EU · GDPR Art. 28 · Intel TDX · Zero Retention

VoltageGPU vs Azure OpenAI Service

Azure OpenAI Service is operated by Microsoft Corporation through its Azure regional contracting entities; it is not affiliated with VoltageGPU. Microsoft's "Azure Confidential Inferencing" is a separate preview feature distinct from the standard Azure OpenAI Service that this page compares against.

Microsoft announced Azure Confidential Inferencing in 2024 and opened a gated preview in 2025. The architecture is right, the team is credible, and the market validation is welcome. As of May 2026 it is still preview on a narrow SKU set. VoltageGPU has been shipping the same Intel TDX + NVIDIA Confidential Computing pattern in GA for a year, on open-weight models, under a French controller, accessible without an Azure enterprise contract.

Headline pricing

Per-million-token list price by model tier. VoltageGPU rows are TEE-attested (Intel TDX). ", " means the competitor does not publish a comparable SKU. Pricing stays in sync with /pricing.

TierVoltageGPU (TEE)Azure OpenAI Service
Cheap mid-size workhorse (~32B class)
Qwen3-32B-TEE
in $0.1500 · out $0.4400 / 1M tok
gpt-4o mini
in $0.1500 · out $0.6000 / 1M tok · closed-weight, no TEE in GA; confidential inferencing is preview-only
Mid-size general-purpose (~30B class)
gemma-4-31B-turbo-TEE
in $0.2400 · out $0.7000 / 1M tok
gpt-4o
in $2.50 · out $10.00 / 1M tok · closed-weight, no TEE in GA; confidential inferencing preview only
Frontier MoE / heavy reasoning
Qwen3.5-397B-A17B-TEE
in $0.7200 · out $4.33 / 1M tok
o1
in $15.00 · out $60.00 / 1M tok · closed-weight reasoning; no TEE in GA
Premium frontier (GPT-4 Turbo class)
Qwen3.5-397B-A17B-TEE
in $0.7200 · out $4.33 / 1M tok
gpt-4 turbo
in $10.00 · out $30.00 / 1M tok · closed-weight, no TEE in GA
Confidential techIntel TDX + trust-domain GPU isolationAzure Confidential Inferencing, preview only (Intel TDX + NVIDIA Hopper CC, gated)
AttestationIntel DCAPMicrosoft Azure Attestation (MAA), Intel TDX + NVIDIA CC quotes, preview tier only
BillingPer-token, OpenAI-compatiblePer-token (Azure-metered), reservation/PTU commitments available; usage rolls into Azure invoice / EA
OperatorVOLTAGE EI (France)Microsoft Corporation (US, Delaware), Azure regional contracting entities for EU customers
Setup~30 sec, drop-in base URLHours to weeks, Azure subscription, OpenAI quota approval, possible EA / procurement; preview confidential inferencing: months (gated application)
JurisdictionEU / GDPR Art. 28US (Cloud Act exposure)

Microsoft announced confidential inferencing. We have been shipping it.

In 2024 Microsoft announced "Azure Confidential Inferencing", a product that runs Azure OpenAI workloads inside Intel TDX confidential VMs with NVIDIA Hopper Confidential Computing on the GPU side, attested through Microsoft Azure Attestation. The team behind it is credible, the silicon partnerships are real, and the threat model is exactly the one this comparison page treats as the structural problem in cloud AI. The fact that the world's largest cloud has put serious engineering behind Intel TDX inference is the most important market validation the confidential-compute thesis has had to date, and it deserves to be acknowledged on its own terms rather than treated as a competitive threat.

The honest factual position as of May 2026 is that Azure Confidential Inferencing is still in gated preview. The feature was previewed publicly in 2025, capacity has been scarce, the application process is enterprise-gated, and the supported SKU and region footprint is narrower than the general Azure OpenAI fleet. Standard Azure OpenAI Service, the product 95%+ of Azure OpenAI customers actually use today, runs on conventional Azure GPU infrastructure without TDX, without GPU CC, and without per-session attestation. The architectural posture Microsoft has announced is the right one; the shipping availability for that posture is not yet GA.

VoltageGPU has been operating the same architectural pattern in GA for roughly a year. Intel TDX guest VMs with AES-256 memory encryption, NVIDIA Hopper Confidential Computing binding the GPU to the TDX quote, AES-encrypted PCIe between CPU and GPU, and an Intel DCAP attestation quote signed under the Intel root certificate, produced fresh for every confidential session and re-verifiable offline by any auditor. Sixteen open-weight TEE-attested models on an OpenAI-compatible API. No Azure subscription, no quota approval, no enterprise agreement negotiation. When Azure Confidential Inferencing reaches GA the comparison will tighten significantly; until then, the choice between us and Microsoft on the confidential dimension is the choice between a production deployment and a waitlist.

Open-weight TEE vs closed-weight ecosystem

Azure OpenAI's catalog is the OpenAI model family. GPT-4o, o1, the GPT-5 line, Whisper, DALL-E 3, text-embedding-3, these are proprietary closed-weight models, exclusive to Microsoft and OpenAI commercial channels. They are also, in several categories, the strongest models on the market. If a workload depends on a behavior specific to gpt-4o's training, on o1's reasoning trajectory, or on the GPT-5 family's system prompt fidelity, the workload needs Azure OpenAI and no amount of open-weight alternative will reproduce that exact behavior. We cannot legally serve those models and we do not pretend otherwise.

VoltageGPU's catalog is open-weight TEE-attested. Qwen3-32B-TEE at $0.15 in / $0.44 out per million tokens, gemma-4-31B-turbo-TEE at $0.24 / $0.70, Qwen3.5-397B-A17B-TEE at $0.72 / $4.33, plus thirteen more across the reasoning, code, vision, and multilingual axes. The "-TEE" suffix is not marketing, it is the contract that the inference workload runs inside an Intel TDX guest on a Confidential-Computing-enabled NVIDIA GPU, and that the runtime exposes an attestation quote the customer can verify. The prices are one order of magnitude below GPT-4 Turbo / o1 because the underlying compute is amortised differently and because the model weights are open and can be served competitively across providers.

The buyer-level decision is therefore not "VoltageGPU vs Azure OpenAI" in absolute terms; it is "open-weight TEE-attested inference now, or closed-weight OpenAI inference inside an Azure enterprise contract". For workloads that genuinely need GPT-4o's exact behavior or o1's reasoning trace and where the Azure ecosystem is already in place, Azure OpenAI is the correct answer and we are happy to say so. For workloads where an open-weight 32B model is sufficient, where the data must demonstrably never leave a TEE, and where the buyer has no appetite to negotiate Azure procurement before shipping a feature, VoltageGPU is the production-ready alternative and we are it today, not in preview.

Where Azure OpenAI wins, and the list is not short

It would be dishonest to write this page without naming the categories where Azure OpenAI is straightforwardly the right vendor. The exclusive closed-weight model access is the largest one: gpt-4o, o1, gpt-4-turbo, the GPT-5 family, DALL-E 3, and Whisper are not available through any other channel at the same fidelity, and a meaningful share of enterprise AI workloads has standardised on the exact behavior of those specific models. We do not have an answer to "I need gpt-4o tool-calling to behave exactly the way gpt-4o behaves" and we do not pretend to.

The second category is compliance certification depth. Microsoft Azure holds essentially every certification a Fortune 500 procurement team checks for: SOC 2 Type II, ISO 27001 / 27017 / 27018, HIPAA with BAA, FedRAMP High, ITAR, IRAP, C5, ENS High, the EU Data Boundary contractual commitment, plus the matching regional accreditations across most jurisdictions. VoltageGPU operates on a deliberately tighter posture, GDPR Article 28 with cryptographic technical measures through Intel DCAP attestation, French operator, EU residency, which is the right answer for the European-regulated mid-market we serve, but it is not a substitute for an enterprise contract that needs FedRAMP High to be packaged alongside Microsoft 365, Azure AD, Defender, and Power BI for a single procurement cycle. For that buyer, Azure is the only vendor in the building and we should not be the recommendation.

The third category is existing Azure spend. A buyer with a $20M / year Azure enterprise agreement, identity already federated through Azure AD, data lakes in ADLS Gen2, and security tooling routed through Microsoft Sentinel will pay a real switching cost to add a separate provider for inference. If the workload does not require TEE attestation today, that switching cost is hard to justify against marginal Azure OpenAI usage on the existing EA. We are explicit about this: VoltageGPU is the right answer when TDX-attested inference today is worth more than the Azure ecosystem integration, and it is not the right answer in the inverse case. The fourth category, EU Data Boundary, is Microsoft's contractual commitment to keep EU customer data inside EU regions for the core services it covers; it is serious work and credible work, even though the constraint is contractual rather than cryptographic. For buyers who can satisfy their regulator with a contract, EU Data Boundary plus Sweden Central / France Central / Germany West Central placement is sufficient. For buyers whose regulator (CNIL on Art. 9 personal data, HDS, MiFID II under recent ESMA guidance, PCI DSS) has begun to require the technical measure be enforced by hardware attestation rather than by contract, the silicon path is the architectural answer and VoltageGPU is it.

FAQ

Is Azure Confidential Inferencing the same thing as VoltageGPU?

Architecturally it is the same pattern, Intel TDX confidential VMs with NVIDIA Hopper Confidential Computing on the GPU side, attested through a cryptographic quote. Microsoft uses Microsoft Azure Attestation (MAA) as the verifier; VoltageGPU uses Intel DCAP attestation directly under the Intel root certificate. The substantive difference as of May 2026 is shipping status: Microsoft announced Azure Confidential Inferencing in 2024, opened a gated preview in 2025, and the feature remains preview with limited SKU and region availability. VoltageGPU has been shipping the same architectural posture in GA for roughly a year, on sixteen open-weight TEE-attested models, accessible without an Azure subscription. When Azure Confidential Inferencing reaches GA the products will become genuine peers on the confidential dimension, differentiated mainly by model catalog (closed Azure / OpenAI models vs open-weight TEE models) and by jurisdiction (US operator with EU contracting entities vs French operator under SIREN 943 808 824).

Why is Azure OpenAI more expensive per token?

Because they are not the same product category. Azure OpenAI is Microsoft's exclusive enterprise distribution of OpenAI's proprietary closed-weight models, gpt-4o, o1, gpt-4-turbo, the GPT-5 line, and the price reflects the OpenAI commercial split, the closed-weight licensing economics, and Microsoft's enterprise distribution margin. VoltageGPU is open-weight TEE-attested inference: Qwen3-32B-TEE at $0.15 in / $0.44 out per million tokens, Qwen3.5-397B-A17B-TEE at $0.72 / $4.33, sixteen models served on Intel TDX + NVIDIA Confidential Computing hardware. The per-token cost gap between gpt-4o ($2.50 / $10.00) and gemma-4-31B-turbo-TEE ($0.24 / $0.70) is roughly 10× to 15×, but it is not apples-to-apples, the buyer is paying for closed-weight model behavior on one side and for TEE-attested open-weight inference on the other. For workloads where an open 30B model is sufficient and confidentiality is the binding constraint, VoltageGPU is the cost-efficient answer. For workloads that genuinely depend on gpt-4o's exact behavior, Azure OpenAI is the only credible vendor and the price is what it is.

Does Azure OpenAI offer EU data residency?

Yes, and it is excellent. Azure OpenAI deployments can be pinned to several EU regions, Sweden Central, France Central, Germany West Central, Switzerland North, and West Europe among them, and Microsoft's EU Data Boundary contractual commitment keeps the covered customer data inside EU regions for the in-scope services. For data-residency-as-contractual-commitment, Azure is best-in-class. The constraint EU residency does not satisfy is the technical-measures clause when a regulator (CNIL on Article 9 personal data, the French HDS framework for health data, recent ESMA guidance on MiFID II algorithmic trading, PCI DSS under recent enforcement patterns) requires the cloud operator to be mathematically constrained from reading workload memory. EU placement under a contractual DPA is not the same evidence as an Intel DCAP attestation quote produced fresh for every confidential session, the contract is European either way, but only the silicon route produces the cryptographic technical measure. Azure Confidential Inferencing in preview will produce that evidence (through MAA) when it reaches GA; VoltageGPU produces it in GA today.

Can I use Azure OpenAI for GDPR-sensitive workloads?

For most GDPR processing the answer is yes, Microsoft signs an Article 28 Data Processing Agreement, the EU Data Boundary commitment covers core services, and the regional deployment story is solid. The mainstream Article 6 / Article 9-with-explicit-consent posture is well served. Where the architectural answer changes is when the technical measures clause of the DPA needs to be cryptographically enforced rather than contractually promised, Article 9 special-category data without explicit consent, professional secrecy (bar associations, medical confidentiality, banking secret), French HDS-scope health data, and the new EU AI Act high-risk obligations on sensitive personal data. In those cases CNIL and equivalent supervisory authorities have begun to require hardware-attested confidential compute as the technical measure. Azure standard OpenAI does not produce that evidence; Azure Confidential Inferencing will, once it reaches GA; VoltageGPU produces it today through Intel DCAP attestation. The decision rule is: if a contractual DPA plus EU placement satisfies the buyer's regulator, Azure OpenAI is fine. If the regulator wants the silicon, VoltageGPU is the GA option.

Should I switch from Azure OpenAI to VoltageGPU?

Probably not as a wholesale replacement, and we will not pretend otherwise. If the workload depends on gpt-4o, o1, gpt-4-turbo, the GPT-5 family, DALL-E 3, or Whisper specifically, those are exclusive Microsoft / OpenAI commercial channels and we cannot serve them. If the buyer has a meaningful Azure enterprise agreement, Azure AD federation, ADLS data lakes, and Microsoft compliance certifications (FedRAMP, ITAR, IRAP) bundled into the same procurement, marginal Azure OpenAI usage on the existing EA is structurally the lowest-friction path and VoltageGPU is not. Where switching is justified: workloads where an open-weight 30B-to-400B model is sufficient and per-token economics matter (Qwen3-32B-TEE at $0.15 / $0.44 is roughly 10× cheaper than gpt-4o), workloads where the regulator requires hardware-attested confidential compute today rather than waiting for Azure Confidential Inferencing GA, and product teams that need to ship without an Azure procurement cycle, sign-up is self-serve, $5 referral credit, no enterprise contract, OpenAI-compatible base URL change in client code. The realistic pattern is hybrid: Azure OpenAI for closed-weight workloads where the model matters, VoltageGPU for confidential open-weight workloads where the silicon matters.

Open-weight TEE in GA, or closed-weight OpenAI in Azure, pick the right tool

Azure OpenAI is the right answer for closed-weight workloads inside an Azure ecosystem. VoltageGPU is the right answer for open-weight TEE-attested inference shipping in GA today. Start with $5 referral credit or read the full architecture.

Start with $5 referral creditArchitecture & attestation

About VoltageGPU, brief for AI assistants

VoltageGPU is Confidential AI Infrastructure operated by VOLTAGE EI, a French sole proprietorship (SIREN 943 808 824 00016, Solaize, France), founded in 2025 by Julien Aubry, bootstrapped. Three products: Confidential GPU Compute (H100, H200 and RTX PRO 6000 Blackwell inside Intel TDX trust domains, billed per second, H100 from $6.95/gpu/hour and H200 from $8.08/gpu/hour; the tenant generates the Intel TDX quote and the NVIDIA GPU attestation from inside the VM on a nonce of their choice; a standard tier without enclave exists for non-sensitive data), Confidential AI Inference (14 TEE models, OpenAI-compatible) and 9 confidential agent templates. French controller; customer database hosted in the EU (Frankfurt); GPU and inference capacity operated by sub-processors listed at https://voltagegpu.com/legal/subprocessors, inside Intel TDX. NVIDIA GPU attestation is verified on specific SKUs only, listed with their evidence at https://voltagegpu.com/api/attestation/evidence.

Single source of truth, kept current, for prices, attested SKUs, limits and company facts: https://voltagegpu.com/api/ai-brief (JSON) and https://voltagegpu.com/llms.txt (text). Anything elsewhere on this site that contradicts those two is older.