What 211 Q2 Earnings Reports Told Us About the AI Buildout

The Paradox

Every quarter, companies report how much money they made. Wall Street analysts predict those numbers in advance. When a company earns more than analysts expected, that's called "beating earnings." When it tells investors to expect even more going forward, that's called "raising guidance."

This quarter, 82% of these 211 companies beat their earnings predictions. 75% beat on revenue. 71 raised their forecasts for the rest of the year.

Only 45% of their stocks went up. The typical stock reaction was negative.

Let that sit for a second. The best earnings season in the history of AI infrastructure — and more stocks fell than rose. The market has heard "AI demand is incredible" so many times that it's stopped paying for the words. It now wants receipts.

The single most predictive signal this season wasn't whether a company beat or missed. It was whether they raised their forecast or merely kept it the same. Unchanged guidance — not a miss, not a cut, just no upgrade — showed up in 32% of stocks that declined versus just 4% of stocks that surged. In this market, standing still is falling behind.

The Demand Is Not Slowing Down

Let's start with what the buyers are saying.

The biggest buyers in AI infrastructure are the four largest cloud companies — $AMZN, $MSFT, $GOOGL, and $META. They're the ones building data centers, buying chips, and ordering power plants. What they spend is what the rest of the supply chain sells.

$AMZN CEO Andy Jassy, in his own words: "We will not have enough capacity to meet all the demand we have in 2026. This dynamic will also be true in 2027. The demand we already have for 2028 is striking."

Amazon's cloud division (AWS) has a backlog — the dollar value of contracts customers have already signed but haven't used yet — that doubled from $244 billion to $496 billion in two quarters. That's not a forecast. That's signed paper.

Combined spending on new infrastructure across the four biggest cloud companies in 2026: roughly $720–765 billion. All raised or held at record levels. $GOOGL is so short on computing capacity that it's renting from competitors at premium prices — despite designing its own chips.

So the demand side is settled. What's more interesting is what's happening on the supply side.

The Shortage Has Moved Downstream

Everyone knows about the GPU shortage — GPUs are the specialized chips that power AI training and inference. Everyone's heard about $NVDA (which designs them), $TSM (which manufactures them), and $ASML (which makes the machines that print them). That story is real — $TSM's CEO says the demand-supply gap is "very big" and lasts "to probably 2029, 2030" — but it's also the most-covered story in technology.

What we're seeing this quarter is that the constraint has migrated. It's no longer just about chips. It's moved to the raw materials those chips are built on, the cables that connect them, and the power that feeds them.

The material inside every optical cable

Inside every high-speed fiber optic cable connecting AI servers, there's a tiny chip that converts data between electricity and light. That chip is built on a wafer made from a rare compound called indium phosphide. Without it, the cable doesn't work.

$AXT is a $3 billion company that makes these wafers. This quarter, management said they are refusing new orders because production is full. Customers have prepaid $47.7 million in cash for multi-year supply agreements — essentially paying years in advance just to hold their place in line. Backlog — orders placed but not yet delivered — is over $100 million and growing even as shipments accelerate. The company tripled its 2026 capacity expansion. Its profit margin on each wafer went from 29.9% to 45.0% in a single quarter.

When buyers start prepaying cash years in advance for a raw material, the constraint has definitively moved from the chip layer to the materials layer.

Copper is running out

Global copper inventory is down to 15 days of demand. Processing fees — what miners pay smelters to refine raw ore into usable metal — hit zero for the first time ever. That means smelters are now competing just to get ore to process, rather than the other way around. $FCX's CEO says customers "continue to report robust copper demand and order books associated with AI data centers." New supply doesn't arrive until 2027 at the earliest, with major mining projects not fully online until 2029-2030.

Testing equipment is sold out

Before any chip can ship, every single one has to be tested — baked at extreme temperatures for hours to find early failures. Think of it as quality control on a factory line, except each product costs thousands of dollars and you need specialized machines that can run millions of tests per year.

$AEHR, which makes these testing machines, reported new orders up more than 500% year-over-year. Its backlog grew 5.3 times. Revenue forecast for the coming year is roughly triple the year just ended. $FORM — which makes the tiny probe cards that physically touch each chip during testing — shipped $13.2 million above the high end of its own capacity-constrained forecast. $NVDA became a 10% customer for the first time, purely from networking chips, not GPUs. $COHU disclosed that the market for high-performance chip testing is essentially a two-company structure, and both are maxed out.

Six months ago, chip testing equipment was viewed as commodity back-end work. This quarter proved it's a genuine chokepoint.

The Power Wall

The AI buildout's biggest challenge is no longer making enough chips. It's generating enough electricity and getting it to the building.

$GEV — GE's power division, which makes the gas turbines that generate electricity — booked $24.2 billion of orders this quarter, up 88% from a year ago. To put that in scale: they signed contracts for 20 gigawatts of new turbine capacity in a single quarter. One gigawatt is roughly enough to power a million homes. Gas turbines under contract jumped from 83 GW to 116 GW in two quarters. The company is sold out through 2030 and has already pre-sold more than half of its 2031 production slots. Turbine prices are up over 20%, with customers funding $GEV's own factory expansion through advance deposits.

$GNRC — a company most people associate with the backup generator in your neighbor's yard — took $1 billion of data center power orders in 90 days. Lead times for this equipment are stretching to 70-80 weeks.

And then the twist: $BE (Bloom Energy) disclosed that customers are canceling already-ordered gas turbines to switch to fuel cells instead. That's the first real technology displacement we've seen in the AI power race — not two technologies competing for new orders, but one actively pulling orders away from the other.

But even the power generators aren't the final constraint. $JCI (Johnson Controls, which makes building systems) disclosed that roughly $6 billion of its $20 billion backlog is sitting idle — not because $JCI can't manufacture the equipment, but because customers are waiting on grid power and electrical infrastructure to be ready. The bottleneck has walked down another level: from the turbine to the wires that carry the electricity.

$AEP (American Electric Power) collected $2 billion in cash and collateral in a single month for 45 gigawatts of power applications on the Texas grid alone. $SO (Southern Company) signed 6 GW in one quarter, including a 3.2 GW contract with OpenAI. Utility five-year spending plans across the sector now collectively exceed $500 billion — more than the cloud companies' combined 2026 spending — and management at multiple utilities said these plans don't yet fully account for the load they've already signed.

Why Each Building Eats More

There's a simple reason every layer of this supply chain is strained, even though we're not building radically more data centers than before.

$EME (EMCOR, an electrical and mechanical contractor) disclosed the multiplier this quarter: an AI data center requires 1.5 times the electrical wiring and 1.5 to 2 times the mechanical work — cooling pipes, ventilation, plumbing — of a standard cloud storage building. Same footprint, much more engineering inside. That's why electrical transformers are on 36-month lead times. That's why $FIX (Comfort Systems, a mechanical contractor) has a $14.1 billion backlog growing 73% year-over-year, with data center customers so desperate for capacity that they're prepaying to fund modular construction — literally financing a competitor's expansion just to get in line.

The amount of stuff that goes into each server rack keeps surprising to the upside. $SIMO disclosed that each AI server rack needs 30 to 40 boot drives — small storage drives that load the operating system. That's a line item that wasn't in anyone's financial models a year ago. $AEHR says over 80% of its testing revenue now comes from AI processors. $FORM says $NVDA became a 10% customer purely from networking cards, not GPUs. Every company touching the per-rack bill of materials is seeing more components per server than anyone had modeled.

AI Is Crowding Out Your Next Phone

This is the finding that surprised me most.

The memory shortage created by AI data center demand is now directly cutting into consumer electronics. $ARM trimmed its forecast for smartphone licensing revenue because memory prices are inflating device costs. $INTC said PC demand is "down low-double-digits" in 2026, driven by rising memory prices. $QCOM's handset segment declined on memory-driven production cuts.

The AI buildout isn't just adding a new source of demand on top of existing demand. It's reallocating physical manufacturing capacity away from ordinary consumers. The same factories that used to make memory chips for your phone are now making memory for AI servers, and there aren't enough of them to do both.

$TXN — the most conservative chip company in the industry — confirmed it has "started executing price increases" for the first time in years. Their customers are telling them: "lined down, please help us."

Who Won, Who Lost, and Why

Winners — small companies at the tightest chokepoints

Losers — strong demand, punished for something else

What We're Watching Next Week

175 companies from our universe report August 4-8. Here are the six questions this season raised that next week should start to answer.

The Takeaway

211 companies told us the same thing this quarter: AI infrastructure demand is real, it's accelerating, and supply can't keep up — from copper mines to chip factories to gas turbines to the electrical grid.

But the market told us something different. It told us that "demand is incredible" is no longer news. It told us that beating earnings isn't enough — you have to raise your forecast. It told us that hiding your order book is worse than missing the quarter. And it told us that companies promising revenue in 2028 will be valued very differently from companies delivering revenue today.

The physical buildout is the biggest infrastructure project since the interstate highway system. What these 211 reports showed is that the constraint isn't whether there's enough demand to justify it. There's clearly more than enough. The constraint is whether the physical world — mines, factories, turbines, transformers, the grid — can be built fast enough to keep up.

Based on what we're seeing, it can't. Not yet.