Newsletter · · Ashutosh Agarwal

Chip Stocks Bleed on Meta Compute Glut Fear as the Power Grid Roars - The AI Capex Tracker - Week of July 3, 2026

The AI Capex Tracker for July 3, 2026. A read that Meta will build a cloud business to resell AI capacity sent the semiconductor ETF down more than 6 percent and neoclouds CoreWeave and Nebius down double digits, while PJM hit a record 166 GW of power demand and frontline operators said ERCOT interconnection is now the binding constraint, and a soft June jobs report cut July Fed hike odds.

The AI Capex Tracker

Week of July 3, 2026: Chip Stocks Bleed on Meta Compute Glut Fear as the Power Grid Roars

Issue: Friday, July 3, 2026

TL;DR

  • The Meta "excess-compute" story is now the whole tape's fault line. The semiconductor ETF fell more than 6% on the read that Meta will build a cloud business to sell AI capacity, neoclouds CoreWeave and Nebius fell double digits. Supply catching up to demand is the one narrative this rally can't survive. (The Rundown, Jul 2)
  • The physical layer says the opposite of a glut. PJM hit a record 166 GW of demand and warned of an "imminent electricity reliability emergency"; day-ahead power ran from $44 to $436/MWh (+~900%) with a 7–8pm peak above $1,200. Frontline developers say ERCOT is now "critical path even just to get studied." (Squawk on the Street, Jul 2; Open Circuit, Jul 2)
  • Markets are shut today for the Fourth, and the macro just softened underneath the trade, June payrolls came in at 57k vs ~115k expected, with 74k of prior revisions, cutting the odds of a July Fed hike. Thin tape into a heavy late-July earnings run. (The Rundown, Jul 2)

What's New

Markets are closed today for Independence Day, so there's no fresh print, but the Wednesday–Thursday tape did the damage, and the operator commentary is the richest in a week. Ranked by dollar P&L impact:

1. The Meta selloff got quantified, and it hit the middlemen hardest. The Rundown, Jul 2, Zaid Admani (daily-markets host; relaying the tape). "The semiconductor ETF fell more than 6%… the catalyst for this wipeout was the news that Meta is planning to build their cloud business to sell off excess AI computing capacity… especially neo clouds like CoreWeave and Nebbius, which were down double digits." His framing is the bear thesis in one line: "if Meta, of all companies, suddenly has more compute than it needs, maybe it means the supply of AI computing power is finally catching up to demand." That's the over-ordering tell this note has watched for since day one, now in price.

2. The power operators are screaming that demand is real. Open Circuit, Jul 2, the sweep's best insider source: Ian Black (SVP & global head of energy, Digital Realty, a ~300-facility public REIT) and Holly Adams (SVP energy, Beal Infrastructure, Blue Owl–backed). Black has been "waiting on Encore for two years to give us a study in Texas… Texas is now becoming critical path even just to get studied." Utilities are culling ~80% of queue applicants with financial-security rules "that look exactly like MISO and PJM in 2013, 2014." His stat for why this permeates the value chain: a data center runs $10–13/watt vs ~$2/watt for renewables, so "for a billion dollars of renewable spend, an equivalent for a data center is $15 billion." Powered land is now a far higher bar, "permits in hand before you can get studied."

3. Behind-the-meter has gone from workaround to default. EnergyCents, Jul 2, Ben (energy analyst). Behind-the-meter projects for data centers have scaled from single-digit MW historically to an average of ~2 GW per proposed project in Q1 2026, predominantly gas because developers "can't get [grid power] immediately" and it's "very costly… to sit on the chips they procured and not generate revenue." The supply wall is real: Baker Hughes turbines "sold out through 2028," Siemens quoting 24–36 months, GE slipping toward 2030 "because of bottlenecks at the EPC level." Fuel cells and recips win the near-term slots on lead time and permitting ease.

The Debate

The Meta headline split the specialist tape into two clean camps this week.

Bear steel-man, this is the over-ordering confession. Prof G Markets, Jul 2's guest (an AI-capex bear known for leaking OpenAI's financials): "Meta has no use for this compute because they don't have an AI story." His demand math is the scary part, "80% of all AI compute is owned by or used by OpenAI or Anthropic, with the rest being Meta… it's a mirage of demand." He cites a UBS study that "60% of organizations are token minimizing," and warns there's "over 100 gigawatts of compute capacity under construction. If we can only sell six… that demand is mostly coming from two unprofitable companies." His contagion path runs through the neoclouds, private-credit funds, and levered Japanese banks. Corroborating tell: OpenAI reportedly delayed its IPO after failing to clear a $1T valuation.

Bull steel-man, reselling is monetization, and the physical layer proves scarcity. Futurum Equities, Jul 2 (bullish PMs) call the neocloud selloff "drastically exaggerated": AWS is "raising GPU rental prices by 20%," "multiple customers are fighting for one GPU," and Meta subleasing ~500 MW at market "would be like $6 of earnings… investors should love that." Their sharpest point: nobody panicked when SpaceX sold excess capacity, yet "when Mark says he might do the same thing, everybody lost their minds." They concede the real long-tail risk, hyperscaler landlords becoming marginal competitors to the neoclouds if they overbuild, but date it to ~2030. And the macro bears stayed loud: Jeremy Grantham called AI "the biggest investment bubble that arguably has ever occurred," a "game of chicken" between staggering infrastructure debt and slow revenue (Tom Bilyeu's Impact Theory, Jul 2).

Sell signals to watch: a hyperscaler confirming at late-July earnings that it's marketing excess capacity (Meta most likely); the share of Meta's fleet it dumps crossing ~50%; GPU rental rates rolling over; a forward-capex cut; and the ERCOT Batch Zero firm-load list (due August) coming in far below the headline queue.

Stocks in Play

NVDA. Bull: still the asset the whole trade routes through, and its $2B stakes across the optics layer (Lumentum, Coherent, Marvell) keep it central beyond GPUs. Bear: first to de-rate on any demand blink, the SOXX –6% shows how fast the complex re-prices on a "supply caught up" headline (The Rundown, Jul 2). Next: TSMC mid-July; Q2 in August.

AVGO. Quiet on operator color; OpenAI's Broadcom custom-silicon program got secondhand airtime but no first-party detail. Next: ASIC ramp cadence.

AMD. Quiet on the roadmap. Next: AMD Advancing AI Day, July 2026 (MI450X/Helios).

MSFT. Bull: even the loudest bear concedes Microsoft holds "almost as much capacity as Meta" with mothball optionality, and Futurum calls it "the best opportunity in the Mag 7 other than NVIDIA" (Prof G Markets, Jul 2; Futurum Equities, Jul 2). Bear: if Meta's glut read is right, Azure's own utilization becomes the next question; still an H1 laggard. Next: FY26 Q4 capex, late July.

GOOGL. Bull: reportedly had the standing to turn down Meta's request for more capacity, supply strength. Bear: the bears re-cast that as "it's all Anthropic" demand, not Gemini's; monetization still unbroken out (Prof G Markets, Jul 2). Next: July capex guide.

AMZN. Quiet, the neutral cloud hosting both OpenAI and Anthropic. The one read-through: AWS reportedly raising GPU rental prices ~20%, a scarcity tell that cuts against the glut narrative (Futurum Equities, Jul 2). Next: late-July earnings.

META. The eye of the storm. Bull: excess capacity is a monetizable asset, ~500 MW at market "≈ $6 of EPS," and the stock rose on the news (Futurum Equities, Jul 2). Bear: reselling is the loudest over-ordering tell yet, "Meta has no use for this compute because they don't have an AI story," with impairment risk if it can't clear the capacity (Prof G Markets, Jul 2). Next: earnings, late July, the single most important print of the cycle.

Read-Throughs

  • Power / thermal (VRT, ETN), the ballgame, and it's roaring. PJM's record 166 GW demand and ~900% day-ahead spike ($44 → $436, peak >$1,200) is a live demo that the bottleneck is physical, not order books (Squawk on the Street, Jul 2). With ERCOT studies stuck two years deep and turbines sold out to 2028–2030, own the onsite-power, switchgear, and fuel-cell enablers (Open Circuit, Jul 2; EnergyCents, Jul 2).
  • Utilities / nuclear (Vistra, Constellation, Talen). Independent power producers are direct heat-wave beneficiaries, capturing the merchant price spike (Squawk on the Street, Jul 2).
  • Data-center REITs. +30% YTD; yields low-double to mid-teens with supply tightness feeding pricing power, though capped by hyperscalers' insourcing option (Nareit REIT Report, Jul 2).
  • Memory / neoclouds (MU, CoreWeave, Nebius). The demand-blink fear hits the levered middlemen first, double-digit drops on the Meta read (The Rundown, Jul 2). Counter-signal: Apple is chasing blacklisted Chinese memory (CXMT, YMTC), you don't do that unless the shortage is acute.
  • Optics / photonics (LITE, COHR, MRVL, GFS, TSEM). A promotional stock-pitch show (The MoneyFlows Show, Jul 2) flagged silicon photonics as "the next memory trade": treat the targets as hype, but the hard facts matter: NVIDIA holds $2B stakes in Lumentum, Coherent and Marvell, and Jensen Huang says needed photonics capacity is "substantially higher than the world has today."

What Changed vs Last Issue

Last issue (Jul 2, "Meta turns seller. The AI air pocket shows.") caught the rumor and the first reaction, Micron ~–10%, Nebius ~–17%. Today the story matured on three fronts:

  • The damage widened and the debate crystallized. No longer one stock's tell, the whole semiconductor ETF fell >6%, CoreWeave joined Nebius down double digits, and the specialist tape split cleanly into "monetization" (Futurum) vs "mirage of demand" (Prof G's bear).
  • The physical layer pushed back hard. New this issue: PJM's record 166 GW / ~900% price spike and frontline-operator confirmation (Digital Realty, Beal) that ERCOT is the binding constraint, the demand looks real even as glut fear grows. The two threads now openly contradict; late-July earnings resolve it.
  • The macro shifted underneath. A soft June jobs report (57k vs ~115k) cut July hike odds, new since last issue. No fresh aggregate-capex print this holiday window; the ERCOT Batch Zero firm-load read (August) remains the next hard number.