Jensen's weekly PR circus finally makes sense: Nvidia was drowning in wafers and RAM
So "fungible, durable and highly rentable" that Nvidia ended up committing 6.3 billion dollars (£4.8 billion) to rent unsold GPU capacity back from its own partners.
For years, watching Jensen Huang show up on a different stage every other week to deliver the same breathless sermon about limitless AI demand started to feel less like a victory lap and more like a nervous tic. After The Wall Street Journal published its post-mortem on Jensen's abrupt resignation on 12 November 2026, following revelations of extramarital affairs including one with a woman suspected of being a Chinese intelligence asset, the non-stop promotional marathon suddenly made complete sense. Behind all the leather-jacket theatre, Nvidia wasn't struggling to keep up with organic customers. It was staring down a wall of its own hardware supply commitments while its biggest buyers quietly looked for the exits.
Sources: The Wall Street Journal via Moomoo | BigGo Finance
The 279 billion dollar avalanche behind the weekly keynotes:
According to The Wall Street Journal, Jensen summoned then-CFO Colette Kress and his senior finance executives to an August meeting to confront a worsening internal mess. Major cloud providers were accelerating plans to move onto alternative chips and cut their dependence on Nvidia, while the enormous stockpile of chip wafers and memory Nvidia had already contracted to buy was barrelling toward the company "like an avalanche". You can see that exact avalanche sitting right inside Nvidia's Q2 fiscal 2027 filings for the period ended 26 July 2026, where supply and capacity purchase commitments more than doubled in a single quarter from 119 billion dollars (around £90 billion) to 279 billion dollars (around £211 billion), driven primarily by memory procurement.
Sources: The Wall Street Journal via Moomoo | Capital & Compute
That context turns Jensen's public sermons into dark comedy. At GTC Berlin in October, the WSJ reported he pressured half a dozen cloud CEOs with a story about a Japanese soup vendor who refused to sell him a second bowl because there was a long queue outside, scolding them that they were not establishing the value of his products. When you are locked into £211 billion of incoming wafers and memory while hyperscalers build custom silicon to escape your fat margins, preaching artificial scarcity stops being swagger and turns into survival.
Sources: The Decoder | Tovima
Nvidia became a bank because real buyers were running out:
If paying customer demand had truly been bottomless, Nvidia would never have needed an internal lending desk handing out credit guarantees just to clear inventory. Yet the WSJ report detailed how Jensen's son-in-law, Nico Caprez, spent months pushing deals that offered upstart cloud providers and data-centre developers financing support so they could afford Nvidia chips. The pipeline ran like this: Nvidia commits 279 billion dollars (around £211 billion) for silicon and RAM > major cloud giants pivot to in-house chips > Nvidia extends equity, loan backstops, and lease guarantees to cash-burning labs and debt-heavy rental startups > those outfits turn straight around and buy Nvidia hardware so Nvidia books the chip revenue.
Sources: The Wall Street Journal via Moomoo | Wccftech | NVIDIA Credit Report
Inside the company, Kress tried to rein in Caprez, questioning how Nvidia was valuing these guarantees and forcing a proposed 250 billion dollar (£189 billion) backstop for OpenAI's Ohio data centre down by roughly half to a capped 105 billion dollar (£79 billion) commitment. Jensen's reaction in that August meeting was to interrupt her presentation, label her "a deer in the headlights", and dismiss her as a mere "reporting CFO". By late October, Nvidia had stacked up 164.5 billion dollars (around £124 billion) in guarantees and procurement commitments, including a 1.5 billion dollar (£1.1 billion) GPU lease-back arrangement with Lambda and a 6.3 billion dollar (£4.8 billion) pledge promising to absorb CoreWeave's unsold capacity through April 2032. Now Jensen is gone, and the "deer in the headlights" is interim CEO left holding the ledger.
Sources: The Decoder | Medium | CoreWeave Credit Risk | NVIDIA Credit Report
I called Oracle the biggest bag holder, and Microsoft just pulled the plug:
While Nvidia pocketed hefty margins upfront before getting dragged into vendor financing, the worst position in this market always belonged to the landlords taking on debt to own physical hardware. On this site, I called Oracle out early as the biggest bag holder of the GenAI bubble precisely because their whole business plan boiled down to borrowing billions to buy depreciating GPUs built for training rather than inference, only for Nvidia's very next chip cycle to turn them into expensive junk. On 6 November 2026, Microsoft obligingly proved it.
Bloomberg reported that Microsoft declined to renew the multi-year deal it signed in 2023 to rent Oracle cloud servers for AI-powered Bing and has spent recent months physically switching that capacity off, because Oracle's older-generation Nvidia chips no longer met Microsoft's specifications. Put simply, the industry shifted from training giant models to inference, which means actually running them for users. Inside three years, inference workloads changed so sharply in memory bandwidth and performance per watt that 2023-vintage training silicon fell off a cliff. On Ornn Exchange's settled-trade index this month, an A100 rents for 1.06 dollars (80 pence) an hour compared to 9.84 dollars (£7.44) for a Blackwell B200, roughly a nine-to-one collapse. Even Jensen joked at GTC 2025 that once Blackwell shipped in volume "you couldn't give Hoppers away", prompting Barclays analyst Ross Sandler to warn that hyperscalers were likely overstating profits by under-accounting for how fast these chips go stale.
Sources: NeoTeo | Winzheng | TokenPost | Business Insider
By renting from Oracle instead of buying those specific machines, Microsoft casually walked away around year three and handed the remaining multi-year depreciation problem back to Oracle, an arrangement Crayhill Capital co-founder Carlos Mendez summed up to the Financial Times as "financing a fax machine the week before someone invented email". Oracle tried to insist demand for older GPUs remained strong, yet ten days before the Microsoft leak it quietly announced plans to add Google TPUs and Cerebras inference chips onto Oracle Cloud. Nobody sitting on a debt-financed mountain of competitive Nvidia silicon scrambles to rent out Google and Cerebras inference processors.
Sources: The Data Center Financing Calculus | Oracle Stock Profile
Oracle's debt farce makes the timing even funnier. Total debt has climbed past 169 billion dollars (around £128 billion), free cash flow plunged to negative 23.7 billion dollars (nearly £18 billion burned) in fiscal 2026, and S&P cut its credit rating to one rung above junk. At its 28 October Investor Day, Oracle swore to bondholders it would cap all new capital raises through fiscal 2030 at 20 billion dollars (£15.1 billion), split evenly between shares and debt. Five days later, on 2 November, Oracle filed to blow through that entire four-year limit in one shot, pricing 10 billion dollars (£7.6 billion) in shares at an 11.2% discount alongside 10 billion dollars (£7.6 billion) of senior notes paying coupons up to 8.65%. Throw in Oracle cancelling an estimated 7 billion dollar (£5.3 billion) Tencent chip deal and serving a force majeure notice on its 18 billion dollar (£13.6 billion) Project Jupiter Stargate campus in New Mexico as syndicate banks quoted Jupiter loans at 89 to 91 cents on the dollar, and Oracle's trophy cabinet as the bubble's premier bag holder is locked in.
Sources: Oracle IR | The Daily Upside | Oracle Leverage Analysis | CNA | Oracle Cash Flow | Project Jupiter Analysis
CoreWeave was an Ethereum miner cosplaying as an AI hyperscaler:
If Oracle is the bloated legacy bag holder, CoreWeave is the pure-play GenAI bubble darling. Founded in 2017 by three New Jersey commodities traders as Atlantic Crypto, an Ethereum-mining outfit that hoarded Nvidia GPUs, it rebranded to CoreWeave after crypto cooled off and pitched Wall Street on being a sleek "Maserati" compared to Amazon and Microsoft's sluggish "minivans". Nvidia played hype man by taking an equity stake, pouring another 2 billion dollars (£1.5 billion) in during January 2026, handing it priority chip shipments, and signing that 6.3 billion dollar (£4.8 billion) capacity backstop. When CoreWeave went public at 40 dollars (£30) a share in March 2025, retail momentum pushed it all the way to 187 dollars (£141) by June, even though Microsoft alone accounted for 67% of its 2025 revenue.
Sources: Fortune | App Economy Insights | Verified Investing | Seeking Alpha
CoreWeave's co-founders clearly understood what sort of ride they were selling. According to Bloomberg, the three co-founders have unloaded 3.3 billion dollars (£2.5 billion) in stock since the IPO as shares plunged roughly 78% from their peak down to 40.58 dollars (around £31) earlier this month. Strip away the flashy backlog headlines from this week's Q3 results and the maths is bleak: CoreWeave booked 3.66 billion dollars (£2.8 billion) in revenue and just 207 million dollars (£156 million) in operating profit against 846 million dollars (£640 million) in quarterly interest expense, posting a 709 million dollar (£536 million) net loss on 46.9 billion dollars (£35.5 billion) in recourse and non-recourse debt. With five-year credit default swaps hitting a record 1,194.5 basis points in October (implying a greater than 60% probability of default) and unsecured bonds trading past 16% yields, CEO Michael Intrator has sworn off unsecured bonds altogether, pinned his hopes on 9 billion dollars (£6.8 billion) in expected customer prepayments, and started shopping a 4 billion dollar (£3.0 billion) loan secured directly by CoreWeave's Meta contracts. The Maserati now runs on customer deposits, pledged contracts, and an Nvidia backstop.
Sources: Felix | Earnings Summary | Techdamentals | CoreWeave Credit Risk | Vector
The rest of the neocloud clown car and who really holds the debt:
Step outside Oracle and CoreWeave and the whole neocloud sector looks like a stampede for the exit before public investors notice the maths. In October, Nvidia-backed Australian operator Firmus pulled its planned 5 billion dollar (around £3.8 billion) IPO within 48 hours after institutional funds balked at a 30.6 billion dollar (£23.1 billion) valuation built alongside roughly 30 billion dollars (£22.7 billion) in debt, triple its private valuation just two months earlier. London-based, Nvidia-backed Nscale updated its IPO filings on 12 November revealing a 1.69 billion dollar (£1.3 billion) third-quarter net loss on 307.5 million dollars (£232 million) in revenue while considering delaying its 35 billion dollar (£26.5 billion) New York float into 2027 after Microsoft exited its 3 billion dollar (£2.3 billion) Essex UK deal over grid delays and walked away from a Texas expansion option. Gartner summed up the wreckage cleanly in October: at current spending rates, more than 3.5 trillion dollars (£2.6 trillion) in AI infrastructure faces impairment by 2030 and over 80% of neoclouds will cease independent operations.
Sources: StartupFox | Nscale Filing | Bharat Express | DQ Channels
Even worse is where the debt landed once banks decided they had seen enough. On 6 November, Bloomberg reported Dutch lender ING quietly put 500 million dollars (£378 million) of data-centre loans up for bulk auction all at once, including a 25 million dollar (£18.9 million) piece of Oracle's Project Jupiter campus. Meanwhile, off-balance-sheet structures like Meta and Blue Owl's 27.3 billion dollar (£20.6 billion) Hyperion bond vehicle pushed AI infrastructure borrowing away from tech balance sheets and straight into conservative institutional funds: Moody's found U.S. life insurers held 85.2 billion dollars (£64.4 billion) in AI-linked debt by June, while state insurance regulators discovered roughly half of life insurers' collateral-coded asset-backed securities were filed simply as "Other". Then on 3 November, ordinary voters stepped in and backed bans or restrictions on data centres in forty out of forty-one local ballot measures across ten states, while Texas elected Gina Hinojosa after she campaigned on an immediate statewide moratorium.
Sources: MLQ.ai | Substack Credit Analysis | Forbes | Business & Human Rights
At least we will miss Jensen's GenAI crack:
Say whatever you like about the circular loans and the spy drama, nobody in tech is going to match the unhinged energy Jensen brought to a stage once the bubble kicked in. Half the time he sounded like a late-night infomercial host high on his own supply. One week he was wheeling out a monstrous server rack and deadpanning "four elephants, one GPU", weighing computing hardware in zoo animals like it was a normal metric before asking if it could run Crysis. The next he was preaching pure MLM maths with "the more you buy, the more you save", a slogan that gets substantially funnier once you find out Nvidia was literally co-signing the loans so customers could afford the "saving" part. Between insisting every firm on earth was an "AI factory" built to churn out "these incredible tokens" and casually torching his own customers' server values by bragging "you couldn't give Hoppers away", watching Jensen freestyle corporate fever dreams in a leather jacket was genuinely top-tier comedy.
Sam Altman looks keen to pick up that comedic baton while Jensen packs his bags. With consumer hardware prices already pushed through the roof, OpenAI's chief executive went out in public and complained it was "quite terrible" that ordinary people still have DRAM sitting inside their own PCs "being used five percent of the time" instead of inside AI data centres, getting roasted so mercilessly online that he briefly renamed himself "Ram Altman" on X. You really have to respect the brass neck of a man burning billions on chatbot losses telling gamers they are being selfish with sixteen gigs of desktop memory.
Sources: PCMag on DGX | PC Gamer | John Furrier Substack | Business Insider | PCMag on Altman | Tom's Guide
Verdict
Once you strip out the stage gimmicks, what Nvidia actually did over the last three years was screw over ordinary hardware buyers and torch real-world resources just to buy extra quarters for an unviable bubble. Gamers, PC builders, and small studios got priced out of basic graphics cards and sticks of memory on the bogus claim that real-world AI demand required every transistor on earth, while Nvidia quietly monopolised foundry wafers and memory pipelines purely to keep the shortage story alive.
Meanwhile, the physical bill got dumped straight onto power grids, rural water supplies, and the environment. Millions of tons of concrete, dedicated gas turbines, and absurd amounts of clean silicon and electricity were poured into cooling sprawling server farms filled with hardware engineered to become expensive e-waste faster than a phone contract, all to keep generating synthetic tokens nobody actually wants to pay full price for. If the entire GenAI hardware craze collapses into a heap of discounted debt and empty buildings, regular people shouldn't shed a single tear for the firms left holding it.