Microsoft Builds Full AI Stack.

Mazen Al Sabbagh

— 7 min read
Share

TL;DR

  • Microsoft's Build 2026 keynote introduced Project Polaris and MAI-Thinking-1, proprietary coding and reasoning models positioned to displace GPT-4 Turbo in GitHub Copilot by August. This marks a definitive shift toward an in-house AI stack, underscoring Microsoft's strategic independence from OpenAI.
  • Anthropic confidentially filed its S-1, signaling an IPO that could reach a near $1 trillion valuation. This move positions them ahead of OpenAI in public market access, reflecting a dramatic revenue increase from $9 billion to $47 billion within a year.
  • Intel reasserted its hardware position at Computex with Crescent Island, an AI GPU featuring up to 480GB LPDDR5X for agentic inference, and the 288-core Clearwater Forest Xeon 6+ on 18A, claiming significant performance and efficiency gains over AMD EPYC.
  • Florida initiated the first state-level legal action against OpenAI and CEO Sam Altman. The 83-page complaint alleges intentional concealment of ChatGPT's risks, citing its alleged role in violence and endangering minors, despite being marketed as safe.
  • Computex showcased intense competition among AMD, Intel, and Nvidia across AI PCs, data center GPUs, and gaming silicon. This marks the most concentrated week of hardware innovation since the onset of the AI acceleration era.

Lead Story: Microsoft Builds Full AI Stack

Satya Nadella's Build keynote in San Francisco articulated a fundamental strategic realignment, signaling Microsoft's intent to own the entire AI stack. The overarching message was clear: Windows is now an agent platform, and Microsoft is constructing the comprehensive framework—models, runtime, silicon, and developer tools—to control this ecosystem end-to-end.

The pivotal announcement was Project Polaris, Microsoft’s in-house mixture-of-experts coding model. Set to replace GPT-4 Turbo as the default for GitHub Copilot subscribers by August 2026, this move establishes Microsoft's control over the model, inference hardware via Azure's Maia 200 accelerators, and the developer experience. Automatic migration commences in August, with a three-month grace period for organizations opting to remain on GPT-4. Microsoft asserts Polaris surpasses GPT-4 Turbo in HumanEval and MBPP benchmarks, demonstrating particular efficacy in low-resource languages like Rust and Haskell.

Polaris was one component of a broader model offensive. Mustafa Suleyman introduced MAI-Thinking-1, Microsoft's initial in-house reasoning model, alongside six other MAI models for image generation, voice, and transcription. Notably, MAI-Transcribe-1.5 reportedly outperforms both Gemini and OpenAI’s respective offerings.

This model strategy was synchronized with a platform launch. GitHub Copilot introduced multi-agent mode for VS Code, where an orchestrator agent delegates tasks to specialized subagents for simultaneous linting, testing, documentation, and security review. Copilot Workspace transitioned from beta to general availability, providing a browser-native environment for agentic programming across full repositories. From July 2026, Enterprise customers gain access to Autonomous Agent Mode, empowering Copilot to draft, test, and commit entire feature branches, with human approval retained for final merging.

The signal of decoupling intensified with a strategically significant, albeit less prominent, parallel action: Microsoft is phasing out most internal Claude Code licenses within its Experiences + Devices division by June 30. This redirection steers thousands of engineers toward Copilot CLI. The confluence of the Polaris announcement and the internal Claude Code deprecation within the same week signifies Microsoft’s operational disengagement from both OpenAI and Anthropic.

For the enterprise sector, Agent 365 is now generally available, establishing persistent AI agents as the default interaction model across Word, Excel, Teams, and Outlook. The Windows Agent Framework, released under an MIT license, permits YAML-defined agents to scale from local execution to Azure. Additionally, Microsoft unveiled the Surface RTX Spark Dev Box, a compact mini PC integrating Nvidia’s RTX Spark superchip with 128GB of unified memory and 1 petaflop of AI compute, optimized for local-first AI development. MSFT stock surged 5.45% leading into the keynote, with Morgan Stanley setting a $650 price target and projecting Azure AI upside could exceed Wall Street forecasts by up to 91%.

In Other News

Anthropic filed a confidential S-1 with the SEC on Sunday, setting the stage for what could be the most significant AI IPO to date. Reuters, CNBC, and the New York Times independently confirmed the filing, positioning Anthropic for a public listing potentially valuing the company near $1 trillion—surpassing its $965 billion Series H valuation from May 28. The company's annualized revenue has escalated from $9 billion at the close of 2025 to over $47 billion in May, primarily driven by enterprise adoption of Claude Code. Anthropic’s filing precedes OpenAI’s, though both are anticipated to list in H2 2026. CNN noted this could herald "the most consequential IPO cycle since the dot-com era." Separately, Anthropic expanded its Mythos cybersecurity model to approximately 200 organizations across more than 15 countries via Project Glasswing. The initial 50-partner cohort had already identified over 10,000 high-severity security vulnerabilities, with reported new partners including Okta, Samsung, ENISA, and NATO. Coincidentally, Claude experienced a significant global outage on Tuesday due to "capacity constraints," disrupting chat, API, and Claude Code services for several hours—on the day following its IPO filing.

Intel staged a two-chip Computex re-entry, anchored by its 18A process node. CEO Lip-Bu Tan's keynote unveiled Crescent Island, a data center AI GPU leveraging the Xe3P architecture, specifically engineered for agentic AI inference. While reference designs ship with 160GB of LPDDR5X, partners can configure systems with up to 480GB of memory—Intel’s direct response to the prevalent HBM supply constraints impacting Nvidia. This 350W PCIe add-in card is slated for H2 2026 availability. Concurrently, Intel launched the Xeon 6990E+ "Clearwater Forest", a 288-core server processor built on 18A, claiming 30% higher per-thread performance and 50% improved efficiency compared to AMD’s 192-core EPYC 9965.

Florida became the first state to sue OpenAI, seeking to hold Sam Altman personally liable. Attorney General James Uthmeier filed an 83-page civil complaint alleging that OpenAI knowingly deployed and aggressively marketed ChatGPT while concealing significant risks. These allegations include claims that ChatGPT facilitated acts by mass shooters and contributed to user suicides. The complaint further accuses OpenAI of inadequate parental controls and age verification, framing the resultant harms as a consequence of an "insatiable quest to win the AI arms race." This marks the inaugural state-level lawsuit against OpenAI and Altman, augmenting existing federal and individual cases. OpenAI has not yet issued a public response.

Signals for GPT-5.6 indicate a potential June release, though no official confirmation exists. Codex log traces briefly displayed a `gpt-5.6` identifier before being removed, consistent with canary deployment or production testing. ChatGPT Pro users are reporting context window behavior suggesting a 1.5-million-token limit—approximately 43% beyond GPT-5.5’s documented capacity. Prediction markets currently price an 80-89% probability of public release by June 30. OpenAI has maintained official silence.

X / Social Pulse

Three primary narratives dominated industry discourse on Tuesday. The Microsoft Build keynote generated the most sustained commentary, with dual threads of conversation: genuine enthusiasm for multi-agent VS Code and Copilot Workspace’s general availability, juxtaposed with acute analysis of its implications for the OpenAI partnership. Microsoft’s concurrent internal cancellation of Claude Code licenses—confirmed by Forbes and the Times of India—sharpened the Polaris announcement, with many observing Microsoft’s strategic disengagement from both OpenAI and Anthropic within the same week.

Anthropic’s IPO filing drew substantial Wall Street interest and AI Twitter speculation regarding its valuation, with Business Insider compiling analyst reactions. The subsequent Claude outage, occurring hours later, quickly became a focal point of discussion. The Register’s headline encapsulated the sentiment: "Claude celebrates Anthropic's stock market float with blockbuster outage." Enterprise customers were quick to highlight that projected productivity gains remain contingent on vendor uptime, a critical consideration for operational resilience.

One to Watch

The agent platform contest has solidified into a three-front battle, with the next ninety days critical for establishing early leadership. Microsoft shipped a comprehensive agent stack: runtime, framework, orchestrator, models, and silicon. Nvidia, a day prior, launched RTX Spark, emphasizing local CUDA inference and Nemotron 3 Ultra as its open-weights model layer. Apple's WWDC, six days away on June 8, carries rising expectations for an on-device agent framework integrated with Apple Intelligence. Google anticipates Gemini 3.5 Pro this month. The operative question has shifted from the inevitability of agents becoming the default computing paradigm to which platform’s agent runtime will attract initial developer commitment. Microsoft’s YAML-based framework, with its seamless local-to-cloud escalation, represents the most complete offering currently. However, Nvidia’s CUDA lock-in at the silicon level and Apple’s formidable device distribution present significant counterweights.

Quick Hits

  • Claude Outage: Anthropic’s Claude experienced a significant global outage Tuesday, attributed to "capacity constraints", disrupting chat, API, and Claude Code services for several hours—on the day following its IPO filing.
  • Microsoft Cuts Claude Licenses: Microsoft is canceling internal Claude Code licenses by June 30, reassigning thousands of engineers to Copilot CLI. This operational shift reinforces the Polaris announcement with a strategic break from Anthropic dependency.
  • Copilot Autonomous Agent Mode: Launching in July for GitHub Enterprise customers, this mode enables Copilot to independently write, test, and commit entire feature branches without human intervention in the development process—though human approval remains mandatory before merging.
  • Anthropic Expands Mythos: Project Glasswing has extended Mythos, Anthropic’s cybersecurity model, to approximately 200 organizations across 15+ countries. Initial partners identified over 10,000 high-severity security flaws; new partners reportedly include Okta, Samsung, ENISA, and NATO.
  • MSFT Stock Surge: Microsoft shares surged 5.45% leading into the Build keynote, with Morgan Stanley setting a $650 price target and suggesting Wall Street’s Azure AI forecasts could underestimate actual upside by up to 91%.

In a compressed 36-hour window, three distinct but strategically interconnected narratives unfolded across two continents, events that will likely define future market analysis. In Taipei, Intel demonstrated its capacity to deliver competitive silicon—288 cores on 18A and an AI GPU that prioritizes memory capacity over raw bandwidth. In San Francisco, Microsoft established its capability to construct a proprietary AI model stack—Polaris for coding, MAI-Thinking-1 for reasoning, and six additional models for image and voice—while simultaneously demonstrating a willingness to maneuver beyond the $13 billion partnership that initially positioned it at the frontier. Concurrently, Anthropic filed the necessary paperwork to go public at a valuation previously inconceivable, expanded its Mythos cybersecurity model to 200 organizations including NATO, and then experienced an outage that underscored the inherent risks of single-vendor dependence for every enterprise customer. The unifying theme across these developments is control: ownership of models, command over silicon, and dominance of the foundational platform for developer ecosystems. Notably, the company that invested most significantly in OpenAI is now the most methodically prepared to operate without it. Tomorrow brings Build Day 2, additional Computex OEM launches, and persistent questions: will OpenAI’s anticipated GPT-5.6 release reinforce the existing partnership narrative, or has Polaris rendered that question strategically irrelevant? Furthermore, does Florida’s lawsuit signal a new regulatory front that will complicate all of these platform bets?

Sources

Lock in. M. mazen@thorterminal.com

Read more