Nvidia’s narrative took a major hit this week due to multiple factors including the emergence of a credible rival, OpenAI’s struggles, and a trade war pincer that has them caught between Trump and China.
Is a Government Backstop the Bull Case?
Nvidia’s narrative, which it kicked off with the launch of Ampere architecture and A100 chip in 2020, subsumed any competing story-lines in the post-pandemic American stock markets in 2022, and engulfed the entire American economy under Trump.
Nvidia’s narrative that Large Language Models (LLMs) like OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini ARE the future of technology and the global economy has made them the world’s largest corporation by market cap.
The Trump administration may be all-in with Nvidia’s narrative and may be signaling its willingness to backstop the industry to prevent the AI bubble popping.
Paypal mafioso and Trump tech czar David Sacks had previously indicated the government would not be backstopping OpenAI earlier this month.
I argued elsewhere that Sacks’ tweet contributed to the downturn in AI stocks we saw in the first weeks of this month.
But on the 24th, Sacks seemed to signal a change in his (very influential) thinking, in response to a Wall Street Journal piece headlined, “How the U.S. Economy Became Hooked on AI Spending and subtitled, “Growth has been bolstered by data-center investment and stock-market wealth. A reversal could raise the risk of recession.”
According to today’s WSJ, AI-related investment accounts for half of GDP growth. A reversal would risk recession. We can’t afford to go backwards.
— David Sacks (@DavidSacks) November 24, 2025
Paired with the White House’s executive order titled “Launching the Genesis Mission” released the same day, many took this as bullish news for Nvidia’s narrative and the larger AI bubble.
The closest thing I could find to an action word in the EO was this promise to:
…build an integrated AI platform to harness Federal scientific datasets — the world’s largest collection of such datasets, developed over decades of Federal investments — to train scientific foundation models and create AI agents to test new hypotheses, automate research workflows, and accelerate scientific breakthroughs.
Critics of the Nvidia narrative will of course point out that when a government backstop is the bull case, the bubble is in trouble.
Maybe that’s why the a handful of naysayers who’ve been fighting the tide, are now being joined by some major investors who are putting their money where the critics’ keyboards have been.
Team Bear Beefs Up
Pioneering Nvidia bears like AI scientist Gary Marcus and journalist Ed Zitron have now been joined by notable investors including the “Big Short” contrarian Michael Burry and billionaire Stanley “I killed the Pound” Druckenmiller who sold all 214,060 of his funds’ Nvidia shares.
Marcus and Zitron remain opinion leaders in the space, however.
Marcus is currently dealing with the emergence of a class of rival AI experts who’ve been on the AGI (Artificial General Intelligence) bandwagon and are now getting off.
AGI is the patent nonsense that LLMs are just a few months away from creating super-intelligent, self-replicating machines.
Naturally, belief in AGI has been the conventional wisdom in Silicon Vally for the last couple of years and continues to be a big part of the bulls’ case for the Nvidia narrative.
Marcus has taken lots of heat for calling bullshit on LLMs as the road to AGI from the get go, and is now expressing mixed feelings about the big names who are joining him on the critical side.
Those names include Meta’s Chief AI Scientist Yann LeCun and OpenAI co-founder Ilya Sutskever.
As for Ed Zitron, his latest “The Hater’s Guide To NVIDIA” is well worth the subscription price and the estimated 54 minute reading time.
It includes an excellent, concise explanation of why Nvidia has become the hero of the Nvidia narrative which has reshaped the U.S. economy:
Back in 2006, NVIDIA launched CUDA, a software layer that lets you run (some) software on (specifically) NVIDIA graphics cards, and over time this has grown into a massive advantage for the company.
The thing is, GPUs are great for parallel processing – essentially spreading a task across multiple, by which I mean thousands, of processor cores at the same time – which means that certain tasks run faster than they would on, say, a CPU. While not every task benefits from parallel processing, or from having several thousand cores available at the same time, the kind of math that underpins LLMs is one such example.
CUDA is proprietary to NVIDIA, and while there are alternatives (both closed- and open-source), none of them have the same maturity and breadth. Pair that with the fact that Nvidia’s been focused on the data center market for longer than, say, AMD, and it’s easy to understand why it makes so much money. There really isn’t anyone who can do the same thing as NVIDIA, both in terms of software and hardware, and certainly not at the scale necessary to feed the hungry tech firms that demand these GPUs.
Anyway, back in 2019 NVIDIA acquired a company called Mellanox for $6.9 billion, beating off other would-be suitors, including Microsoft and Intel. Mellanox was a manufacturer of high-performance networking gear, and this acquisition would give NVIDIA a stronger value proposition for data center customers. It wanted to sell GPUs — lots of them — to data center customers, and now it could also sell the high-speed networking technology required to make them work in tandem.
This is relevant because it created the terms under which NVIDIA could start selling billions (and eventually tens of billions) of specialized GPUs for AI workloads.
…
Because nobody else has really caught up with CUDA, NVIDIA has a functional monopoly…NVIDIA has been printing money, quarter after quarter, going from a meager $7.192 billion in total revenue in the third (calendar year) quarter of 2023 to an astonishing $50 billion in just data center revenue (that’s where the GPUs are) in its most recent quarter, for a total of $57 billion in revenue, and the company projects to make $63 billion to $67 billion in the next quarter.
But never fear, Ziton also puts a stake through the heart of the Nvidia narrative over the course of several thousand words. Here’s a key point:
NVIDIA makes so much money, and it makes it from a much smaller customer base than most companies, because there are only so many entities that can buy thousands of chips that cost $50,000 or more each.
Zitron cites pseudonymous finance poster “Just Dario” as someone who’s provided key insights into the workings of Nvidia and Dario’s latest piece on the company is worth reading in full, but the TL;DR explanation of Dario’s role in the larger Nvidia narrative wars can be grasped from glancing at these tweets about whether or not Enron is a valid comparison point for Nvidia:
Nvidia says it’s not Enron. I actually agree with them, they are Envidia https://t.co/8zN308lZKK pic.twitter.com/IeahPNSd4Q
— JustDario 🏊♂️ (@DarioCpx) November 25, 2025
I should note that this pretty much unknown (and I suspect AI-using Substack) writer Shanaka Anslem Perera got the credit from Yahoo Finance with triggering Nvidia’s now infamous “we’re not Enron” memo.
Yahoo also quoted “Jim Chanos, who is famous for predicting the fall of Enron, (who) thinks the comparison between Nvidia and Lucent bears weight.”
“They’re [Nvidia is] putting money into money-losing companies in order for those companies to order their chips,” Chanos said.
As for “Big Short” Burry, his new Substack is a bit rich for my blood, although serious investors will likely find it a bargain, but his latest contribution to Nvidia’s narrative involves comparing Nvidia to Cisco before the dot.com bust:
Michael Burry's case against Nvidia: pic.twitter.com/XafV3SLYVT
— Nat Wilson Turner (@natwilsonturner) November 26, 2025
This adds to Burry’s previous X.com post alleging that AI industry accounting practices of “understating depreciation by extending useful life of assets artificially boosts earnings (are) one of the more common frauds of the modern era.”
And of course, Burry’s close to a billion-dollar bet against Nvidia’s narrative has impacted our story as well.
There’s also been a spate of bad news for Nvidia in the tech business press.
Two More Problems for the Nvidia Narrative
Two headlines from The Information neatly summarize the latest issues Nvidia is facing in the marketplace, unfortunately, it’s an expensive subscription only newsletter and we’ll have to look elsewhere for explanations:
- Google Further Encroaches on Nvidia’s Turf With New AI Chip Push
Summarized by Reuters as “Meta Platforms is in talks with Google to spend billions of dollars on the Alphabet company’s chips for use in its data centers starting from 2027, The Information reported, a move that would cast Google as a serious rival to semiconductor giant Nvidia.The talks also involve Meta renting chips from Google Cloud as early as next year and are part of Google’s broader push to get customers to adopt its tensor processing units (TPUs) – used for AI workloads – in their own data centers, the report said, citing people involved in the talks.”
Nvidia responded with a statement that is getting mocked online.
- China Is Slowly but Surely Breaking Free From Nvidia
Summarized by Reuters as “Chinese regulators have barred TikTok-owner ByteDance from deploying Nvidia (NVDA) chips in new data centers, The Information reported on Wednesday, citing two company employees.ByteDance bought more Nvidia chips than any other Chinese firm in 2025 as it raced to secure computing power for its billion-plus users amid concerns Washington could curb supply, according to the report.
The reported ban underscores Beijing’s efforts to reduce reliance on U.S. technology, a campaign that has intensified as Washington tightens curbs on exports of advanced semiconductors to China.
The Mid-Wits Weigh In
No debate in 2025 would be complete without one of the Abundance bros weighing in.
Naturally Ezra Klein’s “Abundance” co-author Derek Thomas (co-writing with Understanding AI founder Timothy B. Lee is coming down in the middle with “Six reasons to think there’s an AI bubble — and six reasons not to” and shrewdly saves the bull case for its paying customers. Talk about knowing your audience.
The Real Bulls Include Jim Cramer and AGI Crazytown’s Finest
But I’ll leave the real bull case to the legendary CNBC commentator Jim Cramer (I should note that the Inverse Jim Cramer ETF has been getting killed lately).
His case boils down to growth:
I am a huge believer in growth stocks, and …growth stocks are what draws me to the hyperscalers for my travel trust. Now we own many, many stocks from other universes drug stocks, aerospace materials, data center builders. They have long term growth stories, but they’re not turbocharged with the resources of these gigantic companies.
That’s what always brings me back to the Mag-7. These stocks are prominent because of their success. The companies they represent have bountiful profits, which is why they could rise to their lofty trillionaire status in the first place. It’s why I don’t kick them out when they’re down in fact. Or in fact, it’s why I might buy them for the trust.
But the far more entertaining bull case for the Nvidia narrative is made by Utopia believers like Tomas Pueyo who is helping his 119,000 Substack subscribers prepare for a post-scarcity Utopia brought about by “super-intelligence.”
Admittedly, I have an immediate and utter disdain for anyone pitching imminent Utopia but a couple of gummies and Pueyo’s stuff becomes quite entertaining. Here’s a taste from his latest, “AI: How Do We Avoid Societal Collapse and Build a Utopia?“:
In the previous article, we saw that we’ll eventually live in a utopia, a world of full abundance and without the conflicts over scarce resources that have plagued humanity throughout history.
Well then.
But lest readers think Mr. Pueyo isn’t concerned with the obstacles in our path, there’s more:
intelligence will not be infinite until there’s infinite energy and infinite compute, which will also need plenty of raw materials and land. So the scarcity of intelligence might not even be completely eliminated until more of these inputs are sufficiently abundant.
Assuming AIs still serve humans, how will they prioritize? They will need a signal of what matters most to humans. How will humans convey that? Through money.
…
So it’s very unlikely that we’re going to get rid of capitalism. We need the price signal to convey the optimal allocation of capital.
But in that world, where humans are not working anymore because they’ve been fully automated by AIs more intelligent than themselves, how do you decide how much money each person should have?
I’ll leave it up to readers who are really dying to know just how many angels are dancing on the top of his pinhead to read more of his work.
Just be aware that the Nvidia narrative bears are up against a whole lot of people who really really really want to believe and are clapping as hard as they can to keep Tinkerbell alive the AI bubble inflated.
Unfortunately for the bulls, and Nvidia’s narrative. AI slop is having very real world impacts. The kind people notice.
— Nat Wilson Turner (@natwilsonturner) November 26, 2025
Will Americans take comfort in their 401K wealth effect if they can’t even go into a happy turkey coma after Thanksgiving dinner because Nvidia’s narrative led to AI slop getting served to the table?
Posts Related to Nvidia’s Narrative:
- Flailing OpenAI Follows Meta’s Malignant Model
- How Google Is Winning Struggle Among AI Giants
- Yet More AI Bubble Worries
- Has Ed Zitron Found the Fatal Flaw with OpenAI and Its Flagship ChatGPT?
- OpenAI Laying the Groundwork for Massive Federal Bailout
- OpenAI Slipped Shopping Into 800 Million ChatGPT Users’ Chats
- Meta’s Mark Zuckerberg Loves To Throw Good Money After Bad
- OpenAI as The Money Pit
RE: …intelligence will not be infinite until there’s infinite energy and infinite compute…
It sounds like Mr. Pueyo could use a refresher on the concept of infinity. These AI humpers really are whacked out of their gourds, aren’t they?
Krell machine or bust (h/t Dr. Morbius).
The Krell data center. Pretty prescient.
Morbius to Ostrow and Adams underground, pointing off into the distance:
“Twenty billions,” points in the opposite direction, “twenty billions.”
Adams, “Listen. Trades opening and closing.”
Morbius, in reply, “And they never rest. This is one of their funding vehicles. You can feel the extracted rents rising. Look down here. Look down gentlemen, are you afraid of the deplorables?” Morbius points upwards: “Seventy-eight hundred tranches. And four hundred other funding vehicles like this one.”
Later Morbius is to sum it up as: “The harnessed power of an exploited planetary system.”
Beware the Monsters From the App!
Lol. Gotta say: that is one of my fave scenes in the whole film.
I remember thinking when I really looked at those Krell doorways that the set designer for the film had read Lovecraft. Those doors are shaped like Lovecraft described the unnamed aliens from “At the Mountains of Madness.” They too suffered a terrible doom due to their hubris and overconfidence in creating the shoggoths. Monsters from the Id indeed.
It’s superb. Every time I see it, and listen to it, I marvel.
One thing about tensor processing units is they use way less energy on inference. At least that might help make money on AI. ..Unless we all end up paying for openAI oracle etc. electricity bills.
Well, ok, it’ll bring down, but certainly not eliminate, the overhead going to the utility cos. Won’t do anything about paying for the silicon, the optical networking gear, the programming, the training, the $1M+ salaries, and most importantly, the profit margin needed to pay back the gigantic pile of capital they’re in the process of lighting on fire.
And of course, one only trains the model once (in principle) and when it’s ready tens, hundreds or even thousands of copies are made to distribute the load of inference (for there may be tens of thousands of request per second).
Maybe it’s just me, but even WITH infinite compute and infinite energy, I don’t think we’d be capable of hitting infinite intelligence. I don’t think we have it in us. We’ll probably hit infinite bullshit first, if we’re not almost there already.
The universe may not contain infinite energy and we cannot have infinite compute unless compute is absolutley efficient.
As to BS, well, that is open ended and without physical limits
Based on Zitron and Darion’s summaries it appears Nvidia and AI are headed for the mother of all busts. NVIDIA’s scheme is: Nvidia invests in some company, say two billion dollars. That company then buys a couple billion dollars of Nvidia GPU’s. Now using the GPU’s as collateral, the company borrows 20 billion. It then buys 15 billion of Nvidia GPUS. Now it goes back and borrows another 20 billion using the last twenty billion worth of gpus. or it signs a contract with OpenAI for $50 billion in compute services, and then uses that contract to borrow 40 billion to buy another 30 billion in Nvidia gpus. So, number always goes up for everybody as long as somebody can still borrow money. Some VC said at the current rate, all the VC money will be gone in six quarters ($200 billion ? I forget exactly). So when that day comes and the whole thing collapses, the bankruptcies will be phenomenal. I bet Nvidia drops by half in value, and the NYSE goes down by 30%. OpenAI has announced 1.4 trillion in spending in the next 15 years, when it makes maybe 12 billion a year currently. I think in 2027, if it lives that long, it needs a 100 billion just for that year.
One interesting link I’ve read that I don’t think was mentioned was this article: https://www.theverge.com/ai-artificial-intelligence/822011/coreweave-debt-data-center-ai aka (“I looked into CoreWeave and the abyss gazed back: Meet the company Nvidia is propping up”)
This is an excellent case study for one of the companies Nvidia is propping up. Possibly the best quote that guarantees a bust is as follows:
“These [Nvidia] investments [in CoreWeave] are not circular; they are complementary,” CoreWeave’s Davis wrote. “This is about an entire ecosystem all rowing in the same direction to accelerate the AI economy. There’s nothing circular about it. Rather, these partnerships are about accelerating innovation and adoption. We are, collectively, defining the next-generation operating system for civilization.”
Lemmings one and all…
Another interesting point mentioned was that at the end of July, Nvidia owned $3.9B in non-marketable securities–probably investments in a number of companies like CoreWeave. So I thought, I wonder where Nvidia is now on these securitiesthat they just released a new quarter of data?… $8.2B or over 100% growth in 3 months.
My son is in intro accounting and a major AI user. I asked him if they looked at many corporations’ balance sheets in his class. He said not many. I said, go look at NVDA and tell me if you think they’re a $5T company. Seems like a big bust in on the horizon although I easily could see another “socialism for the rich” scenario in the offing.
The companies building these data centers for companies like OpenAI might be the first to fail. They are taking on huge debt levels to build these things, with often nothing more than the contracted promise to buy compute services as the collateral behind that debt. Like Oracle, who signed a $300 billion dollar contract with OpenAI to supply 4.5 GigaWatts of computing power, deliverable Jan 1, 2027. Oracle hired Crusoe to build the Stargate Abilene data center, which I believe last month the first of eight buildings was turned over to Oracle with 200 MW of power. Another seven have to be built, equipped and made operational in the next 12 months, lol. The 4GW power substation hasn’t even really started construction yet, Oracle needs to come up with $40 billion just for the GPUs, and has to start paying Crusoe $1 billion dollars a month for the next 15 years for the construction. Good question as to where Larry is going to find the money. The other thing is these contracts are worded such that if the data center is not completed to the specs, the counter-party can walk away. So if, as expected, Oracle fails to build out all the gpu power OpenAI contracted, and Altman walks, who is Oracle going to be able to find to buy the compute power ? Nobody. Then there’s the problem that OpenAI won’t have the money to pay Oracle anyway.
Hundreds of billions of dollars are being incinerated, and when the flames go out, the cold will be unreal.
Maybe the IDF is helping him out with some of the jewelry they stole from Palestinians.
I know that gold has risen a lot but not enough to cover this hole completely.
Maybe the precious gems?
A good friend who has been working in the US power utility sector recently left that industry for headhunting. He is specialising in recruiting the people required to design and project manage the power connections for AI datacentres.
He said over coffee, I might need another specialism. I told him he did….
Expect that Nvidia doesn’t make the GPUs themselves. They are a (great) hardware and (very poor) software design company. Their “assets” are the engineers, tools, knowhow and of course IP surrounding these chips.
So, if the topic of a massive bailout comes up, wouldn’t it just be cheaper to liquidate the company, and sell the tools and re-hire everyone into a new venture?
Nvidia’s real lock is CUDA, the software that you use to drive the gpu’s. It is of course proprietary and nobody has anything remotely close. I’m sure anyone who even thinks of trying will run into a thicket of patents surrounded by a moat guarded by an army of 10,000 patent lawyers.
https://www.eetimes.com/after-three-years-modulars-cuda-alternative-is-ready/
CUDA alternatives.
One of the most successful has been the open-source project ApacheTVM. TVM’s main aim is to enable AI to run efficiently on diverse hardware by automating kernel fusion, but generative AI proved to be a technical challenge, since algorithms are larger and more complex than for older computer vision applications. Generative AI algorithms are also more hardware-specific (like FlashAttention). TVM’s core contributors formed a company called OctoAI, which developed a generative AI inference stack for enterprise clusters, but it was acquired recently by Nvidia, casting some doubt on the project’s future.
Another widely known technology, OpenCL, is a standard designed to enable code portability between GPUs and other hardware types. It has been broadly adopted in mobile and embedded devices. However, critics of this standard (including Lattner) point to its lack of agility to keep up with fast-moving AI technology, in part because it is driven by a “co-opetition” of competing companies who generally decline to share anything about future hardware features.
Other commercial projects of this nature are still in the early stages, Lattner said.
Worth noting that CUDA is not for AI only. To give a real life usage egzample, my brother is a CAD specialist contractor, models architecture and interiors for developers and furniture producers etc. He bought his third Nvidia GPU now because there is a lot of software that requires it for speeding up or automating certain operations like cloth physics (very useful in interior modeling). The GPUs return their price manifold saving him a lot of time.
This is true. I currently am involved in high performance computing (undergraduate research) and we often discuss gpu vs cpu for the calculations we are doing. Any time you have a parallelizable problem, ur gonna wanna consider gpu. They originally became big in gaming, because most image processing is inherently parallel. And now that we cant just design smaller faster chips, theres a motivation to find more ways to speedup like parallel processing.
We also optimize for memory in my lab. One of the grad students was talking about how cuda doesnt have any way to parallelize sparse or banded matrices. They actually just pass everything to cpu. Which is something they could work on.
NVDA chip buyers need more years (if not decades) to pay off their notes that capitalized their “hyper scaling” than NVDA PUs will be state of art.
There are efficiencies….
The short term answer to AI is reduce the cost to train your llModles. Then reduce the run cost for compute /inference. Low case “l” not typo. LLM are often too expensive for the answer they provide.
DeepSeek does this with SW.
The other efficiency is in the PU’s. Which TPU may be just one.
OpenAI won’t make any money ever, many NVDA customers go broke and will default a bunch of bonds. The Fed can keep solvent a lot of CDS resulting from a lot of analysts failing to question Huang and Altman. While they ignored macro economics and new disruptive product profit cycles. Bond underwriting.
There are all kinds of disrupters chasing tech that will bring down the TBTF.
No govt bail out, which tech firm will be first Bear Sterns, ENRON?
To put a 2025 technology LLM AI into a humanoid android, approximately 4.3 m³ would be required for the hardware (Computing: 1.5 m³, Cooling: 2.0 m³, Power Supply: 0.5 m³, Connectivity: 0.3 m³).
This AI must be pre-trained in data centers that require about 1600 m³ for the hardware, and between two and five of these data centers are needed for the training phase.
Once trained, several of these data centers will perform the task of fine-tuning and improving the LLM model, and others will handle storage and backups. Additional data centers will surely be required for this latter task.
This AI could perform local inference and fine-tuning, but it would still require connectivity to data centers for model updates (weekly/monthly), data synchronization (optional), specific tasks requiring the cloud, internet searches, complex non-time-sensitive analysis, data backups, among other things.
In other words, self-replicating machines and Terminator-style robots are still a long way from becoming a reality.
Great compendium as always, Nat. What a time to be alive.
In a YouTube video I heard the argument that this certainly looks like a bubble, and there are lots of warning flags, but those firms mostly use their giant pile of cash to do the investing, rather than using (leveraged) debt. I can’t assess it that is totally true, but if it is, then it is a very risky bet, and given that the US economy is driven by those “Magnificent Seven”, it may not hurt the mainstream economy all that much? Of course, there are many investors who will get hurt, so that’s bad. Is this a reasonable point of contention with the bubble-story or not?
It’s the pension funds and retail investors who will lose everything when the bubble pops.
As a horrible old Man who was licensed to run a collection agency for more than a decade ( I actually specialized in second and third placement debts, including corporate debt) I look at the Money flows.
And while NVIDIA is not as blatant a scam as WeWork was, the financial relationships between NVIDIA and its “Customers” are Peculiar, to say the least.
I’m reminded of Keyne’s saying that the market can stay irrational longer than I can stay solvent, regarding the short sellers.
Since Sam Altman’s ultimate contribution to society and business is going to be individualized sex porn, the only real question for the business world is, what is the monetary value as a percentage of GDP?
I assume we can take OpenAI being worth $500B to some investors based on his penultimate contribution of individualized non-sex porn for all ChatGPT users’ needs for companionship and flattery as a starting point for computing the value of his ultimate contribution.