If we’re in a bubble, it’s like the dot-com bubble, not the tulip bulb bubble. Ask yourself whether the Internet turned out to be a big deal. Also, under any circumstances, many current startups will fail spectacularly, and this could easily include some very big bets (cf. Webvan). OpenAI’s International Conundrum — The Information @Rittenhouse Research: Biggest takeaway from this (very good) interview was Altman's outline how OAI's financial model could eventually work. - OAI enterprise is growing faster than consumer - OAI enterprise growth is constrained by lack of compute capacity (e.g. enterprises are coming to OAI asking… https://t.co/RtmIYTst6x https://x.com/i/status/2002038608501301346 https://www.noahpinion.blog/p/the-ai-bust-scenario-that-no-one https://x.com/Oaktree/status/1998430811159409007 https://ceodinner.substack.com/p/the-ai-wildfire-is-coming-its-going https://www.theinformation.com/articles/openai-projected-least-220-million-people-will-pay-chatgpt-2030 https://www.understandingai.org/p/six-reasons-to-think-theres-an-ai From https://thezvi.substack.com/p/ai-142-common-ground: Chris Bryant: The head of Alphabet Inc.’s AI and infrastructure team, Amin Vahdat, has said that its seven- and eight-year-old custom chips, known as TPUs, have “100% utilization.” theinformation.com/articles/anthropic-projects-cost-advantage-openai Meltem Demirors (@Melt_Dem) posted at 8:56 AM on Sun, Nov 09, 2025:reflecting on the recent slew of “it’s an AI capex bubble” w my partner @kellyjgreer - feels like much ado about nothingthere’s some confounding narratives at play:- the magnitude of the capex numbers is enormous and unprecedented - largest infra buildout since WWII- only(https://x.com/Melt_Dem/status/1987565089696694666?t=c2xIrwJRU6pk0uzxpmmHBw&s=03) https://www.theinformation.com/articles/anthropic-projects-70-billion-revenue-17-billion-cash-flow-2028 https://epochai.substack.com/p/introducing-the-frontier-data-centers https://nikolajurkovic.substack.com/p/are-we-in-an-ai-bubble?t=Hr_MPzFeYKIGTDbvWW-dDA&s=03 tae kim (@firstadopter) posted at 11:40 AM on Sat, Oct 25, 2025:Why AI is Underhyped and Isn't a Bubble YetHere's why AI isn't a bubble today. I've distilled the data points and ideas from my previous columns and coverage. If you prefer facts and evidence-based reality over vibes and conflated narratives, enjoy!-Big Tech valuations are(https://x.com/firstadopter/status/1982155280423977243?t=Q-ORfSjvzF7yiM-Gcasb6g&s=03) https://peterwildeford.substack.com/p/ai-is-probably-not-a-bubble https://www.dwarkesh.com/p/thoughts-on-the-ai-buildout Review the stateof.ai report, slides 99-104, 111-112, 131 Share draft with Siméon, he commented on Noah Smith’s post and argued that a pop is plausible https://x.com/ajeya_cotra/status/1977881746797687223 https://x.com/snewmanpv/status/1977931996249928011 https://www.understandingai.org/p/16-charts-that-explain-the-ai-boom https://pluralistic.net/2025/10/16/post-ai-ai/#productive-residue https://epochai.substack.com/p/openai-is-projecting-unprecedented https://www.dwarkesh.com/p/thoughts-on-the-ai-buildout https://x.com/dwarkesh_sp/status/1981074799758921843 https://www.noahpinion.blog/p/should-we-worry-about-ais-circular → the circular deals don’t determine whether this is over-investment, but they may increase volatility / exacerbate any collapse From https://thezvi.substack.com/p/ai-139-the-overreach-machines: Gunjan Banerji: Goldman: “We don’t think the AI investment boom is too big. At just under 1% of GDP, the level of spending remains well below the 2-5% peaks of past general purpose technology buildouts so far.” [see subsequent image] Epoch AI (@EpochAIResearch) posted at 2:19 PM on Fri, Oct 10, 2025:New data insight: How does OpenAI allocate its compute?OpenAI spent ~$7 billion on compute last year. Most of this went to R&D, meaning all research, experiments, and training.Only a minority of this R&D compute went to the final training runs of released models. https://t.co/Jq9768bSB1(https://x.com/EpochAIResearch/status/1976714284349767990?t=UU51tJ2sBPX7dGTHMWK9rw&s=03) Corry Wang (@corry_wang) posted at 6:17 PM on Fri, Jan 01, 2021:1/ Lessons From The Tech Bubble:Last year, I spent my winter holiday reading hundreds of pages of equity research from the 1999/2000 era, to try to understand what it was like investing during the bubbleA few people recently asked me for my takeaways. Here they are - https://t.co/41nTJdrFR1(https://x.com/corry_wang/status/1345192541545766915?t=YtZvbF0fgkf-zBeVWX2wxA&s=03) Benjamin Todd (@ben_j_todd) posted at 6:29 PM on Mon, Oct 20, 2025:The FT is my favourite newspaper, but this seems to be bad reporting.The article "What GPU pricing can tell us about how the AI bubble will pop" points out a rack of 8 A100 chips need to generate $4/hour to cover the capital cost of the chips over 5yr.It then points out that https://t.co/9SEbY5rnLW(https://x.com/ben_j_todd/status/1980446425499726094?t=AJEb5wjbwagwVHE55lNGWw&s=03) https://www.theinformation.com/articles/nvidia-broadcom-amd-face-new-risks-openai-deals https://thezvi.substack.com/p/bubble-bubble-toil-and-trouble https://www.theinformation.com/articles/oracle-assures-investors-ai-cloud-margins-struggles-profit-older-nvidia-chips https://www.theinformation.com/articles/salesforce-ceo-shifts-ai-strategy-openai-threat-looms https://www.derekthompson.org/p/why-ai-is-not-a-bubble https://www.derekthompson.org/p/this-is-how-the-ai-bubble-will-pop Could quote from the first section of “Money Stuff: OpenAI Has a Business Plan” for a funny intro to the question. It really is the greatest business plan in the history of capitalism: “We will create God and then ask it for money.” On a pure science-fiction suspension-of-disbelief basis, this business plan is perfect and should not need any updating until they finish building the superintelligent AI. Paying one billion dollars for a 0.2% stake in whatever God comes up with is a good trade. But in the six years since announcing this perfect business plan, Sam Altman has learned[2] that it will cost at least a few trillion dollars to build the super-AI, and it turns out that the supply of science-fiction-suspension-of-disbelief capital is really quite large but not trillions of dollars. David Shapiro ⏩ (@DaveShapi) posted at 5:38 AM on Fri, Oct 17, 2025: Some people still think that AI is a "bubble" so here's my updated take. They point at a few facts like: The "circularity" of the market The fact that OpenAI and other startups are not profitable It "feels" like the Dot Com revolution ("where there's hype, there's a https://t.co/3WTUQxaTjP (https://x.com/DaveShapi/status/1979165009495171523?t=VNbxocWJqnzeiR6iBkZ2zQ&s=03) https://www.ft.com/content/a169703c-c4df-46d6-a2d3-4184c74bbaf7 [me, on Signal] My two cents: The amounts being invested are eye-watering, but to my understanding, not at all unprecedented – the 1800s railroad build-out was much larger (as a % of GDP), the 1990s telecom build-out was at least as large (again, as % of GDP). "If we stop, the economy will collapse" is a terrible reason to keep investing. If these investments are going to pay off, then there's no economic reason to stop (there may of course be other reasons, relating to safety, disruption, etc.). If these investments are not going to pay off, then inflating the bubble further just delays + worsens the inevitable. None of this is binary. Current investments might yield a 500% profit, or 20% profit, or break even, or lose a bit of money, or wind up being sold off for pennies on the dollar. The economy might collapse, or suffer a brief recession, or a brief slowdown in growth, or do fine. (Granted, it is tricky to gently deflate a bubble, so the distribution of potential economic scenarios may be somewhat bimodal, but we've recently seen that soft landings are possible.) None of this is predetermined; the choices that governments and businesses make, and the skill with which they execute, will play an important role. [Igor] On railroads: I would encourage you to do the math on the capex, opex, lifetime of data centers vs railroads and the implied ROIC required for things to be rational. The essays linked here https://futurism.com/future-society/ai-data-centers-finances might serve as inspiration. Rest I more or less agree with, albeit probably from a more ai critical lens I do think the ai bubble popping will be this generations 2008 Epoch AI (@EpochAIResearch) posted at 9:23 AM on Wed, Oct 15, 2025: One way bubbles pop: a technology doesn’t deliver value as quickly as investors bet it will. In light of that, it’s notable that OpenAI is projecting historically unprecedented revenue growth — from $10B to $100B — over the next three years. 🧵 https://t.co/WCgLhoLb8B (https://x.com/EpochAIResearch/status/1978496866624176507?t=UHaEDOor8F84JrM3JTUPoQ&s=03) Ajeya Cotra (@ajeya_cotra) posted at 4:38 PM on Mon, Oct 13, 2025:I'm interested in more detailed work on how much profit AI companies will capture and whether there will be a bust in the next 2y.Appreciated the emphasis that "market crash" != "low-impact technology" https://t.co/A8SuW9IGB4(https://x.com/ajeya_cotra/status/1977881746797687223?t=9gVZ8D__Q8-HEPjeZ2P_0Q&s=03) Derek Thompson (@DKThomp) posted at 8:24 AM on Wed, Oct 15, 2025:New newsletter: WHY AI IS NOT A BUBBLEEverybody is calling AI a bubble now, even the folks building it.I hate when conventional wisdoms get too conventional.So I read and listened to the smartest cases for why AI is NOT a bubble.My summary —>https://t.co/9gjWvh2WZL(https://x.com/DKThomp/status/1978482238229561816?t=dvmpuBl2JhgiJ3AhZKANJA&s=03) https://www.theinformation.com/briefings/anthropic-said-target-26-billion-annualized-revenue-2026 https://www.transformernews.ai/p/what-happens-when-the-ai-bubble-bursts-crash https://www.slowboring.com/p/the-ai-boom-is-propping-up-the-whole https://x.com/ecommerceshares/status/1978392637682876551?s=46 https://www.linkedin.com/feed/update/urn:li:activity:7384323912786141184/ https://x.com/deanwball/status/1978458776832266669 https://davekarpf.substack.com/p/its-giving-enron https://www.wheresyoured.at/ai-is-a-money-trap/ https://www.theinformation.com/articles/openai-working-softbanks-arm-broadcom-ai-chip-effort Derek Thompson (@DKThomp) posted at 7:06 AM on Tue, Oct 14, 2025:What's the best source -- from an article, bank, analyst note, etc -- for the level of "AI revenue" that hyperscalers are seeing right now?(Yes, I know this can be fudged in a million ways; eg, Meta can always claim that X percent of its ad revenue is thanks to AI.)(https://x.com/DKThomp/status/1978100071896944905?t=e48XnPEWU-DmRQwNBr6VBQ&s=03) https://www.theinformation.com/articles/microsoft-let-openai-play-field https://www.theinformation.com/articles/race-rent-nvidia-chips-cloud-intensifies?rc=jsaoww https://epochai.substack.com/p/the-epoch-ai-brief-september-2025 https://www.bloomberg.com/news/features/2025-10-07/openai-s-nvidia-amd-deals-boost-1-trillion-ai-boom-with-circular-deals https://futurism.com/future-society/ai-data-centers-finances https://paulkrugman.substack.com/p/technology-bubbles-causes-and-consequences https://www.noahpinion.blog/p/americas-future-could-hinge-on-whether https://www.exponentialview.co/p/is-ai-a-bubble Andrew Curran (@AndrewCurran_) posted at 10:02 AM on Tue, Oct 07, 2025:Jamie Dimon told Bloomberg that JPMorgan has now reached AI equilibrium. JPM spends $2 billion a year on developing artificial intelligence technology, and saves about the same amount annually. Their plan is to gain first-mover advantage by incorporating agentic AI at all levels. https://t.co/nLA1z1zFGC(https://x.com/AndrewCurran_/status/1975607594556563555?t=yHaY5hPnM8LkyG-tR2Rm-w&s=03) https://www.theinformation.com/articles/internal-oracle-data-show-financial-challenge-renting-nvidia-chips?utm_campaign=article_email&utm_content=article-15862&utm_medium=email&utm_source=sg&rc=jsaoww https://www.nytimes.com/2025/10/06/technology/openai-amd-chips.html Follow up with Konstantin Pilz from RAND, per email introduction from Matt Chessen after we were matched 1:1 at The Curve but Matt couldn’t attend. Open Philanthropy (from https://www.openphilanthropy.org/research/ai-safety-and-security-need-more-funders/): We’ve supported the Center for Security and Emerging Technology at Georgetown University, which compiles detailed data on AI investment, semiconductors, and governance efforts for use by policymakers, researchers, and journalists. Epoch AI (@EpochAIResearch) posted at 0:04 PM on Tue, Sep 30, 2025:Announcing our new AI Companies Data Hub!We collected key data on frontier AI companies, including revenue run rates, funding, staff, usage rates, and compute spend.This free resource will help you understand the trajectory and economics of AI.Highlights in thread! https://t.co/bk4h5PixbA(https://x.com/EpochAIResearch/status/1973101761964462582?t=4ldNUkTnYFeFUnp6g07usQ&s=03) https://peterwildeford.substack.com/p/openai-nvidia-and-oracle-breaking https://www.theinformation.com/articles/openais-first-half-results-4-3-billion-sales-2-5-billion-cash-burn https://pluralistic.net/2025/09/27/econopocalypse/ Dan Hendrycks (@DanHendrycks) posted at 7:12 PM on Sat, Sep 27, 2025:OpenAI's recent benchmark suggests that AIs are nearing human-level economic capability.But the best indicator of utility is usage. OpenAI also noted that since GPT-4's launch, usage by early adopters has only grown 1.4x.The actual usefulness isn't increasing that sharply. https://t.co/vvxrNUY6KB(https://x.com/DanHendrycks/status/1972122329787593164?t=y53UwIR3qIcecdepfhUq5Q&s=03) https://x.com/lugaricano/status/1965765898540822820 https://pca.st/episode/f3b124e3-e29a-45c7-b373-fa6efef8e676 From https://thezvi.substack.com/p/ai-135-openai-shows-us-the-money: The investment numbers are even more dramatic. AI investment was already responsible for 20-43% of Q2 2025 GDP growth. Heninger’s numbers imply that AI labs (collectively) would be investing $720 billion to $1.2 trillion by 2027 if they remain on trend — that investment alone would generate 2-4% nominal GDP growth. I think it’s unlikely investors will pony up that much capital unless the models surprise significantly to the upside in the next year or two, but even still, 1-2% nominal and 0.5-1% real GDP growth coming from just AI investment in 2026-27 seems entirely plausible. https://www.exponentialview.co/p/is-ai-a-bubble?utm_source=%2Fsearch%2Fai bubble&utm_medium=reader2 Derek Thompson (@DKThomp) posted at 11:49 AM on Tue, Sep 23, 2025:New pod: This is how the AI bubble could burstInvestor and writer @pkedrosky joins to explain- how AI capex is eating the economy- how financial wizardry pays for a zillion data centers without showing up on Big Tech's books- how it could implodehttps://t.co/rnKFwuWDpv(https://x.com/DKThomp/status/1970561295834341438?t=O8dwl-Iltt1uoXKqMHjCsg&s=03) How sustainable are the current economics of the AI buildout? https://x.com/deanwball/status/1970490882047779111?t=obmfFWrMdMAcXEKYciWdug&s=03 https://www.theinformation.com/articles/openai-spend-100-billion-backup-servers-ai-breakthroughs “AI Agenda: OpenAI’s Spending Spree is Reordering the Cloud Market” “Applied AI: Salesforce, Microsoft Find Selling AI to Enterprises Is Easier Said Than Done” “https://www.theinformation.com/articles/openais-350-billion-computing-cost-problem” https://sundaylettersfromsam.substack.com/p/the-amazon-of-thought Look into Ed Zitron’s writings https://www.lesswrong.com/posts/KW3nw5GYfnF9oNyp4/trends-in-economic-inputs-to-ai https://www.theinformation.com/articles/microsoft-hopes-hastened-ai-rollout-price-discounts-can-fuel-office-365-growth From SemiAnalysis, a pointer to what sound like useful resources from them: We’ve forecasted xAI’s CapEx on Core Research, our institutional research service and are now closely tracking the ROIC of AI investments across the hyperscalers and AI labs in our new Tokenomics model. Anecdote from someone who works at a major tech company: The other day I was in a meeting with an internal AI guru/proselytizer/etc, and the guru was showing off the various tools they could use. He also dropped, as an aside, "your team is fairly low utilization, only 20% ... the gold star is HR, they are close to 100%".  Turns out the metric is 'what percent of the team uses AI on any given day". I immediately told my team: "go out, buy a second monitor, always have an AI tool running on it.  Ask it some sort of question, at least once a day. I don't care at all what you use it for, but what I can never have happen is to be told that I don't get more headcount because my team isn't using AI enough". Another: I have a friend who’s a senior software engineer at Amazon who has said that the mandate and metrics internally work just like that.  She told me that all of the seniors she talks to believe that AI is garbage that is going to take their jobs and have no interest in using it. The juniors are happily using AI. Apparently, aside from the mandate, there has been very little training about how to make good use of AI in her org. Cite the thinking machines crazy pitch meeting is an example. We're starting an AI company with all of the best people, but we can't answer any questions and we're looking for 2 billion on a $10 billion valuation as mentioned. Maybe 10 minutes into the Derrick Thompson. What if it's not a bubble podcast? Overall theme is You know bubble may be defined in terms of things like stock price, stock market valuations. I'm going to be less interested in that. A lot of the signals that we have available to look at are levels of investments. You know valuations things like that a lot of that has to do with people's expectations. I'm more interested in the reality of what's actually happening. Expectations are useful to the extent that you know if we if we believe that in the efficient market hypothesis. If we believe that the people who have these expectations that are acting you know based on their revealed expectations to the extent that we, we think they know what they're doing that provides a signal. Now we have a lot of evidence such as that crazy pitch meeting suggesting that you know people are flying pretty blind. That doesn't mean that this is all going to collapse, but it it does suggest that the signal of the efficient market signal you know make have come disconnected here as it generally turns out to have been an impasse bubbles, so it it is perhaps unnecessary. Though not sufficient condition for a bubble that the investors are just piling in based on vibes and momentum rather than hard analysis of some kind of Rich and Rich data on it. And it's clear that that necessary though insufficient condition has obtained here. What I'm interested in is at the you know there's this long valued AI value chain I'm interested in. What's the level of investment that's going in at the front you know spending on chip spending on data center construction. Probably spending on you know labor costs, salary and interesting question. Whether equity labor cost should be included in that? So what's what's the investment going in and what's the actual value coming out at the end user end either consumer or business end user. You know people draw all these diagrams of like overlapping circles and things, but if we just look at what's going in at the front and what's going out at the back then we can eliminate a lot of that complexity because it just factors out of of this analysis. People also talk about the you know another scent of bubble being creative pricing or creative financing mechanisms. My feeling there is that. Again, that's sort of a necessary or at least suggestive but but not sufficient evidence for a bubble. You know, all that really means is that there's an appetite to invest more than can be obtained through conventional methods or or simply that these investments are getting big enough to no longer fit into the historical business model of the entities doing the investment. You know if Google or Oracle or Microsoft or whatever go around and do these creative financings you know I presume that the kinds of financing they're doing are very standard in some industries. It just wasn't standard in their particular industry, but this is perfectly reasonable because they're essentially getting into an industry they weren't in before. So again, that you know the creative financing is it's a nudge. it's it's a bit of evidence that like something funny might be happening, but it doesn't actually prove that this is a bubble and and so again you know what I'm really interested in is you know is the level of investment warranted in the sense that is it going to translate into into returns in a reasonable time frame that gets into the time frame question, which is interesting. You know even let's stipulate that in some long-term future. We will want the amount of gigaflops billed out That's greater than what's being done today. The question is not whether the world is going to want all this computing capacity. It's whether it's going to want it soon enough to make building it today. Be a good investment you know versus waiting 235 10 years and then building it out with you. Know having saved on time, value of money and also you know using more you know more advanced and cost effective chips and and even energy supplies and so forth. Not to mention. Also just avoiding the the rush Factor. You know paying it paying extra to get land and energy and other inputs quickly when you know you could probably get the same inputs for at least a somewhat reduced cost. If you weren't in such a hurry, you know certainly xai is doing things with like trucking in portable, gas, generators and so on that I I presume are not the most cost efficient way to earn money into into kilowatts. So these are the kind of factors I'm interested in exploring and so we need to dig into whatever data we can find about what's actually being infested to build data centers. Just very tricky because you know things get reported from a lot of different angles, so there's potentially a lot of you know double counting and double reporting, but that's what extent can we tease out? What is the actual total investment in AI Data center construction? I'm also someone interested in how that breaks down across the value chain. You know from Nvidia to TSMC and tsmc's own suppliers and and like the various other inputs for power and so forth and then and this will probably the be the first heart very hardest part to investigate is what's the data coming in at the end user end and then how does that break down across application developers like cursor model providers like Open AI You know Data center operator is and and so forth. Note that I'm not going to discuss potential impact on the economy, crowding out other kinds of investment. You know electricity use of data centers, things like that. These are all important questions but they're just out of scope for me. Justpp just prior to halfway through the Derrick Thompson. What if this isn't a bubble episode? He starts talking. Azeem starts talking about the ratio of of capex to revenue. So in the rail railroad construction bubble I believe he said capex with roughly twice revenue in the dot-com bubble. It was forex and currently in the AI build out. It's 6X you know this gets directly to the kind of question I was presently look at. I do wonder you know where he's getting that 6X figure from and whether it's the you know true end user end-to-end kind of figure that I've proposed looking at then a little bit earlier in the episode he discusses depreciation schedule for GPUs. Yeah, I think it'd be interesting to look at. What is the breakdown of of AI Data center CAPEX how much of that is you know building shell power and HVAC networking equipment GPUs and so on. And you know, what is the how quickly do each of those things? Depreciate how long do they physically last? I've seen at least one passing reference suggesting that you know GPUs might actually physically wear out surprisingly quickly. If they're used intensively to how long do they physically last and and how long do they depreciate just in terms of being you know outmoded you know no longer as as cost efficient power efficient. What have you as as as newer hardware? That inbound revenue can be can take multiple forms. It can be you know customers and businesses directly using products that are essentially pure AI like Chet EBT. Even there it gets interesting. Like you know look at like something like cursor or shortwave. You know what freshen of the you know a significant chunk of the revenues from those products is flowing through into revenue for operating AI models, but not you know, unclear exactly what that percentage is. Then you have you know the the big AI providers themselves using AI internally to power their own products or optimize their own products. So that's everything from you. Know. Google selling Gemini services to Facebook using AI to improve AD targeting or to generate content on their beyond their own platforms. There is some AI, advertising and product you know. Purchase affiliate links built into AI products themselves ; I feel like I had one more variation in mind but and now I can't remember what it was. Azeem says that growing into the current valuations would require revenues to double each year for a while and then slow down. I don't quite get that. I would think they would need to more than double you know you. Can't you know if if revenue sour currently won six of capex then doubling each year for a couple of years? Given the type depreciation of you know the the rate at which these assets age I don't see how even that's enough might ask him to review my draft. Then of course there's the question of AGI and you know how much. How much revenue would you know replacing all human labor bring in? I'm not really going to go into that because you know I'm going to assume that that's outside the return on investment window of today's data center. Buildouts you know it just kind of that's kind of just kind of a separate. All bits are are off thing. So you know if the question is is this a bubble or or is that more deadly defined it is? Is the current level investment a good idea? Like you can just have this separate theory if you want that you know. Maybe the investment is is justified because it speeds the path to AGI and AGI will be so valuable. But even that I would book a scans at because most of these chips that are being deployed are not being used to develop AGI. They're being used basically to the purposes that are ultimately about generating revenue along the way to support the further R&d. And so if the revenue on the way isn't going to justify the chips, then you know then that indirect argument about AGI isn't going to justify either Azeem briefly hand waves toward another another benchmark which is you know. Basically look at the total market for for you know digital services. I think you mentioned something like 1 trillion a year, although I'm not actually sure which scope that that was meant to refer to it. Anyway you know if you can imagine that all digital services are going to be are they're going to start? You know needing AI needing to use AI dreaming, competitive or and or AI might enable you know all those markets to grow. Then you could use that as another anchor point estimate revenues will be going. Although I'm not sure that it makes sense to look at that as an anchor point for the next few years. That seems more like a total addressable market kind of a data point [Child Page: AI Financials; potential for an AI Winter / Bubble Bursting] Lauren Wagner (@typewriters) posted at 5:16 AM on Mon, Sep 08, 2025:A few ideas for what can explain this:1. Businesses are adopting AI tools that don’t actually meet their business needsa/they aren’t adequately assessing product value beforehandb/they may be adopting general AI tools when they actually need more specialized systems (like(https://x.com/typewriters/status/1965026517940904151?t=9g4i_5N-k9korSlaWCCj_Q&s=03) https://www.theinformation.com/articles/openai-says-business-will-burn-115-billion-2029 “Applied AI: Census Bureau Says Corporate AI Adoption is Slowing” [Child Page: Does R1 Undermine the Business Case for Aggressive AI R&D?] The concept of a "lead" is funny when it's easy to fast follow. Is there a world where advanced techniques can be kept truly hidden? If anything OpenAI releases can be copied (for much lower cost) within a year, then all they get for their troubles in pushing forward the state of the art is a brief period of exclusive access8. [Ben Thompson of Stratechery: On the positive side, OpenAI and Anthropic and Google are almost certainly using distillation to optimize the models they use for inference for their consumer-facing apps; on the negative side, they are effectively bearing the entire cost of training the leading edge, while everyone else is free-riding on their investment. Indeed, this is probably the core economic factor undergirding the slow divorce of Microsoft and OpenAI. Microsoft is interested in providing inference to its customers, but much less enthused about funding $100 billion data centers to train leading edge models that are likely to be commoditized long before that $100 billion is depreciated. Or from Hacker News: DeepSeek just further reinforces the idea that there is a first-move disadvantage in developing AI models.] Which isn’t even particularly exclusive, as there are at least three frontier labs (OpenAI, Anthropic, and Google DeepMind) who are spending the big bucks to stay relatively close to one another, even if OpenAI seems to be slightly ahead overall and certainly is best at generating headlines. If you believe that AGI is coming within the next few years and will have value measured in the trillions (as the leaders at the frontier labs appear to do), then spending many billions of dollars may make sense even if it only buys you a few months of advantage. But if the pace of progress slackens, or the economic impact takes time to manifest, this logic could change. An issue here is that the collective experience of the tech industry leads everyone to believe that you respond to competition by innovating harder – we’re not used to a situation where it’s consistently 10x cheaper to follow than to lead. (To be clear, it’s not certain that this is the situation in AI either. But we’ve accumulated a fair amount of evidence that it might be.) Eric Gastfriend captures this idea in a meme: (Via Helen Toner, who adds, “it's not a perfect metaphor - the Chinese companies are working super hard and doing real research to keep up”.)