Newsletter · · Ashutosh Agarwal

1.4 Trillion and Counting as Google Reports Tonight - The AI Capex Tracker - Week of July 22, 2026

AI capex podcast intelligence for the week of July 22, 2026. Morgan Stanley raised its hyperscaler spending forecast to $1.2 to $1.4 trillion for 2027 to 2028 even as marquee customers like Delta's CEO and Mark Cuban questioned the returns, and the whole debate now hinges on Google reporting after the close tonight.

The AI Capex Tracker

Week of July 22, 2026: 1.4 Trillion and Counting as Google Reports Tonight

Issue: Wednesday, July 22, 2026.

TL;DR

  • Morgan Stanley just raised the number again: hyperscaler AI spending is now pegged at $1.2–1.4 trillion for 2027–2028, up from "a little over a trillion." But here's the twist the bulls gloss over: about 60% of that money leaves the country, buying chips and gear from Asia, so it adds only ~40 basis points (0.4%) to US growth, the boom is real, it's just partly someone else's boom. (Thoughts on the Market, Jul 21)

  • The people who are supposed to buy all this AI are starting to ask, out loud, who's going to pay for it. Delta CEO Ed Bastian (a huge corporate customer, not a market pundit) said flatly, "the hype is far greater than the actual value," and that if you strip AI out of the economy, "GDP is pretty flat." Mark Cuban warned that borrowing hundreds of billions on 50-year bonds is "planning for perfection," and that some data centers could end up "turned into pickleball courts." (Masters of Scale, Jul 21; All-In, Jul 20)

  • This is the day the argument gets its first real answer: Google reports after the close tonight, the first of the giants, with Microsoft, Amazon, Meta and Apple next week and short bets against US stocks at a 15-year high. If Google reaffirms the spend, trapped short-sellers could be forced to buy back in; if it blinks on capex, the bears finally get their proof. (Bloomberg Surveillance, Jul 20)

What's new

A quick note on the window: this is a Wednesday recap, sweeping podcasts from Tuesday, July 21 (plus a few Monday shows that hadn't been picked up yet). And it lands on the single most important morning of the cycle: Google prints tonight. So the fresh material splits neatly in two: the sell-side just moved the spending number up, and the people who actually buy AI moved their skepticism up to match. Ranked by where the real money sits:

1. Morgan Stanley raised the spending forecast, again, and quietly flagged where the money actually goes. Thoughts on the Market, Jul 21. Michael Gapen (Morgan Stanley Chief US Economist) and Chetan Ahya (Chief Asia Economist), on the bank's quarterly economics roundtable.

The headline: Morgan Stanley keeps having to revise its estimate for US hyperscaler and AI-related capital spending higher. Gapen: "We were thinking a little over a trillion for 2027. Now we're more like 1.2, 1.3 trillion, maybe as high as 1.4 trillion in 2028." That's roughly 3.5% of the entire US economy going into one category of corporate investment.

But Gapen's second point is the one a PM should tape to the monitor. You'd look at 3.5% of GDP and assume it's a massive engine for US growth. It isn't, not directly, because "roughly about 60% of that hyperscaler CapEx spending goes to items like computers and peripherals... that have a very, very high import content." Net of imports, AI capex is adding only about "40 basis points to growth this year" (that's 0.4%, meaningful for an economy growing ~2–2.5%, but not the whole story). In plain terms: America writes the checks, and more than half the money flows overseas to the companies that make the chips and machines.

That's why Ahya's Asia numbers matter so much. Asian semiconductor exports to the US are "growing by 90%," and he names the three big winners in order: "Korea, number one, Taiwan... and also Japan." He also widened the lens in a way worth internalizing: AI is not even the biggest capex story in Asia. For 2026, "AI and semiconductor companies CapEx is about $380 billion," but energy capex is "going to be $900 billion." There's a whole industrial super-cycle (AI, energy, defense, supply-chain reshoring) underneath the AI headline. Why it moves numbers: this both re-anchors the whole thesis higher (the 2027–28 build is bigger than most models assume) and tells you the cleanest way to play the American spend may be the Asian and energy suppliers who capture 60% of it.

2. The customers are getting loud, and Delta's CEO asked the question the whole trade rests on. Masters of Scale, Jul 21. Ed Bastian, CEO of Delta Air Lines.

This is a fresh and valuable voice because Bastian isn't a strategist talking his book, he's a giant enterprise buyer of technology, exactly the kind of company whose spending is supposed to justify all this. And he's a skeptic. His blunt version: "I understand the hype is far greater than the actual value that's ultimately gonna be derived because these technology companies are not about people... they care about winning their race."

Two lines from him are the ones that move the thesis:

"Everywhere I go, I ask the question, who's gonna pay for the trillions and trillions and trillions of dollars of investment that are being committed in CapEx... to this race." (Ed Bastian, Masters of Scale, Jul 21)

And the reality check on the economy itself: "If you strip out AI's contribution to GDP, you realize GDP is pretty flat." His framework for whether Delta actually buys is refreshingly simple: "there's two forms to a return. There's the value we get from it and how much did we spend to get it. And if we're going to spend more than we're going to get from it, we're not going to use it." Crucially, the opportunities he sees so far "are largely about efficiency. They're not necessarily about how do you grow your top line... if it's all about efficiency, well, I'm not sure I have that level of investment appetite just to be a little faster. If you show me how I'm going to triple my business, I'm more interested." Why it matters: the bull case needs enterprises to convert AI capacity into real spending. Here's a marquee customer saying the value is mostly cost-cutting, not growth, and cost-cutting doesn't command unlimited budgets. He also flagged the buck-passing on power directly: the utilities "say, well, the technology guys are going to pay for it. Well, the technology guys aren't the ones selling to customers... we are."

3. Mark Cuban: the bubble won't wipe out Main Street, it'll wipe out the funds, and maybe strand the data centers. All-In, Jul 20. Mark Cuban, investor and entrepreneur.

Cuban's take is more nuanced than a simple "it's a bubble." He argues this is not a dot-com repeat, because in 2000 "there were s*** companies going public, getting crazy valuations... with no revenue, no traffic, no nothing," and "you don't see that at all today." So it won't hurt most people, but "it could just destroy a lot of VCs and a lot of funds and a lot of PE, right? Because they're going all in." The pressure he describes is the one that makes bubbles dangerous: fund managers now "got to outperform the fund next door. And they're all in Anthropic and getting their outcomes at SpaceX and celebrating. But if s*** hits the fan?"

His two hardest points for the buildout:

  • Financing risk. The market leaders ("Google, et cetera, Meta") are "borrowing hundreds of billions of dollars" while "spending all their cash flow on CapEx. And then they're borrowing on top of that. 50-year bonds... That's planning for perfection." (He also nods to "a private credit problem right now" layered underneath.)
  • The stranded-asset risk. "If there's a price performance curve on AI that minimizes the power requirements, there's going to be a lot of data centers that are going to be turned into pickleball courts." His analogy is the fiber glut: everyone laid fiber, then the technology got so much better that "there wasn't a bandwidth problem anymore," and the excess sat as "dark fiber that people bought for pennies on the dollar."

He's blunt on ROI, too: putting "$100 billion to work... has to come back, not just in revenue, but it has to come back in earnings... And you just don't know." Why it matters: this is the clearest articulation yet of the specific mechanism, debt-funded, priced-to-perfection capex meeting a fast-improving efficiency curve, that turns "we can't build fast enough" into "we built too much."

4. Power got two fresh, credible operator voices, and both say the same thing: it's physical, it's slow, and it's lumpy. The Real Eisman Playbook, Jul 21. Steve Eisman with a power-and-utilities equity analyst; and Inevitable, Jul 21, Jake Carter, CEO of Blue Energy.

On the Eisman show, the analyst laid out the grid math cleanly. The US needs somewhere between "100 gigawatts of new capacity... by 2035" and as much as 350 GW, and "at 350, you're more than doubling all of the electricity generation in the United States." Realistically the grid is set to grow "at like a 30-gigawatt-type clip" per year over the next five years, and this is baseload (24/7 firm) power, the hardest kind to add. A single megawatt powers "roughly 1,000 homes," so 30 GW a year is an enormous lift.

The equity read-through is GE Vernova (GEV), which Eisman owns and which has gone from "negative value" inside old GE to a ~$150 billion company. Its gas-turbine business is "booked out until 2030," even "2031," "with price," meaning orders placed today don't even turn on until the end of the decade. And it's cross-selling the grid gear (transformers, high-voltage equipment) alongside the turbines, and the analyst cited a Chevron deal "last week" as the template. One genuinely new wrinkle for the chip names: NVIDIA's move to an 800-volt power architecture with its new chips "basically requires more power" and "more of a grid equipment, like from a GEV," and it introduces a pause risk, because a developer planning a data center now has to decide "do I rush ahead with the old NVIDIA chips or do I wait till the new ones come on?" His conclusion: "it's gonna be very lumpy... under the best of circumstances." He also poured cold water on the "everyone's building gigawatt data centers" story: "the only real gigawatt-scale data center that's been built is by SpaceX," and there's "double counting" because a developer chasing 10 sites "in the end... might be only 2."

On Inevitable, Blue Energy's Jake Carter added the nuclear angle from the operator's chair. His pitch for why hyperscalers want nuclear is not sentiment, it's risk management: "The reason they need nuclear power isn't because it's clean... it's because it's firm capacity that hedges their gas risk. They're building a ton of gas turbines right now to feed the AI growth" and are quietly "taking on a lot of gas commodity risk, a lot of gas supply risk" (imagine a polar vortex knocking out gas supply and "gigawatts of compute" with it), plus emissions-rule risk. Blue Energy's model is a "gas-to-nuclear conversion" at the Port of Victoria, Texas (build half a nuclear plant, splice in gas turbines to start producing power early, then switch the steam over to nuclear), sited next to a Crusoe data-center project of "1.5 gigawatts of load." His sober note on the nuclear-restart trade everyone loves: the easy restarts are basically done. After Palisades, Three Mile Island (now the Crane Clean Energy Center) and NextEra's Duane Arnold, "you probably won't see that many more." His own new-build target: a commercial small modular reactor "as early as 2032." Why it matters: both voices confirm the demand-for-power-gear thesis while puncturing the idea that supply arrives quickly: the bottleneck is real, it favors the equipment makers, and it makes the buildout jerky rather than smooth.

5. From Asia: the memory supercycle is real, but the customers hold the whip. Merryn Talks Money, Jul 20. A bottom-up Asia fund manager, with host Merryn Somerset Webb.

This closed the memory/optics coverage gap from yesterday. The fund manager's core worry is ROI plus substitution: "the ROI of this hyperscaler investment is still questionable," and Chinese models "priced at a fraction of the cost" (he cited a software-company friend who uses "the 90% cheaper Chinese model" for basic tasks and the Western model only for the hardest ones). His line on the demand math: "a lot of the demand assumptions assume that everyone will always want to use the best model, but that's not really the case." On memory specifically, he explained the supercycle's origin: years of underinvestment (NAND in "deep loss-making," SK Hynix so leveraged it "had to go to the market to raise convertible bonds") colliding with a sudden agentic-AI demand surge that "just need a lot of memory." It can last "a couple of years" on the supply side, but the stock prices may not, because "stock market is about expectation versus reality." His warning: with "a handful of hyperscalers paying for all these products," if demand softens even slightly while new supply floods in (Korean and Chinese memory players are raising money, "the Chinese competitor is doing an IPO in the coming days"), "the bargaining chip [goes] back from the memory producers to the payers, and we could see prices start to fall." The tell he flagged: Samsung's Q2 preliminary profit was up nearly "19 times" year-on-year, "but share price is actually down post that result," the same "records everywhere, stocks nowhere" pattern that defined last week.

The debate

Steel-manning both sides of the roughly $1.2–1.4-trillion hyperscaler spending thesis. Last cycle the market's own behavior (selling record supplier results) flipped the burden of proof onto the buyers. This cycle the buyers spoke, and they're not fully convinced. It all funnels into tonight.

Bull: the number keeps going up, the technology is real, and this is a dip inside a decade-long build. Morgan Stanley isn't cutting its forecast; it's raising it, to $1.2–1.4 trillion for 2027–28, with Asian semiconductor exports up 90% (Thoughts on the Market, Jul 21). The cleanest bull framing came from futurist investor Armando Pantoja: "The bubble is in valuations. It's not in the technology," and the dot-com crash "destroyed trillions, but it did not destroy the internet." He dismisses the "seven companies passing money to each other" argument (money circulating in a small group doesn't make an economy fake), notes that unlike 2000 there's no wave of junk companies IPO-ing on hype, and insists the right mental model is physical: AI "is an energy and infrastructure company that looks like a software company to people who don't understand it... intelligence cannot scale without power." His verdict on last week's chip drop: "a valuation correction on compute. It's not a demand collapse... This is a dip. This is an opportunity" (Strategy Sunday, Jul 20). And Citi still expects the giants to guide 2027 capex "over $800 billion" with "no real pushback yet" (Bloomberg Surveillance, Jul 20).

Bear: the buyers are balking, the returns are unproven, and it's debt-funded and priced to perfection. The most damaging new evidence isn't from a short-seller, it's from a customer. Delta's Ed Bastian says the value is "largely about efficiency," not growth, and openly asks "who's gonna pay for the trillions" (Masters of Scale, Jul 21). Mark Cuban adds the financing time-bomb: hundreds of billions borrowed, "50-year bonds," "planning for perfection," and a fiber-style glut that could leave stranded data centers "turned into pickleball courts" (All-In, Jul 20). From Asia, the fund manager warns pricing power can snap back to the hyperscalers the moment demand wobbles, and that cheaper Chinese models let customers dodge the premium tier for most tasks (Merryn Talks Money, Jul 20). And Morgan Stanley's own math is a subtle bear point: 60% of the spend leaves the US, so the domestic payoff is thinner than the headline implies (Thoughts on the Market, Jul 21).

The synthesis. Nobody disputes the spending is enormous and the demand is real right now. The two open questions are unchanged: do the returns justify the next dollar, and is all the good news already in the price? What's new this cycle is who's voicing the doubt, not just professional bears but the enterprise customers and the billionaire operators who lived through the last two bubbles. That raises the stakes for tonight: Google's guide is the first data point that can actually answer "are the buyers still buying?"

Sell signals to watch:

  • Any hyperscaler guiding capex flat or down. Still the master trigger, and Google tonight is the very first test.
  • "Efficiency, not growth" language from big customers. Bastian just said it; if that becomes the enterprise consensus, the demand curve the bulls model gets a lot flatter.
  • Debt and 50-year-bond issuance stepping up. Cuban's "planning for perfection" risk; watch how much of the build is now borrowed rather than funded from cash flow.
  • Memory prices rolling over. The moment new Korean/Chinese supply meets any demand wobble, pricing power flips back to the hyperscalers (Merryn's fund manager).
  • A cheaper-chip/efficiency breakthrough. The "dark fiber" scenario that strands capacity; the 800-volt chip transition already threatens a "wait-and-see" pause in orders.

Stocks in play

NVDA. Bull: the forecast that funds it just went up, not down: Morgan Stanley now sees $1.2–1.4 trillion of AI capex in 2027–28, and the new 800-volt chip generation actually deepens Nvidia's pull on the whole power-and-grid supply chain (Thoughts on the Market, Jul 21; The Real Eisman Playbook, Jul 21). Bear: that same 800-volt transition is a double-edged sword: it can freeze orders as buyers decide whether to "rush ahead with the old NVIDIA chips or... wait till the new ones," making the buildout "very lumpy," and Nvidia remains the crowded global proxy for a trade the customers are now openly questioning. Next: tonight's Google guide (read-through to GPU orders), then the rest of the hyperscaler prints next week, then Nvidia's own August quarter.

AVGO. Bull: the custom-chip (ASIC) thesis is unchanged: the merchant-silicon story that turns hyperscaler capex into Broadcom revenue is squarely in the path of the raised Morgan Stanley number. Bear: no fresh operator color this cycle, and the interconnect/custom-silicon names sit right in the "cheaper Chinese competition" and "returns unproven" crosshairs. Next: custom-ASIC ramp confirmation and the hyperscaler capex guides that fund it, starting tonight.

AMD. Bull: the buyers keep funding Nvidia alternatives, and any credible second source benefits as the overall pie grows toward $1.2–1.4 trillion. Bear: quiet on the podcasts this cycle, no fresh MI-series datapoint to reset the story, so it trades on the group's beta rather than its own catalyst. Next: the next MI-series design win or AI event; read-through from the hyperscaler guides.

MSFT. Bull: a core anchor of the ~$1.2–1.4-trillion build; the "guide higher" camp expects it to reaffirm and extend spending when it reports next week, the outcome that would floor the whole complex. Bear: it sits inside the financing-and-returns squeeze the bears describe, and the "efficiency, not growth" worry (Bastian) lands directly on the Copilot monetization question. Next: earnings next week, the capex guide is the number that matters.

GOOGL. Bull: it reports after the close tonight and is the first real read on the entire thesis; the market wants to see cloud revenue growth, the huge guaranteed-cloud backlog, and core Search holding up while the capex flows in (Bloomberg Surveillance, Jul 20). Bear: consensus already sees gross margin compressing on the capex load, and with short interest at a 15-year high this is a wide, two-sided event: a soft capex line or a bigger-than-feared margin hit confirms the "spend crushes profits" fear. Next: tonight, after the close, the single most important event of the cycle.

AMZN. Bull: AWS carries an enormous cloud backlog and its own Trainium chips are a genuine second path to Nvidia, both leveraged to the raised spending forecast. Bear: same funding squeeze and same unanswered returns question as the group; a giant enterprise customer (Bastian) just said the payoff is mostly efficiency. Next: earnings next week and the AWS capex guide.

META. Bull: it has real optionality to monetize spare capacity (the compute-rental idea from prior cycles) and remains a top-tier spender riding the upgraded forecast. Bear: Cuban named it directly as one of the market leaders "borrowing hundreds of billions" and issuing "50-year bonds," the poster child for debt-funded, priced-to-perfection capex (All-In, Jul 20). Next: earnings next week, watch the capex guide and any update on renting out compute.

Read-throughs

  • Power / thermal: still the strongest, most physical signal, and now with two fresh operator voices. GE Vernova (GEV) is the cleanest expression: gas-turbine orders "booked out until 2030"–"2031" with price, cross-selling grid gear (transformers) into data centers (the Chevron deal), and a direct tailwind from Nvidia's 800-volt shift, which "requires more of a grid equipment, like from a GEV" (The Real Eisman Playbook, Jul 21). The grid needs ~30 GW of new firm capacity a year for five years, and the supply arrives slowly, bullish for the equipment makers, a brake on the buildout. Vertiv (VRT) and Eaton (ETN) were name-quiet again, but the demand-for-gear read only strengthens. Most actionable trade: stay long the power/grid/cooling supply chain; the physical bottleneck is worsening, not easing.

  • Utilities / nuclear: demand intact; nuclear is a hedge, not a quick fix, and the easy restarts are used up. Blue Energy's Jake Carter reframed hyperscaler nuclear demand as risk management, "firm capacity that hedges their gas risk" against a gas-supply shock that could take out "gigawatts of compute," with a gas-to-nuclear conversion in Texas (Port of Victoria) next to a 1.5 GW Crusoe data center, and a first commercial SMR targeted "as early as 2032" (Inevitable, Jul 21). The sober note for the nuclear-restart trade (Vistra, Constellation, Talen): after Palisades, Three Mile Island/Crane and Duane Arnold, "you probably won't see that many more" easy restarts, the next leg is slow new-build. Carry the power positions; the demand is intact, but nuclear supply is a 2030s story.

  • Memory: respect the tightness, but the customer holds the whip. The supercycle is real (years of underinvestment plus an agentic-AI demand surge), and it can run "a couple of years," but the fattest margins are the most cyclical thing you own: with a handful of hyperscalers as the only buyers and both Korean and Chinese suppliers raising capital, "the bargaining chip [goes] back... to the payers" the instant demand softens (Merryn Talks Money, Jul 20). Samsung's Q2 profit up ~19x with the stock down is the warning label. Watch DRAM contract pricing and the imminent Chinese memory IPO.

  • The Asia angle: the American spend is largely an Asian revenue stream. Morgan Stanley's 60%-goes-to-imports point plus 90% semiconductor-export growth means Korea, Taiwan and Japan capture the majority of US hyperscaler capex (Thoughts on the Market, Jul 21). And the second-order Korean beneficiaries the fund manager flags (banks, construction, even luxury retail, funded by outsized memory-company bonuses) are a way to play the theme one step removed from the volatile chip pure-plays (Merryn Talks Money, Jul 20).

  • Optics / networking: still a coverage gap. No podcast in this window carried substantive color on the optical (Coherent, Lumentum, Fabrinet) or interconnect-silicon (Marvell, Astera Labs, Credo) names. Two cycles running now, a gap to close on the next sweep, not evidence the pressure has lifted.

What changed vs last issue

Last issue (Jul 21, "Records everywhere, stocks nowhere. Google reports Wednesday.") caught the supplier blowouts that fell anyway and the deflating China "Kimi" scare, and set up the earnings gauntlet. This cycle two things moved:

  • The forecast went UP. Yesterday the working numbers were ~$725–800 billion for the industry and $920 billion as a high-end 2026 estimate. This cycle Morgan Stanley put a sharper, higher stake in the ground for the out-years: $1.2–1.3 trillion in 2027, up to $1.4 trillion in 2028, explicitly "revised higher." The bull's core fact (the number keeps climbing) got fresh, top-tier sell-side confirmation.

  • The skepticism moved from strategists to buyers. Yesterday the doubt came from macro voices (Boyle, the J.P. Morgan depreciation warning) and market technicians. This cycle it came from the demand side of the economy: Delta's CEO ("who's gonna pay?", "efficiency, not growth") and Mark Cuban ("50-year bonds... planning for perfection," "pickleball courts"). When the customers and the operators start asking the returns question out loud, the burden of proof on tonight's print gets heavier.

  • Power picked up two named operator voices. Yesterday's power story was Lazard's Bilicic (utilities as a 7–9% pick-and-shovel play). This cycle added the equity-analyst detail on GE Vernova (booked to 2031, 800-volt tailwind, "very lumpy" buildout, "only real gigawatt-scale data center... is by SpaceX," real double-counting in the project pipeline) and the nuclear operator's view that nuclear is a gas-risk hedge with a 2032 new-build timeline and no easy restarts left.

  • The memory gap got filled, with a caution. Yesterday optics/memory drew almost no coverage. This cycle an Asia fund manager gave the two-sided read: the supercycle is real and could run a couple more years, but it's a hyperscaler-controlled commodity and cheaper Chinese models threaten the demand assumptions. Samsung up ~19x on profit and down on the stock is the same tell as last week's TSMC/ASML.

  • Deltas on the big numbers: 2027–28 US AI capex framed at $1.2–1.4 trillion (Morgan Stanley, up from "a little over a trillion"); ~40 bps contribution to US GDP with ~60% of spend imported; Asia semiconductor exports +90%; Asia AI/semi capex ~$380 billion vs energy capex ~$900 billion for 2026; US grid needs ~100–350 GW of new capacity by 2035, adding ~30 GW/yr; GE Vernova gas-turbine backlog booked to ~2031; Samsung Q2 profit ~+19x YoY yet the stock fell; short interest against US stocks still at a 15-year high.

  • Still open: no hyperscaler has yet guided capex flat, Google tonight is the first test; the optics/networking names (Marvell, Astera, Credo, Coherent, Lumentum) drew no coverage for a second cycle; and the returns question (do the buyers other than OpenAI/Anthropic show the revenue) is exactly what the next ten days of prints must answer.

Next prints to watch: Google, tonight after the close (the whole cycle's swing vote), then Microsoft, Amazon, Meta and Apple next week plus a Fed meeting, with the one question that decides everything being whether any of them dares to guide capex flat.