Last week we covered the Bloomberg confirmation that Microsoft is routing production Office workloads to its own MAI models. This week, the picture is clearer: Microsoft’s first-party AI stack is no longer aspirational. It is deployed, priced, and architected as a complete vertical from device to cloud. For CTOs and engineering leads, the question has shifted from “are Microsoft’s models ready?” to “how do we architect around a stack where one vendor controls silicon, models, and applications?”

The Complete First-Party Stack

Microsoft now operates the only full-stack AI portfolio outside of Google and Apple. Here is the architecture as of July 2026:

On-device (Aion 1.0): Lightweight inference for summarization, rewriting, and intent classification — running on CPU, GPU, or NPU inside Edge and Windows. The 14B-parameter Aion 1.0 Plan variant handles agentic workflows with tool-calling and sub-agent orchestration via the Windows Agent Framework. Open weights are expected on Hugging Face this month.

Edge (Phi-4 family): Ten MIT-licensed models spanning 3.8B to 15B parameters. Phi-4-mini runs on Raspberry Pi 5. Phi-4-mini-flash-reasoning uses the SambaY hybrid architecture (Mamba + sliding window attention + Gated Memory Units) for 10× throughput gains. Phi-4-reasoning-vision-15B handles multimodal reasoning with a SigLIP-2 vision encoder.

Cloud frontier (MAI family): Seven models announced at Build 2026, covering reasoning (MAI-Thinking-1), agentic coding (MAI-Code-1-Flash), image generation (MAI-Image-2.5 and Flash), speech-to-text (MAI-Transcribe-1.5), and text-to-speech (MAI-Voice-2 and Flash). All trained from scratch on commercially licensed data with zero distillation from third-party models.

Silicon (Maia 200): 3nm process, 140B transistors, 216GB HBM3e. Running MAI inference in production today. MAI-Code-1-Flash will fully migrate to Maia 200 by August 2026.

Foundation models (MAI-1-preview): Microsoft’s first end-to-end trained foundation model, currently testing on LMArena. Limited public information, but it signals Microsoft’s intent to compete at the frontier model level — not just in task-specific models.

This is not a collection of point solutions. It is a deliberate architectural stack where each layer feeds the next.

Inside Copilot’s Routing Architecture

The most strategically important detail from the past two weeks is not a model release — it is the disclosure of how Microsoft 365 Copilot routes traffic. Copilot is not a single-model product. It uses a task categorization layer that classifies each interaction by task type, complexity, and quality requirement before selecting a model.

What routes to MAI (first-party, high-volume):

  • Email reply drafting from thread context (Outlook)
  • Thread and meeting summarization
  • Spreadsheet formula generation from natural language (Excel)
  • Subject line and header generation
  • Structured data extraction from fixed-format documents
  • Task and action item extraction from meeting transcripts
  • Agentic coding in VS Code and GitHub Copilot CLI (MAI-Code-1-Flash)
  • Image generation and editing in PowerPoint and OneDrive (MAI-Image-2.5)

What routes to frontier third-party models:

  • Novel analysis across long unstructured documents
  • Multi-step reasoning with complex dependencies
  • Creative generation requiring stylistic judgment
  • Sensitive or ambiguous tasks where error cost is high

The routing is dynamic. As MAI models improve, the routing envelope expands. Microsoft has every incentive to shift more traffic to its own models — each query routed to MAI on Maia 200 eliminates the per-token margin paid to OpenAI or Anthropic.

Notably, Copilot Cowork — Microsoft’s autonomous agent product launched June 16 — runs on Anthropic Claude (Opus 4.8 and Sonnet 4.6), not MAI. Microsoft’s own Cowork 1 fine-tuned model was announced but is not yet powering the agent. This is an honest signal: MAI models are production-ready for commodity inference, but Microsoft acknowledges where frontier third-party models still hold an edge.

MAI Pricing: The Unit Economics Story

Microsoft has published pricing for its MAI models in Foundry, and the numbers are worth examining for any team doing cost modeling:

  • MAI-Transcribe-1: $0.36/hour of transcribed speech (43 languages, 4.86% WER on FLEURS — lowest of any competitor)
  • MAI-Voice-1: $22/1M characters (15+ languages with voice adaptation and emotional control)
  • MAI-Image-2.5: $5/1M text input tokens, $8/1M image input tokens, $47/1M image output tokens
  • MAI-Code-1-Flash: Positioned cheaper than Claude Haiku while scoring 51.2% on SWE-Bench Pro (vs Haiku’s 35.2%) and using 60% fewer tokens

The pricing pattern is aggressive. Microsoft is not pricing MAI at parity with third-party alternatives — it is pricing below them while claiming better performance on specific benchmarks. For high-volume workloads (transcription, image generation, code completion), the cost differential compounds rapidly.

The distribution story matters here too. MAI models are available on Microsoft Foundry, the MAI Playground (free web testing), and — notably — on OpenRouter, Fireworks AI, and Baseten. This is the first time Microsoft has shipped first-party models on non-Azure inference platforms. Enterprise teams locked into other cloud providers can now access MAI models without Azure commitments.

Florence-2: Stability Over Hype

While Phi-4-reasoning-vision-15B grabs headlines as the multimodal successor, Florence-2 continues as the enterprise workhorse for pure computer vision. There is no Florence-3 on the roadmap. Instead, Microsoft is investing in ecosystem maturity: the C#/.NET NuGet package for local ONNX execution, NVIDIA DeepStream integration for real-time edge video analytics, and Azure AI Vision Image Analysis 4.0 SDK with 164-language OCR.

Florence-2’s unified prompt-based architecture — one set of weights handling 12+ vision tasks via text prompts — remains architecturally elegant. For teams building production CV pipelines, the choice is clear: Florence-2 for lightweight, high-throughput vision tasks; Phi-4-reasoning-vision-15B for multimodal reasoning that requires language generation and chain-of-thought.

The Phi-4 Density Play

The Phi-4 family deserves specific attention from engineering leads evaluating edge AI. The core thesis is density — 14B models competing with 70B-class models on math and code benchmarks.

Phi-4-reasoning-plus scores 82.5% on AIME 2025, competitive with DeepSeek-R1-Distill-Llama-70B at one-fifth the size. Phi-4-mini (3.8B) outperforms many 8B-class models on MMLU with a 200K-token vocabulary and native function calling across 22+ languages. The SimpleQA caveat is important — Phi-4 scores just 3.0 on factual recall, making it unsuitable as a general knowledge assistant without retrieval augmentation.

The architectural innovation in Phi-4-mini-flash-reasoning is the standout. The SambaY architecture — combining Mamba state space models, sliding window attention, and Gated Memory Units — represents a genuine departure from vanilla Transformers. 10× throughput improvement at 2-3× lower latency is not an incremental optimization. It is the kind of architectural shift that changes deployment economics for edge AI.

What This Means for Enterprise Strategy

Three takeaways for CTOs making infrastructure decisions this quarter:

1. The vendor lock-in calculus has shifted. Microsoft’s vertical integration — silicon through models through applications — was historically a lock-in concern. In 2026, it is a cost-reduction strategy. MAI models on Maia silicon eliminate third-party margins. The question is whether the cost savings justify the architectural commitment.

2. Model evaluation should be workload-specific. Microsoft’s routing architecture inside Copilot is the template. Do not standardize on one model for all tasks. Route commodity inference (summarization, structured extraction, formula generation) to cost-efficient models like MAI. Route complex reasoning to frontier models. The savings are in the routing, not the model choice.

3. The open-weight Phi-4 family is the safest edge bet. MIT-licensed, 10+ variants, running on everything from Raspberry Pi to NVIDIA Jetson. No vendor lock-in, no licensing risk, and competitive with models 5× their size. For teams deploying local inference — whether for cost, latency, or data sovereignty — Phi-4 is the default choice.

Microsoft’s AI strategy in 2026 is not about catching up to OpenAI or Anthropic. It is about building a complete, vertically integrated stack where Microsoft controls every layer. The models are ready. The silicon is ready. The question for enterprise buyers is whether the integration premium is worth the flexibility trade-off.

Follow developments on X: https://x.com/kkaminski