Azure OpenAI Service

Microsoft's hosted entry for OpenAI models — physically sharing underlying compute with OpenAI (sometimes the other way around, OpenAI uses Azure), capturing the largest enterprise share of GPT via enterprise compliance + regional deployment + investment relationship.

1. Core Product / Service

Azure OpenAI Service (AOAI) is the only official distribution channel for OpenAI models outside OpenAI's own offering, and Microsoft's flagship product at L3b.

• Model menu: GPT-5 series (GPT-5, GPT-5-Pro, GPT-5-Mini, GPT-5-Nano, synced release with OpenAI), GPT-4o / GPT-4o-mini / GPT-4.1, o3 / o4-mini / o3-mini, DALL-E 3, Whisper, TTS, text-embedding-3-large/small.
• Deployment types: Standard (shared pay-per-token), Provisioned Throughput Units (PTU) (reserved capacity, hourly billing, low-latency guarantee), Global Standard / DataZone Standard (cross-region routing optimizing throughput/price), Batch (async, 50% discount).
• Enterprise features: Azure Active Directory integration, customer-managed keys, private endpoints, VNet integration, Azure Policy compliance, GDPR / HIPAA / FedRAMP High.
• AI Foundry / Azure AI Studio: upper-layer RAG / agent / fine-tuning framework, gluing OpenAI models + Azure data services + Microsoft Graph (Office 365).
• Microsoft first-party models: Phi-4 series, self-trained Mistral etc. also under the same AOAI entry; Azure AI Foundry also hosts Llama / DeepSeek / Mistral / Cohere (this part is similar to Bedrock / Vertex), but the AOAI product name strictly refers to OpenAI / Microsoft models.

2. Target Users & Pain Points

• Microsoft Office / Azure large customers: already running data on Azure + using Microsoft 365 Copilot, AOAI is a contract extension; they won't go sign with OpenAI directly.
• European / regulated-region enterprises: OpenAI direct API historically wasn't fully available in all regions / data residency was inflexible; AOAI offers dedicated deployments + data residency commitments in EU / UK / Australia / Canada and other regions.
• Large government / defense: GPT is available on Azure Government Cloud — no equivalent OpenAI direct API channel.
• Pain points: compliance approval + regional residency + existing Azure contracts → the three main reasons AOAI is preferred over OpenAI direct.

3. Competitive Landscape

Differentiation: OpenAI model exclusive non-first-party distribution + Azure AD/compliance ecosystem + Office 365 Copilot traffic foundation.

4. Unique Observations

• Per-token pricing (Standard, 2026-05):
  • GPT-5: $1.25/M input + $10/M output — parity with OpenAI direct
  • GPT-5-Mini: $0.25/M input + $2/M output — parity
  • GPT-5-Nano: $0.05/M input + $0.40/M output — parity
  • GPT-4o: $2.50/M input + $10/M output — parity
  • GPT-4o-mini: $0.15/M input + $0.60/M output — parity
  • o3-mini: $1.10/M input + $4.40/M output — parity [1]
• vs 1P price gap (take rate): strict price parity, take rate nominally 0%. Microsoft recoups value through other dimensions:
  1. Microsoft holds 49% of OpenAI's economic rights (publicly reported) — OpenAI runs GPT training + inference on Azure, generating large compute spend that flows back to Azure revenue;
  2. Office 365 / Microsoft 365 Copilot packages GPT into $30/month/seat SKUs, where take rate sits at the application layer (not at AOAI);
  3. AOAI customers inevitably incur other Azure service spend (storage / networking / data), where take rate is actually higher.
• vs third parties: OpenAI models have no other legal token API channel outside AOAI — this is different from Bedrock / Vertex's "hard exclusivity." Claude on Bedrock is also available on Vertex / Anthropic direct; but GPT has no plan B outside OpenAI / AOAI.
• Inference engine: not publicly disclosed — treated as OpenAI in-house stack (Triton / self-developed) + Azure NDv5 / NVIDIA H100 / H200 / B200 GPU pool. Microsoft also partially uses its in-house chip Maia 100 (2024 launch) to run some AOAI workload — another case of hyperscaler in-house chips amortizing inference costs.
• Compute sourcing: 100% Azure first-party data centers — Microsoft has invested ~$13B+ in OpenAI, providing massive H100 capacity for OpenAI training + AOAI inference. OpenAI used Azure GPUs to train GPT-5; inference servers are shared between OpenAI direct + AOAI capacity.
• Hidden take rate mechanism: OpenAI's compute spend on Azure ≈ Microsoft's discount on GPU rental ≠ market price. Microsoft + OpenAI's financial structures are intertwined; the "parity pricing" of token API is just appearance — actual profit sharing is determined by private terms between the two companies.
• Lagging launch: After GPT-4o release, AOAI lagged ~1-2 weeks; some niche features (Realtime API, early fine-tuning versions) lag further on AOAI. This is the trade-off for enterprise users vs OpenAI direct: compliance in exchange for speed of access.
• Strategic risk: The OpenAI / Microsoft relationship has been tense multiple times in 2024-2026 (OpenAI exploring other clouds, Stargate project pulling OpenAI toward Oracle / SoftBank); if OpenAI renegotiates to multi-cloud GPT, AOAI's "hard exclusivity" narrative breaks down.

5. Financials / Business Scale

• GA date: 2023-01
• Microsoft investment in OpenAI: cumulatively ~$13B+ (including compute subsidy, cash); Microsoft holds OpenAI economic rights (49% profit share, up to a specific return cap).
• AOAI revenue: Microsoft does not separately disclose, but 2024 earnings mentioned multiple times "AI business surpassing $13B ARR" mainly driven by Copilot + AOAI.
• Customers: 60,000+ Azure OpenAI enterprise customers (self-reported); including Coca-Cola, Mercedes, Moody's, ICRC, KPMG, Unilever, BMW, US Army (CoPilot Air Force).

6. People & Relationships

• Parent: Microsoft Azure, Satya Nadella (CEO Microsoft), Scott Guthrie (EVP Cloud + AI).
• Strategic partner: OpenAI — Sam Altman, Greg Brockman; OpenAI is the major compute consumer mutually bound with Azure.
• Microsoft in-house AI team: Mustafa Suleyman (CEO Microsoft AI, joined 2024; co-founder of DeepMind / Inflection), Microsoft Research / Phi team.
• Competes with: aws-bedrock, gcp-vertex, OpenAI direct API (internal "friendly competition"), together-ai / fireworks-ai (in open-source model tier).
• Hosts models from: OpenAI (full GPT + DALL-E + Whisper + Embeddings), Microsoft (Phi series); AOAI same portal also contains Llama / DeepSeek / Mistral / Cohere / xAI Grok etc. (under Azure AI Foundry, not strict AOAI menu).

Sources

• [1] https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/ (2026-05-10)
• [2] https://azure.microsoft.com/en-us/products/ai-services/openai-service (2026-05-10)
• [3] https://news.microsoft.com/2023/01/23/microsoftandopenaiextendpartnership/ (2026-05-10)
• [4] https://learn.microsoft.com/en-us/azure/ai-services/openai/ (2026-05-10)