---
title: "Who is ultimately paying for AI infrastructure? This round of earnings reports provides a new answer."
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/topics/43139888.md"
description: "In the first few articles this week, we have been looking at the same issue, just approaching it from different companies. SK Hynix showed us that money is now being picked over; Microsoft and Meta make us ask whether the landlords are continuing to build and if tenants are arriving; Amazon delivered a report card stating 'rent is verified, but cash recovery is still under construction.' But putting these earnings reports together, what truly deserves a wider lens is another, larger question: this AI infrastructure..."
datetime: "2026-08-01T22:36:11.000Z"
locales:
  - [en](https://longbridge.com/en/topics/43139888.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/43139888.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/43139888.md)
author: "[熊猫校长Ming](https://longbridge.com/en/profiles/16386316.md)"
generator: "portal-rs"
---

# Who is ultimately paying for AI infrastructure? This round of earnings reports provides a new answer.

In the first few articles this week, we have been looking at the same issue, just approaching it from different companies. SK Hynix showed us that money is starting to be spent selectively; Microsoft and Meta prompted us to ask whether landlords are continuing to build towers and if tenants have arrived; Amazon delivered a report card stating "rent has been verified, but cash recovery is still under construction."

But putting these earnings reports together, **what truly deserves a zoomed-out perspective is another larger question: Has this AI infrastructure actually formed an economic cycle capable of sustaining itself?** Cloud providers can keep expanding their data center, chip, network, and power construction sites; however, if the ultimate payers remain only a handful of frontier model companies, this cycle will still face questions regarding customer concentration and financing closed loops.

This round of earnings reports gives me more confidence, not because Capex has grown again, but because there is a new answer to "who ultimately pays." **The payers are shifting from a small group of frontier model companies to a broader range of enterprises; moreover, enterprises are not just purchasing compute power on the cloud but are also integrating AI into their own workflows, paying per seat and by usage.**

![Image](https://pub.pbkrs.com/uploads/2026/8ff09506328d568c7c75e9b7b92d14ac?x-oss-process=style/lg)

## **I. Payers are spreading from a few model companies to enterprises**

Microsoft provided a signal previously undervalued by the market: **nearly 90% of its annual Microsoft Cloud revenue comes from customers outside of frontier model companies; the incremental portion of commercial RPO in the quarter also came entirely from these customers.** This does not mean that companies like OpenAI and Anthropic are unimportant; they remain key sources for frontier capabilities, compute demand, and product innovation. What has truly changed is that beyond frontier model companies, traditional enterprises are growing into a second pillar of payment.

Amazon provided validation in the same direction. AWS Q2 revenue accelerated from a 28% year-over-year growth in Q1 to 37%; both AWS's AI business and chip business have individually reached annualized revenues exceeding $25 billion. **We cannot attribute all of AWS's growth directly to AI revenue, but it at least indicates that external enterprise customers' demand for cloud, models, and compute capabilities is becoming stronger.**

What changes here is the risk structure. In the past, the market worried that several cloud providers would invest increasingly more, ultimately just renting equipment to a few model companies; these model companies would then finance and procure from each other, making income look lively while payment sources remained insufficiently diversified. Now, the simultaneous signals from Microsoft and Amazon are: **enterprise customers are beginning to become more significant payers, and AI infrastructure no longer relies solely on a few major tenants, with revenue sources starting to broaden.**

## 

## **II. Enterprises are not just renting buildings but also buying AI tools within them**

If enterprises purchasing Azure and AWS prove they are willing to pay for cloud capabilities; then Microsoft's disclosure regarding Copilot allows us to see for the first time more clearly: **enterprises are paying for AI entering their daily workflows. Paid seats for Microsoft 365 Copilot exceeded 20 million last quarter and surpassed 30 million this quarter, with net new seats doubling quarter-over-quarter.**

This number is not "30 million Agents," but over 30 million paid enterprise Copilot seats. Its value lies not just in sales volume, but in the change of payment form: enterprises are no longer just procuring cloud resources for their IT departments but are beginning to integrate AI into Office, code development, data analysis, and business processes, becoming tools employees actually use daily.

**Furthermore, Microsoft is moving its business model from purely per-seat pricing to "seats plus usage." Cowork has already started usage-based billing, with thousands of clients paying for its use.** GitHub Copilot has also introduced usage-based billing; the company stated that after the new model takes effect, Business/Enterprise seats continue to grow, and significant consumption revenue has emerged, with Copilot revenue accelerating over 60% quarter-over-quarter. For cloud providers, this means AI applications not only add a layer of subscription revenue but may also bring more elastic consumption revenue as actual workload increases.

Here, another figure needs clarification. Agent 365 has nearly 40 million registered Agents within two months, distributed across tens of thousands of companies; this shows enterprises are incorporating Agents into their governance, identity, and security systems, but registered Agents are not paid seats and cannot be directly counted as revenue. It currently serves more as a scale and deployment signal, and future observation will focus on whether it can convert into more disclosable commercial results.

## 

## **III. Why are tenants willing to continue paying rent? Because they are calculating costs themselves**

Previously, we used "landlords, buildings, and tenants" to understand cloud providers' capital expenditures: landlords build, tenants move in, rents increase, and finally, breakeven is discussed. But this round of earnings reports added the most critical link: why are tenants willing to rent? Because they are calculating costs themselves.

Enterprises are willing to buy seats for Copilot and pay for Cowork by usage not to fill an AI budget. They expect less repetitive labor, faster product iteration, smoother internal processes, and ultimately better operational results. A series of signals disclosed by Microsoft—increased large client deployments, faster formation of high usage rates, and more clients starting to pay by usage—all indicate that enterprises are moving from pilot projects to deeper workflow deployments.

However, boundaries must be maintained here. Layoffs, organizational downsizing, and automation can show that enterprises are recalculating operational accounts and that the motive of "cost reduction first, efficiency improvement later" is real; but it cannot be deduced backwards that "the number of people laid off equals the number of people replaced by AI." Why enterprises pay and whether they have already achieved long-term, stable ROI are two different levels of issues.

Level one: Enterprises have started paying—cloud revenue, RPO, Copilot seats, and usage fees are direct evidence. Level two: Why enterprises are willing to renew and expand—cost reduction, speed-up, and higher usage rates are the ROI signals we are seeing. Level three: Whether AI can stably translate into revenue growth and margin improvement across most industries still requires verification industry by industry.

## 

## **IV. AI infrastructure is starting to transform from a line into a verifiable loop**

In the past, the market mostly saw a one-way industrial chain: cloud providers invest Capex, and GPUs, HBM, networks, and power receive orders. Now, we can start drawing it as a loop: Enterprise-verifiable ROI → Enterprise AI seats, usage, and cloud spending → Cloud provider revenue and RPO → AI Capex → GPU / HBM / Network / Power → Stronger, cheaper AI capabilities → Enterprise ROI.

This is the AI Positive Flywheel—the virtuous cycle of AI infrastructure. The first three links are the evidence this round of earnings reports is providing: spread of enterprise payers, acceleration of Copilot seats, emergence of usage billing, and growth in AWS external demand; the middle Capex and upstream orders are the familiar AI Factories; the final link is whether enterprises truly continuously obtain ROI, which determines whether this chain can return to the start.

**Therefore, who ultimately pays determines how long this round of AI infrastructure can run.** If enterprise-side ROI continues to be validated by more earnings reports, more workflows, more renewals, and expansions, then cloud provider revenue will in turn support Capex, Capex will support the upstream industrial chain, and the entire system will no longer be just unilateral capital investment but will begin to possess its own economic returns.

## 

## **V. This also gives us a clearer condition for admitting mistakes**

Being bullish on AI infrastructure does not mean just staring at how much Meta will spend next year, nor immediately negating all investments upon seeing free cash flow turn negative in a single quarter. What truly needs continuous observation is whether this loop continues to turn.

Next quarter, I will focus on four things: whether Copilot paid seats and Cowork usage payments continue to accelerate; whether enterprise RPO and cloud revenue continue to prove payer diffusion; whether Agent 365's registration scale can convert into more disclosable commercial signals; and whether high enterprise usage rates can translate into more specific ROI, while cloud provider revenue and operating cash flow can gradually accommodate high Capex.

**Conversely, if enterprise AI ROI does not continue to appear, enterprise payments will slow down first, followed by cloud revenue and RPO slowing down, cloud provider Capex cooling down, and finally, orders for GPUs, HBM, networks, and power coming under pressure.** This is the path along which this positive cycle might reverse in the future, and where we should be ready to correct our judgments at any time.

However, at least this round of earnings reports has allowed us to see something previously unclear: **AI infrastructure is no longer just cloud providers unilaterally throwing money into construction sites. Enterprise payers are increasing, and AI is beginning to move from cloud capabilities into workflows that enterprises are willing to pay for daily.**

In previous articles, we finally saw the tenants; in this article, we begin to see why tenants are willing to continue paying rent.

![Image](https://pub.pbkrs.com/uploads/2026/8a59475a1fc1134ba17d5dcae5e0dd46?x-oss-process=style/lg)

Real-market observation, does not constitute investment advice.

-- Panda Principal Ming | AI Infrastructure Research

Related research targets:

$Amazon(AMZN.US) $Microsoft(MSFT.US) $Alphabet - C(GOOG.US) $Meta Platforms(META.US)

### Related Stocks

- [SKHY.US](https://longbridge.com/en/quote/SKHY.US.md)
- [SKHYV.US](https://longbridge.com/en/quote/SKHYV.US.md)
- [AMZN.US](https://longbridge.com/en/quote/AMZN.US.md)
- [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md)
- [AMZU.US](https://longbridge.com/en/quote/AMZU.US.md)
- [MSFL.US](https://longbridge.com/en/quote/MSFL.US.md)
- [MSFO.US](https://longbridge.com/en/quote/MSFO.US.md)
- [MSFD.US](https://longbridge.com/en/quote/MSFD.US.md)
- [MSFU.US](https://longbridge.com/en/quote/MSFU.US.md)
- [MSFX.US](https://longbridge.com/en/quote/MSFX.US.md)

---
> **Disclaimer: This article is for reference only and does not constitute any investment advice.**