=== THE OPEN-WEIGHT FIGHT ===
Jensen Huang used his first-ever X post last week to enter Washington's fight over open-weight AI. He shared an open letter initially signed by 25 companies, including Microsoft, Meta, and Palantir. The coalition urged policymakers to preserve the development and deployment of open-weight models.
Satya Nadella quickly amplified the message, arguing that openness is essential to a healthy AI ecosystem.
[Embedded tweet — Satya Nadella, Jul 24 2026: "Open-weight models are essential to a healthy AI ecosystem. Together with others across our industry, we are outlining a path for open-weight models to strengthen American competitiveness and expand economic opportunity, while protecting national security."]
To understand why the industry is clashing over this policy, it helps to distinguish how AI software is delivered:
- Closed models (proprietary): The model's internal weights remain locked behind an API. Companies such as OpenAI and Anthropic control the security guardrails, computing infrastructure, and pricing. - Open-weight models: The pre-trained weights can be downloaded and inspected. Businesses can customize and host the software on their own infrastructure without paying every query fee to a single vendor.
The immediate catalyst is model distillation, where developers use outputs from leading closed models to train cheaper alternatives. It is the same concern behind US accusations that Chinese lab Moonshot AI used a leading American model to improve Kimi. Closed-model providers argue that unrestricted distillation amounts to intellectual-property theft. Supporters of open weights see it as part of the competitive process that lowers costs and broadens access.
OpenAI, Anthropic, and Google were initially absent from the letter. OpenAI later joined after Sam Altman said he wanted the US to lead in both open and closed models, while Google's Sundar Pichai endorsed the effort on the company's behalf. Anthropic remained the clearest holdout, arguing that it does not want to ban open models but supports stricter chip controls, action against industrial-scale distillation, and mandatory safety testing for all capable systems.
NVIDIA reinforced the campaign this week by launching the Open Secure AI Alliance, a coalition of nearly 40 companies building tools to defend against AI-powered cyberattacks. Microsoft, SpaceX, and IBM are founding members. Anthropic, OpenAI, and Meta are notably absent.
The policy fight matters because Microsoft is hedging both outcomes. It maintains a multibillion-dollar closed-model alliance with OpenAI while positioning Azure as the indispensable platform for open-weight deployment. Microsoft does not need to predict which model architecture wins. It wants to provide the compute, governance, and security layer underneath all of them.
Today at a glance: 1. Microsoft's Q4 FY26. 2. The sovereignty play. 3. Earnings call takeaways. 4. What moves the needle?
=== 1. MICROSOFT'S Q4 FY26 ===
Income Statement: - Revenue +18% Y/Y to $90.0 billion ($2.4 billion beat). - Gross margin 67% (-1pp Y/Y). - Operating margin 45% (flat Y/Y). - Non-GAAP EPS $4.74 ($0.50 beat).
Core business segments: - Productivity and Business Processes grew 14% Y/Y to $37.8 billion, supported by M365 Copilot, E5, and early E7 momentum, alongside stronger M365 commercial products revenue. - Intelligent Cloud grew 32% Y/Y to $39.3 billion, driven by 43% Azure growth as Microsoft improved fleet efficiency and brought new capacity online faster. - More Personal Computing declined by 4% Y/Y to $12.9 billion, as weaker Windows OEM and Xbox revenue outweighed 10% growth in Search advertising.
Cash flow: - Operating cash flow grew 30% Y/Y to $55.4 billion. - Free cash flow declined by 23% Y/Y to $19.6 billion.
Balance sheet: - Cash, cash equivalents, and investments: $76.8 billion. - Long-term debt: $31.1 billion.
So what to make of all this?
Azure broke out of its range: Azure grew 43%, above management's 39%-40% constant-currency guidance. Intelligent Cloud reached $39.3 billion, more than $1 billion above the high end of the outlook. Microsoft delivered a genuine reacceleration, though it still trailed Google Cloud's extraordinary 82% segment growth.
Copilot added 10 million seats in one quarter: Microsoft 365 Copilot surpassed 30 million paid seats, up from 20 million in Q3 and 15 million in Q2. Penetration remains below 8% of Microsoft's commercial base, but the seat count has doubled in six months.
Xbox's reset is showing up in the numbers: Xbox content and services revenue fell 10%, its third consecutive quarterly decline, and decreased 5% for the year. With additional impairment charges and a disclosed margin near 3%, Microsoft is acknowledging that years of acquisitions have produced inadequate returns.
CapEx consumed the cash-flow upside: Cash CapEx more than doubled to $35.8 billion, taking the full-year total to $115.9 billion, up 80%. Infrastructure spending is growing far faster than operating cash flow, pushing free cash flow lower.
The backlog is enormous but concentrated: Commercial RPO rose 84% to $678 billion, more than twice Microsoft's annual revenue. OpenAI remains the largest disclosed component, providing exceptional visibility with a lingering concentration risk.
AI is pressuring gross margins: Microsoft Cloud gross margin fell three points to 65%, while Intelligent Cloud gross margin dropped to roughly 57%. Yet the segment's operating margin held near 41% because operating expenses increased only 10% against 32% revenue growth.
Adjusted EPS still included Anthropic: Microsoft excluded a $480 million OpenAI investment gain from non-GAAP earnings but left a $3.2 billion Anthropic gain included. Adjusted EPS therefore still benefited materially from non-operating investment accounting.
=== 2. THE SOVEREIGNTY PLAY ===
Microsoft and Mistral just expanded their partnership by several billion dollars. Mistral's Medium 3.5 and OCR 4 models are now available in Microsoft Foundry, while Medium 3.5 is also available in Copilot Studio. More importantly, Azure supports deployments ranging from the public cloud to fully disconnected environments.
The deal in plain terms: - Open weights with an enterprise wrapper: Customers get control over Mistral's models without assembling the infrastructure, governance, and security layers themselves. - Run anywhere: Azure supports cloud, cloud-connected, and fully disconnected deployments, allowing regulated customers to keep sensitive data off the public internet. - Billions behind it: Microsoft made a multibillion-dollar commitment to use Mistral's expanding European GPU capacity while distributing its models through Foundry and Copilot Studio.
Sovereignty is the sales pitch. Banks, governments, healthcare providers, and defense customers increasingly want frontier AI without sending every query to a closed, US-hosted API. Microsoft can offer that control while keeping the deployment inside Azure.
The Mistral deal confirms a broader pattern. Microsoft Foundry now lists more than 11,000 models spanning OpenAI, Anthropic, Meta, Mistral, DeepSeek, xAI, Cohere, NVIDIA, and Hugging Face.
Customers may choose radically different models, but Microsoft keeps the interface consistent. The same APIs, governance, billing, monitoring, and safety tools can wrap around GPT, Claude, Llama, Mistral, or DeepSeek.
Microsoft does not need to own the model that ultimately wins. It needs to own the environment where enterprises deploy it.
The harness is the moat:
Microsoft made that strategy tangible this week with a new cybersecurity stack. The underlying model matters, but the orchestration layer around it may matter more.
What launched: - A first-party cyber model: MAI-Cyber-1-Flash is Microsoft's first in-house cybersecurity model, trained to identify vulnerabilities across large codebases and connect them to known exploits and defenses. - A model router: The system handles roughly 90% of tasks itself and escalates only the hardest 10% to OpenAI's GPT-5.4. - A better cost curve: Microsoft says the combination scored 96% on its CyberGym benchmark while costing nearly 50% less than its current MDASH configuration. - Agents on top: Project Perception coordinates red-, blue-, and green-team agents to simulate attacks, investigate threats, and drive remediation.
[Embedded tweet — Satya Nadella, Jul 27 2026: "Today, we are announcing a series of updates that give customers frontier-grade security at half the cost. MAI-Cyber-1-Flash is our first cybersecurity model, built ground up to find the most challenging vulnerabilities in complex code bases. When combined with MDASH, it …"]
The moat sits in the surrounding data and operational context. Microsoft processes more than 100 trillion security signals each day across 1.6 million customers, giving it a history of exploits, vulnerabilities, and fixes that rivals would need years to reproduce.
Models will keep improving, and inference prices will keep falling. Microsoft's harder-to-replicate advantage is the harness around them: proprietary security context, enterprise distribution, governance, and agents that can turn a model's answer into an action.
=== 3. EARNINGS CALL TAKEAWAYS ===
On backlog diversification: "All sequential commercial RPO growth was driven by commitments from customers outside of Frontier Model companies, and RPO increased 25% when excluding OpenAI." — OpenAI still dominates the headline backlog, but the concentration is no longer worsening at the margin. Microsoft added no sequential RPO from frontier labs, while the rest of the commercial backlog continued to compound rapidly.
On Azure's next leg: "Customer demand continues to exceed available capacity. […] That additional in-quarter capacity for Azure was quickly monetized. […] In Azure, we expect revenue growth of approximately 45% in constant currency […] Even with the strong close to Q4, we continue to expect H1 growth to accelerate." — The Q4 beat came from extracting more throughput from existing hardware and bringing new capacity online faster. With Azure guided to accelerate again in Q1, 43% growth does not look like the peak.
On Copilot moving beyond seats: "Average weekly engagement is on par with Outlook and Teams. […] The number of customers with more than 50,000 seats increased over 7x year-over-year […] We now have this per seat business model, we also now have the usage business model. It's seat plus usage." — Copilot is reaching the engagement of Microsoft's daily communication tools while expanding from a fixed subscription into a second, consumption-based revenue stream.
On the AI cost curve: "We are seeing 40% better performance per watt when running MAI models on Maia 200. […] We are seeing significant efficiency gains, including 89% reduction of GPU costs in Dynamics 365 with MAI-Voice-2-Flash and up to 84% reduced GPU costs in PowerPoint with MAI-Image-2.5." — Microsoft's first-party models and silicon are already producing dramatic unit-cost reductions. The strategy is not merely to reduce dependence on external vendors, but to protect margins as AI usage scales across the portfolio.
On the CapEx accounting change: "Effective at the start of FY27, we are extending the estimated useful life of our data centers and office buildings from 15 to 25 years […] The greater impact is on capital expenditures, as more of our future data center leases will shift from finance leases to operating leases […] Finance leases are included in capital expenditures, while operating leases are not." — FY27 CapEx will become harder to compare with prior years. Some infrastructure commitments will shift outside the reported CapEx figure because of lease classification. The spending remains, but the accounting optics improve.
=== 4. WHAT MOVES THE NEEDLE ===
Can efficiency outrun depreciation? Microsoft is already cutting unit costs through Maia and smaller first-party models, but Intelligent Cloud's cost of revenue still rose 43%. Can those savings scale fast enough to stabilize cloud gross margins?
Will open weights remain deployable? Microsoft's strategy assumes enterprises can keep choosing among American, European, and Chinese models. A narrow crackdown on industrial-scale distillation would be manageable. Broader restrictions on open weights or Chinese models would weaken Foundry's breadth and the sovereignty pitch.
Can Microsoft remain the gateway? The AI thesis depends on Windows, Entra, Microsoft 365, and Azure reinforcing one another. Regulators increasingly view that integration as potential self-preferencing. Forced interoperability would not destroy demand for Copilot, but it could weaken the distribution moat surrounding it.
================================================================
Disclosure: I own AMD, AMZN, GOOG, NVDA, META, MSFT, and PLTR in App Economy Portfolio. I share my ratings (BUY, SELL, or HOLD) with App Economy Portfolio members.