Newsletter · · Ashutosh Agarwal

Nvidia Turns Chips Into Collateral as Google Runs Out of Cash - The AI Capex Tracker - Week of August 20, 2026

The AI Capex Tracker for the week of August 20, 2026. Nvidia's $500 billion financing plan wraps GPUs into an SPV-collateralized investable asset class while capping its own risk at 25 percent, Alphabet just printed the first negative free cash flow since its IPO with a record 7 percent kangaroo bond, and SK Hynix unveiled a $29 billion buyback.

The AI Capex Tracker

Week of August 20, 2026: Nvidia Turns Chips Into Collateral as Google Runs Out of Cash

Issue: Thursday, August 20, 2026

TL;DR

  • We finally know how Nvidia's $500 billion financing plan actually works , and the punchline is that the risk now sits with everyone except Nvidia. The plan pools money from Wall Street's biggest names to lend to "NeoClouds" (new companies that build data centers only to rent out AI computing power). Those loans are wrapped in special vehicles backed by two things: the Nvidia chips themselves, and the rental contract from a customer like OpenAI. Nvidia only backstops up to 25% of each deal , a ceiling of roughly $125 billion , so most of the downside is handed off to pension funds, insurers and private-credit firms. Jensen Huang's framing: this is "the first time that technology chips have become an investable asset class." (Facts vs Feelings, Aug 19)
  • Google just did something it has never done since going public: it ran out of cash to fund its own spending. For the first time since its IPO, Alphabet posted a quarter of negative free cash flow , its ~$200 billion capex bill is now bigger than the cash the business throws off. To plug the gap it has raised more than $50 billion of debt and ~$85 billion of new stock this year, its once-4% buyback is gone, and it just paid its highest-ever interest rate , just under 7% , on a long-dated Australian-dollar bond. The stock trades at more than 70 times free cash flow. (The Intrinsic Value Podcast, Aug 19; Bloomberg Tech, Aug 19)
  • The other side of the memory trade is minting so much cash it can't give it back fast enough , SK Hynix unveiled a $29 billion buyback to be finished by mid-November. That is a huge sum on a very short clock, and management pledged to return more than half of free cash flow to shareholders , potentially ~$170 billion over two years. Jim Cramer argued the whole memory group (Micron, SanDisk, Western Digital, SK Hynix, Samsung) could do $175 billion of buybacks "without a problem," and bet Samsung does "a gigantic one" next. The shovels are still out-earning almost everyone. (Bloomberg Tech, Aug 19; Squawk on the Street, Aug 19)

What's new

A note on the window: this is the Thursday issue, sweeping roughly the last 24 hours (Aug 19). Yesterday's issue was about the size of the hidden bill , the Wall Street Journal's ~$3 trillion of off-balance-sheet AI commitments, and Ben Thompson asking "where does the money come from next?" This cycle we got the answer to that question in unusual detail: the plumbing of the money , how Nvidia's $500 billion plan is wired, what it costs Google to borrow now that its own cash has run out, and where the risk is being quietly parked. Ranked by where the dollars land hardest for a book. One caveat: this window skewed toward market-desk shows, macro commentators and value investors, with one genuine operator quote (AMD's Lisa Su) , I've flagged who's an insider and who's an opinion.

1. The blueprint for Nvidia's $500 billion financing plan is out , and it's a machine for pushing the risk off Nvidia's balance sheet onto everyone else's. Facts vs Feelings, Aug 19 , Ryan Detrick and Sonu Varghese of Carson Group (market strategists , analysts, not operators) walk through the mechanics step by step. This is the clearest explanation yet of the deal our last few issues kept referring to as "Nvidia's $500 billion plan."

Here is how it actually works. On August 11, Jensen Huang announced a memorandum of understanding , note: just an MOU, nothing committed yet , with five heavyweights: BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR (as Detrick put it, "about as major as it gets… the 1927 Yankees" , and only Goldman is actually a bank). Together they set up an "independent compute financing platform" to raise $500 billion of third-party money , from sovereign wealth funds, insurers, pensions, private-credit investors. That money lends to "NeoClouds" , a new bit of jargon worth learning: these are non-traditional cloud companies that exist only to build and rent out AI computing power (think CoreWeave), as opposed to Amazon or Microsoft, who run broad cloud businesses. The NeoClouds borrow, then spend the money mostly on Nvidia chips.

The clever part is the collateral. The debt sits in a special-purpose vehicle and is backed by two things at once: the Nvidia GPUs and the customer's rental contract (e.g. OpenAI agreeing to lease the computing power). As Varghese explained, the NeoCloud borrower can even be junk-rated , what matters is who is on the other end of the lease:

"The NeoCloud can be rated as junk. That's okay… but what matters is who's leasing compute capacity? If it's OpenAI, well, maybe that's a problem. If it's Microsoft, that's okay."

Nvidia's own role is deliberately capped. It offers a residual-value guarantee of up to 25% of a deal , a backstop that only kicks in if the NeoCloud goes bankrupt, the chips can't be re-leased, and the hardware can't be sold. On the headline $500 billion that works out to a ceiling of ~$125 billion of Nvidia exposure. Jensen's line: this is "the first time that technology chips have become an investable asset class."

Why it matters: the market likes this, and you can see it in the price of insurance against a Nvidia default. Nvidia's five-year credit-default-swap spread , essentially the cost to insure its bonds , had doubled to 82 basis points by late July after it agreed to backstop $250 billion of OpenAI leases plus $350 billion of chip-purchase financing; it eased back to ~70 basis points after this MOU spread the risk around. The stock fell ~10% from July 22–29, then rallied ~18%. But as Varghese warned, the tail risk "hasn't gone away… it's transferring to the rest of the market… to all of us." The bull read: Nvidia has professionalized its vendor financing and handed underwriting to firms that can say no. The bear read: it has built a system where a NeoCloud failure now radiates into private credit and insurance balance sheets , and Nvidia is, only half-jokingly, "the lender of last resort."

2. Google's cash finally ran out , the clearest hard-numbers confirmation yet that the AI build has outgrown even the richest balance sheets. The Intrinsic Value Podcast, Aug 19 , Daniel Mahncke and Shawn O'Malley (long-term value investors; Google is their largest position, so treat this as a committed bull being honest about the risks).

The number that stopped them: Alphabet reported its first quarter of negative free cash flow since its IPO. Free cash flow is simply the operating cash a business generates minus what it spends on plant and equipment , and Google's capex, running at ~$200 billion a year, is now eating more cash than the business produces. To fund the gap, Google has raised more than $50 billion of long-term debt and issued ~$85 billion of new shares this year, and the ~4% buyback yield that made it look like a "share cannibal" a year ago is gone: "that part of the shareholder return is something we'll no longer see." (This is the same shift Berkshire's recent equity stake was implicitly helping to fund.)

The valuation math is sobering even from a fan. Google now trades at more than 70 times free cash flow, precisely because capex is crushing the denominator. Run the bearish case forward , capex rising toward the $300 billion some estimate for next year (they guess $250–260 billion), split the usual two-thirds servers / one-third chips, layer on depreciation that is already ~$60 billion (about 12% of revenue) , and operating margins that sit at 34–35% today on ~$475 billion of revenue "could go down to as low as 20% in 2030."

The bull's rebuttal is equally concrete, and it is the demand side: Google Cloud growth accelerated from the low-30s to more than 80% year over year, and the backlog (signed-but-not-yet-recognized business) now stands at ~$500 billion, up 375% year over year, with about half expected to convert to revenue within 24 months. The catch: ~40% of that backlog comes from Anthropic , real demand, but concentrated in a single, still-unprofitable customer. Why it matters: this is the bull-and-bear of the whole hyperscaler trade in one company , genuine, accelerating demand versus a business that has quietly transformed from an asset-light software company into a capital-hungry industrial one that can no longer self-fund. And the same story rhymes at Amazon: also guiding ~$200 billion of capex, having raised close to $90 billion of debt, with Q2 free cash flow also turned negative , though its stock rose ~15% because the market judged the AWS demand more kindly. As the hosts put it, $200 billion each is "roughly the size of the economy of Greece."

3. The cost of all this borrowing is now visible , and rising. Google just paid its highest interest rate ever, and the bond market is quietly getting picky. Bloomberg Tech, Aug 19 and Bloomberg Surveillance, Aug 19 (Lindsay Rosner, fixed-income PM at Goldman Sachs Asset Management , a bond-desk professional).

Alphabet sold $3.9 billion of debt in its debut Australian-dollar ("kangaroo") bond, and paid just under 7% on the 20-year piece , the highest rate it has ever paid on long-dated debt, a reminder that even the best credits are now financing AI into multi-decade-high government yields. Rosner put the flow in context: US investment-grade issuance is $1.6 trillion so far this year, up 34%; the hyperscalers alone account for close to $200 billion, and the "total AI complex" more like $400 billion. The market is absorbing it "in stride" , deals are still oversubscribed , but the bonds of these issuers now trade at wider spreads than similar-quality names, and Goldman is deliberately underweight hyperscaler investment-grade paper, choosing to play AI through structured products and high yield instead.

Her most important structural point is about where the money runs out: the regular investment-grade market has a limit, and when it saturates, the next buyer is private credit, which is sitting on ~$4.5 trillion of dry powder. Why it matters: this is Ben Thompson's "funding staircase" from yesterday, now with price tags attached. The equity market and the debt market are two different conversations , as Rosner said, "buying the bonds of these issuers is very different than owning the equity" , and the bond desks are already demanding more to hold the risk while the equity keeps setting records.

4. AMD's CEO gave the single cleanest bull quote of the cycle: the world needs 100 times more computing power. Business Daily, Aug 19 , a BBC World Service explainer that, usefully, carried a genuine operator on tape: AMD chief executive Lisa Su, speaking in Las Vegas.

Su's framing of the demand: "We see the adoption of AI growing to over 5 billion active users… We don't have nearly enough compute for everything that we can possibly do. So to enable AI everywhere, we need to increase the world's compute capacity another hundred times over the next few years." The same segment stacked up the scale numbers now defining the trade: JPMorgan reckons hyperscalers will invest ~$697 billion this year (bigger than Argentina's entire economy, and bigger than Austria and Finland's combined); Amazon has signaled ~$220 billion for the rest of this year on AI infrastructure against $42 billion of quarterly AWS revenue; and Nvidia plus OpenAI unveiled a $105 billion data-center project in Ohio earlier in the week.

But the episode's real value was the backlash , the constraint that keeps recurring in this trade. Australia's Prime Minister Anthony Albanese announced mandatory standards requiring new large data centers to "underwrite new power supply… pay their full share of grid connection… and put at least as much energy into our grid as they take out of it," plus minimize water use. Ireland's data centers already consumed 23% of all metered electricity in 2025. Why it matters: Su's "100x" is the bull case in one breath , but the political and physical limits (power, water, grid connection, local opposition) are now hardening into law, which is exactly what keeps pushing the build behind the meter and off the public grid.

5. The bears sharpened the "circular financing" critique into a specific, dot-com-era warning , and pointed at the depreciation trick underneath it all. Money Tree Investing, Aug 19 (Phil Weiss of Apprise Wealth Management and Kirk Chisholm , value-oriented advisors) and DHUnplugged, Aug 19 (market commentators).

Weiss put a name to the structure everyone is circling: it is vendor financing, the same maneuver that blew up in the dot-com bust. "Like Cisco and Lucent , which is no longer in existence because it went bankrupt… these companies were financing the purchase of their product. And then when their customer couldn't pay… the items got returned to them." His broader tell: Google issuing stock at all is a warning, because "you issue stock because your company is really, really overvalued" , you would rather borrow if the returns were as certain as claimed. DHUnplugged zeroed in on the accounting fault line: because these chips depreciate fast and Nvidia controls the depreciation rate (it can make yesterday's chip worth less simply by launching a better one), the financiers lending against that collateral have complained , so Nvidia is now reportedly slow-walking its release cadence to protect the collateral value. Their memorable frame for the "compute as collateral" pitch: a token sells for about "$1 per million tokens… no matter what the underlying hardware is," and you can still profit on six-to-eight-year-old chips because electricity and maintenance run in "sub-pennies per million tokens" , unless someone leapfrogs Nvidia, at which point "this entire model collapses." Why it matters: the bear case has moved from "the spending is too big" to "the spending is being financed with a structure that has failed before, resting on an asset whose value the seller can change at will."

The debate

Steel-manning both sides of the hyperscaler capex thesis (~$700 billion in 2026 , JPMorgan's latest tally is $697 billion , heading toward $1 trillion-plus in 2027). Yesterday the fight was about the size of the off-balance-sheet bill ($3 trillion) and whether anyone could point to what closes the gap. This cycle we got the mechanism and the price: the money is coming from private credit and third-party investors, routed through Nvidia's $500 billion platform and the hyperscalers' record bond issuance , and the cost of it is climbing. The narrowed question: does that machine hold together long enough for free cash flow to come back?

Bull , the demand is real and accelerating, the financing has been professionalized, and the cash machines (memory) are printing. The single best demand data point this cycle is operator, not pundit: AMD's Lisa Su says the world needs 100x more compute for 5 billion users (Business Daily, Aug 19). Google's cloud backlog of ~$500 billion, up 375% year over year with half converting inside two years, is not a number on paper , it is signed business (The Intrinsic Value Podcast, Aug 19). Nvidia's financing MOU actually reduces single-company risk by handing underwriting to Goldman, KKR, Blackstone, Brookfield and BlackRock , and the credit market rewarded it by narrowing Nvidia's default insurance from 82 to ~70 basis points (Facts vs Feelings, Aug 19). And the bond market is absorbing $400 billion of AI-complex issuance "in stride," with deals still oversubscribed and $4.5 trillion of private-credit dry powder waiting behind it (Bloomberg Surveillance, Aug 19). Meanwhile the memory suppliers are so flush they are handing back tens of billions , SK Hynix's $29 billion buyback is the tell (Bloomberg Tech, Aug 19).

Bear , the richest company in tech just ran out of cash, borrowing costs are rising, and the financing structure is one that has failed before. Google , the best capital allocator in the group , posted its first negative free cash flow since IPO, is funding itself with $50 billion of debt and $85 billion of stock, killed its buyback, and paid its highest-ever ~7% rate on a long bond; Amazon's free cash flow went negative too (The Intrinsic Value Podcast, Aug 19; Bloomberg Tech, Aug 19). The "compute as collateral" pitch rests on an asset that depreciates fast and whose value Nvidia itself controls, wrapped in vendor-financing structures that bankrupted Cisco's and Lucent's customers in the last bubble (DHUnplugged, Aug 19; Money Tree Investing, Aug 19). And the tail risk Nvidia "took off its balance sheet" did not vanish , it moved to the pensions, insurers and private-credit funds now holding NeoCloud paper (Facts vs Feelings, Aug 19).

The synthesis. The two camps still agree the demand is real; they now disagree about the financing, and this cycle both sides got harder evidence. The bull's demand proof got better (Su's 100x, Google's $500 billion backlog). The bear's funding worry got concrete (Google's negative FCF, the ~7% kangaroo bond, wider hyperscaler spreads, and a named dot-com analog). What is genuinely new is that the abstract "where does the money come from?" now has a diagram , NeoClouds, SPVs, GPU-plus-lease collateral, a 25% Nvidia backstop , and a spreading credit signal (Nvidia, Meta, Alphabet and especially Oracle all seeing default-insurance costs rise off the lows). The debate narrows to timing and to the referee: does Nvidia's late-August quarter show demand it is not itself financing , and does free cash flow start returning , before a NeoCloud, an SPV, or a private-credit mark cracks?

Sell signals to watch: - A NeoCloud or SPV wobble. The new structure , junk-rated NeoCloud, GPU-and-lease collateral, Nvidia's 25% backstop , is the fault line; the first repriced compute-SPV or a NeoCloud that can't re-lease its chips is the first crack (Facts vs Feelings, Aug 19). - Credit-default-swap spreads widening again. Nvidia (82→70bp), Meta, Alphabet and Oracle are all off their lows; a renewed move up , especially at Oracle , is the bond market pricing real default risk (Facts vs Feelings, Aug 19). - Financing costs climbing further. Google's ~7% kangaroo bond was a record; watch for a hyperscaler forced to pay up again, or the investment-grade market saturating and pushing paper into private credit at wider spreads (Bloomberg Tech, Aug 19; Bloomberg Surveillance, Aug 19). - A regulatory or grid-connection defeat. Australia-style mandatory standards, Pennsylvania's scrutiny (100 proposed data centers, "most probably never get built"), and local opposition all slow the grid-tied build (Business Daily, Aug 19; Squawk on the Street, Aug 19).

Stocks in play

NVDA. Bull: its $500 billion financing platform is now a professionalized machine that turns chips into an "investable asset class," offloads most of the risk to outside capital (Nvidia caps its exposure at ~25% per deal, ~$125 billion total), and the credit market applauded , default insurance narrowed from 82 to ~70 basis points; demand backdrop remains overwhelming (Lisa Su's "100x more compute") (Facts vs Feelings, Aug 19; Business Daily, Aug 19). Bear: it is increasingly "the lender of last resort" creating its own demand via vendor financing; its collateral depreciates fast and it is reportedly slow-walking new chips to protect that collateral; and the hyperscalers keep diversifying away from it (see Google's new Marvell deal) (DHUnplugged, Aug 19; Bloomberg Tech, Aug 19). Next catalyst: its quarter in late August (~Aug 26–27) , the demand referee for the whole complex and the test of whether growth is real or self-financed.

AVGO. Bull: it remains the lead designer of Google's TPU and, per Bloomberg's semiconductor desk, "did not lose the contract" , the new Marvell deal augments rather than replaces the relationship, and Broadcom still sits in the two legs the market is chasing (custom silicon and optics) (Bloomberg Tech, Aug 19). Bear: Google formally bringing in a second custom-silicon partner (Marvell) signals that even the crown-jewel ASIC relationships are being multi-sourced , a commoditization risk , and Broadcom shares traded down on the news (Bloomberg Tech, Aug 19; Squawk on the Street, Aug 19). Next catalyst: hyperscaler ASIC ramps and read-through from Nvidia's late-August quarter.

AMD. Bull: CEO Lisa Su delivered the cycle's cleanest demand statement , 5 billion AI users and a need to "increase the world's compute capacity another hundred times over the next few years" , reaffirming a market far bigger than current supply (Business Daily, Aug 19). Bear: no fresh MI450X/Helios product catalyst surfaced this cycle, and if Nvidia is the fungible default while hyperscalers build their own silicon (now with Marvell as well as Broadcom), AMD is squeezed from both sides. Next catalyst: its second-half data-center step-up; Nvidia's late-August read-through.

MSFT. Bull: it is quietly the linchpin that makes the new financing work , in the NeoCloud/SPV structure, a Microsoft lease is the "good" collateral that lets junk-rated borrowers raise money ("If it's Microsoft, that's okay"), a testament to its balance-sheet strength (Facts vs Feelings, Aug 19). Bear: it remains inside the aggregate off-balance-sheet tally, and no fresh Microsoft-specific print appeared this cycle , the open question is still whether its cloud demand converts capex to cash. Next catalyst: evidence cloud demand broadens enough to keep free cash flow positive as capex climbs.

GOOGL. Bull: cloud growth accelerated to >80% year over year with a ~$500 billion backlog (+375%), roughly half converting within two years , genuine, signed demand from the company that owns the whole stack (models, TPUs, distribution, power); Berkshire's equity stake helps fund it (The Intrinsic Value Podcast, Aug 19). Bear: first negative free cash flow since IPO, ~$200 billion capex, >$50 billion new debt and ~$85 billion new equity this year, buyback gone, a record ~7% kangaroo-bond rate, >70x free cash flow, margins that could fall toward 20% by 2030, and ~40% of the backlog concentrated in unprofitable Anthropic (The Intrinsic Value Podcast, Aug 19; Bloomberg Tech, Aug 19). Next catalyst: whether cloud growth converts to real free cash flow rather than more debt and equity.

AMZN. Bull: AWS is still compounding , ~$42 billion of quarterly revenue and management guiding ~$220 billion for the rest of the year on AI infrastructure , and the market rewarded the print with a ~15% move, a friendlier reception than Google got (Business Daily, Aug 19; The Intrinsic Value Podcast, Aug 19). Bear: Q2 free cash flow turned negative, it has raised close to $90 billion of debt this year, and ~$200 billion of capex (the size of Greece's economy) has to convert to durable third-party demand, not just internal AI spend (The Intrinsic Value Podcast, Aug 19). Next catalyst: evidence AWS AI demand is broad and durable.

META. Bull: it retains the most ad-funded optionality of the group to absorb its capex line. Bear: the bond market is starting to notice , Meta's credit-default-swap spreads have widened off their lows alongside Alphabet's and Oracle's , and it remains the face of the off-balance-sheet build (the Hyperion campus) (Facts vs Feelings, Aug 19). Next catalyst: whether ad-revenue growth keeps outrunning capex , otherwise quiet on company-specific news this cycle beyond the widening spreads.

Read-throughs

  • Memory (Micron, SK Hynix, SanDisk, Samsung, Western Digital) , the cash machines of the build, and they can't hand it back fast enough. SK Hynix announced a $29 billion buyback to be completed by mid-November , an unusually fast clock , and pledged to return more than half of free cash flow, potentially ~$170 billion over two years, after its shares fell more than 50% in two months; it is also arbitraging its own listings, having sold higher-multiple US-listed shares and buying back cheaper Korean shares (Bloomberg Tech, Aug 19). Cramer's read: the memory group could do $175 billion of buybacks "without a problem," and he expects Samsung to do "a gigantic one" next (Squawk on the Street, Aug 19). The actionable point: while the hyperscalers burn cash, the memory suppliers are generating so much that shareholder returns , not capex , are their problem. Structurally long, cash-rich, and now returning capital aggressively.
  • Optics/networking (Marvell, Coherent, Astera Labs) , the custom-silicon and interconnect leg got a marquee deal. Google gave Marvell warrants to buy up to $12.2 billion of Marvell stock (58.97 million shares at an exercise price of $206.58, already in the money with the stock near $241) as part of a deeper multi-chip design partnership , crucially including an AI accelerator, the Nvidia-style chip, plus memory augmentation for the TPU. Marvell's market value now tops $200 billion; Cramer flagged the eye-watering gap between its ~$120 billion valuation and ~$11 billion of revenue (Bloomberg Tech, Aug 19; Squawk on the Street, Aug 19). The read-through cuts two ways: bullish for the custom-silicon/optics theme and for Marvell specifically, but a warning shot for Nvidia and a commoditization signal for the ASIC designers, since it shows hyperscalers will happily multi-source their most strategic chips.
  • Power/thermal (VRT, ETN) and bridge generation , the bottleneck is the wire, not the megawatt. The founder of Critical Loop, a new industrial-power startup, made the point sharply: it is now "faster today to build a factory than… to run power to it," and the binding constraint is the transmission and distribution system, with grid-connection waits "two to three years… I've heard of like crazy things like 15 years" , solvable near-term only with on-site storage and software that taps off-peak grid capacity to serve peaks (Supercool, Aug 19). Layer on the tightening politics , Australia's mandatory grid-cost-and-water standards, Pennsylvania's executive order on data-center non-disclosure agreements and its ~100 proposed centers that "most probably never get built" (Business Daily, Aug 19; Squawk on the Street, Aug 19) , and the demand for transformers, switchgear, storage and thermal gear stays captive and sold-forward. No direct Vertiv (VRT) or Eaton (ETN) commentary by name this cycle , a coverage gap, not a signal.
  • Financing/private credit , the newly visible systemic read. The single most actionable structural development is not a chip or a turbine , it is that the AI build is now being financed through NeoClouds and SPVs, collateralized by GPUs and lease contracts, backstopped only up to 25% by Nvidia, with the residual tail risk sitting in private credit, insurance and pension books (Facts vs Feelings, Aug 19). With investment-grade issuance already at $1.6 trillion (+34%) and private credit's $4.5 trillion of dry powder queued as the next buyer (Bloomberg Surveillance, Aug 19), the first sign of stress will almost certainly show up in a repriced compute-SPV, a widening credit-default-swap spread (watch Oracle), or a private-credit mark , not in a hyperscaler's headline earnings.

What changed vs last issue

Yesterday's issue (Wednesday, Aug 19, "Big Tech's AI bill: $3 trillion off the books.") had the bear case sizing the hidden liabilities , the Wall Street Journal's ~$3 trillion of off-balance-sheet commitments and Ben Thompson's "where does the money come from next?" This cycle moved from the size of the bill to the plumbing and the price of the financing:

  • Nvidia's $500 billion plan got its blueprint. New this cycle: the mechanism , an MOU (Aug 11) with BlackRock, Blackstone, Brookfield, Goldman and KKR to raise $500 billion of third-party capital that lends to "NeoClouds," wrapped in SPVs collateralized by GPUs plus a customer lease, with Nvidia offering a residual-value backstop capped at 25% (~$125 billion). Jensen: chips are now "an investable asset class." This is the concrete answer to yesterday's open "funding staircase" question.
  • The funding-gap thesis got hard numbers at Google. New: Alphabet's first negative free cash flow since IPO; ~$200 billion capex; >$50 billion new debt and ~$85 billion new equity this year; buyback gone; >70x free cash flow; margins possibly toward 20% by 2030; Cloud +80% YoY; ~$500 billion backlog (+375%, ~40% Anthropic). Amazon's FCF also negative, ~$90 billion of new debt.
  • The cost of borrowing became visible and is rising. New: Alphabet's record ~7% kangaroo bond ($3.9 billion); investment-grade issuance $1.6 trillion (+34%), hyperscalers ~$200 billion, AI complex ~$400 billion; hyperscaler spreads trading wider than peers; Goldman underweight hyperscaler IG; private credit's $4.5 trillion of dry powder as the next buyer.
  • A named custom-silicon deal reshaped the chip read. New: Google's Marvell partnership , warrants for up to $12.2 billion of Marvell stock plus an AI accelerator , the first clear sign a hyperscaler will multi-source even its flagship TPU (Broadcom "did not lose the contract," but now shares Google).
  • The credit market started pricing the risk. New: Nvidia's default-insurance spread (82→70bp after the MOU), with Meta, Alphabet and especially Oracle also off their lows , the bond market's first quantified nervousness.
  • The bear case got a specific historical analog. New: the "vendor financing" label and the Cisco/Lucent/Nortel dot-com precedent, plus the depreciation wrinkle , Nvidia controls how fast its collateral loses value and is now "slow-walking" new chips to protect it.
  • An operator demand quote arrived. New: AMD's Lisa Su , 5 billion AI users, "increase the world's compute capacity another hundred times" , a rare insider bull data point in a cycle otherwise dominated by desks and commentators.
  • Fresh, concrete data points: Nvidia MOU with BlackRock/Blackstone/Brookfield/Goldman/KKR to raise $500B; NeoCloud/SPV collateral = GPUs + lease; Nvidia backstop ≤25% (~$125B ceiling); Nvidia CDS 82bp late July → ~70bp; Nvidia backstopping $250B OpenAI leases + $350B chip financing; Nvidia-OpenAI $105B Ohio project; Google first negative FCF since IPO; Google ~$200B capex 2026 ($250–300B est. 2027); >$50B new debt + ~$85B new equity; >70x P/FCF; D&A ~$60B (~12% of revenue); revenue ~$475B, op margin 34–35%; Cloud +80% YoY; backlog ~$500B (+375%, ~40% Anthropic); Alphabet kangaroo bond $3.9B at ~7% (record); Amazon ~$200B capex, ~$90B new debt, FCF negative, AWS ~$42B/qtr, $220B rest-of-year AI infra; JPMorgan hyperscaler capex $697B in 2026; SK Hynix $29B buyback by mid-November, >50% FCF payout ($170B/2yr); memory group $175B buyback capacity; Marvell warrants $12.2B (58.97M shares @ $206.58, stock ~$241), Marvell market cap >$200B; IG issuance $1.6T (+34%), hyperscalers ~$200B, AI complex ~$400B; private credit $4.5T dry powder; 30-year yields >5.25–5.30%, real rates >3%; Ireland data centers 23% of metered electricity (2025); grid-connection waits 2–3 years (up to 15); PA ~100 proposed data centers.
  • Quiet vs yesterday: no fresh Vertiv (VRT) or Eaton (ETN) commentary by name; no new MI450X/Helios (AMD) or Broadcom-specific product print beyond the Marvell read-through; no Microsoft-specific print beyond its role as "good collateral"; no update on Meta's Hyperion beyond widening CDS spreads.
  • Still open: whether free cash flow returns before the financing strains; whether Nvidia's late-August quarter shows demand it is not itself financing; whether a NeoCloud, an SPV, or a private-credit mark cracks first; whether credit-default-swap spreads (especially Oracle's) resume widening; and whether regulators and grids keep slowing the build.

Next catalysts to watch: Nvidia's quarter in late August (~Aug 26–27) , the demand referee for the whole complex; any repriced compute-SPV, NeoCloud stumble, or private-credit mark as the first sign the financing machine is straining; credit-default-swap spreads on Nvidia, Meta, Alphabet and Oracle; Jackson Hole and Chair Powell's reaction function into the September Fed meeting; and any further record-rate bond deal from a hyperscaler as the cost of the build keeps climbing.