Newsletter · · Ashutosh Agarwal

The AI Capex Tracker - Week of June 18, 2026: Google Raises $85B and the Buyback Era Is Over

How the hyperscalers stopped self-funding the AI buildout: Google raises $85B of equity, Nvidia issues $28B of bonds, and capex-to-operating-income hits 135% across the top 5 hyperscalers, for the week of June 18, 2026.

The AI Capex Tracker

Week of June 18, 2026: Google Raises $85B and the Buyback Era Is Over

Issue: Thursday, June 18, 2026

TL;DR

  • The hyperscalers have stopped self-funding. Google is raising $85B of equity, its first equity raise since the 2004 IPO, to cover 2026 AI capex of $180–190B, up from $80B in 2025. Meta and Microsoft are reportedly next. The single biggest regime change of the cycle. (The Real Eisman Playbook, Jun 12)

  • Capex-to-operating-income across the top 5 hyperscalers has hit 135%. They are now spending more than they earn, and the price increases are flowing straight to memory makers as an "income transfer," not new capacity. Aggregate 2026 capex is closing on ~$1T. (Monetary Matters, Jun 15; Making Sense / J.P. Morgan, Jun 12)

  • Memory PTs keep ripping, but the smart money is starting to ring the register. TD Cowen took Micron to $1,500; a PM who has been long since $90 just sold ahead of the Jun 24–25 print, citing 30%-single-day downside if bottlenecks dent the quarter. (Power Lunch, Jun 15; Full Signal, Jun 17)

What's New

1. The funding model flipped. This is the story. (The Real Eisman Playbook, Jun 12, Steve Eisman; Excess Returns, Jun 16, Andy Constan, Damped Spring, macro PM).

Google upsized an $80B raise to $85B, all equity, its first equity issuance since the 2004 IPO. 2025 capex was $80B, funded from cash flow; 2026 will be $180–190B, "too much for its cash flow." Oracle added $20B to its plans (FY27 now $40B) and fell 10% after hours; Supermicro dropped 28% on a $7B raise. Constan's macro: buybacks ran ~2% of GDP in 2023–24, and with Google and Meta canceling repurchases that is "$600, $700 billion of shift from share reduction to no share reduction." His one-liner: "As long as CapEx is running at a trillion a year and growing... they're gonna have to keep selling." The tailwind that carried this tape for three years is now a supply headwind.

2. 135%: the hyperscalers are spending more than they earn. (Monetary Matters, Jun 15, David Woo, strategist).

"If you actually look at CapEx relative to operating income... for the 5 hyperscalers, the CapEx to operating income is now 135%." (David Woo)

Woo's sharper point: most of the increase is price, not capacity, the same number of data centers at higher prices, some with delayed completion. The result is "an income transfer from Microsoft to basically Micron," flattering aggregate index earnings since hyperscalers book only a sliver as cost while memory makers book 100% as revenue. J.P. Morgan's Eloise Goulder sized it: MAG7 capex ~$700B this year, >2% of US GDP (vs ~$100B in 2021), and "nearly 100% of operating cash flow from the MAG7 goes on this capex." (Making Sense, Jun 12)

3. Broadcom co-signed Anthropic's chips: a $35B residual-value bet. (The Information's TITV, Jun 16, Dakin Campbell, The Information).

Broadcom anchored a $35B financing with Apollo/Blackstone, 1 GW for Anthropic now, 20+ GW through 2028, by providing a backstop on the chips' residual value, stepping in if they are worth less than expected at lease-end. With $65B debt and <$20B cash, that dents Broadcom's credit profile, but "they were at risk of being left behind" versus Nvidia's playbook. Same show: The Information's math found Nvidia's inference share grew from 66% to 74% over the past year, inference is now ~60% of AI workloads, headed to 80%.

4. Bernstein's Rasgon: still taking capex up, watch CoWoS, not the headline. (TechSurge, Jun 16, Stacy Rasgon, Bernstein, specialist).

The hyperscalers "all reported on top of each other on Wednesday... and still take CapEx up." Rasgon's tell: "by the time you see [capex] go down, it's probably too late," so watch CoWoS packaging reservations and wafer orders in Asia, and whether cloud revenue is accelerating. He sized Broadcom's AI line at $100B+ next year and ASICs at mid-teens % of revenue today, potentially 25–30% of a bigger pie. On GPU-vs-ASIC: wrong question. "If [the opportunity] is big, they both thrive. If it's not, they're both screwed."

5. FERC's large-load ruling lands any day: "buckle up." (POLITICO Energy, Jun 12, FERC Chair Laura Swett).

Swett confirmed FERC delivers its large-load interconnection proposal by end of June and held the line on what matters for power-stack P&L: ratepayers will not subsidize hyperscaler connections, "subsidizing anyone's connection that you don't benefit from is not part of that answer." It reprices who pays across the entire grid build-out.

The Debate

Bull steel-man: the spend is rational and the bottlenecks are real. Demand is genuine: Anthropic's annualized revenue went ~$9B (Dec) to ~$14B (Jan) to ~$30B (April) (TechSurge, Jun 16). Memory has re-rated on long-term agreements with floor margins, "this cycle is definitely much bigger," HBM sold out into 2028 (Power Lunch, Jun 15). Perplexity's Aravind Srinivas: whatever is the bottleneck commands the price, memory up 5x in COGS, CPUs now a fresh bottleneck as agent loops run on them (20VC, Jun 15).

Bear steel-man: late-cycle, commoditizing, and now diluting you. Woo: memory pricing power "is going to be very short-lived" once 2027 capacity lands and China ramps DRAM/NAND (Monetary Matters, Jun 15). Eisman: "Trillions are being spent for a product with no moats and prices are already being cut" (The Real Eisman Playbook, Jun 12). Forward Guidance: a $6T company cannot trade at a growth multiple, and the only way to stop the share-price underperformance is to curtail "your drunken spending" (Forward Guidance, Jun 12). Rasmussen likens it to fracking/shipping, overbuild, then demand goes "from infinite to zero" (Macro Hive, Jun 12).

Sell signals to watch: a hyperscaler curtailing 2-year capex guidance; CoWoS/wafer reservations rolling over; cloud revenue decelerating while capex rises; memory -20% from highs; a beat-and-fade on the Jun 24–25 Micron print; equity issuance failing to clear (SoftBank already could not get a $6B loan against its OpenAI stake).

Stocks in Play

NVDA. Bull: inference share rising 66% to 74%; Rubin ramps to full data-center capacity in 2027. Bear: even Nvidia is now issuing debt ($28B of bonds, first IG deal since 2021), and its bigger buyback is offset by matching RSU grants, net shares flat. Next: Rubin ramp; Q2 print. (Excess Returns, Jun 16; The Information's TITV, Jun 16)

AVGO. Bull: $100B+ AI revenue guided for next year; the $35B Anthropic deal keeps it in the financing game. Bear: the backstop is residual-value risk on a $65B-debt/<$20B-cash balance sheet, and a "negative announcement from Broadcom a couple weeks ago" already hit the tape. Next: the 20 GW Apollo/Blackstone platform. (The Information's TITV, Jun 16; Power Lunch, Jun 15)

AMD. Bull: the fresh CPU bottleneck, "agents are using CPUs more than humans," puts AMD back in the demand story, plus rising inference share. Bear: roadmap quiet on the tape; MI450X/Helios milestones unconfirmed. Next: AMD Advancing AI Day, July 2026. (20VC, Jun 15)

MSFT. Bull: still the only hyperscaler repurchasing shares, "not coming this month" for an equity raise. Bear: swept into the 135% capex/operating-income group; Visser sees multiple compression for spenders turning into "hardware companies." Next: FY26 Q4 capex commentary at July earnings. (Excess Returns, Jun 16; Full Signal, Jun 17)

GOOGL. Bull: best Mag-7 price-performer and the only one in the top-10 S&P contributors; $85B raised to fund a $180–190B build. Bear: the issuance is itself the overhang, and "no moats / prices being cut" hits Gemini economics directly. Next: cadence of further issuance; July capex guide. (The Real Eisman Playbook, Jun 12; Power Lunch, Jun 15)

META. Bull: still funding from relative strength versus the pure model labs. Bear: "not far behind" Google on an equity raise, buyback being cut, and the Street questions capex that buys only a "6–8% bump" in ad accuracy. Next: a raise announcement; July earnings. (Monetary Matters, Jun 15; 20VC, Jun 15)

AMZN. Bull: Trainium/Graviton integration; does not repurchase shares, so no buyback to cut. Bear: "they're coming," already issuing corporate bonds, equity next; ad revenue is cyclical if a slowdown hits. Next: issuance watch; July earnings. (Excess Returns, Jun 16)

Read-throughs

  • Memory (MU, SK Hynix, Samsung): TD Cowen's Krish Sankar took Micron to $1,500 on durable LTA-backed earnings; but Jordi Visser, long since ~$90, sold because "the risk of it being down 30% in a day is very high" if July's print reveals shipment bottlenecks. Two-sided into Jun 24–25; structural bear is Chinese DRAM/NAND entry. (Power Lunch, Jun 15; Full Signal, Jun 17)

  • Power & thermal (VRT, ETN): no direct VRT/ETN tape, but the read-through is loud, Joshua Rhodes (Austin Energy commissioner) flags transformers +200%, wires +180%, switchgear 2–3x. Electrical-equipment pricing power is the cleaner long than the data centers themselves. (Renewable Rides, Jun 16)

  • Grid / nuclear (VST, CEG, TLN): the ERCOT queue swelled to 435 GW, 90% data centers, against an all-time ERCOT peak of 85.5 GW ("a bubble... more bubbly and more silly"). PJM is in open crisis: wholesale +76%, capacity costs +400%, Gov. Shapiro warning of "the largest unjust wealth transfer in the history of US energy markets" and threatening a PJM exit. Constellation's TMI restart for Microsoft is the nuclear template; FERC's ruling lands end-June. (Renewable Rides, Jun 16; The Banker Next Door, Jun 16; HilltopTalks, Jun 17)

  • Networking / optics (AVGO, MRVL, Credo): ASIC/networking remains the share-gain story; Credo guided FY27 revenue +80% at 50% adjusted margins. Digestion flag: J.P. Morgan analysis cited this week found >60% of data-center capacity planned for 2027 is not yet under construction, the air-pocket risk if silicon outruns the buildings. (Chip Stock Investor, Jun 12; The Banker Next Door, Jun 16)

What Changed vs Last Issue

Cadence resumed after a gap; last issue on file is June 1 ("Memory at $1T. Bears find their voice. Broadcom Tuesday."). The deltas:

  • The financing pivot is brand new and dominant. June 1 was memory PTs and a loud-but-not-short bear chorus. Now the hyperscalers themselves are issuing: Google $85B equity (first since 2004), Nvidia $28B bonds (first IG since 2021), Oracle +$20B, Supermicro -28% on a $7B raise, buybacks canceled at Google/Meta. Capex/operating-income at 135%.

  • Aggregate capex re-rated up, the $737B framing gave way to ~$1T for 2026, Google alone $180–190B (from $80B in 2025).

  • Broadcom's June 3 print is rear-view; the new datapoint is the $35B Anthropic backstop and $100B+ AI revenue guide, with explicit residual-value risk.

  • Memory: from a parade of PT hikes to first real register-ringing, Visser sold ahead of Jun 24–25. The watch flipped from "how high do PTs go" to "does the quarter survive the bottlenecks."

  • ERCOT queue 410 to 435 GW (90% DC); binding catalyst shifted from the July 15 SB6 cliff to FERC's end-June large-load ruling, with PJM a second front (Shapiro complaint, +400% capacity costs).