Microsoft AI investments — 11,000 models on the shelf

Azure, Foundry, Copilot and enterprise data let Microsoft profit while OpenAI, Anthropic, Mistral and its own models compete task by task.

Microsoft turns AI investments into enterprise software rivalries with one ruthless move: every model becomes replaceable inventory. OpenAI, Anthropic, Mistral, xAI and Microsoft’s own MAI models can fight for each task while Microsoft controls the platform and sends the bill.

I heard Satya Nadella explain this during Microsoft’s July 29 earnings call, which I was half-listening to because apparently investor transcripts now count as summer reading. Somewhere, my younger self in Ivrea closed the laptop and went outside.

Microsoft spent years investing in AI labs so that the eventual winner would matter less to Microsoft.

The strategy now has serious weight behind it. Azure has passed $100 billion in annual revenue. Microsoft 365 Copilot has more than 30 million paid seats. Microsoft Foundry offers over 11,000 models, a catalog with strong Costco energy.

Microsoft can monetize a model, sell the compute beneath it, connect it to company data and charge for the app where the answer appears. My nonna would understand immediately. Own the espresso machine and the café lease; let everybody else argue about the beans.

The $41 billion tollbooth

Microsoft reported $90 billion in revenue for its fiscal 2026 fourth quarter, up 18% year over year, in its July 29 earnings release. Microsoft Cloud generated $59.3 billion, up 27%, while Azure and other cloud services grew 43%.

AI infrastructure spending has looked like a national space program with worse merch. Microsoft finally has visible revenue arriving from both sides.

Activate Consulting CEO Michael J. Wolf told the Associated Press on July 29 that Microsoft was “winning on both fronts.” Enterprises pay for the Azure infrastructure running their AI, then pay again for Copilot inside software their employees already use.

After 20 years of building and shipping technology products, I’ve seen distribution humiliate technical elegance more times than I can count. A better product can win a demo. The product already sitting on every employee’s laptop wins procurement.

Nadella put the scale plainly in Microsoft’s July 29 earnings release:

“This year, Azure revenue surpassed $100 billion for the first time, and Microsoft 365 Copilot reached over 30 million paid seats, reflecting the confidence customers are placing in us to power their AI transformation.”

Those 30 million seats give Microsoft an enormous laboratory for testing features, prices and model substitutions. A standalone AI company still has to convince procurement that its shiny new tool belongs inside the company. Microsoft can tuck AI into an enterprise agreement already covering Outlook, Teams, Excel, Windows and enough security products to break the spirit of one exhausted IT manager.

Then there is the spending.

Microsoft spent $41 billion on capital expenditures during the quarter, according to AP. CFO Amy Hood said an accounting change would put calendar 2026 guidance near $175 billion, while the underlying investment expectations remained unchanged. AP also cited Microsoft’s previous expectation of roughly $190 billion, including about $25 billion caused by higher component prices.

I underestimated how aggressively Microsoft would keep spending. I expected discipline to arrive sooner, especially once inference economics began resembling a restaurant where everybody orders lobster and pays for a sandwich.

Microsoft can tolerate the feast because its enterprise contracts buy time. Commercial remaining performance obligation reached $678 billion, up 84%, according to the earnings release. Those signed commitments help turn GPUs and datacenters into recurring consumption.

Nadella told investors how quickly the physical footprint is growing:

“We added 31 new datacenters across 5 continents this quarter, bringing the total to 88 this year, as we expand our footprint in response to accelerating demand.”

The AI labs provide the intelligence. Microsoft owns a growing chunk of the building where it reports for work.

OpenAI and Anthropic become inventory

Microsoft Foundry now carries more than 11,000 models, including products from OpenAI, Anthropic, Mistral and xAI alongside Microsoft’s MAI family.

Nadella described the catalog during Microsoft’s fiscal 2026 fourth-quarter call:

“We offer the broadest model catalog in the cloud, with over 11,000 models, including the latest from OpenAI, Anthropic, Mistral, xAI, as well as our own MAI family.”

Customers genuinely benefit from that choice. A bank can use one model for document extraction, another for coding and a smaller local model inside a regulated workflow. Quality matters. So do latency, cost and the physical location of the data.

Foundry becomes more valuable with every model added because Microsoft handles the evaluations, deployment, billing and policy enforcement. Once customers can swap models task by task, the model starts behaving like an ingredient.

Microsoft says the number of customers building with models from multiple providers increased fivefold during the first half of 2026.

Nadella gave investors the figure directly:

“Since the start of the year, we have seen a 5X increase in the number of customers building with models from multiple providers.”

Levi Strauss & Co. is using OpenAI and Anthropic models in Foundry while bringing more than 1,000 domain-specific agents into one enterprise AI platform. Levi’s can change model suppliers. Microsoft remains embedded in the architecture.

The financial relationships make this wonderfully awkward. Microsoft recorded a $3.2 billion gain from its Anthropic investment during the quarter. Its full-year statements separately showed a $4.963 billion increase to net income from the impact of OpenAI investments.

Microsoft benefits when a lab appreciates in value while engineering its products to survive without that lab.

Cold. Also brilliant.

The infrastructure follows the same playbook. Microsoft says its Maia 200 chips support OpenAI and MAI models, delivering 30% better performance per dollar than the latest-generation hardware already in its fleet. Cobalt virtual machines power Microsoft workloads and systems run by customers such as Adobe, Arm, Elastic, OpenAI, Sprinklr and TomTom.

Every workload gives Microsoft more scale. Every new model gives Foundry another option. Partners can keep competing without anybody dramatically storming out of the alliance, although I assume several lawyers are permanently typing.

Excel knows what the model labs don’t

Microsoft’s nastiest advantage lives inside ordinary user behavior.

A general model company can train on code repositories and spreadsheets. Microsoft can measure whether a developer accepts a GitHub Copilot suggestion in VS Code, returns two days later or repeatedly rewrites the output while whispering words unsuitable for the stand-up meeting.

Excel shows whether an agent completed a workflow or sent somebody back to manual formulas and quiet despair.

Benchmarks rarely capture that.

According to a July 23 post from Microsoft AI’s Superintelligence team, MAI-Code-1-Flash achieved an approximately 10% higher code acceptance rate than GPT-5.4 Mini and Claude Haiku 4.5 in VS Code. Developers were 6% more likely to return across multiple days than with GPT-5.4 Mini, and 11% more likely than with Claude Haiku 4.5.

Microsoft AI stated the acceptance result directly:

“It has an approximately 10% higher code accept rate than GPT 5.4 Mini and Claude Haiku 4.5 in VS Code.”

Acceptance rate beats a benchmark trophy because it comes from somebody using the model under actual deadline pressure, with a pull request waiting and Slack already becoming hostile.

Microsoft then took the MAI-Code-1-Flash checkpoint and trained it in an Excel reinforcement-learning environment. Production feedback showed quality comparable to GPT-5.6 on common Excel tasks. The specialized model could also run on Nvidia H100 and A100 GPUs instead of requiring the latest accelerators.

That transfer should make model suppliers nervous. Microsoft owns the product-specific evaluation harness, so it can see where a frontier model earns its price and where a smaller MAI model can take over.

Nadella made the ambition explicit in comments covered by ITPro on July 23:

“We are now seeing MAI models outperform general-purpose frontier models in many use-cases while using a fraction of the tokens.”

The savings are already substantial. Microsoft reported an 89% reduction in GPU costs inside Dynamics 365 using MAI-Voice-2-Flash. In PowerPoint, MAI-Image-2.5 cut GPU costs by as much as 84%.

I learned a smaller version of this lesson while building a connected espresso machine for Pascucci. We had the physical machine, cloud services and brew telemetry inside one product. The ugly problems always appeared where those layers touched. Whoever could see the full loop had a diagnostic advantage over every isolated supplier.

Yes, I have used coffee telemetry as a business lesson. I was born in Italy. I’m legally required to make caffeine sound strategic.

GitHub and Excel give Microsoft that loop at gigantic scale. The labs build increasingly capable intelligence. Microsoft writes the exam and watches millions of people take it.

A visual showcasing various AI models, highlighting Microsoft's extensive investment in artificial intelligence technology.

Image alt text: How Microsoft turns AI investments into enterprise software rivalries across Azure, Foundry, Copilot and MAI.

Caption: Partners provide models. Microsoft owns the production loop that decides which ones keep the job.

Permission to press “execute”

Enterprise AI becomes valuable when it can change a purchase order, approve access, resolve an IT ticket or contact a customer. An elegant answer in a chat window is cute. Permission to alter a live business process is money.

The platform needs to know who the employee is, what the policy allows and which customer record is current. Every action must also survive the auditor who arrives six months later with the emotional warmth of airport security.

Microsoft’s expanded Databricks partnership reaches directly for that control. The July 23 agreement extends into the 2030s and integrates Databricks Genie and Unity AI Gateway across Entra, Power BI, Purview, Foundry, Microsoft 365, Teams and Copilot.

Databricks says more than 20,000 organizations use its platform, including 70% of the Fortune 500. It plans to expand its own use of Azure Databricks for core business operations and adopt Cobalt 200, which Microsoft says offers up to 50% better performance than its predecessor.

Foundry competes for developers. Databricks brings Microsoft closer to governed enterprise data. Copilot owns the employee interface, while Dynamics pushes into the business-application layer.

ServiceNow sees the same control point. Its AI business crossed $1 billion in annual contract value during Q2 2026, and production agentic deployments increased ninefold in nine months, according to the company’s July results. ServiceNow ended the quarter with 658 customers generating more than $5 million each in annual contract value.

Chairman and CEO Bill McDermott opened with characteristic restraint:

“ServiceNow’s exceptional Q2 results solidify our position as the fastest-growing major enterprise software and cybersecurity company.”

ServiceNow calls itself the “AI control tower” and has expanded the product to govern agents wherever they run. Its Action Fabric lets ServiceNow and third-party AI execute work through ServiceNow workflows, with Anthropic as the first design partner.

That puts ServiceNow against Microsoft’s attempt to govern agents through Foundry, Entra and Microsoft 365. The two companies can announce partnerships while fighting over the same customer. Enterprise software has always been very civilized that way.

SAP has an equally credible attack because it owns systems of record. In its July 23 results, SAP reported a current cloud backlog of €22.9 billion, up 27%, while cloud revenue rose 22% to €6.28 billion.

CEO Christian Klein said SAP’s momentum comes from AI grounded in customers’ most critical business processes and data. He has the premise right. A supply-chain agent with access to live inventory and approval rules will beat a brilliant chatbot guessing from a PDF.

Salesforce is pushing through observability. Agentforce’s Session Tracing Data Model records user input, planner decisions, retrieved sources, action flows, errors and final outputs. Its waterfall traces follow work across multiple agents, while citations take reviewers to the source material used.

Salesforce has also made its standard Agentforce observability stack unmetered. Builders can inspect production traces without burning additional Data Cloud credits on the default tooling.

Model quality wins the demo. Execution rights win the account.

Europe needs to own more than the model

I’m genuinely happy about Microsoft’s expanded agreement with Mistral. Europe needs AI champions capable of building serious models, and Mistral is one of the few European labs operating at globally relevant scale.

The July 21 agreement includes a multibillion-dollar Microsoft commitment to use Mistral’s expanded European GPU infrastructure. Mistral is adding thousands of Nvidia Vera Rubin GPUs, while Mistral Medium 3.5 and OCR 4 are entering Microsoft Foundry. Medium 3.5 is also available inside Copilot Studio.

The deployment options matter for regulated industries. Customers can run Mistral models in Azure’s public cloud, inside customer-controlled environments or fully disconnected through Azure Local. A disconnected system can support defense and critical-infrastructure workloads where an external API is unacceptable.

Molto bene.

The deal also exposes Europe’s weakness. Mistral gains compute and worldwide distribution. Microsoft retains the main enterprise doorway through Azure, Foundry and Copilot Studio. Europe can produce a superb model and still rent access to the customer.

Mistral CEO Arthur Mensch said in the July 21 Microsoft-Mistral announcement that the partnership gives Mistral access to enterprises and public institutions at global scale. He is right to take the distribution. I would take it too.

Europe now needs to build beyond laboratories. We need European cloud capacity, orchestration products, enterprise channels and infrastructure that customers can actually deploy. A European model inside an American platform’s dropdown gives buyers useful choice. Technological sovereignty requires European ownership across much more of the stack.

European Commission President Ursula von der Leyen set the direction when she announced InvestAI on February 11, 2025:

“We want AI to be a force for good and for growth. We are doing this through our own European approach, based on openness, cooperation and excellent talent.”

The European Commission said InvestAI would mobilize €200 billion for AI investment, including €20 billion for European AI gigafactories.

I support that approach passionately. Public funding still has to produce companies that keep the customer relationship. Otherwise, Europe funds research, trains talent and supplies strategically valuable models while an American hyperscaler handles distribution and captures the product data that compounds over time.

Microsoft understands sovereignty as a product feature. European founders and policymakers need to understand it as market structure.

Model choice can deepen platform lock-in

An 11,000-model catalog reduces dependence on a single AI lab. It can also deepen dependence on Microsoft’s routing system, identity layer, governance tools and application stack.

Swappable ingredients do not make the restaurant portable.

Nadella told investors that every organization should build its own “continuous learning loop” and avoid outsourcing core intellectual property. Buyers should apply his advice to Microsoft too.

Before committing a production agent to Foundry or Copilot, I’d ask five questions:

Where do our product-specific evaluations live?
Can we export agent memory and execution traces?
Who controls identity and permissions?
Can we replace the model without rebuilding the workflow?
Are we buying a measurable outcome, or financing a migration wearing an AI costume?

The same questions belong in SAP and Oracle negotiations. A TechRadar Pro analysis by Chad Stewart noted that SAP has promoted more than 200 specialized agents coordinated through roughly 50 domain-specific assistants. Oracle’s Fusion agents live inside Fusion Cloud, so customers running E-Business Suite must re-platform before using them.

The upgrade can swallow the budget before an agent proves its value. TechRadar cited Americas’ SAP Users’ Group research in which 61% of members named budget as their biggest challenge.

I’ve shipped systems across mobile apps, IoT hardware, machine learning and cloud infrastructure. Vendor lock-in usually enters through the component that accumulates operational knowledge. The original technology remains replaceable on paper; its history becomes painfully sticky.

Enterprises should negotiate ownership of evaluations, traces, memories and permission mappings with the intensity they once reserved for database exports. Those assets form the learning loop. Whoever controls them improves faster and collects rent longer.

By 2028, enterprise buyers will spend less time debating benchmark leaders and more time asking who controls an agent after it joins the org chart.

OpenAI, Anthropic and Mistral will keep producing extraordinary intelligence. Microsoft is building the workplace where that intelligence gets hired, evaluated, permissioned and eventually replaced.

My bet is simple: the most powerful AI company in enterprise software will decide when the smartest model is worth paying for.

Right now, Microsoft is writing that decision into Excel.

Frequently asked questions

How does Microsoft make money from enterprise AI?

Microsoft monetizes enterprise AI at several layers: Azure compute, Foundry model deployment and governance, connections to company data, and Copilot inside workplace software. Enterprises can pay for infrastructure running AI and then pay again for AI features within Microsoft applications already covered by existing agreements.

Does Microsoft Foundry’s model choice reduce vendor lock-in?

Offering more than 11,000 models reduces dependence on any single AI lab, but customers can become more dependent on Microsoft’s routing, identity, governance, evaluation and application layers. Models remain swappable while execution traces, permissions, memories and product-specific learning loops accumulate inside the Microsoft platform.

What advantage does Microsoft have over independent AI model labs?

Microsoft can observe how models perform inside products such as VS Code, Excel, Dynamics 365 and Microsoft 365 Copilot. That product feedback provides evaluations based on acceptance, repeat use, workflow completion and cost, helping Microsoft decide when frontier models justify their price and when smaller specialized models can replace them.