Newsletter · · Ashutosh Agarwal

Capex Forecasts Leap Toward 37 Percent but the Chip Rally Stalls - The AI Capex Tracker - Week of July 8, 2026

The AI Capex Tracker for the week of July 8, 2026. Hyperscaler capex forecasts get marked up toward 37 percent growth, but rising capex no longer lifts the stocks: Samsung's record print sold off, money rotated into memory and HBM, and SK Hynix passed Samsung as Korea's most valuable company ahead of its US listing.

The AI Capex Tracker

Week of July 8, 2026: Capex Forecasts Leap Toward 37 Percent but the Chip Rally Stalls

TL;DR

  • The number that matters just got revised up, hard. Gene Munster says the Street was modeling 17% hyperscaler capex growth for next year a few weeks ago, then 23%, and after Amazon's fresh $25B bond raise and Google's capital-raise comments, "that probably suggests the numbers are going to be up around 37%, so much higher." His read: the buildout is in its "early innings." (Closing Bell, Jul 7)
  • But rising capex no longer buys you a higher stock: the reaction function has flipped. Samsung printed a record (revenue "more than doubled," profit up "19-fold") and the stock fell; Bloomberg's Ed Ludlow: "a stock would go up when a company would take its CapEx numbers up… That game seems to be out of fashion." Money is rotating out of the spenders and into the memory/infra beneficiaries. (Bloomberg Intelligence, Jul 7)
  • Q2 hyperscaler capex hit ~$168B, +74% YoY, and the "excess compute" resellers multiplied. XAI is renting Anthropic compute at $1.25B/month and Meta is opening a cloud business; bulls call it monetizing overbuild, bears call it an admission of it. (The Financial Exchange, Jul 7)

What's new

Yesterday's note said demand answered the credit scare with signatures. Today the tape asked a harder question: if the capex forecasts keep going up, why are the stocks going down? Ranked by dollar P&L impact:

1. The 2027 capex forecast is being marked up in real time, from 17% to a possible 37%. Closing Bell, Jul 7. Gene Munster, Managing Partner, Deepwater Asset Management.

Munster used Amazon's $25B raise as a tell: "The street's looking for 23% capex growth from the hyperscalers for next year. And now a few weeks ago that was 17%… that probably suggests that the numbers are going to be up around 37%." He frames physical AI (robotaxis, automation) as a further "30 to 50 percent positive impact on the overall AI spend" that "obviously hasn't started," i.e., the demand runway is longer than consensus. The catch he keeps flagging: "There is a question about what the multiple investors are willing to pay for that." Higher numbers, more multiple anxiety. That is the entire trade in one sentence.

2. Samsung's record print got sold, and that's the signal, not the noise. Bloomberg Intelligence, Jul 7. Ed Ludlow, Bloomberg Tech anchor.

Numbers were great; the stock (up ~150% into the print) fell anyway. Ludlow: "nothing in the memory markets changed from when you and I went to bed last night." His caution on margins is the one to internalize: "in those periods where the memory names have had very high margins, they've never been maintained. They ebb and flow, they crash to a floor." BI's Amazon capex path: $200B to $300B in full-year '27, funded by a $25B, 8-tranche IG deal (3–40yr). The behavioral shift matters more than any single number: capex-up is no longer a buy trigger.

3. Q2 hyperscaler capex ~$168B (+74% YoY), and the "excess compute" resale trend spread. The Financial Exchange, Jul 7 (market-recap show; Visible Alpha–sourced).

Google, Meta, Amazon and Microsoft spent ~$168B in the June quarter per Visible Alpha, up 74% YoY, with Meta "potentially risking going into negative cash flow." Two fresh resale prints: XAI is renting Anthropic capacity at $1.25B a month, and Meta is entering cloud, selling compute in a market long owned by AWS/Azure/GCP. Bull read: monetizing spare capacity (the AWS origin story). Bear read: "is this Meta's acknowledgement… that we are spending a lot of money," the late-90s broadband overbuild analogy, live.

4. Memory is where the money went: SK Hynix is now Korea's most valuable company, HBM sold out. Last Week in AI, Jul 7 (news-discussion show; corroborated facts).

Micron's "entire calendar year 2026 is just gone for HBM supply… all spoken for," including HBM4, and it struck a supply deal with (and invested in) Anthropic. SK Hynix overtook Samsung as South Korea's most valuable company on 61% of the global HBM market: HBM has gone from commodity to moat. Its US ADR debuts Friday, July 10 (~$29B raise); watch SMH/SOXX inclusion as the institutional access valve. And the custom-silicon front widened: Amazon's Trainium/Graviton/Nitro business "crossed $20 billion of annual revenue run rate" in Q1 and is in talks to sell chips to external operators, while OpenAI (3nm, eight HBM stacks) and Microsoft's Maya both tape out on TSMC 3nm, so Nvidia "is going to be feeling a lot of heat," and 3nm allocation is now a four-way knife fight. (Market Mondays, Jul 7)

The debate

The $700B+ thesis didn't move; today the fight was about what a higher number is worth.

Bull steel-man: the forecasts are being revised up and the runway is long. Munster: capex growth marked from 17% toward ~37% for next year, physical AI barely started, "early innings" (Closing Bell, Jul 7). HSBC's Max Kettner called the ~20% mid-May-to-June hyperscaler drawdown "unjustified… a little bit overblown," this being the "third wave" of the AI-overspend narrative, and each wave "can last for a couple of weeks. But after that, you're kind of done… the numbers are just there" (Closing Bell, Jul 7). HBM sold out through 2026 is hard evidence of demand, not hope (Last Week in AI, Jul 7).

Bear steel-man: a bubble with thin revenue durability and a 20% credit tail. VC Rob Ward (CMO Confidential): "There's no doubt this is a bubble… something like 80% of all gains this year are coming from" AI, with private rounds getting a 2x markup "before that Series B round has closed." His durability worry reads straight across to enterprise-software revenue: a marquee legal customer told him he'd unplug the AI vendor after the trial, "my lawyers won't even know the difference," switching costs near zero (CMO Confidential, Jul 7). On structure: ~80% of the buildout is mega-tech equity, but the ~20% funded by private credit is where "endowment funds… are going to lose money" if it turns (Bite-Sized Business Law, Jul 7). And the market flashed the tell itself: capex-up no longer pays.

Sell signals to watch: a hyperscaler cutting forward capex (none did, the opposite); memory margins rolling off the "90%" peak toward Ludlow's "floor"; the SK Hynix / TSMC / ASML prints next week (Jul 10–16) confirming demand or marking the top; "excess compute" reselling turning from a trickle into a habit; private-credit AI paper widening to distressed spreads.

Stocks in play

NVDA. Bull: still the platform: TSMC (reports ~Jul 16, ~50% revenue jump expected) is the demand tell, and Nvidia made itself TSMC's number-one client over Apple. Bear: the custom-ASIC front is widening on every side, Amazon selling Trainium externally, OpenAI and Microsoft on 3nm, "Nvidia is going to be feeling a lot of heat." Next: TSMC print Jul 16; Q2 in August. (Last Week in AI, Jul 7; Market Mondays, Jul 7)

AVGO. Bull: the Apple "Baltra" partnership runs through 2031, the arms dealer to Google, Meta, OpenAI and Apple. Bear: the in-house threat is the counter: the debate is "how many companies can afford to keep the entire stack in-house," and only ~3 can, but each one that does is a design loss. Next: ASIC ramp cadence. (Market Mondays, Jul 7)

AMD. Bull: quietly up ~$20 to ~$570 on the day; specialists still rate AMD (and Intel) a likelier next-$1T name than Marvell. Bear: no fresh operator catalyst, shipment timelines still aspirational until proven. Next: AMD Advancing AI Day, July 2026 (MI450X/Helios). (Market Mondays, Jul 7; Chip Stock Investor, Jul 7)

MSFT. Bull: IG issuer with the balance sheet to fund the build; Maya inference chip on 3nm gives it an internal-silicon lever. Bear: the relative loser of the group: "Microsoft has really struggled this year because people aren't sure about its AI position or what it's really getting out of the tens of billions" it's spending. Next: FY26 Q4 capex, late July. (Bloomberg Intelligence, Jul 7)

GOOGL. Bull: the relative winner: "Alphabet's doing pretty well because people remain pretty positive about its AI position," and TPU-rack rentals extend the captive-silicon story. Bear: Google's own capital-raise commentary is part of what pushed Munster's capex-growth number to ~37%, the spending is real and rising. Next: July capex guide. (Bloomberg Intelligence, Jul 7; Closing Bell, Jul 7)

AMZN. Bull: the day's clearest demand tell: a second $25B IG raise, capex path to $300B in '27, and a $20B+ custom-silicon run rate that's now a profit engine for AWS. Bear: the same raise is what has investors "getting jittery" about the multiple; leaning on debt to fund capex that grows "much faster than revenue." Next: late-July earnings. (Bloomberg Intelligence, Jul 7; Closing Bell, Jul 7)

META. Bull: the XAI/Anthropic-style compute resale reframed as building the next AWS off spare capacity. Bear: still the FCF-risk name of the group ("potentially risking going into negative cash flow"), and the compute reselling reads to skeptics as an overbuild admission. Next: late-July earnings, the referee print of the cycle. (The Financial Exchange, Jul 7)

Read-throughs

  • Memory (MU, SK Hynix, Samsung): the rotation trade. Money moved out of hyperscalers and into "Micron, SanDisk, Western Digital, Seagate." Micron's 2026 HBM is fully sold, SK Hynix owns 61% share and lists Friday (~$29B). The trade to respect both ways: demand is real and Samsung's sold-off record print is a live warning that peak memory margins "crash to a floor." (Bloomberg Intelligence, Jul 7; Last Week in AI, Jul 7)
  • Networking / optics (MRVL). Specialists pushed back on Jensen's "$1T Marvell" call: FY28 revenue guided to $16.5B (data center +55%), but at ~$245 the reverse DCF implies ~47% annual per-share profit growth for five years, "we would pass." The Celestial AI buy keeps Marvell in the optical-interconnect fight. Priced for perfection is the read. (Chip Stock Investor, Jul 7)
  • Power / thermal (VRT, ETN, nuclear). Quiet on the specialist tape this cycle, no fresh utility/ERCOT/Vertiv/Eaton operator color surfaced. The only power thread was the reminder that AI's true ceiling is physical: "the costs are not technology costs. They're power, water… a planetary constraint." Carry the prior stance; watch for the next ERCOT/SB6 print. (Bite-Sized Business Law, Jul 7)

What changed vs last issue

Yesterday (Jul 7, "Anthropic's $19B check answers the credit scare") the question was whether the financing holds. Today the question changed: the financing is flowing (Amazon's second $25B raise), the forecasts are being marked up (Munster: 17% → 23% → ~37%), and yet the stocks won't reward it. The delta this cycle is in the reaction function, not the numbers:

  • Capex estimates got revised UP. New anchor: ~$168B Q2 hyperscaler capex (+74% YoY, Visible Alpha); 2027 growth now pegged as high as ~37% vs 23% consensus a few weeks ago; Amazon's own path $200B → $300B for '27.
  • But "capex-up" stopped being a buy signal, Samsung's record print sold off, hyperscalers underperformed, and money rotated into memory/storage. Ludlow: that game "is out of fashion."
  • "Excess compute" reselling spread from Meta's Jul 6 headline to XAI renting Anthropic capacity at $1.25B/month, bulls say monetization, bears say overbuild.
  • HBM leadership crystallized: SK Hynix passed Samsung as Korea's most valuable company (61% HBM share); its US ADR lists Jul 10.

Next real-time prints: SK Hynix (Jul 10), ASML (Jul 14), TSMC (~Jul 16). Three demand tells in one week, the tape will either confirm the 37% or start pricing the top.