The Lyceum: AI Daily — Jul 30, 2026

AI is reaching the income statement in two sharply different ways. Microsoft sells the infrastructure and the assistant, while Meta uses AI to improve an advertising machine that must finance everything else. Meanwhile, an independent developer is trying to squeeze a 26-billion-parameter model through a MacBook’s solid-state drive—a small experiment with large implications for local agents.

Recency note: Reuters’ reports on DeepSeek’s chip development, Meituan’s domestic-chip training, President Donald Trump’s planned oversight order, and Anthropic’s access restrictions fall outside this edition’s strict 24-hour window. Microsoft’s fresh results reflect Reuters’ broader infrastructure-spending theme rather than recycling it as a separate story; The Washington Post’s Pentagon press-policy report is outside this newsletter’s AI remit.

Microsoft’s AI investment is beginning to show up in revenue—and in paid seats. The company reported $90 billion in quarterly revenue on Wednesday, with Azure revenue growing 43% on the session. It also said Microsoft 365 Copilot has passed 30 million paid seats and that new subscriptions more than doubled from the previous quarter.

Those are Microsoft’s figures, but the business model behind them is unusually legible. Microsoft sells computing capacity, models, the workplace interface and, increasingly, the agents operating inside corporate systems. It does not wait for AI to improve someone else’s product; it charges at several layers of the stack. (Microsoft’s AI Spending Is Starting to Look Like a Flywheel)

The bill remains extraordinary. Microsoft said it spent $41 billion on infrastructure during the quarter, roughly two-thirds of it on shorter-lived equipment such as processors. Yet the company generated $19.6 billion in free cash flow. That does not prove AI spending pays for itself, but it shows Microsoft can absorb an enormous buildout without surrendering cash generation. (Microsoft’s AI Spending Is Starting to Look Like a Flywheel)

The next test is straightforward: Copilot renewals and agent use must rise while Azure maintains its momentum. Failure would look quieter—rising infrastructure spending, slowing paid-seat growth and steadily narrowing free cash flow. (Microsoft’s AI Spending Is Starting to Look Like a Flywheel)

Meta’s AI is sharpening its core business, but the accounting is getting uglier. Quarterly revenue rose 28% on the session to $60.8 billion, according to the Associated Press. Meta CEO Mark Zuckerberg said AI is accelerating the company’s core business and creating enterprise opportunities, although Meta did not disclose a separate AI revenue figure.

The costs were harder to hide. AP reported that profit fell 14% on the session, expenses rose 55% on the session and free cash flow dropped 91% on the session to $784 million. The quarter also included $2.4 billion in legal charges and $1.18 billion in severance costs, so AI cannot fairly be blamed for the entire decline.

That leaves Meta with a strategic problem. It mostly monetizes AI indirectly: better recommendations keep people scrolling, and better targeting helps advertisers spend. Microsoft can place a price on Copilot seats and Azure capacity; Meta must prove that smarter feeds can carry the cost of frontier-model development. (Microsoft’s AI Spending Is Starting to Look Like a Flywheel)

If Meta’s enterprise products or “personal superintelligence” strategy produce a measurable revenue line, the spending will look like construction. If AI revenue remains indistinguishable from ordinary advertising gains while cash generation stays compressed, it will look more like an open-ended research subsidy.

A MacBook may be able to run a model that would ordinarily demand far more memory. Independent developer Andrey Mikhaylov released TurboFieldfare 0.3 on July 29, a specialized inference engine for running Google’s Gemma 4 26B-A4B model on Apple Silicon. Instead of holding the full model in memory, TurboFieldfare keeps a small shared core and working cache in RAM while streaming other weights from the Mac’s solid-state drive.

Mikhaylov reports roughly five to six tokens per second on an 8 GB M2 MacBook Air and substantially higher speeds on newer hardware. Those numbers have not been independently reproduced, and the software is specialized for one text-only Gemma checkpoint rather than being a general local-model platform.

If the technique transfers to other mixture-of-experts models—which activate only part of a model for each task—consumer laptops could become credible hosts for private, tool-using agents. That would reduce cloud costs and weaken safety systems built around centralized API access.

The downside is equally clear. Failure would look like a clever demo trapped by storage speed, battery drain and narrow model support. Watch for independent benchmarks, support for additional models and sustained performance during long agent sessions rather than short prompts.

Microsoft is feeding GPUs into a furnace that prints Copilot invoices. Meta is shoveling legal bills into an advertising engine, while a MacBook sips Gemma weights through an SSD straw.

Forty million agents apparently have corporate badges now; whether any of them know where the meeting room is remains unreported.

Keep an eye on the cash flow.

Forward this to the colleague who still thinks “AI adoption” is one thing.