AI news videos

NVIDIA Says Six Year Old GPUs Still Earn

NVIDIA has published its own investment case for AI factories, and it is worth reading as exactly that. There is no launch, no pricing and no product announcement in this post. It is a marketing argument, written by NVIDIA about hardware NVIDIA sells, and it closes by directing readers to Jensen Huang's GTC Berlin keynote scheduled for October 21.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• Backs the hook ('NVIDIA says chips it sold years ago still earn money. This is NVIDIA's own investment case.'). Article text: 'Durable: NVIDIA GPUs and systems keep earning years after they ship, extending useful life.' Editorial decision: the hook's second sentence names the piece as NVIDIA's own investment case so the provenance caveat lands inside the first five seconds rather than at second forty, as required for a no-event vendor essay.(PRIMARY AND SOLE SOURCE. NVIDIA corporate blog post, 'Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment', blogs.nvidia.com, section 'AI Infrastructure', dated October 1, 2026, byline Shruti Koparkar (an NVIDIA product-marketing byline). Captured 2026-10-01T13:06:12Z. First-party vendor publication, not journalism.)
• Backs scene 1 ('NVIDIA says each megawatt of an AI factory costs roughly sixty million dollars, and useful life shapes the return.'). Article text: 'AI factories are built by the megawatt, even by the gigawatt. Each megawatt factory costs roughly $60 million, and AI factory operators will only commit capital on that scale with a clear view of the return on investment.' And: 'Useful life: How long its AI hardware keeps earning.' Editorial decision: narrated with the attributor 'NVIDIA says' because the dollar figure is NVIDIA's own unsourced estimate; spelled as words per the no-digits rule.(NVIDIA's own cost figure, stated in the article's opening paragraph with no citation to any outside party.)
• Backs scene 2 ('NVIDIA says old hardware keeps earning after new hardware ships. It says a chip shipped six years ago is still in service.'). Article text: 'Not every workload needs the newest system. The right fit depends on a workload's complexity and shape, which is why the last generation keeps earning after the next one arrives.' And: 'The NVIDIA A100 GPU shipped in 2020 and is still in commercial service six years later, demonstrating its continued economic value'. Editorial decision: the model name A100 is a banned glued brand-plus-digit token for the text-to-speech engine, so narration says 'a chip shipped six years ago' and the name A100 appears only on the on-screen card, where it is safe. The ship year is rendered as 'six years ago' rather than a spoken calendar year.(NVIDIA's 'Durable: The Installed Base Keeps Earning' section. The useful-life claim is NVIDIA's, about NVIDIA's own installed base.)
• Backs scene 3 sentence one ('NVIDIA cites outside analysis showing operators keep extending depreciation schedules.'). Article text: 'Over the years, every major operator has extended the depreciation schedule on its servers, which is a guess about when hardware stops earning, and one that keeps moving out. A September 2026 Sprout analysis, "The Productive Life of a Data Center GPU," tracks how that schedule has shifted across every major operator.' Chart caption: 'Every major operator has extended server life. Source: Sprout, "The Productive Life of a Data Center GPU," September 2026. Data based on company disclosures and press reporting compiled by Sprout.' Editorial decision: narrated as 'NVIDIA cites outside analysis' rather than in our own voice, because the only access we have to Sprout is NVIDIA's selection and framing of it.(Sprout, 'The Productive Life of a Data Center GPU', September 2026, as cited and characterised by NVIDIA. We did not obtain the Sprout analysis itself; only NVIDIA's description of it and its chart caption are available.)
• Backs scene 3 sentence two ('Silicon Data says that chip is still worth a quarter of its original price.'). Article text: 'Silicon Data shows a six-year-old A100 GPU is still worth a quarter of what it cost, where a five-year depreciation schedule had it at zero more than a year ago.' Editorial decision: the firm is named out loud so the economic claim is never in our voice. The second half of the source sentence (the five-year schedule writing it to zero) was dropped to stay inside the thirty-word scene cap and to avoid putting a second figure in the same sentence as 'a quarter'.(Silicon Data, a GPU market-pricing firm, as cited by NVIDIA. No methodology, date, sample or link detail for the Silicon Data figure is given anywhere in the post.)
• Backs scene 4, the mandatory honesty scaffold ('NVIDIA published this argument about its own hardware, and nothing in it is independently audited. It points readers to NVIDIA's own keynote.'). Article text: 'NVIDIA AI factories are engineered to be productive, durable and fungible: more profitable tokens, longer useful life and deeper demand. That's what maximizes their return.' And: 'Learn more about NVIDIA AI factories by tuning in to NVIDIA founder and CEO Jensen Huang's GTC Berlin keynote on Wednesday, Oct. 21, at 11 a.m. CEST.' Editorial decision: the keynote date is deliberately not narrated (no spoken calendar dates), so the narration says 'NVIDIA's own keynote' without the date.(The article's own closing call to action and byline, establishing that the piece is vendor marketing that terminates in a vendor event promotion.)
• The YouTube transcript pull returned zero of five results (the installed yt-dlp is too old to fetch captions) and the timeline probes returned only irrelevant noise about consumer GPU mining return on investment and gaming hardware. No second outlet, no filing, no analyst note and no reporting on this post was obtained. Consequence: this single first-party NVIDIA blog post is the sole source for every narrated claim in the script, and no independent corroboration of any figure exists. Every narrated sentence therefore names its attributor out loud.(CORROBORATION STATUS: NONE. Recorded once for the whole script.)
• We read the captured article in full. It contains no independent audit of any claim, no methodology, sample size, date range or link for any of the third-party figures (SemiAnalysis AgentX, Sprout, Barkr, Silicon Data, Ornn Data are each named and then asserted in a single sentence), and no event of any kind: no product launch, no availability date, no price, no pricing change, no customer win announcement, no partnership, no earnings disclosure and no regulatory filing. The only forward-looking item in the entire post is an invitation to an NVIDIA keynote. The piece is an argument, not news of an occurrence.(NEGATIVE CITATION: full-text search of article.txt for audit, methodology and event material.)
• This is NVIDIA product marketing about NVIDIA's own products, carrying an NVIDIA product-marketing byline (Shruti Koparkar) on NVIDIA's own blog. Its thesis is that buying NVIDIA hardware is a good return on investment. The supporting figures come from named outside firms (SemiAnalysis, Sprout, Barkr, Silicon Data, Ornn Data) but are selected, ordered and presented by NVIDIA, with no methodology given for any of them and no access provided to the underlying analyses. The post closes by directing readers to NVIDIA's own GTC Berlin keynote. Editorial consequence: every scene in the script names its attributor out loud, no comparative or economic claim is narrated in our own voice, and the final scene states plainly that NVIDIA published this about its own hardware, that nothing is independently audited, and that the post points readers to NVIDIA's own event.(PROVENANCE PROBLEM, recorded explicitly as a condition of producing this piece.)
• Article text not used: 'SemiAnalysis AgentX data shows NVIDIA Vera Rubin NVL72 systems deliver over 30x higher throughput per megawatt than NVIDIA GB300 NVL72, and up to 45x lower cost per million tokens on the DeepSeek V4 Pro model.' Also unused: 'Power is the binding constraint on an AI factory. This makes tokens per second per megawatt the number that governs earning capacity.' Reason: tokens per megawatt, cost per token and performance per watt are burned ground for us, already published as nvidia-ai-factory-tokens-per-watt, nvidia-perf-per-watt and nvidia-power-flexible-ai-factories-grid. The script was built on the 'Durable' pillar instead. Secondary reason: Vera Rubin NVL72 and GB300 NVL72 are banned glued brand-plus-digit tokens on mic.(DELIBERATELY NOT NARRATED: the 'Productive' pillar, because we have already published it.)
• Article text not used: Lilly on 'a 1,016-GPU, on-premises cluster', Pinterest 'across 14,000 GPUs', Revolut, Runway, 'Texas A&M University: Running molecular simulation and AI drug discovery on its supercomputer, at 95-98% utilization across 26 projects and seven institutions', Dassault Systemes, and 'Unilever builds product imagery from digital twins rather than photo shoots, cutting production costs in half'. Also unused: 'More than 1,000 ready-made CUDA-X libraries' and 'more than 10 million developers'. Reason: narrating any of this requires a spoken multi-item list, which measurably slows the read and has broken the sixty-second ceiling before. Fungibility was available as supporting colour but was cut entirely to hold the word budget on the durability angle.(DELIBERATELY NOT NARRATED: the customer roster and the 'Fungible' pillar detail.)
• Four usable durability figures were cut so that only three figures are spoken in the whole script and no two land in the same or adjacent sentences. Dropped: 'CoreWeave recently extended bookings for units first introduced in 2020 through 2029'; the chart caption's 'Microsoft's NVIDIA V100 fleet ran 8.4 years against a six-year book life'; 'Barkr puts useful life at five to six years for an eight-GPU H100 system and nine to 10 years for GB300 NVL72, based on what those systems resell for'; and 'Ornn Data finds the market paying 80% as much to rent an A100 GPU on a five-year contract as on a one-month contract'. The three figures that survived, in spoken order, are sixty million dollars (scene 1), six years (scene 2) and a quarter of its original price (scene 3), each separated by at least one figure-free sentence. The Barkr and Ornn Data lines additionally carry banned brand-plus-digit tokens (H100, GB300 NVL72, A100) that cannot be spoken.(DELIBERATELY NOT NARRATED: figures dropped purely for the two-figure spacing rule and the three-figure budget.)