Newsletter · · Ashutosh Agarwal

Nvidia Gives Back a Trillion as the AI Trade Migrates to Memory and Power - The AI Capex Tracker - Week of July 9, 2026

AI capex and semiconductors newsletter for July 9, 2026. Nvidia has shed roughly $1 trillion of market value in under two months as the AI trade migrates from GPUs toward memory and power, with SK Hynix's $28B US listing and an Apple-Broadcom chip commitment in focus.

The AI Capex Tracker

Week of July 9, 2026: Nvidia Gives Back a Trillion as the AI Trade Migrates to Memory and Power

TL;DR

  • Nvidia has shed roughly $1 trillion of market value in under two months and now trades at its cheapest since before the AI boom. BI's Mandeep Singh: profit-taking plus a custom-silicon overhang, every hyperscaler building its own inference chip is "a sign that there will be more competition for Nvidia down the line." (Bloomberg Intelligence, Jul 8)
  • The trade lost its north star. The Silicon Data token-price index is "down almost 20% from its May high" even as hyperscalers spend "94% of operating cash flow on CapEx," so MAG7 is +1.1% YTD vs. the S&P +10%. The AI trade "hasn't really ended, it's just migrating" to memory and power. (InvestTalk, Jul 8)
  • Where the money is going lists this week: SK Hynix's ~$28B US share sale, with Situational Awareness and Baillie Gifford anchoring up to $7B. HBM leader, ~$1.1T cap, yet 7x forward "because memory is brutally cyclical." (TBPN, Jul 8)

What's new

Yesterday the tape shrugged at rising capex. Today it started selling. Ranked by dollar P&L impact:

1. Nvidia round-tripped a trillion, and is now the "cheap" AI name. Stock is "at its cheapest since before the AI boom after losing roughly $1 trillion in market value over less than two months." Singh reads it as risk-off "indiscriminate selling" plus a structural worry: Amazon, Microsoft, Meta, OpenAI and Anthropic are all chasing their own inference silicon, inspired by Google's TPU. Nothing competes inside six months, "the timelines are already 2 years out," "but the market is anticipatory." A report that China will let its top firms buy the lower-grade H200 (Hopper) nudged the stock back green intraday. Training stays Nvidia's; inference is the fight. (Bloomberg Intelligence, Jul 8)

2. The monetization "north star" cracked. The token-price index, "the cleanest real-world measure of AI monetization that exists," is down ~20% from its May high (DeepSeek's 75% cut, open-source pressure), while infrastructure cost rises. Demand isn't the issue: GPUs and HBM are "sold out through all of this year… no real relief until 2028." The tell is behavioral: 94% of operating cash flow now goes to capex, "stopping their buybacks," and "the market is no longer willing to pay up for the spending without seeing the return." (InvestTalk, Jul 8)

3. Equity is bleeding, but credit is wide open: that divergence is the signal. Amazon's latest $25B deal "reads like it was received terribly… I don't think that was the case." It has issued "almost $100 billion of debt this year across currencies" at ~10–15bps of extra concession, "a rounding error" on 30-year money at +110bps. "Amazon could raise as much money as they want whenever they want," and SpaceX and Oracle paper still trades with a "deep bid." The build is funded; the multiple is what's under attack. (Bloomberg Intelligence, Jul 8)

4. SK Hynix's blockbuster US listing hits a tough tape, and gets bought anyway. ~$28B NASDAQ sale this week (Seoul shares +750% in a year); 2025 revenue +47% to $63B, profit doubled to $28B, Q1 revenue tripled YoY. Singh on the "tough tape to sell that": "investors who believe in this memory cycle… will want a piece of it." (TBPN, Jul 8; Squawk on the Street, Jul 8)

5. Apple-Broadcom, >$30B, US-built. A chip commitment "expected to exceed $30 billion," with the real prize the chatter (via Mark Gurman) of Apple-Broadcom AI server chips. AVGO +3.5% on the day, +41% over twelve months. (Bloomberg Intelligence, Jul 8)

The debate

The ~$700B ("three quarters of a trillion") 2026 capex number didn't move. The fight is again about what a higher number is worth, and the bears landed the cleaner punch.

Bull steel-man: migration, not the end. Hardware is genuinely scarce (GPUs/HBM sold out into 2028); memory has "the best pricing power" of any sector, with customers rushing to lock in, "I don't see a reason why you would want to sell it"; and credit stays wide open, so the build is funded. (Bloomberg Intelligence, Jul 8; InvestTalk, Jul 8)

Bear steel-man: the returns aren't showing up. Token pricing down 20% while capex eats 94% of cash flow; the five hyperscalers' debt jumped "$228 billion in the six months to March… nearly five times more than any previous quarterly increase," drawing central-bank warnings that "AI projects may not make enough money to repay the debt," with an explicit 1873-railway-default analog (InvestTalk, Jul 8). Cory Doctorow on the unit economics: AI firms are "selling $100 bills for a dollar apiece," and Anthropic's attempt to go "from $1 to $5" "lost most of its customers" (TechStuff, Jul 8). And the circular-financing tell desks are chewing on: Google's $920M/month ($11B/yr) compute deal with SpaceX/xAI, struck days before SpaceX's IPO at "twice the going rate" (Rebel Capitalist News, Jul 8).

"The market is no longer willing to pay up for the spending without seeing the return." (InvestTalk, Jul 8)

Sell signals to watch: the token index breaking further below its December base; memory gross margins rolling off the ~80% peak; a hyperscaler cutting forward capex (none has); credit concessions widening from a "rounding error" to something real; "excess compute" resale deals multiplying (Google-SpaceX, xAI-Anthropic).

Stocks in play

NVDA. Bull: still the training platform; China H200 access reopens a market; the "cheapest since the boom" tag is itself a setup. Bear: ~$1T of value gone in two months, and every major customer is taping out its own inference chip. Next: Q2 in August. (Bloomberg Intelligence, Jul 8)

AVGO. Bull: Apple deal ">$30B," US-built, with AI-server-chip optionality, the arms dealer keeps winning designs. Bear: the same in-housing wave eventually points at merchant ASICs too. Next: Apple AI-server-chip confirmation. (Bloomberg Intelligence, Jul 8)

AMD. Quiet on the tape this cycle, no fresh operator color. Next: AMD Advancing AI Day, July 2026 (MI450X/Helios).

MSFT. Bull: balance sheet to fund anything; a credible inference chip would cut "the Nvidia tax." Bear: the group's relative loser: capex "closer to $190 billion by calendar 2026… about 54% of revenue" and rising, valuation down "almost 39%, 40%" over twelve months to <20x earnings, and "still behind in that chip war." Next: FY26 Q4 capex, late July. (Bloomberg Intelligence, Jul 8)

GOOGL. Bull: the one hyperscaler analysts wave through on capex "because they have their own chips," TPU is the template the industry is copying. Bear: its own ~$11B/yr SpaceX/xAI compute commitment is Exhibit A in the circular-financing critique. Next: July capex guide. (Bloomberg Intelligence, Jul 8; Rebel Capitalist News, Jul 8)

AMZN. Bull: unlimited funding: "$100 billion of debt this year," 30-yr at just +110bps, plus its own Trainium silicon. Bear: leaning ever harder on debt to fund capex growing faster than revenue. Next: late-July earnings. (Bloomberg Intelligence, Jul 8)

META. Bull: shipping GPU-hungry consumer product, the new Muse Image/Video models give the spend a demand outlet. Bear: still the FCF-risk name, building its own inference chips like everyone else. Next: late-July earnings, the referee print. (TBPN, Jul 8)

Read-throughs

  • Power / grid: the Texas trade sharpened. Even haircutting large-load requests 50–70%, ERCOT still faces "over 100 gigawatts of demand by 2030," with the reserve margin projected negative in 2029–30 and top large-load figures now "about 445 gigawatts." The Texas Energy Fund has committed "something like $4 billion" of its $7.5B to dispatchable-gas projects run "mainly by the incumbents: NRG, Vistra, Calpine, Constellation," alongside a 765kV Permian build. Actionable: regulated Texas utilities plus dispatchable-gas incumbents own the buildable pipeline. (Energy Capital Podcast, Jul 8) A pundit basket adds nuclear (Constellation, Vistra), uranium (Cameco, Centrus), gas midstream (EQT, Williams, Kinder Morgan) and grid gear (Eaton, Vertiv, GE Vernova). (Follow the Money, Jul 8)
  • Memory: the rotation's landing zone. SK Hynix's ~$28B listing is the cleanest institutional access valve into HBM; Singh calls memory's record margins the wrong thing to sell into an AI-wide flush. The counter is cyclicality: 7x forward exists for a reason. Respect it both ways. (TBPN, Jul 8)
  • The chip-war read. Singh's throwaway is the durable one: "sometimes half of the CapEx goes to Nvidia." Hyperscalers with a working chip (GOOGL, AMZN) escape that tax; the one that can't yet (MSFT) wears it. That spread is the multi-year trade. (Bloomberg Intelligence, Jul 8)

What changed vs last issue

Yesterday (Jul 8, "Capex forecasts leap to 37%. The tape shrugs.") the puzzle was why rising forecasts weren't lifting stocks. Today the shrug became a sell, and the numbers moved:

  • Nvidia's drawdown got a headline number: ~$1 trillion gone in under two months, valuation back to pre-AI-boom levels. "Capex-up no longer pays" hardened into "AI-infrastructure-down, indiscriminately."
  • The rotation is now quantified: MAG7 +1.1% YTD vs. S&P +10% / NASDAQ 100 +18%, money has left the spenders for memory and power.
  • A monetization crack appeared: the token-expenditure index is -20% from its May high, a cleaner bear tell than last cycle's Samsung-sold-the-print read.
  • SK Hynix moved from "lists Jul 10" to "prices this week," ~$28B, Situational Awareness + Baillie Gifford up to $7B.
  • New ERCOT anchor: large-load figures up to ~445 GW (vs. last cycle's 144 GW of TTM requests), reserve margin negative 2029–30.
  • Credit stayed open (Amazon "$100B this year," +110bps) even as equity bled, the equity-vs-credit divergence is the cleanest tell of the cycle.

Next prints: SK Hynix pricing this week; late-July hyperscaler earnings (MSFT, AMZN, META) that test whether $190B-scale capex buys any revenue.