Claude-specific pricing per region is not consolidated in public sources. The generic Vertex AI pricing documentation does not provide token-level rates for Claude models. To get exact pricing for Sonnet 4.6 or Opus 4.7 in your target region, check the Model Garden console within your Vertex AI workspace, or request a formal quote through your account manager. Anthropic's standard pricing for Claude serves as a reference point, and Vertex typically matches that baseline, but regional variations and volume discounts require confirmation.

Second, regional availability is documented at the Vertex AI product level but not per Claude model. Some models may not be available in all 35+ regions. The authoritative source is Model Garden. Open your Vertex AI project, navigate to Model Garden, find the Claude model card you plan to use, and check the region selector. If a model is not available in your required region, ask your account manager about the rollout timeline.

Third, VPC Service Controls and CMEK support for Claude is confirmed at the product level but lacks Claude-specific configuration guidance. When you're at the point of implementing these controls, request a security deep-dive from your Google Cloud sales team. They can walk through the mechanics of locking Claude endpoints behind service perimeters and integrating CMEK.

Fourth, Google's Data Processing Addendum does not explicitly state whether Google uses Claude inference requests for model training or improvement. The DPA limits processing to "service delivery" but defers detailed ML-training terms to appendices and product-specific clauses. Ask your account manager to clarify whether Claude requests are ever retained, used for model improvement, or subject to any opt-out mechanism.

None of these gaps are unique to Vertex. They are standard enterprise procurement activities. Every CTO evaluating Vertex should expect to ask these four questions and should expect direct answers from their account manager.

Sovereign cloud and data-boundary options

Google Cloud offers three sovereign deployment tiers: Data Boundary, Dedicated (with partners like Thales and S3NS), and Air-Gapped. These offerings exist for organisations with stringent data sovereignty requirements such as GDPR right-to-erasure mandates, China or Russia data-localisation restrictions, or France's SecNumCloud certification. For these tiers, the question for your procurement team is simple: is Claude available?

Google's documentation on sovereign cloud does not explicitly state Claude model availability in Data Boundary, Dedicated, or Air-Gapped configurations. If your compliance requirements push you toward one of these tiers, this is a blocking question to resolve with Google Cloud sales before moving further. The sovereign cloud products are real and mature, but Claude coverage in each tier is undocumented.

Cost considerations

Vertex AI pricing for Claude typically aligns with Anthropic's published rates but varies by region. No consolidated pricing matrix exists in public documentation. Your cost comparison with other routes, including Bedrock, Foundry, or data-platform routes, depends on regional availability and whether you qualify for committed-use discounts. Chapter 6 of this hub covers cost modelling across routes in detail. For Vertex specifically, the process is straightforward. Get regional per-token rates from Model Garden or your account manager, and plug them into your capacity planning model. One additional hedge belongs in this section. With the 22 April 2026 announcement of Gemini Enterprise Agent Platform, the billing surface for Claude on Google Cloud may consolidate over time under a new SKU path rather than the current Vertex AI line items. This has not been confirmed in public documentation. For any quote you receive from Google Cloud today, ask your account manager to state explicitly whether the pricing is stable through the Agent Platform transition and whether committed-use discounts you agree on now transfer across that transition unchanged.

When Vertex AI is the right choice

Vertex AI is your clear winner if four conditions hold. First, your infrastructure is or will be on Google Cloud. Vertex integrates with your existing GCP services, IAM, networking, and billing. Moving to a different cloud or bringing in a dual-cloud architecture weakens this advantage. Second, you require FedRAMP High certification. If you work with U.S. government clients or operate under federal contracts, Vertex eliminates a major procurement obstacle. Third, you need data residency in a specific geography. Vertex's 35+ regional footprint and data-residency guarantee make region-locked deployment straightforward. Fourth, you want managed, serverless Claude inference with no infrastructure operational burden. Vertex is the definition of that model.

When Vertex AI is not the right choice

Vertex AI is not the right choice if your constraints pull in a different direction. If your infrastructure lives on AWS, Bedrock is simpler and more integrated with your existing platform. If you have legacy Azure commitments and need to consolidate cost and vendor relationships, Foundry may be the pragmatic choice despite its current limitations on the Anthropic guarantee. If your data resides in a data warehouse like Snowflake or Databricks and your teams prefer warehouse-native inference, data-platform-native Claude deployments such as Snowflake Cortex Code may offer better integration. If you have extremely demanding latency requirements, multi-region deployment introduces trade-offs that require cost and architecture analysis.

What to ask your Google Cloud account manager

Prepare these nine questions before your next account-management conversation. Account managers expect them. They are not overhead.

First, provide a detailed regional pricing matrix for Claude Sonnet 4.6 and Opus 4.7 in the specific regions where you plan to deploy. Second, confirm which Claude model versions are available in each of your target regions. Third, clarify the current Google Cloud Data Processing Addendum on training-data use and request explicit confirmation that Claude inference requests are not retained for model improvement. Fourth, request a security deep-dive on configuring VPC Service Controls and CMEK specifically for Claude inference endpoints. Fifth, confirm availability of Claude in Google Cloud's sovereign cloud options, meaning Data Boundary, Dedicated, and Air-Gapped, if your compliance requirements point toward those tiers. Sixth, clarify Google's audit logging capabilities and whether Cloud Audit Logs capture all Claude API calls. Seventh, define the incident response SLA for Claude-related security events. Eighth, confirm contract terms including whether you can negotiate a Data Processing Agreement specific to Claude, whether you can limit sub-processors, and whether commitment discounts are available. Ninth, ask what the migration path looks like from current Vertex AI deployments to Gemini Enterprise Agent Platform, whether existing pipelines, service accounts, VPC Service Controls perimeters, and CMEK configurations carry across unchanged, whether the data-residency guarantee in your contracted region is preserved without reassessment, and whether existing FedRAMP High authorisation extends to Claude accessed through the Agent Platform surface or requires re-scoping.

Security and compliance quick reference

Google Cloud's compliance footprint is comprehensive. Vertex AI inherits those certifications. ISO 27001 (information security management), ISO 27017 (cloud security), ISO 27018 (personal data protection), SOC 2, and SOC 1 are foundational. FedRAMP High applies across Americas, EMEA, and Asia Pacific. Additional regional certifications span GDPR and HIPAA at the baseline, and include TISAX (Germany), NCSC Cyber Essentials Plus (UK), HITRUST CSF (healthcare), CMMC (defence contractors), DISA, and 20+ additional standards across regions. This is not a substitute for the consolidated controls matrix in Chapter 8, which maps each route's capabilities against specific security domains. But it confirms that Vertex meets the compliance baseline for most regulated industries without additional certification work.

What this chapter does not cover

Provider comparison with Bedrock, Foundry, and data-platform routes lives in Chapter 8. Per-provider security controls and the governance implications of MDM-delivered configuration live in Chapters 8 and 9. Cost modelling and contract negotiation guidance live in Chapter 6. The mental model in this chapter is enough to judge whether Vertex fits your constraints at a high level. The later chapters are where the detailed procurement work happens. Chapter 5 covers data-platform-native Claude deployments in full.

Primary sources

• Google Cloud. Vertex AI overview. Retrieved 22 April 2026.
• Google Cloud. Anthropic Claude models on Vertex AI documentation. Retrieved 22 April 2026.
• Google Cloud. Vertex AI regions and locations reference. Retrieved 22 April 2026.
• Google Cloud. Google Cloud compliance certifications. Retrieved 22 April 2026.
• Anthropic. Cowork on 3P overview page. Retrieved 22 April 2026.
• Google Cloud. Google Cloud sovereign cloud offerings. Retrieved 22 April 2026.
• Google Cloud. Introducing Gemini Enterprise Agent Platform. Published and retrieved 22 April 2026.

Nothing in this article is legal advice. It names regulatory frameworks that apply to enterprise AI deployment and outlines at a high level what each framework requires. Compliance interpretation for your specific regulatory context, jurisdiction, and client contracts must be reviewed with qualified legal counsel. Verify current Anthropic documentation at https://claude.com/docs/cowork/3p/overview before making a procurement decision.