On January 23, 2023, Microsoft announced an additional ten-billion-dollar investment in OpenAI. ChatGPT had been out for exactly two months and five days. There was no public certainty yet that it was more than a viral end-of-year toy. Satya Nadella, Microsoft's CEO, had seen something others hadn't quite seen — or had decided to place the bet before seeing it confirmed.
A few weeks later, in a Financial Times interview about the launch of Bing with ChatGPT, Nadella delivered the line that defined the year: "I want people to know that we made them dance." "Them" was Google. The line was about who was setting the tempo of the industry.
Three years on, that ten-billion-dollar bet turned into Copilot — the most widely distributed AI integration in the world. And into a structural dependence Microsoft is quietly trying to reduce.
What Copilot is, precisely
Copilot isn't a model. It's a brand Microsoft uses to group every AI integration inside its products. Underneath, the primary engine is OpenAI's GPT-4o. There are a few variants (smaller models for simple tasks, a modified Codex for GitHub) but the workhorse is the same one running ChatGPT.
The Copilot product line as of April 2026 has several layers.
Microsoft 365 Copilot (November 2023, $30 per user per month). The enterprise tier. Integrates into Word, Excel, PowerPoint, Outlook, and Teams with access to your organization's internal data via Microsoft Graph — the index Microsoft keeps of your email, files, and calendars. That connection to your own data is what makes it useful for companies: Copilot can summarize a meeting while referencing the documents shared beforehand, or draft an email based on the prior thread.
Copilot Pro (March 2024, $20 per month). The consumer tier. Less integration with internal data, but priority model access and Copilot in Office apps.
GitHub Copilot (shipped in 2021, before ChatGPT). Originally built on Codex — a model OpenAI trained on public code — it was Microsoft's first successful generative-AI integration. It crossed a million paid subscribers in 2024.
Microsoft Copilot (the app/browser, formerly Bing Chat). The general-purpose consumer chatbot, free with a lighter model and paid with GPT-4o.
Copilot in Windows 11. A taskbar button that opens a Copilot pane for OS-level tasks.
Copilot+ PCs (May 2024). A hardware category requiring an NPU (Neural Processing Unit) of at least 40 TOPS, built to run small models locally. It introduced the controversial Recall feature, which periodically snapshots your screen so Copilot can search your history — a feature that generated enough privacy backlash that Microsoft had to delay and redesign the rollout.
The scene behind the scene: dependence and plan B
The question that defines Microsoft in AI isn't "which model do they use?" It's "what happens if OpenAI fails?"
That question stopped being abstract over the weekend of November 17 to 21, 2023, when OpenAI's board fired Sam Altman on Friday and had to reinstate him by Tuesday. Over those five days, Microsoft offered Altman and Brockman an internal lab and hinted at absorbing the OpenAI team if needed. In the end it didn't have to. But the episode exposed the obvious: if OpenAI implodes, Microsoft has a serious problem.
Microsoft's answer to that risk came in two moves.
First, on March 19, 2024, Microsoft announced the hire of Mustafa Suleyman as CEO of a new division called Microsoft AI. Suleyman had co-founded DeepMind (sold to Google in 2014) and later Inflection AI. The hire came through an unusual deal: Microsoft paid roughly $650 million to Inflection for licenses and for taking nearly the whole team — a way to acquire without formally acquiring, sidestepping regulatory scrutiny.
Second, on April 23, 2024, Microsoft Research published Phi-3. It's a family of small models (3.8 billion to 14 billion parameters) trained to run locally on-device. They're not direct competitors to GPT-4 on raw capability, but they're good enough for many tasks and don't require calling OpenAI.
Both moves point to the same goal: reduce exclusive dependence on an outside supplier without breaking the commercial relationship that's paying off.
The real strength: distribution no one replicates
Now the other side. What Microsoft got brilliantly right.
Copilot's competitive advantage isn't technical. It's locational. Microsoft 365 has hundreds of millions of corporate seats. When a twenty-thousand-employee company decides to turn Copilot on, it doesn't have to pick a tool or migrate data — everything is already in Microsoft's ecosystem.
That single detail rewrites the enterprise adoption equation. To adopt ChatGPT Enterprise, a company needs a process: security review, SSO integration, user training, legal sign-off. To activate Copilot, the CIO signs an update to the Microsoft 365 contract they already have. There's a real difference between "onboard a new vendor" and "flip a feature at the vendor you already use."
GitHub Copilot is the clearest example of that dynamic. It shipped in October 2021 — more than a year before ChatGPT — built on Codex, when most of the world hadn't heard of generative AI. By 2024 it crossed a million paid subscribers. Productivity studies from GitHub suggest 20 to 30 percent gains in development speed, with the caveat that the data comes from the vendor. Real adoption among professional developers runs around 30 to 40 percent per 2024 Stack Overflow surveys.
GitHub had 92 million developers. The distribution was already built. AI came on top.
The honest part: where Copilot is unbeatable and where it isn't
Worth taking this apart carefully, because any pro-Copilot-on-everything or anti-Copilot-on-everything read is a simplification.
Where Copilot is unbeatable today: everyday work inside Office for someone who lives there. Summarizing a cluttered inbox, drafting a first pass of a formal email, building a pivot table in Excel from a description, converting a Word doc into a PowerPoint deck. In those cases the integration wins on pure convenience.
Where Copilot isn't first choice: high-stakes work where reliability outweighs convenience. Legal contract analysis, reviewing code that's going to ship, synthesizing research documents with verifiable citations. Here Claude — my main tool in consulting — has a consistent edge because it follows literal instructions with less drift and flags what it doesn't know more often. For that kind of work, the habit of opening a new tab and switching to Claude, even uncomfortable, pays better than staying in Copilot out of ease.
The practical reality for many professionals is multi-vendor. Copilot for the routine inside Office. Claude or ChatGPT for work that demands a tighter review bar. Nobody forces a single choice.
To close, and to keep going
Microsoft executed the most lucrative strategy of the post-ChatGPT era: bet hard and early on the best model on the market, buy priority access, and use a distribution that already existed to amplify it. It's the same move it made with DOS-Windows in the eighties, Internet Explorer in the nineties, Azure in the 2010s. Microsoft knows how to convert distribution into victory.
The open question — and it's worth asking — is how stable a strategy is when it depends on a partner you don't fully control. OpenAI isn't Microsoft. It has its own governance, its own internal tensions, its own incentives. Hiring Suleyman and launching Phi-3 say Microsoft is already thinking about that risk.
If you want to dig into how model capability gets measured and compared, How AIs are measured is the next link. For the broader competitive picture, The AI race.
Do you use Copilot inside Office, or do you jump to another AI when the work gets serious?
You open Word to write a long email. The page is blank and the cursor blinks. Up on the right side of the menu you spot a tiny icon that looks like a braided ribbon: that's Copilot. Out of curiosity you click. A little box asks what you want to write. You type two rough sentences. Copilot hands back three reasonable paragraphs. You tweak a bit and you're done.
That click — that tiny icon next to the menu — is Microsoft's biggest play in AI.
It isn't the strongest model on the market. It isn't the most impressive one in a demo. But it sits exactly where your cursor was already parked.
What Microsoft Copilot actually is
Copilot is the name Microsoft put on the AI it stuffed into its products. It isn't a model Microsoft built — underneath, it's running GPT-4o, the same engine ChatGPT uses. Microsoft licenses it because it invested a lot of money in OpenAI (roughly 13 billion dollars between 2019 and 2023).
What you see when you open Word, Excel, PowerPoint, Outlook, or Teams is the same AI, placed in seven different spots:
- In Word it helps you draft and rewrite.
- In Excel it builds formulas and analyzes data.
- In PowerPoint it generates slides from a document.
- In Outlook it summarizes long emails.
- In Teams it summarizes meetings and gives you transcripts.
- In GitHub (for developers) it completes code.
- In Windows 11 there's a Copilot button on the taskbar.
Same tool, seven different walls of your office.
The advantage that matters: you don't move
Think about how much effort it takes to use ChatGPT while you're writing a report in Word. You have to open the browser, go to the site, copy what you want, paste it, ask for a rewrite, copy the response, switch back to Word, paste. Seven steps.
With Copilot: click the button, ask for a rewrite, see the result right there. Two steps.
Sounds like a small difference. It isn't. Convenience wins the adoption war. A lot of people will never sign up for ChatGPT but are already using Copilot without thinking about it — because it comes inside the Word they open every day.
What Microsoft is winning and what it owes
The strength is obvious. Office has hundreds of millions of corporate users. Windows is on more than a billion machines. GitHub is where most of the world's code lives. If Microsoft slots AI into all of those at once, it doesn't need to convince anyone to install anything.
The weakness is just as obvious, though Microsoft prefers not to shout it. The engine belongs to OpenAI. If OpenAI hits a problem — political, technical, governance, whatever — Microsoft inherits the problem. That's why in March 2024 they hired Mustafa Suleyman (co-founder of DeepMind) to lead a new division called Microsoft AI, and a month later shipped their own family of small models, Phi-3. They're building an emergency exit.
What to take away
Three things to keep in mind:
- Microsoft didn't make the AI, but it placed it where you work. The engine is OpenAI's. The intelligence isn't Microsoft's. What Microsoft pulled off was dropping it inside Word, Excel, and Outlook, which you already have installed.
- Convenience wins. If your entire day runs through Office, Copilot is going to feel unbeatable for sheer ease. For email, spreadsheets, and decks, there's no time to jump tabs.
- Microsoft depends on OpenAI, and it knows it. Hiring Suleyman and launching Phi-3 in 2024 are clear signs the company is working to depend less. If OpenAI stumbles — and the Sam Altman firing weekend of November 2023 showed it can stumble — Microsoft needs a plan B. It's already wiring one.
On January 23, 2023, Microsoft announced an additional ten-billion-dollar investment in OpenAI. ChatGPT had been out for exactly two months and five days. There was no public certainty yet that it was more than a viral end-of-year toy. Satya Nadella, Microsoft's CEO, had seen something others hadn't quite seen — or had decided to place the bet before seeing it confirmed.
A few weeks later, in a Financial Times interview about the launch of Bing with ChatGPT, Nadella delivered the line that defined the year: "I want people to know that we made them dance." "Them" was Google. The line was about who was setting the tempo of the industry.
Three years on, that ten-billion-dollar bet turned into Copilot — the most widely distributed AI integration in the world. And into a structural dependence Microsoft is quietly trying to reduce.
What Copilot is, precisely
Copilot isn't a model. It's a brand Microsoft uses to group every AI integration inside its products. Underneath, the primary engine is OpenAI's GPT-4o. There are a few variants (smaller models for simple tasks, a modified Codex for GitHub) but the workhorse is the same one running ChatGPT.
The Copilot product line as of April 2026 has several layers.
Microsoft 365 Copilot (November 2023, $30 per user per month). The enterprise tier. Integrates into Word, Excel, PowerPoint, Outlook, and Teams with access to your organization's internal data via Microsoft Graph — the index Microsoft keeps of your email, files, and calendars. That connection to your own data is what makes it useful for companies: Copilot can summarize a meeting while referencing the documents shared beforehand, or draft an email based on the prior thread.
Copilot Pro (March 2024, $20 per month). The consumer tier. Less integration with internal data, but priority model access and Copilot in Office apps.
GitHub Copilot (shipped in 2021, before ChatGPT). Originally built on Codex — a model OpenAI trained on public code — it was Microsoft's first successful generative-AI integration. It crossed a million paid subscribers in 2024.
Microsoft Copilot (the app/browser, formerly Bing Chat). The general-purpose consumer chatbot, free with a lighter model and paid with GPT-4o.
Copilot in Windows 11. A taskbar button that opens a Copilot pane for OS-level tasks.
Copilot+ PCs (May 2024). A hardware category requiring an NPU (Neural Processing Unit) of at least 40 TOPS, built to run small models locally. It introduced the controversial Recall feature, which periodically snapshots your screen so Copilot can search your history — a feature that generated enough privacy backlash that Microsoft had to delay and redesign the rollout.
The scene behind the scene: dependence and plan B
The question that defines Microsoft in AI isn't "which model do they use?" It's "what happens if OpenAI fails?"
That question stopped being abstract over the weekend of November 17 to 21, 2023, when OpenAI's board fired Sam Altman on Friday and had to reinstate him by Tuesday. Over those five days, Microsoft offered Altman and Brockman an internal lab and hinted at absorbing the OpenAI team if needed. In the end it didn't have to. But the episode exposed the obvious: if OpenAI implodes, Microsoft has a serious problem.
Microsoft's answer to that risk came in two moves.
First, on March 19, 2024, Microsoft announced the hire of Mustafa Suleyman as CEO of a new division called Microsoft AI. Suleyman had co-founded DeepMind (sold to Google in 2014) and later Inflection AI. The hire came through an unusual deal: Microsoft paid roughly $650 million to Inflection for licenses and for taking nearly the whole team — a way to acquire without formally acquiring, sidestepping regulatory scrutiny.
Second, on April 23, 2024, Microsoft Research published Phi-3. It's a family of small models (3.8 billion to 14 billion parameters) trained to run locally on-device. They're not direct competitors to GPT-4 on raw capability, but they're good enough for many tasks and don't require calling OpenAI.
Both moves point to the same goal: reduce exclusive dependence on an outside supplier without breaking the commercial relationship that's paying off.
The real strength: distribution no one replicates
Now the other side. What Microsoft got brilliantly right.
Copilot's competitive advantage isn't technical. It's locational. Microsoft 365 has hundreds of millions of corporate seats. When a twenty-thousand-employee company decides to turn Copilot on, it doesn't have to pick a tool or migrate data — everything is already in Microsoft's ecosystem.
That single detail rewrites the enterprise adoption equation. To adopt ChatGPT Enterprise, a company needs a process: security review, SSO integration, user training, legal sign-off. To activate Copilot, the CIO signs an update to the Microsoft 365 contract they already have. There's a real difference between "onboard a new vendor" and "flip a feature at the vendor you already use."
GitHub Copilot is the clearest example of that dynamic. It shipped in October 2021 — more than a year before ChatGPT — built on Codex, when most of the world hadn't heard of generative AI. By 2024 it crossed a million paid subscribers. Productivity studies from GitHub suggest 20 to 30 percent gains in development speed, with the caveat that the data comes from the vendor. Real adoption among professional developers runs around 30 to 40 percent per 2024 Stack Overflow surveys.
GitHub had 92 million developers. The distribution was already built. AI came on top.
The honest part: where Copilot is unbeatable and where it isn't
Worth taking this apart carefully, because any pro-Copilot-on-everything or anti-Copilot-on-everything read is a simplification.
Where Copilot is unbeatable today: everyday work inside Office for someone who lives there. Summarizing a cluttered inbox, drafting a first pass of a formal email, building a pivot table in Excel from a description, converting a Word doc into a PowerPoint deck. In those cases the integration wins on pure convenience.
Where Copilot isn't first choice: high-stakes work where reliability outweighs convenience. Legal contract analysis, reviewing code that's going to ship, synthesizing research documents with verifiable citations. Here Claude — my main tool in consulting — has a consistent edge because it follows literal instructions with less drift and flags what it doesn't know more often. For that kind of work, the habit of opening a new tab and switching to Claude, even uncomfortable, pays better than staying in Copilot out of ease.
The practical reality for many professionals is multi-vendor. Copilot for the routine inside Office. Claude or ChatGPT for work that demands a tighter review bar. Nobody forces a single choice.
To close, and to keep going
Microsoft executed the most lucrative strategy of the post-ChatGPT era: bet hard and early on the best model on the market, buy priority access, and use a distribution that already existed to amplify it. It's the same move it made with DOS-Windows in the eighties, Internet Explorer in the nineties, Azure in the 2010s. Microsoft knows how to convert distribution into victory.
The open question — and it's worth asking — is how stable a strategy is when it depends on a partner you don't fully control. OpenAI isn't Microsoft. It has its own governance, its own internal tensions, its own incentives. Hiring Suleyman and launching Phi-3 say Microsoft is already thinking about that risk.
If you want to dig into how model capability gets measured and compared, How AIs are measured is the next link. For the broader competitive picture, The AI race.
Do you use Copilot inside Office, or do you jump to another AI when the work gets serious?
GitHub Copilot launched in technical preview on June 29, 2021, built on a model called Codex — a GPT-3 variant OpenAI had fine-tuned specifically on public code from GitHub. At that point, ChatGPT didn't exist. DALL-E was an academic paper. The "consumer generative AI" category hadn't formed. And yet Microsoft, through its subsidiary GitHub, was already shipping a commercial generative-AI product integrated into a tool 60 million developers used daily.
That launch deserves to be the starting point of any serious analysis of Microsoft's AI strategy. Because Microsoft Copilot in Office (November 2023) was narratively the mass-breakthrough moment, but GitHub Copilot in 2021 was the operational dress rehearsal where the company tested the thesis that defines its whole current position: existing distribution + external model + deep integration = structural advantage. When ChatGPT exploded, Microsoft already knew how to do this. It just had to scale.
The economic architecture of the OpenAI relationship
What conventional analyses call "Microsoft's investment in OpenAI" is, in practice, a layered commercial agreement. Worth unpacking precisely, because the structure explains much of both companies' strategic behavior.
First tranche, 2019: one billion dollars. Not pure capital. A significant portion is Azure credits — OpenAI pays with access to Microsoft infrastructure. Microsoft gets the right of first negotiation on model licensing.
Second tranche, January 2023: ten additional billion. Here the structure gets more complex. According to Bloomberg and The Information reporting, the deal includes Microsoft receiving 75 percent of OpenAI's profits until it recovers the investment, then a significant share (49 percent per some sources) up to a threshold, then a minority position. In parallel, OpenAI uses Azure as its exclusive compute provider — meaning much of the money returns to Microsoft as cloud billing.
Third tranche, further 2024-2025 investments. Later reporting puts the cumulative total near thirteen billion. Exact numbers aren't public.
What matters for understanding the strategy: Microsoft didn't buy OpenAI. It bought a preferential license over OpenAI's technology, with clauses protecting its access even if OpenAI's governance changes, and a financial structure where operational risk sits mostly on OpenAI's side (it has to generate enormous profits to "exit" Microsoft's privileged position).
The exposure moment: November 2023
The Altman-firing weekend exposed the structural fragility of that architecture. When OpenAI's board moved to remove him on Friday the 17th, Nadella got the call minutes after the public announcement. Per later reporting from The New York Times and The Atlantic, Microsoft had less than two hours of notice — a clear governance-relationship failure.
Nadella's response was immediate: he offered Altman and Brockman an internal Microsoft lab with unlimited resources. The implicit message was more interesting than the explicit one: Microsoft was demonstrating it could, if it had to, absorb OpenAI's technical team and keep development going in-house. That latent threat was a factor in Altman's reinstatement the following Tuesday.
But the episode, though "resolved" from outside, left an operational lesson. Microsoft walked away with concrete evidence that OpenAI could make decisions affecting it without prior consultation. Microsoft's internal response to that lesson was the diversification plan it deployed through 2024.
Microsoft AI and Phi-3: building plan B
On March 19, 2024, Microsoft announced the formation of Microsoft AI as a consumer-facing division, with Mustafa Suleyman as CEO. The Inflection AI deal was unusual: Microsoft paid approximately $650 million in licenses without formally acquiring the company, and took nearly the entire team — roughly 70 people including Karén Simonyan as chief scientist. This structure sidestepped traditional acquisition review and gave Microsoft an in-house model-research capacity it hadn't previously had.
On April 23, 2024, Microsoft Research published the Phi-3 technical paper (arXiv:2404.14219). The family includes Phi-3-mini (3.8B parameters), Phi-3-small (7B), and Phi-3-medium (14B). What's technically notable isn't the size — these are relatively small models — but the training strategy: the authors (Microsoft Research) report Phi-3-mini, with 3.8B parameters, achieving performance comparable to Mixtral 8x7B on several standard benchmarks using a fraction of the compute. The reported trick is "aggressive training-data curation" — less data, but higher quality.
Phi-3's strategic position is distinctive: it doesn't compete with GPT-4o on raw capability. It competes on another axis: the ability to run locally on a Copilot+ PC without sending the query to OpenAI. That means Microsoft can build on-device AI features (advanced autocomplete, local-file summarization, real-time transcription) without depending on the OpenAI API for every call.
Read together with Copilot+ PC hardware architecture (40+ TOPS NPU minimum requirement), the strategy gets clearer. Microsoft is building an on-device AI layer that can execute autonomously — and that, if OpenAI is ever unavailable, keeps working. It doesn't cover every use case (complex tasks will still need large cloud models) but it covers the mass-consumer use cases where volume and latency matter more than absolute capability.
The Recall controversy and what it reveals
In the May 2024 Copilot+ PC announcement, Microsoft introduced a feature called Recall: the operating system periodically captures snapshots of your screen, indexes them locally using on-device models, and lets you search your history in natural language. The product pitch was strong: "remember what you did three weeks ago."
The backlash was immediate and severe. Security researchers (notably Kevin Beaumont) demonstrated that the Recall database was accessible without proper encryption in preview versions, and that any malware with user-level access could extract years of visual history including passwords, private documents, and messages. Microsoft delayed the launch, redesigned the security architecture (encryption, explicit opt-in, biometric authentication for access), and shipped a more restricted version later.
The episode reveals a structural tension inside Microsoft worth naming. The company has internal pressure to "ship AI everywhere" that stems directly from committing 13 billion to OpenAI — it needs to justify that investment with visible products. That pressure produces product decisions prioritizing speed over review, and Recall is the clearest case of a feature reaching real users before maturing on critical security dimensions.
Competitive positioning: distribution versus capability
Worth assessing Microsoft's position in AI against the rest of the field using a precise analytical frame.
In-house model capability. Microsoft trails OpenAI, Anthropic, and Google on frontier models. Phi-3 competes in the small-model layer, but there's no in-house GPT-4o or Claude Opus. The OpenAI dependence at the frontier is structural.
Enterprise distribution. Microsoft leads everyone. Office 365 has around 400 million paid seats (financial-analyst estimates 2024-2025). No competitor has anything comparable in direct enterprise reach.
Developer distribution. GitHub Copilot dominates its category. The main competitors — Anthropic's Claude in terminals, Amazon CodeWhisperer, Codeium — capture minority fractions. Cursor and other AI-first IDEs are gaining ground, but GitHub Copilot has the edge of being integrated where the code already lives.
Cloud infrastructure. Azure is the second-largest in the market (behind AWS). The Azure + OpenAI API combo gives Microsoft a unique position: it can offer enterprise customers access to OpenAI models under enterprise privacy agreements (Azure OpenAI Service) that OpenAI directly doesn't offer with the same guarantees.
Direct consumer. Here Microsoft trails. The Microsoft Copilot app has far less use than ChatGPT for general consumer queries. The Copilot brand in consumer didn't gain traction equivalent to what it gained in enterprise.
The pattern is clear. Microsoft dominates where existing distribution converts to advantage. It trails where it has to build the category from scratch.
Editorial thesis
I'll close with my own read.
The thesis Microsoft implicitly proposes — that distribution outweighs model capability — is correct in the short and medium term. Over the next three to five years, the company will keep dominating mainstream enterprise AI because Office is where the corporate world works, and Copilot is what comes with Office. That's a win already paying off.
But that thesis has a structural blind spot, one the Suleyman + Phi-3 episode confirms. Microsoft depends on OpenAI for the frontier engine. OpenAI isn't Microsoft. It has its own governance, its own tensions (the Altman weekend, the Sutskever and Leike departures, the superalignment team dissolution), and can move in ways Microsoft doesn't control. The commercial architecture Microsoft negotiated protects a lot — but it doesn't protect against an institutional collapse of OpenAI or against OpenAI deciding unilaterally to compete head-on.
Plan B (Mustafa Suleyman, Phi-3, Copilot+ PCs with NPUs) is built precisely for that risk. It's an implicit acknowledgment that the dependence is the Achilles' heel. The question defining Microsoft's next five years in AI isn't "can it win the enterprise segment?" It's "can it win the enterprise segment without depending on a supplier it doesn't control?"
My read: Microsoft won't break with OpenAI as long as OpenAI produces the best model. But it will build, in parallel, an in-house layer that lets it — if ever needed — substitute the dependence progressively. The pace of that build will be the most important leading indicator. If Phi-5 or Phi-6 in 2026-2027 reach parity with frontier models, the power relationship between Microsoft and OpenAI gets redrawn. If not, the dependence stays structural — and the Achilles' heel stays there, under the surface of an unbeatable distribution.
What's your empirical test for deciding whether an enterprise AI integration is mature: when it does the task well, or when it also doesn't leave you exposed if the model provider wobbles?
