Microsoft: Is the tech giant sleeping, or is it actually building an empire for the AI ​​era?

How can we explain a loss of value on the stock market of around 34% below its all-time high ($555.45 in July 2025 compared to end of March 2026), and is stock market afraid that Microsoft is transforming from a software company (high profit) to an infrastructure company (with erosion of low profit margins)?

While Microsoft's stock is being punished for crazy CAPEX, the AI ​​market is converging on a battle between four different empires: Copilot, ChatGPT, Gemini and Claude. The strategic question is no longer "who is smarter", but who owns the value chain - from the chip to the document. Therefore, it seems that Microsoft is not an "old giant" but a giant that is waking up strong, with tremendous cash flow, a clear strategy and a platform that cannot be replaced. In terms of a zero-sum game for the user's attention and time - the four giants (Microsoft, OpenAI, Google, Anthropic) are less concerned with being the "best model" and more with being the middle layer between the person and the information, the code, the document, the transaction.

The stock market sees the expenses (CapEx) that are immediate, but it has difficulty pricing the cumulative value (LTV - Lifetime Value) of an enterprise customer who is locked into the system. From this perspective, the decline in Microsoft's stock looks more like a "tax on impatience" than an expression of strategic distrust - but this is already an analytical conclusion, not an empirical fact. If the AI ​​ROI is realized (as happened in the past with the bet on the cloud), the stock will grow again. This is not a risk - it is an investment phase.

Microsoft has built a very wide "fishing net" (450 million Office users) and even if only about 3.3% of them switched to a paying copilot, every additional percentage of conversion is worth billions of dollars in net revenue with zero marketing costs (because the customer is already there). This is an advantage that no other LLM provider has.

Want to comment? We're waiting for you on X

Is Microsoft a "giant that didn’t wake up" - or running a marathon while the market is looking for a sprint?

Microsoft is in a critical transition phase: it is investing tens of billions in AI (CapEx of about $30 billion in the last quarter, an increase of 89% YoY), but the market is reacting with pressure. The stock has fallen about 15-18% since the beginning of 2026 (from ~$540 at the peak in 2025 to about $370–400 today), with sharp declines of 10-14% after the Q2 FY26 reports. The main reason: Investors are concerned that the huge investments in AI infrastructure (GPUs, data centers) are not yet translating into proportional growth in Azure revenue (39% growth - slightly below expectations) and improving margins. However, the company is still a cash generator (strong operating cash flow), with ~$89 billion in liquidity (cash + short-term investments). But the high CapEx is compressing margins (down to ~68% margin) and affecting free cash flow. Azure + Microsoft 365 are the growth engines, but AI is still "burning" money before it generates full ROI.

On the surface, the raw data paints a picture where:

  • Microsoft has a significantly smaller Copilot user base than Google Gemini or ChatGPT.
  • AI CAPEX is weighing on the stock in the short term.

In practice, it maintains a technological advantage over China (through OpenAI + Azure), but is exposed to supply chain risks (TSMC in Taiwan). Therefore, building less dependence on Nvidia through independent chips, which strengthens US national security in AI. However, US and Chinese regulation of chips may slow down the pace.

But from a geostrategic perspective, the picture is different:

  1. Control of the work layer - Microsoft holds a very high power position (Shapley/Banzhaf) in the family of operators: M365, Windows, Azure, Security. Copilot is an attempt to turn this layer into a "convenient monopoly" of AI agents within the organization.
  2. Partial control of the chip, but not the assembly line - avoiding a possible acquisition of chip manufacturer Intel is not only a weakness; it is also a choice to leave the risk of the factories in the hands of others, and to concentrate on system design based on Maia + Cobalt + Azure.
  3. Diffusion of dependence on models - Alongside a deep dependence on OpenAI, Microsoft is establishing an internal Superintelligence team under Suleiman, to develop models frontier, including the use of Maia 200.

If we apply the equilibrium in political systems of Shapsley Wingest - Microsoft does not strive for a model monopoly, but for a dominant combination of several "power centers": models (OpenAI + internal), cloud infrastructure, work infrastructure (M365) and the Copilot agent layer. And even if ChatGPT seemingly wins in terms of user volume (about 900M weekly vs. ~33M in Copilot), Copilot generates higher ARPU (~$30 per month in corporate-business licensing vs. $20 in personal-business for ChatGPT) and long-term contracts, and high loyalty ("stickiness") in the Office integration (in total, Copilot generates about $2.5-3.5 billion in estimated ARR and with huge potential from 450 million M365 seats).

Microsoft's advantage is based on three principles:

1. ARPU (Average Revenue Per User) is the winning statistic

While most ChatGPT users, for example, are freemium users, every enterprise Copilot user is a "cash machine":

  • ChatGPT: Revenue is based primarily on Plus subscribers ($20) and casual users. The conversion rate to paid subscribers is relatively low for all users.
  • Microsoft Copilot: A Copilot for Microsoft 365 subscription costs $30 per user per month, which is added to the base license price (E3/E5). That is, Microsoft receives an amount ranging from $60 to $90 per month for each such user.
  • Bottom line: 15 million paying users at Microsoft generate annual revenue (ARR) of over $5.4 billion (an amount that OpenAI needs hundreds of millions of free users to approach).

2. "Sticky" revenue (Retention)

The big difference is in the risk of abandonment:

  • ChatGPT: is a consumer product. A user can cancel the subscription as soon as they feel "satisfied" or when a better competitor model comes out (like Claude or Gemini).
  • Copilot: is an infrastructure product. Once an organization like PwC or Accenture embeds Copilot into their Outlook, Teams and SharePoint, they cannot easily "cancel" it. The AI ​​becomes part of the work processes, which ensures a stable cash flow for years to come.

3. The Upsell Model: From "Assistant" to "Agent" (Copilot Studio)

Here lies the real horizon. Microsoft is not just selling "Chat":

  • Copilot Studio: Organizations pay extra to build autonomous agents. This is a Consumption-based pricing model. The more actions the agents perform in the ERP or CRM, the more Microsoft earns.
  • The revenue here is not limited to the number of employees in the organization, but to the number of transactions that the AI ​​​​performs. This is a paradigm shift - Microsoft begins to profit from "work" and not just from "licenses".

Ironically, Microsoft is following the model of Amazon in its early days, building a physical and technological infrastructure so expensive and extensive that no other player (except perhaps Google or Amazon itself) can afford to compete with it.

Have you read this far? You're done!
The expansion is intended for those interested in the in-depth analysis on which the above summary is based and constitutes an academic-scientific expansion of it.

What's happening to Microsoft?

  1. Microsoft under stock market pressure, but with an unusually strong base:
    As of 3/27/26, MSFT shares are trading around $366 - about 34% below the all-time high ($555.45 in July 2025) and down ~21% since the beginning of 2026, against the backdrop of huge AI CAPEX, a relative slowdown in Azure, and the escalation of the Iran-US-Israel war that weighed on the entire tech sector.
  2. Four different strategic models:
    • Copilot (Microsoft) - about 6 million daily active users and about 15 million paying users in Microsoft 365 Copilot, but only ≈3% of the M365 business base; deep integration with Office/Teams/Windows, and now the Copilot organization is also unified under Jacob Andreu.
    • ChatGPT (OpenAI) - ~900M weekly users, 50M+ consumer subscribers, and ~9M paying enterprise customers; a significant revenue driver with ~$10B in annual recurring revenue by 2025.
    • Gemini (Google) - 750M+ Gemini app users, 2B+ AI Overviews searches per month, and 8+M paid Gemini Enterprise seats; deep integration across Search, Android, Workspace, and Cloud.
    • Claude (Anthropic) - Less dominant in consumer mass, but very deep in enterprise: Claude for Enterprise product on AWS Marketplace (≈$40/user/month), Opus 4.6 model with 1M token context window, Marketplace platform, and launch of Claude Cowork with deep integrations with Google Drive, Gmail, and DocuSign.
  3. Chips and Infrastructure: Microsoft vs. Google vs. AWS+Anthropic:
    • Microsoft - Maia 100/200 accelerators and Cobalt 100 CPUs, alongside continued conscious dependence on Nvidia/AMD.
    • Google - Ironwood TPU family (7th generation) and Axion CPU, as part of an AI Hypercomputer that is directly challenged by Nvidia systems.
    • Anthropic - relies on AWS Trainium/Inferentia as a main partner, while simultaneously locking up huge TPU capacity in Google Cloud - a distinct multi-cloud model.
  4. Theoretical picture:
    • Nash equilibrium + Tragedy of the Commons - Each giant invests tens of billions in CAPEX on chips and infrastructure, because exiting the game is riskier than continuing the investment. ‎‎‎
    • Rational Choice & Selectorate - Microsoft, Google, and Anthropic target high-value enterprise users (Copilot, Gemini for Workspace, Claude for Enterprise), while OpenAI has a broad consumer base and a premium business tier.
    • Agenda-Setting & Veto Players - The organizational changes at Copilot and Claude (Marketplace, Cowork) are moves to create a "work operating system" controlled by a few power nodes, rather than a collection of gadgets.

Recently, Microsoft has made a transition from the leadership of a "visionary" like Mustafa Suleyman to the leadership of a "product person" like Jacob Andreou, thus marking the end of the era of trial and error and the beginning of the era of industrialization and consolidation. What are the strategic implications of this move?

1. Ending the organizational "split personality"

Until now, Microsoft has suffered from dissonance: the consumer copilot tried to be experiential and light, while the business copilot (B2B) was productivity-focused and rigid.

  • Implication: Nadella understands that in the age of AI, the user is the same user. The interface experience (UX) must be uniform. The unification of the teams is intended to prevent duplication of development (R&D) resources and create a single “knowledge graph” that accompanies the user from home to the office.
  • In terms of game theory: This is a move to reduce internal friction. Instead of two divisions competing for budgets and attention, Microsoft is creating a "united front" against Apple and Google.

2. The Snap Effect: UX is King

The choice of Jacob Andrew, who came from a company that lives and breathes user engagement like Snapchat, is a statement of intent.

  • The challenge: Copilot’s problem today is not the "brain" (the LLM), but the "hands" (the interface). It is often cumbersome, slow, or unintuitive.
  • The solution: Andrew is supposed to bring the lightness and availability of consumer apps into the heavy M365 workspace. The goal is for the AI ​​agent to be as accessible as a "story", and not as complex as a SQL query.

3. The Four Pillars: The New Roadmap

The division into four pillars shows a systemic approach:

  1. The Copilot Experience: The Front-end. How it looks and feels.
  2. AI Assistant Platform: The "pipeline" that allows organizations to build their own agents.
  3. Productivity Apps: Deep integration into Word, Excel, and Teams - that’s where the big money is.
  4. AI Models: Model development (including the small SLMs).
    This separation allows Microsoft to change the "engine" (model) without changing the "body" (user experience).

4. The Geo-Business Implication

This move is a direct response to Apple Intelligence. Apple has shown that AI succeeds when it is "transparent" and seamlessly integrated into the operating system. Microsoft, which was engaged in a technological arms race with OpenAI, now understands that the real battle is for screen time and user habits.

5. And what about Suleiman?

The move of Suleiman, who arrived as a "rock star" from Inflection, indicates that Microsoft is past the "hype" and visionary phase. Suleiman is a man of big models and futuristic predictions; Andrew is a man of profit and usability. Nadella chose execution over promises.

A Macro Look at the Four AI Giants - Google, OpenAI, Anthropic, and Microsoft

1. Microsoft and the Stock Market Shock: CAPEX Politics in an Age of War

From a capital market perspective, Microsoft is in a "huge - but show me the money" phase.

After an exceptional AI rally in 2024-2025, the stock reached a high of $555.45 and fell by about 34% to about $366. 2026 opened as the worst since 2008 - down about 21% YTD, mainly due to a quarterly report that presented a combination of:

  • Strong revenue growth (double-digit increases in revenue and profits),
  • but a jump of tens of billions in the quarter in CAPEX for AI,
  • and a relative slowdown in Azure growth compared to expectations.

Added to this is the Iran-US-Israel war, which increases volatility, supports high energy prices and creates "collective punishment" for investment-heavy tech stocks. Just one day of geopolitical aggravation wiped about $500 billion off the value of five tech giants, including Microsoft, Nvidia and Meta.

In terms of economic political theory and Public Choice, the capital market behaves like a short-sighted electorate:
It "punishes" a leader (Nadella) for investing heavily in the short term, even if the investment is intended to create structural superiority in the long term - as long as no immediate ROI is seen in EPS and free cash flow.

2. Copilot, ChatGPT, Gemini, Claude - Four Strategic Philosophies

2.1 OpenAI / ChatGPT - Mass First, Monetization Later

According to February 2026 data, ChatGPT has approximately 900 million weekly users, over 50 million consumer subscribers, and approximately 9 million paying business customers. ‎‎‎
OpenAI's model, in Rational Choice terms, maximizes reach: first, they create infrastructure - global discourse, daily usage habit (2+ billion prompts per day) - and only then deepen layers of business collaboration and premium revenue.

2.2 Google / Gemini - Harnessing the Power of the Network

Google is playing a different game: Gemini is not just a chatbot, but an AI layer embedded within Search, Workspace, Android, and Cloud.

  • 750 million monthly users per app.
  • AI Overviews on Google Search reach ≈2 billion users per month.
  • About 8 million+ paid Gemini Enterprise seats, thousands of companies and business workflows.

In terms of Shepley and Benzaff (coalition power measures): Google already has a huge concentration on the global information node - Search and Android, so the marginal value of Gemini uses is extremely high: each AI unit connects to an existing huge network.

2.3 Anthropic / Claude - Judicial quality, organizational depth

Anthropic positions itself as a "judicial player": Claude Opus 4.6 is positioned as a premium model for code, agents and enterprise workflows, with a context window of 1M tokens.

The company has built:

  • Claude for Enterprise on the AWS Marketplace, about $40/user/month, with dedicated solutions for finance, education and life sciences.
  • Claude Marketplace - an ecosystem of business products (GitLab, Harvey, Snowflake and more) that work with Claude.
  • Claude Cowork - a "digital colleague" agent platform, with deep integration with Google Drive, Gmail and DocuSign - an explicit attempt to "eat" middleware SaaS layers.

This is a classic move in terms of selectorate theory: Anthropic builds around Claude an "alliance of stakeholders" - AWS, Google Cloud, Marketplace partners who all benefit from its success, thereby expanding the base of supporters (selectate) of its model.

2.4 Microsoft / Copilot - an execution layer above the "work infrastructure"

Microsoft takes a different approach, however, with Copilot built on the M365 infrastructure: Word, Excel, Outlook, Teams, Windows.

  • About 6 million daily active users,
  • about 15 million paying users on Microsoft 365 Copilot - only ≈3% of the M365 customer base.

The downside: significantly smaller mass compared to ChatGPT or Gemini.
The advantage: each user is a very high value unit - enterprise ARPU, deep connection to the enterprise data (Microsoft Graph) and severe lock-in: leaving Copilot means, in practice, leaving significant parts of the work infrastructure.

Using Agenda-Setting Theory, Nadella rewrites the decision-making structure: unifying all Copilot (consumer + business) under Jacob Andreu, and diverting Mustafa Suleiman to focus on Superintelligence and the models themselves. ‎‎‎‎
Thus, "the model is the product", but Copilot is the "operating system" that organizes the work around it.

3. The Chip Competition: Maia vs. Ironwood vs. Trainium and the "Intel Question"

3.1 Microsoft - Maia, Cobalt and Conscious Dependence on Nvidia

Microsoft announced Maia 100 as an internal AI accelerator, and later Maia 200 ("Braga") - a new generation of AI accelerators designed for large-scale inference. Alongside this, it is launching Cobalt 100, an Arm-based dedicated CPU for the Azure cloud.

Nevertheless, Nadella clearly states: Even after Maia 200, Microsoft will continue to purchase chips from Nvidia and AMD - that is, this is a strategy of diversification, not disengagement.

3.2 Google - TPU Ironwood and Axion

Google, for its part, is launching TPU Ironwood - a seventh-generation TPU, with FP8 ExaFLOPS performance in pods of up to 9,216 chips, and Axion - an Arm-based CPU for cloud VMs. ‎‎‎‎
The Ironwood+Axion combination is the direct equivalent of Maia+Cobalt - two giants building a vertical "hyper-AI supercomputer".

3.3 Anthropic - Deep partnership with AWS and Google

Anthropic chooses not to build its own chips, but to forge two super alliances:

AWS as a lead partner: $8 billion Amazon investment, commitment to use Trainium/Inferentia as a training and inference base, and Claude as a flagship model on Amazon Bedrock.
Google Cloud: Commitment to use up to 1 million TPUs as part of a multi-gigawatt AI factory by 2026.

This is a different kind of Nash equilibrium: Anthropic does not aim to control the physical supply chain, but to be the "preferred model" of two giant infrastructure providers – thereby creating market power for itself thanks to dual access to cheap hardware.

3.4 Why doesn’t Microsoft buy Intel?

Why doesn’t Microsoft simply buy Intel, and end the dependency? Directly touches on Veto Players and Public Choice Theory:

  • Intel is a national "strategic asset", and a Microsoft takeover would have run into a regulatory wall in the US, Europe, and Asia.
  • Owning FABS would have shifted Microsoft from a "software+cloud" game to a heavy game of long-term capital investments, with location risks (Taiwan, geopolitics) and weakened its flexibility in dealing with different suppliers.

In terms of destructive Nash equilibrium - such an aggressive move could provoke a worldwide backlash (regulators, competitors, countries), and cause greater overall damage than the profit it would receive from monopolizing chip production.

In terms of Tragedy of the Commons, Microsoft is trying to turn the "common pasture" of AI chips into a field where it has an escape route (Maia/Cobalt), without collapsing the entire system by becoming the majority owner of the herd.

4. Comparing business models: Who benefits more from business customers?

The business philosophies can be summarized as follows (inspired by recent comparative analyses):

  • ChatGPT Enterprise: Strong brand, flexible product, suitable for a huge range of uses. In short: Mass → Data → Revenue.
  • Gemini for Workspace / Cloud: Tailored for organizations living within Google Workspace and Google Cloud; takes advantage of the depth of integration with Search and Android.
  • Claude for Enterprise: Aimed at high-stakes reasoning, very long-term contextual, and tied to the secure environments of AWS (and partly Google Cloud).
  • Microsoft 365 Copilot: Turns M365 into a "semi-autonomous" layer: not just text, but agents (Copilot Tasks, Agent 365, Copilot Cowork) that execute entire business processes.

Against competitors like ChatGPT which wins in terms of user volume (about 900M weekly vs. ~33M active on Copilot), Copilot wins in terms of quality and since 15 million paid M365 Copilot seats (160% YoY growth, as of January 2026), compared to ~50 million paying users on ChatGPT (mostly consumers) generate higher ARPU (~$30 per month in corporate-business licensing vs. $20 in personal-business), long-term continuous contracts, and high loyalty ("stickiness") because it is integrated into Office. The income from a single subscription is higher and more guaranteed. In total, Copilot (including GitHub ~4.7M paying users) generates an estimated $2.5-3.5 billion in ARR. Although less than OpenAI, but with huge potential from 450 million M365 seats.

In Pareto equilibrium terms:
For an organization that already subscribes to M365, adding Copilot, even if it’s not cheap, can improve productivity in a way that doesn’t require switching providers. It’s an internal “Pareto-improving” change. For Google, the same principle applies to Workspace; for AWS+Anthropic, to Bedrock customers.

From a Rational Choice perspective, each player maximizes something different:

  • OpenAI - mass adoption.
  • Microsoft - share of wallet and increased customer lock-in to Azure and M365.
  • Google - centralization of information (Search + Workspace + Android).
  • Anthropic - quality of judgment and context, partnering with cloud providers as superdistributors.

Appendix - Sources and Mirrors (Concise)

Microsoft – Stock, Performance, Copilot and Chips:

OpenAI / ChatGPT:

Google / Gemini:

Tech‑Insider, TheWorldData, LinkedIn, 9cv9 - MAU, Enterprise Seats, AI Overviews. ‎‎‎‎‎[theworlddata.com] [linkedin.com] [aws.amazon.com] [sqmagazine.co.uk] [secondtalent.com]
Google Cloud Blog, CNBC - Ironwood, Axion, AI Hypercomputer. ‎‎‎‎[tomshardware.com] [cnbc.com] [blog.google] [aboutamazon.com]

Anthropic / Claude:

Cross-platform comparisons:

IntuitionLabs, CopilotConsulting, FieldGuideToAI, BlueHeadline - ChatGPT/Claude/Gemini/Copilot comparisons for the enterprise market.‎‎‎‎‎ [copilotcon...ulting.com] [businessinsider.com] [digidai.github.io] [blueheadline.com] [vife.ai]

Methodological note: The analysis is based on theories from the fields of economics and political science, including the theory of "economic politics" and relying on game theory in combination with rational analysis (Rational Choice), internal political constraints (Public Choice; Audience Costs), coalition dynamics (Minimal Winning, Veto Players), and sensitivity analysis to extreme events (Monte Carlo), and more. They are not deterministic predictions but rather a chance assessment under the latest public information. See the academic background in detail "Academic theoretical background - theories, metrics, and algorithms in behavioral analysis from game theory and political science".