{"id":200885,"date":"2025-11-26T08:48:19","date_gmt":"2025-11-26T08:48:19","guid":{"rendered":"https:\/\/www.europesays.com\/ie\/200885\/"},"modified":"2025-11-26T08:48:19","modified_gmt":"2025-11-26T08:48:19","slug":"neural-dispatch-the-ai-maths-is-broken-and-a-bubble-that-many-pretend-to-not-see-is-deflating-ht-tech","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ie\/200885\/","title":{"rendered":"Neural Dispatch: The AI maths is broken, and a bubble that many pretend to not see, is deflating \n(HT Tech)"},"content":{"rendered":"\n<p>The numbers simply don\u2019t add up. And that is enough to warrant a temporary departure from the standard Neural Dispatch format, for this week. On November 20, Nvidia reported number for the third quarter fiscal 2026. Record revenue of $57.0 billion, up 22% from Q2 and up 62% from a year ago, the tech giant says. \u201cBlackwell sales are off the charts, and cloud GPUs are sold out. Compute demand keeps accelerating and compounding across training and inference \u2014 each growing exponentially. We\u2019ve entered the virtuous cycle of AI. The AI ecosystem is scaling fast \u2014 with more new foundation model makers, more AI startups, across more industries, and in more countries. AI is going everywhere, doing everything, all at once,\u201d says Jensen Huang, founder and CEO of NVIDIA. If you look at just this summary (as most of you would, with limited attention spans, you\u2019d be impressed).<\/p>\n<p>     <img decoding=\"async\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/11\/Main-Nvidia_1764107185160_1764107198195.png\" alt=\"Nvidia has almost $19.8 billion worth of unsold chips, and that\u2019s up significantly in 3 months\" title=\"Nvidia has almost $19.8 billion worth of unsold chips, and that\u2019s up significantly in 3 months\"\/>   Nvidia has almost $19.8 billion worth of unsold chips, and that\u2019s up significantly in 3 months    <\/p>\n<p>But it was the Wall Street that read between the lines, and the Nasdaq as well as S&amp;P 500 slid\u2026and then kept sliding. A snapshot \u2014 Dow was down 0.8%, S&amp;P 500 down 1.6%, and Nasdaq down 2.2% on the day. There was much more to Nvidia\u2019s earnings, which I\u2019ll attempt to summarise to save you time (and re-emphasise that to call this AI conversation a \u201cbubble\u201d wouldn\u2019t exactly be out of place).<\/p>\n<ul>\n<li>Nvidia has almost $19.8 billion worth of unsold chips (these would be primarily GPUs) sitting in warehouses, and that number is up significantly in three months \u2014 the inventory amount was around $10 billion then. There is a reason why I put Huang\u2019s entire quote in the previous paragraph. Apparently, the official line is that demand is through the roof and supply is limited. You know it as well as anyone, shortage claims cannot be true, if there\u2019s so much inventory sitting around. It\u2019s common sense. The only question to be asked is \u2014 are customers simply not buying, or are they buying without actually having the money to buy? Nvidia though explains this as a build-up for future demand.<\/li>\n<li>Nvidia has also reported that there are $33.4 billion in unpaid bills, or accounts receivable as the terminology goes, and that number is up89% in a year. This means customers who bought chips haven\u2019t paid for them yet. The average payment window is now 53 days, instead of 46 days a year prior. That extra week represents a few more billion, and when they may arrive is anyone\u2019s guess. A more concerning scenario for the chipmaker emerges if many of Nvidia\u2019s customers start to behave like under-capitalised AI startups struggling to pay for hardware in case venture funding dries up or if the AI bubble deflates before these companies achieve profitability.<\/li>\n<li>Over the past month, we\u2019ve seen absolutely brash attempts at a circular economy to keep the AI bubble pressurised somehow. The latest in that chapter also featured Nvidia which is participating with its own $2.5 billion \u2018investment\u2019 in Elon Musk\u2019s xAI which wants to raise $20 billion. I was reading somewhere, this would involve an interesting financing structure with a special purpose vehicle (SPV) that will purchase the Nvidia hardware, which xAI will then lease for a five-year term. This isn\u2019t the first <a rel=\"nofollow noopener\" href=\"https:\/\/httechnology.substack.com\/p\/neural-dispatch-sir-demis-hassabis?triedRedirect=true\" target=\"_blank\" class=\"backlink\" data-vars-page-type=\"story\" data-vars-link-type=\"Manual\" data-vars-anchor-text=\"circular AI\">circular AI<\/a> deal and neither will it be <a rel=\"nofollow noopener\" href=\"https:\/\/httechnology.substack.com\/p\/neural-dispatch-openai-needs-a-favour?triedRedirect=true\" target=\"_blank\" class=\"backlink\" data-vars-page-type=\"story\" data-vars-link-type=\"Manual\" data-vars-anchor-text=\"the last\">the last<\/a>.<\/li>\n<\/ul>\n<p>The bottomline is, the same pile of dollars are being circulated between different AI companies, hand holding each other in the hope that the bubble isn\u2019t discovered, and gets counted as revenue at each corporate stop this wad of cash makes on its journey.<\/p>\n<p>American investor and hedge fund manager Michael Burry made a rather blunt post X after Nvidia\u2019s earnings release. He wrote, \u201cThe idea of a useful life of depreciation being longer because chips from more than 3-4 years ago are fully booked confuses physical utilisation with value creation. Just because something is used doesn\u2019t mean it is profitable.\u201d He pointed out that airlines keep old planes around and in service, which come in handy during the festive period rush, but they are only marginally profitable. The reality is, Nvidia\u2019s CFO had pushed back on the GPU accounting (which I\u2019ve explained above) and in a statement said the useful life of Nvidia\u2019s GPUs is a significant total cost of ownership advantage over rivals \u2014 and points to A100 GPUs that were shipped 6 years ago still being utilised at full capacity by customers. But it isn\u2019t that simple. The A100s consume as much as 3x more power per compute (the unit is FLOP, or Floating-Point Operations Per Second) than the H100s that followed it. And that in itself is approximately 25x less power efficient than Blackwell generation chips. A debate is raging \u2014 should depreciation be 3 years, 5 years, or 7 years? Compulsion more than choice?<\/p>\n<p> <strong>THINKING<\/strong>   <img decoding=\"async\" src=\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2025\/10\/quoteText.png\" alt=\"\"\/>  <\/p>\n<p>\u201cWe\u2019re doing a 500 megawatts, gigawatts\u2026It\u2019s going to cost eight bazillion trillion dollars.\u201d &#8211; Elon Musk, at the U.S.-Saudi Investment Forum 2025<\/p>\n<p>With Elon Musk, it is difficult to know if he was genuinely confused, or it was just an artificially induced fog. But this was Musk, introducing xAI\u2019s planned 500 MW AI data centre partnership with Saudi Arabia. And of course, this is powered by Nvidia, which is why CEO Jensen Huang was almost sweating when he said \u201cstop it\u201d as Musk stumbled between megawatt and gigawatt. Not a casual occasion to stumble, for the man who many believe is the saviour of humanity (of course that mission will also be powered by Nvidia, but I digress).<\/p>\n<p><strong>The Context:<\/strong> That is the whole AI bubble, condensed into one <a rel=\"nofollow\" href=\"https:\/\/x.com\/AutismCapital\/status\/1991248050015371656?s=20\" target=\"_blank\" class=\"backlink\" data-vars-page-type=\"story\" data-vars-link-type=\"Manual\" data-vars-anchor-text=\"beautifully unhinged\">beautifully unhinged<\/a> exchange. They thought no one would notice in the cloud of big numbers and excitement. A CEO who\u2019s raising tens of billions for compute doesn\u2019t know (or pretends not to know) the difference between megawatts and gigawatts. Understandably, the CEO of the world\u2019s most valuable semiconductor company visibly panics because the quiet part \u2014 that no one really knows where this is going or how much it will cost \u2014 has just been said out loud, at an event filled with sovereign wealth funds. Has \u201cfake it till you make it\u201d morphed itself into \u201cbuild it till the grid collapses and hope the ROI eventually materialises\u201d for the AI era?<\/p>\n<p><strong>A Reality Check: <\/strong>We find ourselves amidst a moment where AI companies are committing trillions of dollars to data centres without a clear business model beyond \u201cAGI will pay us back at some point.\u201d Power costs worldwide are going up, as is the demand for water. Something simply has to give, at some point. AI companies and startups are being funded with billions to be ready to buy GPUs that don\u2019t exist yet, to train models nobody knows how to monetise, for customers who aren\u2019t sure why they need them.<\/p>\n<p>Musk\u2019s quote isn\u2019t a joke \u2014 it\u2019s as close as we\u2019ll ever get to an accidental confession from the AI bros. The AI boom today is powered by physics, marketing and spreadsheets that print whatever number keeps the funding round alive. Nobody has any idea what true power requirements will be (that\u2019ll after all depend on usage, and no one knows that too), how many chips are actually needed, or what the returns look like. And yet everyone keeps buying compute because everyone else is buying compute. This is how bubbles form, through collective delusion wrapped in technical jargon. And you can\u2019t blame Jensen for sweating, because this is a market built on curating expectations, and the worst possible thing is someone admitting they\u2026 don\u2019t know what they\u2019re talking about.<\/p>\n","protected":false},"excerpt":{"rendered":"The numbers simply don\u2019t add up. And that is enough to warrant a temporary departure from the standard&hellip;\n","protected":false},"author":2,"featured_media":200886,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[261],"tags":[291,88337,110976,289,290,47932,110977,18,19,17,292,864,82],"class_list":{"0":"post-200885","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-artificial-intelligence","8":"tag-ai","9":"tag-ai-developments","10":"tag-ai-maths","11":"tag-artificial-intelligence","12":"tag-artificialintelligence","13":"tag-blackwell","14":"tag-cloud-gpus","15":"tag-eire","16":"tag-ie","17":"tag-ireland","18":"tag-nvidia","19":"tag-software","20":"tag-technology"},"share_on_mastodon":{"url":"https:\/\/pubeurope.com\/@ie\/115615138524345577","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts\/200885","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/comments?post=200885"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/posts\/200885\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/media\/200886"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/media?parent=200885"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/categories?post=200885"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ie\/wp-json\/wp\/v2\/tags?post=200885"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}