This will be the week the numbers lie. “Earnings” will dominate headlines, but they won’t tell you what matters. At Meta (META: NASDAQ) and Microsoft (MSFT: NASDAQ), the most important figure is buried in the cash flow statement. Capital expenditures—not revenue, not margin—reveal how serious these companies are about building for the AI age.

There was lots of news on that front last week: Amazon (AMZN:NASDAQ) and especially Comfort Systems (FIX: NYSE), a mechanical contractor most investors have never heard of — which has outperformed NVIDIA! (NVDA:NASDAQ)-- showed us the ferocity of the AI buildout.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

Companies Discussed

In This Note:

• Companies Discussed

• The HVAC Co. Beating Nvidia

• Tech Earnings Preview: Burn, Build, Repeat

• Tweet O’ The Week

• Epistrophy In The News

• 📆 of Epistrophy Events

• Availability This Week

The HVAC Co. Beating Nvidia

The story of Comfort Systems USA, Pipefitters of the AI Boom

Houston-based Comfort Systems USA (FIX:NYSE) was once seen as a boring HVAC company, a sleepy mechanical contractor. But Friday’s results told us more about the AI boom than any chipmaker’s report.

FIX has outperformed nearly every peer in industrial construction and tech infrastructure. Over the last five years, Comfort Systems has returned 787 percent to shareholders, including dividends. Over ten years: 2,234 percent. That’s more than Nvidia (NVDA: NASDAQ), which returned 764 percent and 1,972 percent over the same periods, respectively. Comfort doesn’t build chips—it builds the buildings that power them. In an economy obsessed with generative AI, that’s enough.

(Full disclosure, it’s a stock we’ve owned for most of a decade.) 

It’s less a regional HVAC installer than a national data center arms dealer. What was once a fragmented rollup of commercial plumbing and electrical businesses now provides liquid-cooled modules for hyperscalers and semiconductor fabs, delivering high-efficiency thermal and power systems faster than stick-built construction allows. FIX, once seen as a boring construction stock, is quietly shaping the AI economy’s physical foundation—duct by duct, watt by watt.

The company’s shift toward high-specification technical infrastructure has accelerated in the last two years. In Q2 2025, Comfort Systems reported record revenue of $2.17 billion, a 20% year-over-year increase, driven by technology-sector demand. “This is the first time that our quarterly revenue has exceeded $2 billion,” CEO Brian Lane said on the July 25 earnings call. Gross margin reached 23.5 percent, an extraordinary level in a labor- and materials-intensive business, and EBITDA jumped 50 percent year-over-year to $334 million. But the most telling figure wasn’t on the income statement. Backlog hit a record $8.1 billion, up $2.4 billion from a year ago. Roughly 40 percent of that revenue is tied to the technology sector, up from 31 percent the prior year. Customers aren’t just booking Comfort Systems for cooling coils—they’re reserving capacity years in advance to ensure power and thermal delivery for data centers still on the drawing board.

Comfort Systems specializes in mechanical, electrical and plumbing (MEP) systems—the circulatory and nervous systems of modern buildings. Its high-margin growth stems from three interlocking competencies: modularization, liquid cooling and design-build integration. Roughly 18 percent of revenue this year has come from off-site modular construction, up from 12 percent last year and 4 percent five years ago. The company now maintains over 2.7 million square feet of modular capacity, expanding to 3 million in 2026. These facilities allow Comfort to assemble complex mechanical racks and liquid-cooled units in controlled environments, dramatically compressing jobsite timelines and limiting labor risk. “We’re trying to do something together,” CFO William George said of the company’s most loyal data center clients. “And we’re choosing who we give our unbelievable and scarce resources to.”

Liquid cooling, once an experimental edge case, is becoming standard for AI data center thermal loads. Comfort Systems isn’t a parts supplier or OEM—it integrates building-scale heat rejection systems that keep racks running within spec at 100+ kilowatts per cabinet. These aren’t commodity installs. The company’s engineering teams work upstream with developers to specify air-handling units, cooling towers, heat exchangers and piping systems suited for the unique thermal profiles of each build. By coupling design and installation—so-called “design-build” delivery—Comfort achieves better margins and more control than traditional plan-and-spec contractors. “The best opportunities right now are presenting themselves more often than not on the technology side,” COO Trent McKenna said. “And companies are choosing to put their skilled workforce in the best possible circumstances to be successful.”

As of June 30, Comfort Systems had 7,935 active projects, with an average contract value of $1.8 million and a pipeline that extends into 2027. Modular construction remains labor-intensive, but offers scheduling certainty that customers increasingly prize. That certainty commands a premium. “Our pricing is obviously very good,” Lane said. “We’re getting paid for the risk and services that we’re delivering.” Gross margins in the electrical segment reached 25.3 percent this quarter, while mechanical margins hit 22.9 percent. Those figures reflect strong execution, favorable mix, and what the company politely calls “good working conditions” with customers who respect timelines, scope and safety.

The economics are beginning to compound. For the first six months of 2025, Comfort generated $400 million in net income on $4 billion in revenue. That’s a 10 percent net margin in a business that once posted mid-single digits. Service work—usually considered a steadier, less exciting profit center—grew more than 10 percent and now accounts for $1.2 billion of annual revenue. Free cash flow reached $222 million in Q2 alone, despite rising working capital needs. As of June 30, Comfort had no net debt and more than $250 million in cash after buybacks and the acquisition of Right Way Plumbing.

To be sure, Comfort Systems does not operate in a vacuum. Large general contractors, unionized MEP firms and hyperscaler-dedicated suppliers are all vying for data center work. Private equity is funding modular challengers, and some customers have encouraged rivals to build capacity. “We don’t think what we do can’t be done by someone else,” George said. “We just try to be so good at it that you’d be crazy to buy it from anybody else.” The moat isn’t IP—it’s speed, precision and trust. But as data center developers scale up gigawatt campuses, the demands on Comfort’s project management, logistics and labor pipelines will grow just as fast. The company’s continued ability to command premium pricing depends not only on skill but also on capacity discipline and customer selection. For now, management is leaning into both.

Comfort Systems no longer fits neatly into legacy construction categories. Its clients are no longer just hospitals and hotels but also cloud providers and chip foundries. Its product is no longer just labor—it’s uptime. “We just make sure we’re the best people to build it,” Lane said. As the world of technology is learning, increasingly, they are.

Tech Earnings Preview: Burn, Build, Repeat

Tech earnings used to be about the income statement—revenue growth, margins, earnings per share. But this is the quarter when the most important number isn’t found on the P&L. It’s buried in the cash flow statement – capital expenditures. For companies building artificial intelligence infrastructure, CapEx has become a direct measure of strategy – both measurement stick and divining rod. This week, at Microsoft (MSFT: NASDAQ), Meta Platforms (META: NASDAQ) we’ll see how it now defines their business models.

And Apple (AAPL: NASDAQ)’s cap ex numbers might be showing their hand as well.

Microsoft spent $11.4 billion on capital expenditures last quarter, up from $9.5 billion a year earlier and up sequentially from $10.7 billion in the prior quarter. That spending supported a broad infrastructure expansion, but much of it on AI. As a result, Chief Executive Officer Satya Nadella said GPU lead times dropped 20 percent, cost per token fell by more than half and performance per watt increased 30 percent. The company opened data centers in 10 countries and processed 100 trillion tokens during the quarter.

That infrastructure is already feeding product usage. Microsoft Fabric, its unified data platform, now has more than 21,000 paying customers, up 80 percent year over year. GitHub Copilot has 15 million users, up fourfold. Foundry, Microsoft’s AI agent service, is used by 70,000 organizations and supports models from OpenAI, Meta, Cohere, Mistral, and others. Power Platform now counts 56 million monthly active users.

Meta Platforms spent $13.7 billion on capital expenditures in the first quarter of 2025, outpacing its $10.3 billion in free cash flow. The company raised its full-year CapEx guidance to between $64 billion and $72 billion, up from $60 billion to $65 billion. Chief Financial Officer Susan Li attributed the increase to additional data center buildout and rising hardware costs. The spend is being used to construct the infrastructure stack Meta says it needs for AI agents, multimodal assistants, and hardware.

Chief Executive Officer Mark Zuckerberg outlined five product categories: AI-enhanced advertising, content recommendation, business messaging, Meta AI and devices. A new Reels ad model increased conversions by 5 percent. Threads reached 350 million monthly active users. Meta AI, they say, now serves nearly 1 billion users. The company launched a standalone Meta AI app and began scaling LLaMA 4, with a larger model still pending.

Business messaging is also infrastructure-led. In countries like Vietnam and Thailand, businesses already use WhatsApp to sell directly to consumers. Zuckerberg said AI will enable the same model in high-wage economies. “I expect that every business will have an AI agent,” he said. ‘Natch, they expect that capability will use the infrastructure Meta is building now.

Apple is in a different spot. 

But Apple disclosed “only” $6.0 billion in capital expenditures for the first half of fiscal 2025. That figure has remained flat for years— even as the company announces new U.S. investments and AI ambitions. It’s entirely likely the savings from abandoning “Project Titan”, the Apple Car, have been replaced in the shift towards AI.

Apple has announced plans to spend $500 billion in the United States over four years, including server manufacturing in Texas and chip sourcing from Arizona. But there’s little evidence of that in its filings. 

Apple Intelligence, the company’s new generative AI initiative, emphasizes privacy. Many features run on-device via the Apple Silicon neural engine; others rely on a new system called Private Cloud Compute. Chief Executive Officer Tim Cook said the M4-powered Mac Studio can run 600-billion parameter models in memory. But Apple has yet to disclose which models it uses, what infrastructure supports them, or how they scale globally.

Together, these three companies are showing how CapEx now signals operational direction. Capital expenditures used to support the business – now they divine the future of the business. 

Tweet O’ The Week

Epistrophy In The News

It was cool to be quoted in this Forbes Australia about the Australian fintech startup Airwallex. The innovation coming from Aussie tech firms is something to behold.

I also joined the Schwab Network to preview big tech earnings and explain why infrastructure spend is now the best forward indicator of strategy. Over at NewsNation, I broke down the Trump administration’s new AI policy agenda—what’s real, what’s not—and later returned to discuss the rise of AI-enabled consumer devices, including why an AI-enabled smart device that spies on you might actually be useful.

📆 of Epistrophy Events

Availability This Week

I’m in San Francisco this week and available for briefings, meetings or even coffee. Particularly interested in connecting with folks thinking seriously about energy constraints in AI infrastructure—or anyone building the things that build the future.

Written reports are available to clients, with video summaries on YouTube and, of course our popular summaries of the summaries (yes, the second derivative) on Instagram and Tiktok.

Are these notes are helpful to you? Suggestions? I’d love to discuss them further and, as always, comments, questions and ideas are appreciated. If you have a friend or even a frenemy whom you think might benefit from this note, have them reach out and I’ll put them on the list.