Newsletter · · Ashutosh Agarwal

Private Credit Gates AI Data Center Financing as Zuckerberg Admits Demand Stall - The AI Capex Tracker - Week of July 6, 2026

The AI Capex Tracker for the week of July 6, 2026. Over the holiday the bear case gained a financing leg as private-credit funds gated redemptions and Blackstone walked from data-center projects, Zuckerberg conceded agent demand has stalled despite Meta's up to $145B spend, and the physical grid layer kept printing hard scarcity.

The AI Capex Tracker

Week of July 6, 2026: Private Credit Gates AI Data Center Financing as Zuckerberg Admits Demand Stall

TL;DR

  • The bear case got a new, more dangerous leg over the holiday: the money. The private-credit shadow banks that funded the buildout are gating redemptions. Blue Owl's two flagship funds took the largest withdrawal requests in the industry (one hit for 38.1% of shares), Bloomberg pegs ~$14B "trapped," and Blackstone is selling data-center stakes and abandoning a Virginia megaproject through QTS. The marginal lender now wants "show me the signed lease and the power." (Eurodollar University, Jul 5)

  • Zuckerberg said the quiet part out loud. In an internal town hall he admitted "AI agent development over the last 4 months has not accelerated in the way we expected," even as Meta spends up to $145B this year, and the read is that real demand now concentrates into two money-losing labs, OpenAI and Anthropic. (Big Technology Podcast, Jul 3)

  • The physical layer still screams the opposite of a glut. PJM's grid chief laid it out cold: data centers are 94% of projected peak-load growth to 2030, wholesale capacity ran from $28 to $270/MW-day in a year, and a data center takes 1–2 years to build but its power plant takes 5–7. AWS quietly raised GPU rents 20% on July 1, "the first open, on-record pricing signal that compute scarcity is structural, not cyclical." (TED Tech, Jul 3; Telltales, Jul 5)

What's new

Markets were shut Friday for the Fourth and reopen today into a thin tape. The holiday didn't quiet the trade, it deepened it. Ranked by dollar P&L impact:

1. The financing channel started to tighten, the highest-consequence new signal. Eurodollar University, Jul 5, Jeff Snider. This is the vector this note hasn't had to worry about, because money was unlimited. No longer. Blue Owl's credit fund saw ~$4.2B of redemption requests in Q1 and ~$3.6B in Q2; its tech-focused fund took requests equal to 38.1% of shares, capping withdrawals both quarters. Bloomberg reports ~$14B trapped across private-credit funds; Blackstone is "selling data center stakes and abandoning a massive Virginia data center project through QTS." Snider's mechanism is the tell: once the gate closes, the manager "stops funding the dream and starts underwriting the downside," wants "the signed lease... the power... the exit," so "projects get delayed, marginal sites get abandoned... valuations stop floating upward on narrative alone." The AI trade was always half debt story. The debt half is where the crack is.

2. Zuckerberg quantified the demand problem himself. Big Technology Podcast, Jul 3, Alex Kantrowitz with Ranjan Roy (Margins). Per a recording of an internal town hall, Zuckerberg conceded agent development "has not accelerated in the way we expected," while Meta spends up to $145B, part of Big Tech's $700B+ outlay. Kantrowitz's framing distills the bear thesis: the industry is collapsing "toward two points of failure, OpenAI and Anthropic," both "kept afloat by financing and... losing a lot of money." When the biggest spender says its own products can't absorb its own compute, "matching demand" stops being a risk and becomes an admission.

3. The best insider of the sweep says the bottleneck is physics, not orders. TED Tech, Jul 3, Asim Haque, SVP at PJM Interconnection (grid operator for 67M people across 13 states) and a former Ohio utility regulator. His numbers are the cleanest scarcity evidence on the tape: data centers = 94% of PJM's projected peak-load growth to 2030; one capacity auction ran $28 to $270/MW-day in a single year, "a different phenomenon than anything we have experienced." The killer asymmetry: 1–2 years to build a data center, 5–7 to build its power plant, with 50,000+ MW of ready resources stuck on "permitting and siting." Guest Varun Sivaram (Emerald AI) adds the release valve: flexible data centers could unlock up to 100 GW / $4 trillion on existing grids.

4. Compute got priced as a scarce good, and Microsoft put money behind deployment. Telltales, Jul 5 (AI-generated show, used only for corroborated facts). AWS raised GPU instance prices 20% on July 1, the first on-record signal "that compute scarcity is structural, not cyclical." Microsoft committed $2.5B and 6,000 people to a new "Frontier Co." deployment unit and cut up to 5,000 legacy roles to fund it, a bet (at ~38x FCF, capex near $100B TTM) that deployment revenue shows up before the depreciation does. Data-center builds now run $65–75B per gigawatt, memory 30–40% of that.

5. The OpenAI-Broadcom custom ASIC got a name. Artificial Developer Intelligence, Jul 3. OpenAI and Broadcom announced "Jalapeno," a ~9-month inference ASIC design (LLMs in the loop), a compute chiplet ringed by six HBM modules (HBM, not cheaper DDR), claiming a "substantial" cost-per-watt beat versus incumbent (read: Nvidia) gear. Detail is scant, but it's another marker that every lab is racing to vertically integrate silicon.

The debate

The holiday sharpened both sides of the $700B+ thesis rather than resolving it.

Bear steel-man, the funding tightens while demand concentrates. The new leg is credit: gated private-credit funds and Blackstone walking from projects mean the marginal AI dollar just got selective (Eurodollar University, Jul 5). Layer on Zuck's admission and the "two points of failure" read (Big Technology Podcast, Jul 3). Dan Niles on ROI: hyperscalers "don't necessarily make any money on the compute if they give it away... it hasn't been a great return on investment," and memory "has no moat," a glut-vs-scarcity cycle where Micron's cheap-looking 9x forward sits on earnings it will plow back into capacity (Excess Returns, Jul 3; The Disciplined Investor, Jul 5). Alex Karp's needle: "if it was so valuable... why are they charging for tokens?" (Memes and Markets, Jul 4).

Bull steel-man, the physical layer proves scarcity, and reselling is monetization. Haque's PJM numbers and AWS's +20% GPU rents are hard scarcity prints, not order-book optimism (TED Tech, Jul 3; Telltales, Jul 5); only 20–30% of planned data centers are built (Money On Tap, Jul 4); and agentic demand compounds as AI starts using AI at machine speed, the case for more tokens, not fewer (David Senra, Jul 5). One All-In segment framed the counter-move as Anthropic pushing vertical integration and an on-prem shift, another sign the labs are hardening their own stacks (All-In, Jul 3).

Sell signals to watch: private-credit gating widening or more data-center projects abandoned; a hyperscaler confirming it is marketing excess capacity at late-July earnings (Meta most likely); memory spot pricing rolling over as new capacity lands; a forward-capex cut; AWS/GPU rental rates reversing off the +20% mark.

Stocks in play

NVDA. Bull: still the platform the whole trade routes through, even LPU challenger Groq's rumored $20B partnership is with Nvidia, making its silicon available to all Nvidia customers (David Senra, Jul 5). Bear: first to de-rate if the "two-labs-of-real-demand" read is right and tightening financing chokes the marginal buildout. Next: TSMC mid-July read-through; Q2 in August.

AVGO. Bull: OpenAI's "Jalapeno" ASIC, co-developed with Broadcom on a 9-month cycle with a cost-per-watt claim, validates the custom-silicon franchise (Artificial Developer Intelligence, Jul 3). Bear: details thin, ramp still a promise. Next: ASIC ramp cadence.

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

MSFT. Bull: put $2.5B and 6,000 people into a "Frontier Co." deployment unit and made Copilot a permanent $21/user SKU, a declared bet deployment revenue arrives (Telltales, Jul 5). Bear: ~$100B capex TTM against ~$76B FCF at ~38x; if Azure demand is mostly OpenAI/Anthropic, utilization is the question. Next: FY26 Q4 capex, late July.

GOOGL. Bull: "Alphabet is winning [the vertical-integration] race by far," silicon to distribution (Artificial Developer Intelligence, Jul 3). Bear: bears re-cast Google's demand as Anthropic's, not Gemini's. Next: July capex guide.

AMZN. Bull: AWS raised GPU rents 20% on July 1, a scarcity tell that cuts against the glut narrative (Telltales, Jul 5). Bear: the neutral cloud hosting both OpenAI and Anthropic, exposed if either wobbles. Next: late-July earnings.

META. The eye of the storm. Bull: excess capacity is monetizable if it builds the cloud business it's now designing. Bear: Zuckerberg's own words, agent development "has not accelerated in the way we expected" despite $145B, make reselling the loudest over-ordering tell yet, with impairment risk if the capacity doesn't clear (Big Technology Podcast, Jul 3). Next: earnings late July, the single most important print of the cycle.

Read-throughs

  • Power / thermal (VRT, ETN), still the cleanest long. Haque's PJM math ($28 to $270/MW-day, the 5-to-7-year plant gap, 94% of load growth) says the binding constraint is physical, and Vertiv remains "the primary name" in liquid cooling with little competition (TED Tech, Jul 3; Money On Tap, Jul 4). Watch Emerald AI's up-to-100 GW flexibility unlock, it would soften the "must-build" scarcity case if it scales.

  • Utilities / nuclear (Vistra, Constellation, Talen). Nuclear is Haque's 95%-reliability "gold standard" and captures merchant price spikes, direct beneficiaries of the same congestion (TED Tech, Jul 3).

  • Memory / networking (MU, MRVL). Rotation into the "recipients" continues, but the risk is now explicit: Micron, SK Hynix, Samsung plus China's CXMT are all adding capacity, which "will have the effect of lowering these really, really sky-high profit margins" over 3–5 years (Chip Stock Investor, Jul 4).

  • Optics (LITE, COHR). Lumentum is the pure-play leader with OCS systems shipping in H2; Coherent's indium-phosphide supply is a chokepoint, Nvidia holds ~$2B stakes in each (Chip Stock Investor, Jul 4).

  • Neoclouds / data-center REITs (CoreWeave, Nebius, QTS-type assets). The private-credit gate and Blackstone's pullback hit the levered middlemen first, they re-price before the chips do (Eurodollar University, Jul 5).

What changed vs last issue

Friday's issue (Jul 3, "Chips bleed on Meta's glut. The grid roars.") framed the fight as a Meta glut headline (SOXX –6%) colliding with a roaring physical layer, with late-July earnings as referee. Three things moved over the weekend:

  • The bear case gained a new, higher-stakes leg: financing. Private credit, the shadow-bank plumbing behind the buildout, is now gating redemptions (Blue Owl, ~$14B trapped) and Blackstone is selling data-center stakes and abandoning a Virginia project. Absent from the Jul 3 note, it's the "who funds the marginal project" tell, the most actionable new development of the cycle.

  • Meta's confession got a source and a number. No longer a rumor, Zuckerberg himself said agent development stalled despite $145B, and the recipient tape kept bleeding (Micron down another ~14% intraday).

  • The physical-scarcity counter got its most credible voice in PJM's Asim Haque ($28 to $270/MW-day; the 5–7-year build gap), reinforced by AWS's +20% GPU price. Meanwhile the soft June jobs number now shows in Fed futures, year-end hike odds fell to ~45–47% from ~60–70%.

Numbers to track: aggregate 2026 capex still ~$700B+ (Meta ~$145B); data-center build cost $65–75B/GW; the ERCOT Batch Zero firm-load read (August) remains the next hard number.