AI Spending Needs Receipts Now

The Short Version

The AI capex story is splitting in two.

There is spending with receipts.

And there is spending with a vision.

Both can be reasonable.

Only one is easier to defend when the bill arrives.

On July 29, Microsoft reported quarterly revenue of $90 billion, Microsoft Cloud revenue of $59.3 billion, Azure and other cloud services revenue up 43%, and more than 30 million paid Microsoft 365 Copilot seats. On the earnings call, Satya Nadella also said Agent 365 has nearly 40 million agents registered across tens of thousands of companies, Foundry has 100,000 customers, and Microsoft cut latency and infrastructure costs across several AI workloads.

Microsoft also spent a lot: $41 billion in quarterly capital expenditures.

But it still produced $19.6 billion in free cash flow.

Meta told a different story. In its Q2 results, revenue rose 28% to $60.8 billion, but costs and expenses rose 55%, net income fell 14%, free cash flow dropped to $784 million, and capital expenditures were $31.1 billion. Meta now expects 2026 capex of $130 billion to $145 billion.

On Meta's earnings call transcript, Mark Zuckerberg laid out the bet: AI is improving the core ads and recommendations business now; personal agents will become the next product wave; business agents, APIs, compute, and enterprise services may become new revenue lines.

That is not nonsense.

Meta has real distribution.

Meta has real ad gains.

Meta has more than 1 million businesses using Meta Business Agents every week.

But the useful contrast is this:

Microsoft's AI story is starting to look like an operating model.

Meta's AI story still looks like a portfolio of promises.

Very expensive promises. The spreadsheet has entered the chat.

The market is asking a better question

For the last few years, the AI question was:

Who has the best model?

Then it became:

Who has enough compute?

Now it is becoming:

Who can turn compute into measurable customer behavior?

That is the right evolution.

A company can spend $40 billion on infrastructure and still be rational if the spending turns into revenue, retention, workflow lock-in, lower unit costs, and products customers use every day. A company can spend $30 billion and still make investors nervous if the return path depends on a future assistant category that is not yet proven.

This is not because investors suddenly became anti-AI.

Microsoft is not being punished for spending heavily. It is being rewarded because it can attach the spending to visible demand:

• Azure passed $100 billion in annual revenue.
• Azure and other cloud services grew 43%.
• Microsoft 365 Copilot passed 30 million paid seats.
• Copilot net seat adds more than doubled quarter-over-quarter.
• Foundry revenue more than doubled year-over-year.
• Agent 365 has nearly 40 million registered agents.
• GitHub Copilot revenue accelerated after usage-based billing.
• Free cash flow remained very large despite capex.

That does not prove every Microsoft AI product is perfect.

It proves the story has receipts.

There are seats, contracts, usage, cloud demand, margins, and named enterprise deployments.

That is what an AI investment case starts to look like after the demo phase.

Copilot is no longer only a product

The most interesting Microsoft detail is not only the 30 million paid Copilot seats.

It is the stack around them.

Nadella described Microsoft as separating the harness, context, memory, and action space from any one model family. That sounds like earnings-call poetry until you realize it is the product strategy.

Microsoft does not want Copilot to be a wrapper around one model.

It wants Copilot to be the enterprise surface where models can be swapped, governed, grounded in company data, priced through seats and consumption, and distributed through Microsoft 365, GitHub, Dynamics, Fabric, Foundry, Security, and Azure.

That is why the model-choice line matters. Microsoft says customers are increasingly building with models from multiple providers, and that Foundry includes models from OpenAI, Anthropic, Mistral, xAI, and Microsoft's own MAI family.

The product is not:

"Use our model."

It is:

"Bring your workflows here, and we will route intelligence through the system."

That is a much stronger enterprise position.

If models get cheaper, Microsoft can benefit.

If customers want Claude for one task and OpenAI for another, Microsoft can still be the control plane.

If agents need identity, logs, memory, sandboxes, evals, data connectors, and compliance, Microsoft already owns many of the boring enterprise surfaces.

Very unfair. The boring surfaces were the good surfaces all along.

Meta's core business is working

The easy version of this story is:

Microsoft good, Meta bad.

Too simple.

Meta's core business is clearly not broken.

Revenue grew 28%. Family of Apps ad revenue grew 27%. Zuckerberg and Susan Li both pointed to AI improving recommendations, ad ranking, creative tools, and internal product development. Meta says 9 million small businesses are using at least one of its AI ad creative tools. It is processing public Reels and Feed posts through LLMs to improve content understanding. It is using Muse models in recommendations and creative advertising.

That is real AI monetization.

Not a science project.

Not a chatbot demo.

AI is making the ad machine better.

The problem is that the ad machine may not be enough to justify the size and shape of the next investment wave.

Meta is not only saying:

"AI improves ads."

It is also saying:

"We are building personal agents for billions of people, business agents for advertisers, an API business, productivity tools, glasses, and maybe a compute business for large customers."

That is a lot of story.

Some of it may work.

Some of it probably will.

But each layer has a different proof requirement.

Better ad ranking can be measured in revenue and conversions.

Business agents can be measured in paying businesses, message volume, sales completion, retention, and support cost.

Personal agents need trust, repeated use, cross-app context, permissions, privacy, and a reason to exist beyond "Meta also has an assistant."

Compute rental needs enterprise-grade contracts, isolation, margins, support, reliability, and strategic trust from customers who may compete with Meta.

One capex line is trying to serve all of that.

That is why the market squints.

Personal agents are the hardest promise

Zuckerberg's personal-agent pitch is ambitious:

agents that work 24/7 on your behalf to help with goals, health, relationships, finances, and whatever else you want.

That is a huge consumer claim.

It is also much harder than coding agents.

Zuckerberg said this directly: engineers are technical and willing to spend time making coding agents work. Personal agents need to work out of the box for billions of people.

Exactly.

Consumer agents have less tolerance for setup.

Less tolerance for weird failures.

Less tolerance for "please connect these seven systems and review these logs."

Less tolerance for a model that misunderstands a relationship, a budget, a health goal, a calendar, or a private conversation.

Meta does have an advantage: WhatsApp, Messenger, Instagram, Facebook, Marketplace, ads, groups, creators, and glasses. That is an enormous amount of social and commercial context.

It also has the scar tissue.

The more personal the agent, the less forgiving the trust problem.

Meta says it launched incognito mode for WhatsApp and the Meta AI app, and that privacy and security will be fundamental to its agents. Good. Necessary.

But consumer trust is not established by a sentence in an earnings call.

It is established by product defaults, controls, data boundaries, recovery paths, and the repeated feeling that the assistant is helping rather than extracting.

That will take time.

The capex bill will not wait politely.

Compute is a hedge, not proof

Meta also keeps pointing at compute as a possible business.

That makes sense. I wrote recently that Meta's new AI product might be compute, and the logic still holds. If Meta builds more capacity than it can use internally at a given moment, and other labs or enterprises are desperate for capacity, selling compute can make the infrastructure story less binary.

But compute rental is not a magic refund button.

Susan Li said Meta is getting offers for compute at a significant premium over what it paid and is evaluating opportunities. That is useful demand evidence.

It is not yet proof of a durable cloud business.

External compute customers need isolation, support, compliance, security reviews, contracts, region choices, uptime guarantees, billing, migration help, and confidence that Meta will not become a strategic headache. Microsoft, Amazon, Google, Oracle, CoreWeave, Crusoe, and others already live in that world.

Meta can enter it.

But "we have capacity" is not the same as "we are a cloud platform."

The useful investor question is not whether compute demand exists.

It does.

The question is whether Meta can convert that demand into high-quality, recurring revenue without distracting from the consumer-agent and ad businesses that are supposed to justify the same infrastructure.

Again: receipts.

What builders should take from this

For builders, the lesson is not to become an equity analyst.

The lesson is that AI products are moving out of the demo economy.

If you are buying, building, or pitching AI, the question is no longer:

"Can the model do something impressive?"

It is:

• Who uses it every week?
• Who pays for it?
• Does usage expand after deployment?
• Does it connect to the systems of record?
• Does it reduce latency, cost, errors, or labor in a measurable way?
• Can customers govern it?
• Can the vendor run it profitably?
• Can the vendor keep capacity available?
• Does pricing match value, or only token burn?
• Does the product survive when the novelty fades?

The more agentic the product, the more these questions matter.

An agent that runs once is a feature.

An agent that runs every week is an operating expense, a risk surface, a governance object, and maybe a new workflow.

That needs a better justification than "AI is the future."

It needs evidence.

The bottom line

Microsoft and Meta are both spending like AI is the next platform.

The difference is that Microsoft can now point to more signs that AI is already flowing through its enterprise machine: paid Copilot seats, Azure demand, Foundry customers, GitHub usage, Agent 365 registrations, and cash generation alongside capex.

Meta can point to a very strong ads business and credible AI improvements inside it.

But its bigger story still asks for belief: personal agents, business-in-a-box agents, glasses, APIs, rented compute, and a distributed superintelligence philosophy wrapped around one of the most expensive infrastructure plans in tech.

Maybe that works.

Meta has surprised people before.

But the AI market is entering a less forgiving phase.

The demo is not dead.

The benchmark is not dead.

The vision is not dead.

They just have to bring receipts now.