Four CEOs, One Bottleneck: Big Tech's $700 Billion AI Bill

Four Calls, One Complaint

On Wednesday afternoon, within a few hours of each other, the four companies funding most of the world's AI buildout - Microsoft, Alphabet, Meta, and Amazon - all reported earnings. Four CEOs, four CFOs, four hour-long calls.

For three years, the argument about AI has been the same argument: bubble or not. Nobody can settle it, because it's a question about the future. But capital expenditure guidance - the money a company commits to spending on data centers, servers, and chips - is not a debate. It's a number, stated on a recorded call, that analysts will hold against you next quarter.

So we used Barebone to pull all four transcripts and read every word, the way we do every earnings day. Two things fell out.

First, the bill: the four companies now guide to roughly $700 billion of capital spending for 2026 - raised, again, mid-year.

Second, and more interesting: all four management teams spent their calls complaining about the same component. Microsoft put a dollar figure on it. $25 billion. That's the thread this article pulls.

The $700 Billion Bill

Here's what each company told investors on April 29:

Company	2026 capex guidance	What changed	The detail that matters
Amazon	≈$200B	Pointed to roughly $200B for the year	Q1 cash capex alone was $43.2B
Microsoft	≈$190B	~$25B of it is higher component pricing	June quarter guided above $40B
Alphabet	$180 - 190B	Raised from $175 - 185B	2027 expected to "significantly increase"
Meta	$125 - 145B	Raised from $115 - 135B	Raise attributed to memory prices and capacity

Sum the midpoints and you get about $710 billion - for four companies, in one year.

The trajectory is the story. In February, after the previous round of earnings, analysts tallied the five largest US cloud and AI infrastructure providers at $660 - 690 billion of committed 2026 capex - nearly double 2025 levels. This week, the four biggest alone guided past the top of that range. Alphabet, which spent $52.5 billion on capex in all of 2024, is now guiding to three and a half times that for 2026 - and its CFO told investors to expect 2027 to "significantly increase" from there.

This is the part of the AI debate that isn't a debate. Whether the spending should happen is an opinion. That it is happening is now guidance, on the record, four times over. The money will get spent. The only open question is who collects it.

The $25 Billion Line Item

Microsoft CFO Amy Hood gave the most precise answer of the week. Microsoft expects about $190 billion of capital spending for 2026 - and roughly $25 billion of that, she said, comes from higher component pricing.

Read that again. About 13% of Microsoft's AI bill isn't more data centers or more compute. It's the same hardware at higher prices. The component doing the inflating? Memory.

Meta said the same thing in almost the same words. CFO Susan Li attributed the capex raise to surging memory prices and the need to lock in data center capacity for future years; Mark Zuckerberg pointed to component costs, "particularly memory pricing." And Amazon's Andy Jassy was the bluntest of the three:

"There's just not enough capacity for the amount of demand."

Jassy told investors the cost of components - particularly memory - has "skyrocketed." The trade press backs him up: memory and storage prices have in some cases more than tripled since last autumn, and TrendForce projected first-quarter contract prices - the bulk prices hyperscalers negotiate, as opposed to the spot market - rising 55 - 60% quarter-over-quarter for conventional DRAM and 33 - 38% for NAND flash.

Three CEOs, one earnings day, one bottleneck. When the largest buyers of computing hardware on earth all volunteer the same complaint within hours of each other, that's not color commentary. That's a demand signal with a supply problem attached.

Three Companies Control the Chokepoint

The specific memory that AI runs on is HBM - high-bandwidth memory - stacks of DRAM chips bonded directly next to the AI processor so data can move fast enough to keep the chip fed. Without it, a $40,000 GPU idles, waiting on data. Every serious AI accelerator, Nvidia's included, is built around it.

Exactly three companies in the world make it at scale: SK Hynix, Samsung, and Micron. All three have said their HBM output is effectively spoken for - SK Hynix and Micron told investors months ago that 2026 supply is sold out, and both Samsung and SK Hynix have warned the shortage could persist into 2027.

The oligopoly is stable; the pecking order isn't. SK Hynix's share of HBM supply is projected to fall from roughly 59% to ~50% this year as Samsung - the 2025 laggard, which only recently cleared qualification for Nvidia's next-generation HBM4 - claws back from 20% to ~28%. Micron holds roughly a fifth of the market and is the only one of the three you can buy on a US exchange without ADR gymnastics; SK Hynix and Samsung trade in Seoul.

The market has noticed. When we covered the memory complex in late March, Korean coverage put trailing-year gains at roughly +200% for Samsung and +300% for SK Hynix and Micron. The April 29 calls poured fuel on it: Micron rose roughly 40% in April alone, closing the month at $517.16. These are not undiscovered stocks. What the hyperscaler calls changed is the duration of the story - Hood saying Microsoft expects to "remain constrained at least through 2026" is a CFO telling the memory industry its pricing power has at least that long to run.

The Quiet Winner: Broadcom

The second bottleneck in the transcripts is less obvious, because it hides inside a different story: every hyperscaler is designing its own AI chip to reduce its dependence on Nvidia.

Hyperscaler	Custom chip	Who designs it
Google	TPU	Co-designed with Broadcom (a decade-long partnership)
Meta	MTIA	Co-developed with Broadcom
Microsoft	Maia	In-house
Amazon	Trainium	In-house (Annapurna Labs)

A custom AI chip - an ASIC, silicon designed for one owner's specific workload rather than sold generally the way Nvidia's GPUs are - still has to be designed, networked, and shepherded through manufacturing. Google and Meta outsource that to the same company. So does OpenAI, which signed a deal with Broadcom last October to co-develop and deploy 10 gigawatts of custom accelerators.

Zuckerberg made the Meta side explicit on Wednesday's call: Meta is rolling out "more than a gigawatt" of its own custom silicon, developed with Broadcom. For scale, a gigawatt is the output of a large nuclear reactor, all of it feeding one company's homegrown chips.

This is why analyst estimates put Broadcom's share of the custom AI accelerator market around 70%. The numbers are already visible: Broadcom's AI revenue more than doubled year-over-year to $8.4 billion in its January quarter, with $10.7 billion guided for the April quarter. And the demand side keeps confirming - Jassy said Trainium2 is sold out and Trainium3 is nearly fully subscribed. Custom silicon was supposed to be the threat to the AI chip trade. It turned out to be a second lane of it, and one company collects tolls on most of that lane. Broadcom closed April at $417.43.

Is Any of It Converting?

A $700 billion bill only makes sense if the spending generates revenue. Wednesday's calls carried the bulls' best evidence yet:

Microsoft's AI business now runs at a $37 billion annualized revenue rate, up 123% year-over-year. Google Cloud grew 63% to roughly $20 billion in the quarter. AWS - the laggard of 2025 - reaccelerated to 28% growth at a $37.6 billion quarterly scale. Meta's revenue grew 33% to $56.3 billion, which is what funds its entire buildout.

But watch how the market graded the papers. Alphabet rose about 7% after hours - capex raise and all - because cloud growth made the spend look like cost of goods sold for a product people are buying. Meta fell about 6% on a smaller absolute budget, because its AI revenue is a promise rather than a line item. Microsoft, in between, closed roughly flat.

The market is no longer paying for capex announcements. It's paying for conversion. That repricing - spend alone earned applause in 2024 and 2025 - is quietly the most important shift of this earnings season.

The Bear Case Is in the Same Transcripts

Now the section the bulls should read twice, because every piece of it comes from the same week of disclosures.

Part of the "spending boom" is just inflation. Microsoft's $25 billion component-pricing line cuts both ways. If memory prices normalize - and memory has spent four decades as the most violently cyclical commodity in technology - a chunk of hyperscaler capex "growth" deflates mechanically, with no change in demand at all. Some of the $700 billion is capacity. Some of it is paying triple for the same gigabyte.

Supply is already responding. Micron raised its own capex about 25%; Samsung is reportedly planning a ~50% HBM capacity increase for 2026. "Sold out" describes today's gap between supply and demand, not a permanent property of the product. Every prior memory boom ended the same way: with the new fabs arriving just as buyers finished double-ordering.

The cash math is getting heavier. Jassy himself warned that capex ramps of this speed pressure free cash flow in the early years, before the infrastructure monetizes. Meta sold $30 billion of bonds last October - one of the largest corporate debt deals on record - to help fund a buildout it used to fund from ad cash flow. And the depreciation debate hasn't gone away: Michael Burry spent last fall publicly arguing that hyperscalers understate the true cost of this hardware by stretching server useful lives. Amazon, notably, went the other direction in early 2025 and shortened the assumed life of some of its servers. If the bears are right about useful lives, today's reported profits are borrowing from tomorrow's income statements.

Guidance is an intention, not a contract. Alphabet raised its 2025 capex guidance three times on the way up. Guides can be cut with exactly the same speed if token demand disappoints. A "digestion year" - hyperscalers pausing to absorb what they've built - wouldn't even be a scandal. It would be normal. But the supply chain that captures this spend trades as if the ratchet only turns one way, after a year in which the memory names tripled or better.

If 2027 brings new HBM supply and a capex pause in the same quarter, the chokepoint premium unwinds fast. That's the trade's real risk - not whether AI is a bubble, but whether the bottleneck stays bottlenecked.

What This Means

The AI bubble argument will not be settled this year, and you don't need it to be. The capex guides are the hard numbers in a soft debate, and as of April 29 they say: roughly $700 billion in 2026, raised mid-cycle, with 2027 guided higher still. Committed spending flows to whoever holds the chokepoints, and this week the CEOs named the chokepoints themselves - memory first, custom silicon design second.

The watch list writes itself from here. DRAM and HBM contract prices, because the moment they roll over, the $25 billion inflation line starts running in reverse. The suppliers' own capex, because that's where every memory cycle has historically died. The conversion metrics - Google Cloud growth, Microsoft's AI run-rate, AWS's reacceleration - because the market just demonstrated it will punish spend without revenue attached. And Broadcom's AI guide, which has become the cleanest quarterly read on whether the custom-silicon lane keeps widening.

Earnings calls are where CEOs, under disclosure rules and analyst cross-examination, tell you where the next year of money is going. This week, four of them pointed at the same place. The question is whether you're positioned before that capital lands - or after it's already in the price.

---

Data: Barebone | Sources: Microsoft FY2026 Q3 earnings call, Alphabet Q1 2026 earnings call, Meta Q1 2026 earnings call, Amazon Q1 2026 earnings call, Broadcom Q1 FY2026 earnings release, OpenAI - Broadcom announcement (Oct 2025), company investor relations | Data as of April 30, 2026