Newsletter · · Ashutosh Agarwal

Records Everywhere, Stocks Nowhere as Google Reports Wednesday - The AI Capex Tracker - Week of July 21, 2026

AI capex podcast intelligence for the week of July 21, 2026. Taiwan Semiconductor and ASML posted records and their stocks fell anyway, the China Kimi scare deflated as a memory-hungry model, and the whole thesis narrowed to the earnings calendar with Google reporting Wednesday and short interest at a 15-year high.

The AI Capex Tracker

Week of July 21, 2026: Records Everywhere, Stocks Nowhere as Google Reports Wednesday

Issue: Tuesday, July 21, 2026.

TL;DR

  • Every company that actually makes this boom possible just reported blowout numbers, and their stocks fell anyway. Taiwan Semiconductor (the factory behind Nvidia's chips) posted record profit up about 77%, beat, and raised its 2026 build budget to $62 billion; ASML (which makes the machines that make the chips) said it's adding 30% more capacity in both 2027 and 2028. Both stocks dropped after the news. As Deutsche Bank's Ozan Tarman put it, "SK Hynix delivered. Price action didn't. TSMC delivered. Price action didn't." (The Morning Filter, Jul 20; Bloomberg Surveillance, Jul 20)
  • The China "Kimi" panic that knocked chips into a bear market on Friday is already being walked back, and may even be good news for memory. The new Chinese model, Kimi K3, is genuinely strong at coding but still trails the best US models overall, and, unlike last year's DeepSeek, it's a huge model that needs roughly 1.4 terabytes of memory just to run, which is bullish, not bearish, for Micron and SK Hynix. Macro strategist Andreas Steno's verdict: "I basically think the market got this wrong." (Real Vision, Jul 20; The Rundown, Jul 20)
  • This is now a binary set up on the calendar: Google reports Wednesday, then Microsoft, Meta, Amazon and Apple next week, and short bets against US stocks just hit their highest level since 2010. If the giants reaffirm their roughly $800 billion-plus spending plans, trapped short-sellers could be forced to buy back in and spark a squeeze; if even one guides spending flat, the bears win the argument. (The Rundown, Jul 20; The Circuit, Jul 20)

What's new

A quick note on the window: this is a Tuesday recap, sweeping podcasts from Sunday and Monday (roughly July 19-20). Monday's issue caught the weekend crash: the chip index into a bear market, Chanos's bear case, the funding scare. Over the next 48 hours the story didn't get more bearish; it got more interrogated. The marquee bears made their case last week. This week the specialists and the operators pushed back, the China scare started to deflate, and the whole argument narrowed to one thing: the earnings that land over the next ten days. Ranked by where the real dollars sit:

1. The picks-and-shovels crowd delivered records, and the stocks fell anyway. That's the tell. The Morning Filter, Jul 20. Dave Sekera, Chief US Market Strategist at Morningstar.

Two of the most important companies in the entire AI supply chain reported last week, and both blew the doors off. Taiwan Semiconductor, the contract factory that physically builds Nvidia's chips, raised its 2026 revenue guidance to over 40% growth and lifted its capital-spending budget by $8 billion to $62 billion, a 15% jump. Morningstar raised its fair value on the stock by 27% (to $534 per US-listed share) on the news. ASML, the Dutch company whose machines are the only tools on earth that can print the most advanced chips, said it's increasing capacity by 30% in 2027 and likely another 30% in 2028, with 2027 essentially sold out already. Morningstar took its ASML fair value up to $2,050 a share and lifted its 2030 revenue forecast from $60 billion to $70 billion.

And both stocks fell after reporting. Sekera's explanation is the single most important idea in this issue: the market has already priced the strong growth for this year, and even for 2027. "At this point, I think it's now starting to becoming about trying to forecast 2028." When companies this good get sold on record results, it means investors have stopped paying for the present and started worrying about the price of staying ahead. Economist and podcaster Patrick Boyle framed the same TSMC reaction bluntly: the market "has started asking a question it spent 3 years refusing to ask: What does all of this actually earn?" (Patrick Boyle On Finance, Jul 19)

2. The China "Kimi" scare that triggered the crash is deflating, and it's the wrong stock to short over it. Real Vision, Jul 20. Andreas Steno (Steno Research), on Macro Mondays.

Friday's sell-off was blamed on Kimi K3, a new model from the Chinese lab Moonshot. A weekend of analysis has cooled the panic. Steno's read: yes, Kimi K3 took the lead in one narrow category, front-end coding, but on the broad measures it "is also behind both Claude Fable and ChatGPT 5.6." His chart of the best Chinese model versus the best US model shows the same pattern that's held for over a year: China closes to within about 5-7 percentage points of the frontier, the US ships a new flagship, and the gap reopens. "I basically think the market got this wrong." His counter-intuitive conclusion for a book: "I think we need to price hardware higher as a consequence of this", because if China really can serve intelligence more cheaply, cheaper intelligence gets used more, which means more chips, not fewer.

The most useful nuance came from The Rundown, Jul 20, host Zaid Admani: Kimi K3 is different from last year's DeepSeek shock. DeepSeek scared everyone because it needed less computing power. K3 is the opposite: it's the biggest open-source model ever at 2.8 trillion parameters, and it needs "about 1.4 terabytes just to run the model locally." In plain terms, running it takes a mountain of memory chips. "So the memory makers like SK Hynix and Samsung and Micron might actually be fine here." The real worry Admani flags isn't compute demand, it's that AI intelligence itself "might be getting commoditized way faster than investors expected," which threatens the profits of the model-makers (OpenAI, Anthropic), not the chip-makers.

3. Meta started renting out spare computing power, a small crack in the "everything is scarce" story. RiskReversal Pod, Jul 20. Guy Adami and Dan Nathan.

The whole bull case for chips rests on scarcity, there's never enough computing power. So it matters that, per a New York Times report over the weekend, Meta is looking to lease out computing capacity from its data centers to outside customers like Anthropic. As Dan Nathan put it, "you have these two huge companies [SpaceX and Meta] that are spending shit tons of money and they have excess compute. Isn't the whole story in this space about scarcity?" Adami's answer: "what you just described blows a hole in the whole bull thesis of scarcity." The caveat that keeps recurring: these rental deals are short-term: Elon Musk himself flagged that the SpaceX/xAI compute leases are "very short-term contracts," which undercuts the "durable long-term demand" narrative the whole build rests on.

Citi's Ron Josey, interviewed on Bloomberg Surveillance, Jul 20, was more measured (he doesn't like the phrase "excess capacity" because "everyone we talk to talk about the term insatiable demand"), but he confirmed the mechanics: Meta has roughly 8 gigawatts of power today, "maybe going to 14 over the next year," and while it waits for its next big AI products to launch, "maybe we can actually get some dollars back or subsidize some of the costs" by renting the gap. Watch whether "rent the spare compute" spreads: if it does, scarcity pricing weakens.

4. The power reality check: utilities are the pick-and-shovel play, and the cheap-turbine land-grab is a mirage. Tech Disruptors, Jul 20. George Bilicic, Vice Chairman and Global Head of Power, Energy & Infrastructure at Lazard.

This is the freshest operator-grade voice of the sweep, the banker who structures these deals, not a pundit. His headline: the boring utility has become a growth stock. A large group of them is now "growing at 7% to 9% a year, if not more," with a few regional names growing 10-15%, and it's "absolutely a real change," not a passing fad. His advice to anyone tracking AI: treat power and utility companies as "a pick and shovel play... thinking about the gold rush reference is a good strategy", but be selective, "some companies are a good play on this and some are not."

Three things a PM should write down from Bilicic:

  • Geography is destiny. The fastest places to get electricity are Texas (ERCOT) and the Southeast, where there's slack in the system. The Northeast and Pacific Northwest are "very, very difficult": you can't build new generation there.
  • The cheap-turbine boom is temporary. Developers are pouring concrete slabs in Texas and bolting on small gas turbines ("basically big jet engines... super high heat rate, super expensive... a lot of emissions") purely for speed. These "will just be displaced when the larger equipment's more available." And crucially: "A peaking gas plant is much more expensive than a renewables asset partnered with batteries. I don't think the world appreciates that."
  • The bottleneck is who can build. Only a handful of firms can build large gas plants ("Bechtel, Fluor, Kiewit, Quanta"), so "they have the ability to garner premium pricing." Lazard's just-published cost data shows the price of a combined-cycle gas plant rising sharply. "It's a great time to be in the construction business." He also flagged that transmission lines are now "impossible to build in this country right now", the real physical brake on the whole theme.

5. The whole story now rests on four American buyers, and one country is the leveraged bet on them. Patrick Boyle On Finance, Jul 19. Patrick Boyle, former hedge-fund manager, citing economist Brad Setser.

Boyle's framing is the cleanest way to understand why a South Korean stock drop is now a US market event. Setser calls it "DRAM dollars": like the petrodollar recycling of the 1970s, Samsung and SK Hynix sell memory chips for dollars and park those dollars offshore instead of bringing them home. The punchline: "stripped of the national flag, the entire South Korean stock market has become a leveraged bet on the server budgets of about 4 American corporations", Amazon, Google, Meta and Microsoft, who between them have laid out around $376 billion in near-term capital spending, with the wider industry closer to $725 billion.

"What if AI turns out to behave less like high-margin software and more like electricity, a commodity where the models are largely interchangeable and nobody has much pricing power?" (Patrick Boyle, Patrick Boyle On Finance, Jul 19)

His warning mechanism is worth internalizing: if the hyperscalers ever suspect the returns aren't there, "they won't put out a press release, they will simply reduce their spending... push the next data center back a quarter or two." Memory is "much closer to an industrial commodity" than a subscription business: prices "don't drift down gently... they tend to fall off a cliff." And Korea's market is "two chip stocks wide" and strapped to leveraged retail funds that sell mechanically into any fall. Roughly 1.2 million Korean retail investors already got margin calls in the recent drop.

The debate

Steel-manning both sides of the roughly $800 billion-a-year hyperscaler spending thesis. Last cycle the question shifted from "is demand real?" to "do the returns work?" This cycle it narrowed again, to a single, testable question: when the giants report over the next ten days, do they keep spending, and can they show the money coming back?

Bull: the numbers are real, the scare is overdone, and the earnings will reaffirm the build. The suppliers just proved demand: TSMC record profit and a raised budget, ASML sold out through 2027, SK Hynix listing a $26.5 billion stake into hot demand (RiskReversal, Jul 20). The China scare is a category error: Kimi is behind on the broad benchmarks and its memory-hungry design is a positive for the memory names, not a negative (Real Vision, Jul 20; The Rundown, Jul 20). Ben Bajarin's "gigawattonomics" math says the economics work: at $6/hour and 75% utilization a rack pays back its Nvidia hardware cost "in a little over two years," and the hyperscalers get far more than $6/hour on long-term Blackwell contracts, so "they can pay that back in less" (The Circuit, Jul 20). Citi's Ron Josey sees no real pushback and expects 2027 capex guidance "over $800 billion" (Bloomberg Surveillance, Jul 20). And with short interest at a 15-year high, a confident set of prints could squeeze the market higher (The Rundown, Jul 20).

Bear: records are already priced, the returns question is unanswered, and the money never comes home. When TSMC and ASML post blowouts and fall, the market is telling you the good news is spent and 2028 is a question mark (The Morning Filter, Jul 20). Ben Bajarin's own partner Jay Goldberg concedes the ROI problem: "we're seeing Anthropic and OpenAI... immense revenue gains, but we're not seeing that from the other five people who are spending on all this CapEx" (The Circuit, Jul 20). Depreciation is the slow-motion risk: J.P. Morgan Asset Management notes business fixed investment hit 14.1% of GDP in Q1, its biggest share in 25 years, and warns it "could of course be very negative for corporate profits if for some reason AI revenue growth were to slow even as depreciation expense rises" (Notes on the Week Ahead, Jul 20). Deutsche Bank's Ozan Tarman sees crowded positioning "a bit like the October to February periods" and warns "August may be a bit more nastier for crowded positioning than people thought" (Bloomberg Surveillance, Jul 20). And the commodity nightmare, value flowing to users, not to those who footed the bill, hangs over everything (Patrick Boyle On Finance, Jul 19).

The synthesis. Both sides now agree on the same facts, the spending is enormous, the supplier earnings are real, and the demand is there. They split on two things: whether the returns justify the next dollar, and whether records already priced in leave any upside. The market's own behavior (selling record results) says the burden of proof has flipped to the buyers. Which is why the entire debate collapses onto the earnings calendar below.

Sell signals to watch:

  • Any hyperscaler guiding capex flat or down. This is still the master trigger, and Google on Wednesday is the very first test.
  • More "excess compute for rent" headlines. Every one of these chips away at the scarcity thesis (Meta already started).
  • Google's gross margin. Consensus already sees it falling from ~71% to ~65% on the capex load (RiskReversal, Jul 20); a bigger drop confirms the "spend crushes profits" fear.
  • A real crack in memory prices. The commodity cliff Boyle describes; watch the Korean names and DRAM contract chatter.
  • Usage-based pricing killing demand. The "water meter" effect, where metered AI bills push customers onto cheaper (often Chinese, open-source) models.

Stocks in play

NVDA. Bull: the unit economics still work, Ben Bajarin's model has a rack paying back its Nvidia cost in ~2 years or less, and he argues Nvidia's token economics are "very, very good, if not better or on par with any costs you'd have for custom ASICs," with the company set to ship north of 10 gigawatts of compute (~65,000-70,000 racks) this year (The Circuit, Jul 20). Bear: it's the epicenter of the China commoditization fear, and, with poetic irony, Korean retail investors are piling into Nvidia even as their own SK Hynix/Samsung crash, making it a crowded global proxy for the whole trade (Patrick Boyle On Finance, Jul 19). Next: the hyperscaler capex guides over the next ten days (read-through to GPU orders), then Nvidia's own August print.

AVGO. Bull: the custom-chip story got a concrete number: Broadcom is set to sell "more than $20 billion of its TPUs that it designed for Google to be used by Anthropic for Claude," a marquee win for the merchant-ASIC model (The 7investing Podcast, Jul 20). Bear: the interconnect and merchant-silicon names were among the hardest hit as the chip index rolled into a bear market, and they sit squarely in the "cheaper Chinese competition" crosshairs. Next: custom-ASIC ramp confirmation and the hyperscaler capex guides that fund it.

AMD. Bull: the industry still wants a credible second source to Nvidia, and the wave of huge custom/alternative-silicon orders (OpenAI's reported ~$10 billion Cerebras deal, Amazon's ~$20 billion) shows buyers actively funding Nvidia alternatives (The 7investing Podcast, Jul 20). Bear: largely quiet on the podcasts this cycle: no fresh operator color, and no specific MI-series datapoint to reset the story. Next: the next MI-series design win or AI event; watch for read-through from the hyperscaler guides.

MSFT. Bull: a core anchor of the ~$800 billion build; the "guide higher" camp expects it to reaffirm and extend spending when it reports next week, the outcome that would floor the whole semi complex (Bloomberg Surveillance, Jul 20). Bear: it sits right in the funding-and-returns squeeze the bears describe, and any hint of slowing capex would be read as the first domino. Next: earnings next week, the capex guide is the number that matters.

GOOGL. Bull: it reports Wednesday after the close and is the first real read on the whole thesis; Citi's Ron Josey wants to see three things: Google Cloud revenue growth, the backlog of guaranteed cloud commitments (already reported around $460 billion), and core Search, which he thinks can still grow ~17% "because of these AI investments" (Bloomberg Surveillance, Jul 20). Morningstar pegs fair value at $433, a discount to where it trades into the print (The Morning Filter, Jul 20). Bear: consensus sees gross margin compressing from ~71% to ~65% on the capex load, with earnings growth decelerating to ~7% next year and a 7% implied move on the print, a wide, two-sided event (RiskReversal, Jul 20). Next: Wednesday after the close, the single most important event of the cycle.

AMZN. Bull: AWS carries a reported ~$360 billion cloud backlog (Wealthion, Jul 20), its own Trainium chips are a genuine second path to Nvidia, and it's placing its own large alternative-silicon orders (the ~$20 billion Cerebras deal for AWS) (The 7investing Podcast, Jul 20). Bear: same funding squeeze and same unanswered returns question as the group. Next: earnings next week and the AWS capex guide.

META. Bull: it has the optionality to monetize spare capacity, it's moving to rent computing power to outside customers like Anthropic, potentially turning cost into revenue while it waits for its next AI products (Bloomberg Surveillance, Jul 20). Bear: that very move "blows a hole in the whole bull thesis of scarcity," and Meta remains the poster child for off-balance-sheet, short-contract financing risk (RiskReversal, Jul 20). Next: earnings next week, watch the capex guide and any confirmation of the compute-rental business.

Read-throughs

  • Power / thermal: the strongest signal of the sweep, and it's a supplier tailwind. Lazard's George Bilicic gave the clearest operator picture yet: utilities growing 7-9% a year (some 10-15% regionally), Texas and the Southeast as the fastest places to plug in, and a genuine construction bottleneck that hands pricing power to the few firms that can build large plants (Bechtel, Fluor, Kiewit, Quanta) while Lazard's own cost data shows combined-cycle gas plant costs rising sharply (Tech Disruptors, Jul 20). The thermal-and-electrical names (Vertiv, Eaton) were name-quiet again, but the demand-for-gear read is only getting stronger. One fresh reliability wrinkle: the grid regulator NERC issued its highest-severity alert (a rare "level-three" warning) back on May 4 after multiple incidents where "more than a gigawatt of data center load dropped off the grid in seconds," threatening frequency stability: the fix, per the discussion, is simply more on-site batteries (Volts, Jul 20). Most actionable trade: stay long the power/grid/cooling supply chain; the physical bottleneck is worsening, not easing.
  • Memory: the panic overshot; respect the tightness, mind the cyclicality. The best insight this cycle is that the Chinese K3 model actually needs enormous memory (~1.4 TB to run), so the Friday memory sell-off was likely a mistake, SK Hynix, Samsung and Micron "might actually be fine here" (The Rundown, Jul 20). But the structural risk Boyle names is real: memory is an industrial commodity, and Korea's market is "two chip stocks wide" and mechanically leveraged, so any hyperscaler spending pause hits Seoul first and hard (Patrick Boyle On Finance, Jul 19). Morningstar adds the technical overhang: over 50% of the Korean market's value is just Samsung and SK Hynix, the index is down ~25% from its high (still up ~50% year to date), and margin calls have forced selling (The Morning Filter, Jul 20). Most actionable: the tightness is real and the K3 scare was overdone, but treat the fattest memory margins as the most cyclical thing you own.
  • Semiconductor equipment / "picks and shovels": still the softer-landing exposure, but no longer cheap. TSMC's raised $62 billion budget and ASML's back-to-back 30% capacity adds are the hardest confirmation yet that the 2027 build is funded (The Morning Filter, Jul 20). The catch: both stocks fell on the news, so the market has already paid for it. Bilicic's separate point reframes the whole category: power and utilities are the other "pick and shovel play" on the AI gold rush, and arguably the less-crowded one (Tech Disruptors, Jul 20).
  • Optics / networking: a coverage gap this cycle, not a signal. No podcast in the July 19-20 window carried substantive color on the optical (Coherent, Lumentum, Fabrinet) or interconnect-silicon (Marvell, Astera Labs, Credo) names. Given these were the group that broke first after Broadcom's June guide, the silence is a gap to close on the next sweep, not evidence the pressure has lifted.
  • Utilities / nuclear: demand intact; nuclear is real but large-scale and slow; politics is the growing gate. Bilicic expects "a lot of nuclear, but it's going to be large-scale nuclear" (the AP1000 and GE Hitachi designs), with maybe one or two small-modular-reactor technologies that actually validate and scale, a reality check on the SMR stock enthusiasm (Tech Disruptors, Jul 20). Vistra, Constellation and Talen were name-quiet. The louder risk is political: on Bloomberg Surveillance the panel flagged rising "not in my backyard" resistance and even the prospect of a windfall tax on the winners; and a governance panel noted the wealth-and-jobs backlash is now drawing proposals like Bernie Sanders's "50% of the equity must go to the United States" and Sam Altman floating giving away 5% (Bloomberg Surveillance, Jul 20; Boardroom Governance, Jul 20). Carry the power positions; the demand is intact, the fresh risk is permits and politics.

What changed vs last issue

Last issue (Jul 20, "Chip stocks hit bear-market territory; Chanos does the math") caught the weekend crash and the arrival of the marquee bears. This cycle the mood didn't get worse, it got scrutinized, and in places walked back. Here's what actually moved:

  • The China scare deflated. Monday's issue treated the Chinese "Moonshot" model as a live threat to memory margins. This cycle the model got a name (Kimi K3) and a closer look: Steno argues the market "got this wrong" because K3 leads only in narrow front-end coding and trails on the broad benchmarks, and The Rundown pointed out K3 is a memory-hungry 2.8-trillion-parameter model (needs ~1.4 TB to run), which makes it bullish, not bearish, for Micron/SK Hynix. Net: the specific memory-crack fear from last issue is partially reversed.
  • The tell shifted from "the math" to "running to stand still." Last issue's frame was returns on capital (Chanos). This cycle the market gave a cleaner signal: TSMC (+77% record profit, budget raised to $62 billion) and ASML (30% capacity adds in 2027 and 2028) delivered blowouts and fell. Ozan Tarman's line captures it: "SK Hynix delivered. Price action didn't. TSMC delivered. Price action didn't." The good news is priced; the fight is now over 2028.
  • A scarcity crack appeared. New this cycle: Meta moving to rent out spare computing power (NYT, over the weekend), which even the bulls concede "blows a hole in the whole bull thesis of scarcity." Watch whether it spreads.
  • The catalyst clock started ticking loudly. Last issue the late-July prints were "days away." Now they're here: Google Wednesday, then Microsoft, Meta, Amazon and Apple next week, plus a Fed meeting. And the setup sharpened: short bets against US stocks are at their highest since 2010, so a confident set of guides could force a squeeze, while a single flat capex number confirms the bears.
  • Power got an operator voice. Last issue's power story was the "off-grid fantasy" rebuttal. This cycle Lazard's Bilicic added the banker's-eye detail: utilities as the 7-9% growth pick-and-shovel play, the cheap-turbine land-grab as temporary and inefficient, construction capacity (Bechtel/Fluor/Kiewit/Quanta) as the real bottleneck, and transmission as "impossible to build" right now.
  • The IBM tell got quantified. Last issue we flagged IBM as the first company to name AI-hardware hoarding as a drag. This cycle Morningstar confirmed IBM's preliminary Q2 was bad and the stock fell ~25%, with management blaming clients shifting spending "away to other areas like servers, storage, memory", the cannibalization thesis showing up in a real print (The Morning Filter, Jul 20).
  • Deltas on the big numbers: the four biggest hyperscalers' near-term capex framed at ~$376 billion (Boyle/Setser), industry ~$725 billion (Boyle) to $800 billion-plus for 2027 (Josey), with a 2026 estimate as high as $920 billion (Wealthion); Google Cloud backlog ~$460 billion, AWS ~$360 billion; TSMC 2026 budget lifted to $62 billion (+$8B); the chip index down ~20% from its June high (bear market) but still ~65% up year to date; Korea down ~25% from its high but still ~50% up on the year; short interest in the S&P at its highest since 2010; and business fixed investment at 14.1% of GDP, a 25-year high.
  • Still open: no hyperscaler has yet guided capex flat: Google Wednesday is the first test; the optics/networking names (Marvell, Astera, Credo, Coherent, Lumentum) drew no coverage this cycle; and the "does the other five spenders' revenue show up?" question that Jay Goldberg posed remains the crux the earnings must answer.

Next prints to watch: Google, Wednesday after the close (the first hyperscaler and the whole cycle's swing vote), then Microsoft, Meta, Amazon 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.