Prime Is (Mostly) Right About AI  🔗 **Source:** [Prime Is (Mostly) Right About AI](https://www.youtube.com/watch?v=VDPMXSAxiWk) So, I'm sure many of you have noticed Ikind of like AI for development work. Ithink it's pretty solid overall. This isa big change from how I used to feelabout it. And over time, I find myselfreaching for AI in my development workmore and more. But not every developerfeels the same way. One in particular isan old friend of mine, Primagen, and hejust put up this video about how the AIeconomy is about to change. This isn'tgoing to be a usual reaction video because Iactually already watched it. But I wantto break down some of his takes becausethere's so much truth to them that Ifeel like the nuance of it beingmissed. But on the other hand, there aresome things that I feel like aren'twell understood around the economics ofall of this. In particular, once Microsoft and Google get mentioned, Ifeel like the details get a bitlost. I want to try to find a middleground that makes sense here, because Primealready is kind of in the middle betweenThe haters and the full AI psychosisfolks. I also like to think I'msomewhere in the middle there. I knowpeople think I'm much more on thepsychosis side, and I can see why. It'sa great video, and I want to do my bestto find the nuance within it. So,we're going to dive right into thisvideo and talk all about Prime's takesright after a quick word from today'ssponsor. You've probably already heardme talk about today's sponsor. It'sBlacksmith. Wait, they ship MacRunners. Okay, I am acting here, but I was thatexcited, if not more so, when I heardabout this because one of the slowestparts of the T3 code build process rightnow is the Macr runners. In order tobuild a Mac app, you need to use a Mac.And since we were building our app forMac, we kind of got stuck on GitHub'sMacr runners, which are really, reallybad. Our whole build process used totake 16 minutes. And a lot of this wasThe Mac version, which would take eight plus minutes. They now take 3 minutes and 34 seconds. And the ARM 64 builds,you know, the Apple silicon ones, would take 6 minutes and 11 seconds on GitHub. And now they take under 4 minutes, 3.32. This made us go from annoyed aboutreleases to kind of excited to do them.And if you're curious why we now ship anightly with T3 code, it's because ofhow much cheaper and faster our buildprocess has become. I'm going to be soreal with you guys. If you're building aMac app on GitHub and you're not usingBlacksmith yet, you're wasting so muchtime. And if you're not building Macapps, but you're still using GitHubactions, you're still wasting timebecause Blacksmith is 2x faster in lessthan half the price. Stop wasting timeand money and build faster atsoyv.link/blacksmith.Let's see what Prime has to say. Thecracks are starting to show. Thefoundations looking a little bit shaky. I'm talking about the AIeconomy, the token economy. Uh, thefirst one comes courtesy of, ofAnthropic, everybody's favorite, thegood guys. Am I right? I love theirlawyers. What ended up happening just acouple of days ago is Anthropic didsomething called a painted door test. Atleast, that's the term I've always heard.I think I've seen other people use fakedoor test where you actually showalternative pricing on your page to see,hey, how much more money could we makeif we charged people more? How many morepeople would leave the pricing page? So,let me give you a hypothetical. So, he'stalking about here is a change thathappened a few weeks ago, where Anthropicwanted to figure out if they could moveClaude code out of the $20 tier so thatyou had to be on the hundred or $200tier to get Claude code usage. The reasonthey wanted to make this change isbecause Claude code is expensive to run,and they wanted to try and save compute. There are pieces he's missing here,though. The first thing he misses isthat this is not the first time theeconomy started to break down here. Iwould argue the first time actually goesall the way back to July of last year.Cursor got hit hard in this case. I'mbringing up cursor because previouslythey build by a number of messages. Soyou would pay an amount of money andthen get a certain number of messages amonth. That sucks because some messagescost a few cents to run and other onescost many dollars to run. Fun fact,GitHub at this point in time stillprices based on number of messages. Theyare also coming around and changing thisat the start of next month. I'veactually had a terminal running tryingto solve a cryptography challenge and ithas been going for almost an hour and ahalf on one message. You get 1,500messages a month on the $40 plan. Entirely unrealistic. And we'll comeback to this, don't worry. But this was,I would argue, where the start of theend was. When Cursor, who doesn't havethe ability to subsidize because theyhave to pay the labs for the inference,like they have to pay OpenAI andAnthropic cash for each request, theycouldn't eat the subsidization becausesome users would do $10 of inferencewith their 200 messages. Others would dothousands of dollars of inference withtheir 200 messages. And Cursor couldn'teat that. So they made changes where your pricing is now more reflective ofwhat it cost to run your requests, nothow many requests you did. So this wasthe first start of the end here. Thenext really big one was when Enthropicstarted changing when you could use theinference. This might not seem like itfits cuz it seems like more inference,but hear me out because this is a reallybig change when Enthropic made it inMarch of this year. A small thank you toeveryone using Claude. We're doublingusage outside of peak hours for the next 2 weeks. This was the first experimentthat Enthropic ran in order to figureout how to price things based on thetime of day. They have a fixed number ofGPUs available, a fixed amount ofcompute that can run requests. So whenthere's a bunch of users during theworkday that are trying to get a ton ofusage out, that means that people whoare running things 24/7, it's moreexpensive for them during that window,effectively. Because if you have 100GPUs and during the workday, 95 are takenup, the last five are going to be seenvery differently than off-peak hours,when you have 50 available. So,Enthropic was trying to steer the powerusers towards off-hours because theywant that compute available during theworkday. This did not go as well as theyhoped. Just a week and a half later,Thoric announced publicly that in orderto manage the growing demand for Claude,they would adjust their 5-hour sessionlimits for free pro and max plans during. Peak hours. Your weekly limits remainunchanged. So, between 5:00 a.m. and11:00 a.m. PT or 1 to 7 GMT, you'll movethrough your 5-hour session limitsfaster than before. This is the keypiece I wanted to highlight. Theprevious thing was them seeing if theycould push us to using off-peak hoursmore. It wasn't enough. So, they had todo the opposite and reduce the amount ofusage we get during specific hours. Togo back to cursor for a second becauseagain, cursor can't subsidize the sameway the other labs can. They don't ownthis nearly infinite amount of computeand GPUs. Cursor is the customer ofanthropic and OpenAI the same way weare. So, they're much more sensitive tothese things. That's why they had tostop subsidizing early. And this is alsowhy they do a lot of internal auditingfiguring out how much compute you getper dollar on these other platforms. So,internally, they measure how much you canget out of a $200 cloud code sub. Previously, you could do up to $2,000 incompute. Now, they're seeing even craziernumbers where you can get up to $5,000of inference in that $200 a month planbecause Anthropic is subsidizing soaggressively. So I just wanted to give abit more history, because Prime startedwith a recent drama which was themtrying to move Claude Code out of the$20 plan. I see that less as trying topush people up to the $100 plan and moreas Anthropic trying to get all of theselow-tier users to stop using Claudecode. Their goal is to claw backcompute, not just to increase profithere. And I really want to highlightthat because Anthropic's growth is crazyin enterprise, which is where they makemost of their revenue. They don't carethat much about people going from the$20 plan to the $100 plan.What they care about is much more specificallythat they are able to get their computeavailable so that they can use it forall the other enterprise customers. Just want to make sure I jump on that, because the enterprise revenue is whatAnthropic cares about. The subscriptionrevenue is mostly just a marketing thing,and the reason they do this crazysubsidization isn't because they thinkthey can make the money back or they'rejust trying to crush and win. It'smostly a marketing play. The issue isthat they're trading compute formarketing effectively, where they own theGPUs, giving it away for these priceshelps them be seen better and be seenmore often. But now that they're low oncompute, it's harder for them to getaway with that. But this is much moreabout the compute than just spikingrevenue. Skipping a bit because I wantto get to the next core point, which is how the costs actually work out as auser and as a business for a companythat is doing inference. Well, they'remaking money off of inference. Likeevery time they do a request call,of course they make money. And it'sLike, yeah, I guess if you measured howmuch a shoe cost by only paying theemployee to sell the shoe, then you'relike, yeah, you're making money off ofevery shoe. They're making money off ofevery request. Yeah, that makes sense,but you're not actually considering thereal cost. Remember, people that usedOpus 45 are now using Opus 46. Nobody'susing Opus 47. A kind of tarted model. No, no one really likes that one. Butthat means all the cost that went into45 if they didn't recoup that ininference cost 45 just cost them money.And this is obviously what's happening.Open AI got a 100 talk about the open AIstuff in a sec. I want to talk about theexamples here with anthropic because forthe most part he is correct. If you'relooking at the cost in simple numberslike you spent $15 in API cost that costAnthropic $1.50 in electricity. They didmake money on that, but there's a couplelayers where that breaks down. The firstis the subscription model where someUsers are using $5,000 of inference for200 bucks. That's a hard enoughsubsidization that they're probablylosing money on it even just from theelectricity costs. Not counting the costfor the GPUs, not counting the cost totrain or the employees or any of theother things, just counting electricity. There are users who are costingAnthropic money purely on electricitydue to the sheer level of subsidizationthey are doing. Estimates for how muchthe compute costs relative to the APIrequests are as low as 15 to 20%. So theraw cost that Anthropic or OpenAI eatsis likely 15 to 20% of what you spend onthe API requests. But again, when you'resubsidizing 20x plus, that washes outand these companies are spending moneyeffectively to give you that level ofsubsidization. And if those numbers forthe cost and electricity sound crazy, Ihave a 5090 doing some fun trainingtasks on the side. And it alone raisedmy electric bill by about $1,000. LikeMy electric bill is around $1,800 a month across my AC, my computers, andeverything here in San Francisco. It'sinsane. Electricity is expensive, andit's cheaper in other places, but it's likehalf the price, maybe. Running a high-endGPU at a full 1,000 watts 24/7 in even the cheapest places in the US, you'retalking $500 plus dollars of electricitya month. Like, it's crazy how expensivethat gets. So, there are other parts here Iwant to break down. Specifically, theamortization of the price over time,where you have to make the amount youspent on the model. There are two pieceshere I want to talk about. I'm going tobreak this down into pre-training andpost-training because what Prime isreferring to here is largely the case inthe pre-training part. There's a reallygood Daario interview where he talksabout this. There's kind of like twodifferent ways you could describe what'shappening in the model business rightnow. So, let's say in 2023 you train aModel that costs $und00 million, um, andthen you deploy it in 2024, um, and itmakes $200 million of revenue. Meanwhile, because of the scaling lawsin 2024, you also train a model thatcosts a billion, um, and then in 2025, youget $2 billion of revenue from that, um.$1 billion, and, um, you spend $10billion to, uh, train the model. So, if youlook in a conventional way at theprofit and loss of the company, um, youknow, you've, you've lost $100 million,the first year, you've lost $800 million,the second year, and you've lost $8billion in the third year. So, itlooks like it's getting worse and worse. Um, if you consider each model to be acompany, um, you know, the, the model thatwas trained in, in 2023, um, was, wasprofitable. You, you paid a hundredmillion, and then it made $200 million ofrevenue. There's some, um, you know, costto inference with the model. Um, but youknow, let's just assume in thiscartoonish example that even ifYou add those two up, you're, you know,you're kind of in a good state. So ifevery model was a company, the modelis actually, you know, in this, in thisexample, is actually profitable. So thisis roughly what Prime was talking aboutthere. If you look at each model and howmuch it costs, and then look at how muchmoney it makes, it comes out as profitable. Usually, if, hypothetically,Sonnet 35 cost 100 million, andthey made back 200 million in revenue. Thatmodel was profitable. Even if they spent400 million to train the next one, thecompany might have lost money, but themodel itself, the money in to the moneyout, worked out profitably. But they haveto keep scaling up the cost of makingthe new model, which only works if theirrevenue scales alongside it. The thingPrime said that I don't agree with,because it doesn't fit this model. Correctly,is that Opus 47 and Opus 46,even to an extent, lost money becausethey weren't as big of a jump as Opus 45. Was and 45 and to an extent 46 are theones people are using. It's possiblethat they're losing money on these modeldrops that get replaced quicker, and theymove on from faster or that don't get asmuch adoption. That is not the case. Thereason why that's not the case is thedifferent types of training that aredone. I'm sure you guys have noticed bynow that sometimes a new model drop justfeels like the last one, but slightlysmarter, slightly faster, slightly morecapable in specific domains. And thenother times, it feels entirely different. Like going from GPD 52 to 53, codeex waslike an oh, it talks to me more, but it'sroughly as smart. And then 53 codeex to54 was like oh yeah, it overreaches abit, but they felt similar. 55 feelsentirely different to the point where itprobably should have been given adifferent name. Either somethingseparate from the GPT line or GPT6,honestly, because the model is sofundamentally different. The reason it'sSo, the reason why it's different is because of thepre-training. There's two main types oftraining with models. You have thepre-training and the post-training. Thepre-training is how you take all of theinformation, all the data you've collected, and bake it into this pile ofparameters that is a model. Then youwant it to behave certain ways. So, thepre-training is how you get theknowledge in effectively, and thepost-training is how you get thebehavior out of it. This is a grossgross simplification of all of the stuffhere. If you're into it, I highlyrecommend looking into it more. It'sreally cool stuff. I'm just trying toget the point across here thatpre-training is the big expensive thingthat costs hundreds of billions, if notbillions of dollars of compute to try and take these terabytes of data andcompress them into this particular shapethat allows you to generate textprogrammatically off it as an LLM. ThePost-training is how you refine the wayit behaves. And post-training has been waymore rewarding than people expected,especially RLVR and RHF, which arereinforcement learning strategies. Turnsout those are really good at making themodel good at things like long tool callruns, long agent runs, doing code, andall those types of tasks. You can RLmodels into getting way better at that. We know this because models likeComposer 2 by Cursor aren't actually anew model. They are fine-tuned versionsof older models like Kimmy K25. KimmyK25 from the Moonshot guys over inChina's open weight model. And sincethe weights are out there, you can RL itand you can fine-tune it externallyyourself. This isn't the thing that likeanyone can do if the model weights areopen. It is also generally speakingoften much less expensive than thepre-training is. So Opus 45 was almostcertainly new pre-training. It was a fundamental change in how Opus works andBehaves. That's why it got three timescheaper than previous Opus models,because they made it smaller than it wasbefore, probably, and ended up making itway smarter than they expected to. Andit is likely from that point on thatfocus 4, 6, and 7 were post training, whichwas less expensive. You can usually seethis just by the cadence of releasecycle. But there are exceptions to this,too. Like some of my research friends in chat are pointing out that GRO 4 spentmore time in compute on RL than they did onpre-training. Even something likeComposer 2 from Cursor allegedly spent4x more on the compute for post-trainingthan Kimmy did on the pre-training inthe first place. So, there areexceptions here, but generally speaking,especially with the big labs, once theydo a new pre-training, the futureiterations aren’t as expensive. They’lloften have five or six of them, and they take the one theylike the most and put that one out. TheIndividual updates to the model oncethey have that pre-training are notthese billion-dollar investments likethe new models were before. So, 55probably cost hundreds of millions, ifnot billions of dollars. 56 Probablywon't be that big of a cost incomparison. Opus 45 probably cost near abillion dollars. Opus 46 is probably a lotless. Just wanted to make that pointbecause it's important to understand. Let's see what Prime has to say aboutOpenAI here. OpenAI got 122 or $120billion investment, and that's enoughmoney to run for 18 to 24 months. Thatis like 5 to 7 billion every single monthin the hole for the next 18 to 24months. They had to do this test becausethey have to know how much money theycan make, because if they don't make somesort of change, they're going tocontinue to lose billions of dollars.Now, if this just happened by itself,I'd say, "Hey, Claude just needs to makemore money. They need to start becomingMore competitive cuz they're competingagainst OpenAI." And then I'd be like,"Okay, that's just that." Butthis isn't the only case of thishappening, because just a couple daysago, guess who else decided to make abit of a price change? Microsoft. Oh,beautiful Microsoft. I love Satya. Ilove Co-Pilot. Isn't Co-Pilot this? Is this the greatest? You're probablyconfused when I say Co-Pilot becauseyou're probably thinking of like one oftheir 50 services. I’m talking aboutthe GitHub, the GitHub Co-Pilot in thiscase. See, the GitHub Co-Pilot, you usedto pay some sort of amount of money, andthen you could perform actions onbehalf. You had some sort of amount ofactions you could execute. Well, what'sthe problem with that? Not every modelcosts the same. That's not the problemwith this. Yes, different models costdifferent amounts, but you can makecertain models two or three times asmany messages per message. The problem, And this is actually the point thattriggered me enough to make my responsevideo. It’s not that some models areexpensive and some are cheap. It’s thatpricing the number of messages makes nosense at all. Imagine if Walmart pricedby the amount of items you had in yourcart. Or if they priced by how manypeople walked in. They charge you whenyou walk in. You go take whatever youwant and leave. Some people go in andtake two pieces of candy. Some people goin and grab five TVs. If you’re pricingby the number of items, that doesn’t work. And the worst part is that Copilotalready is doing this. If you go intoCopilot and you pick a different model,it’ll tell you what the multiplier rateis for different models. For example,here 54 is 1x, but 55 is 7.5x. That’sthe number of messages it uses. So onthe $40 plan, which is what I’m on withCopilot, I believe I’m being compedbecause of open source. So, thank you toGitHub for that. They give you 1,500Messages a month. The problem is thatsome messages cost pennies to respondto. Whether it is Opus or 54 orwhatever, doesn't matter. What matters is what it is doing and how long does ittake? And remember earlier I said I wasrunning this request. This request isnearing 2 hours. I sent one message andit has been going for 2 hours. This onemessage is going to use upwards of $100of compute by itself. This makes nosense. You cannot charge based on thenumber of items in somebody's cart. Youcharge based on how expensive the thingsare. The problem with Copilot and thereason they're making this change nowisn't because the subsidization is over. It's because Microsoft is just a slowermoving company and it took them longerto fix this. That is all that mattershere. And the 7.5x part actuallyemphasizes the point I want to make,which is that these numbers and thesecosts are not based on what the APIprices are or what the labs areCharging. These prices are based on howmuch compute they have available to runthis model and how expensive it is forthem to give that compute up to thesubsidized co-pilot users. Historically,C-pilot users were light enough usersthat they could just eat the cost. Ifsome of those $40 a month users cost them 800 bucks, whatever, it'sMicrosoft. They make money. They canafford that. The problem is there'senough of them now. And it's not justmoney that they are spending. It'scompute they are losing. And they arecompute constraint. They are fighting toreserve as much compute as they can onAzure for all of the other deals for allthe other companies they're doing. Theyneed that compute. It's not just likeyou're costing them too much money whenyou use co-pilot. It's that they need theGPUs that Copilot is using available. The bigger deal about their announcementisn't that they change the pricing. Thebigger deal is that they paused signups. Entirely. You don't pause signupsbecause you want to make more money. Youpause signups because you don't havecapacity. And this is the really bigpoint I want to drive home. The problemisn't that there isn't enough money tospend. The problem isn't that they arespending too much on these users ofthese $20 plans, and they want to get youto spend more. This is not thetraditional up to the right problem thatit is being made out to be. This isn'tthese companies trying to squeeze morepennies out of you as the user. Theydon't care about you. All they careabout is their enterprise customers thatthey make actual money off of. And theydon't have enough Nvidia graphics cardsin their [ \_\_ ] server farms to affordthose customers and to sell them whatmakes them actual money. Do youunderstand how crazy it is that GitHuband Microsoft had to pause signups? Thatis not revenue clawing. That is not adesperate attempt to make more money. That is trying to reserve compute so youstop losing enterprise customers.That is the real story here, and I want tomake sure that is emphasized. You'rethinking about composer too fast fromcursor. That thing costs like nothing.When you're thinking about Opus 47, thatcosts a lot more. We're talking like 20times more the cost. So, obviously,co-pilot, they had to make some sort ofreduction, and that's what they did. Youno longer just get some amount ofexecutes. >> I hate to skip over this part, but it'sjust entirely wrong. Like I, again, thecore points Prime's trying to make hereare right. He's just not in the weedsenough to know these details. And theusage base change isn't, they're tryingto claw money. It's, they're trying tomake sure that users of a $200 a monththing, or a $40 a month thing, or a $20 amonth thing, aren't using so much computethat they don't have any available fortheir customers doing enterprise [\_\_]. As my research friends in chat arepointing out, capacity is going to be ahuge problem. I am very scared. Chipscan't be made fast enough. That is theproblem here. The next section is wherePrime starts comparing Microsoft to theother companies in the space and hebrings up one that's very interestingthat is also very different. See thedifference between Microsoft andAnthropic is Microsoft makes money. Iknow crazy concept for Anthropic, butMicrosoft makes a lot of money. Whetheryou like it or not, they're one of the biggest companies in the world. Andtherefore,they can actually kind of take like abit of a nose dive for a whileaccumulating users. But even at thispoint, they're going, "Hey, we can't wecan't do this. We got to make more. Thisthis is this is silly." This is the samemistake as earlier. I also would arguethis is a misunderstanding of theAnthropic change before, too. TheyDidn't change the potential availability of Claude Code on the $20 a month plan to push people up. What they would do is heavily restrict how much cloud code you get on the $20 a month plan,so you're incentivized to upgrade.They don't care about those subscriptions.All they want is more enterprise customers.That is what the chain should be seen as.The removal of cloud code on the $20 a month cloud plans isn't them pricing you up.It is them trying to slow down the firehose of compute being wasted as more and more users are signing up every day.It's the exact same thing as the C-pilot pause. Think of those the same way. The $20 month change for Claude Code and the pausing of signups for Copilot are the same.This isn't because Microsoft was burning too much money, and they're clawing it back and trying to make more now.This is because they both have the same problem.They're both out of compute because neither Microsoft nor Anthropic are Nvidia.They don’t have the infrastructure to support unlimited demand. Access to unlimited compute. Even Nvidiadoesn't really. They can justmanufacture it, and they're limited bythe companies they rely on to actuallyfab the silicon, and [ \_\_ ] there is notenough compute in the world right now.That is the problem. We are working witha limited resource, and the limitedresource isn't money.And if we keeptrying to make it about money and notthe amount of compute available, you'regoing to start going down conspiracytheory paths that just aren't true. These behaviors do represent the end ofsomething, but this isn't the end of thesubsidy economy. This is the start ofcompute restrictions really affectingthe way companies think. That is the keydifference here. Kic just linked me thistweet from two years ago from Sam. It'sactually almost three years now. We arepausing new chat GBT plus signups for abit. Brown frowny face. The surge inusage post-dev day has exceeded ourEveryone has a great experience. Youstill sign up to be notified within theapp when subs reopen. Again, OpenAI wascompute limited, so they bought literallyall of the compute they possibly could. It started coming in last year and thisyear, and now they don't have theseproblems. Not because they have moremoney than the other companies orthey're more willing to subsidize thanthe other companies. They just have morecompute. This all comes down to computein the end. All of this, even Opus 47being dumber, is arguably because ofcompute. What's going on here?This doesn't even make sense according toMicrosoft. Now, the real winner, honestly,from all of this, is Google.Classic Google. They are pouring like a hundredplus billion dollars a year into AI,and they can just do that. And, guess what?After they pour a hundred billiondollars, two hundred billion into AI, theystill make money. That's wild. Like, theycan do so much money for the next year. After a year, and they don't have to worryabout will investors still find meattractive. That's probably why you'renot getting the same level of hypecoming from Google that you get fromthese companies. I just think this wholepart is just entirely wrong. That's nothow it's working for Google. Google isproviding more free compute than anyoneright now by far. By [ \_\_ ] far. I'llgive an example. I'm going to Googlesearch for the primogen and it justshows me his links. But watch whathappens if I ask who is the primogen. Now we get an AI overview. This is freecompute. Watch what happens when I do iton a signed-out account. Oh, it stilldid it. It still did it. These AIoverviews are free compute. How did heget popular? Now I am doing literal freecompute on google.com. Signed out incognitobrowser doing this for free. The peoplesaying, "Oh, isn't it cached? Why am Iable to do follow-ups? This is a fullbuilt-in chat interface as part ofGoogle. They do a shitload of freecompute. And I have been told by somany people that they're usinganti-gravity because of how aggressivethe subsidizations were. In fact, theGoogle models are so bad that theyincluded Opus 45 in your Googlesubscription that you could use inanti-gravity. And I knew a number ofpeople that were using anti-gravity justfor the subsidized opus usage within it. And that ended up exploding so hard thatGoogle was the first company to startdoing the restrictions for where you canuse the inference, where like people whowere even just building plugins to tracktheir anti-gravity usage started gettingbanned. People who were doing plugins tolink up anti-gravity with open code and[ \_\_ ] all started getting banned as well. They've been aggressive with thatbecause they were too generous initiallyand have had to walk it back since. Thereason you don't think of Google assubsidizing is because they did it soAggressively, initially, they had to clampback first. And Google models are so[ \_\_ ] trash that nobody talks aboutit. So, this is just an entire like blindspot for many people in the developerworld. Google is doing more free computethan anybody else. They were subsidizingharder than anybody else. They had toclamp back faster than anybody else andmore aggressively than anybody else. Google's arguably the most extremeexample of all of this,and they maketheir own compute because they maketheir TPUs,and they are really behind onthat to the point where there are rumorsthat they use CPUs both for training andfor inference because that's what theyhave around. Google is as computed ifnot more so than the other companies,and has acted accordingly. I don't blameanyone for not being in the weeds onthis [ \_\_ ] because you have to pay a lotof attention to notice these thingshappening, but this is just a big miss. Google is as good of an example as youCan get of the things I'm talking about here. The subsidized compute era isn't over, but it is closing because theamount of compute available is less andless. The real story isn't that yoursubscriptions are being limited or the$20 a month tiers are getting morerestricted. The real story is that RAMprices went up. That is what this reallyshould be focused on. The cost problemhere isn't one related to what you'regetting for 20 bucks. It's one relatedto how expensive the compute has gotten. Think about that for a second. Google'salso competing on the frontier. Google'salso attempting to win the market.Google's also trying to convinceeverybody that their AI is the best AI,that they're going to be able toshepherd it and take it into the future,despite the fact of inventing the T andGPT and somehow fumbling the bag and notbeing the first one to market. Nobodyknows the answer to that one. And AI isreally expensive. Uber just got done. Claiming that within 4 months they spenttheir entire year's budget on AI. Gee, Ican't believe this is happening. Howcould you tell every employee tomaximally use AI? By the way, we'rejudging you on AI usage. Oh my gosh,you're using too much AI. How'd thathappen? What the hell? You're notsupposed to be using a year's budget in 4months. What are you even doing? Whatare people doing with all those tokens?And here is where the point I've beenmaking the whole time comes in. Uber isnot using the $200-a-month subscriptionson Cloud Code. They're certainly notusing the $20-a-month ones. Uber isusing inference through the API. So theyare paying directly to Anthropic andOpenAI. The API costs are what are beingtalked about here. They are paying fullrate. My usage of these tools is usingthe personal subscriptions which are waymore generous. I am probably doing similar amounts of inference to theaverage employee at Uber. The differenceIs, I pay 200 bucks a month for it. Theypay 2,000 bucks for it. That's the keypart here. I have heard crazy numbersfrom other companies that I'm hopeful Ican share in the near future from insidesources in many different places. I knowcompanies where there are engineersspending more on compute and spendingmore on inference than they are paid insalaries, like pretty often, because thecompanies have to be under differentterms because a lot of thesesubscriptions explicitly say you can'tuse these when we should be negotiatingan enterprise contract. By the way, thatone message I sent earlier is stillgoing. We're now over 2 hours of thismessage going. And if you're wonderingwhy you can't use the sub for yourenterprise job, I just asked outright. Here you go. The consumer terms excludecommercial offerings by design. Theyopen with a notice clarifying thecommercial terms govern API keys. Theconsole and other offerings thatReference them explicitly cover thatdoes not include Claudia or Cloud Profor individuals. Cloud Code's own legalcompliance page confirms the split. Commercial terms apply to teams.Consumer terms apply to free. Theordinary individual usage assumption.The Cloud Code legal page states thatadvertised usage limits for Pro and Maxplans assume ordinary individual usageof the cloud code in the agency. This isa soft cap on what counts as legitimateuse. There’s also a whole section aboutbusiness domains and non-commercialevaluation. Sign up with an email ownedby your employer. Your account may belinked to your org's enterprise yadayada yada. Where Enthropic permitslimited evaluation. The evaluation useis restricted to personal non-commercialuse only. They also prevent you fromusing it to make competing products,which basically everyone is at thispoint, like everyone is. And the biggestdeal is that when you're on theSubscription tiers, your data can beused for training even after opt out. Flagged content and feedback can stillbe used. There is no ZDR. There is noproper isolation of your data. And mostbusinesses will never let you get awaywith that. BAAS only extends to cloudcode if the customer has executed a BAAand has zero data retention enabled.Both of which require an enterprisecontract. Max plans cannot be HIPAAcovered. They want you on enterpriseplans. If you're doing enterprise work,just be realistic here. So again, ifyou're at an enterprise, you're notpaying the 200 bucks a month. You'repaying for your tokens, and your tokensare much more expensive. Just want tomake sure you understand that because a lotof people don't. >> I don't even know. Like I don't thinkwe're going to be like, "Well, back itup everybody. We're not going to use AIanymore. You're going to have to go backto hand coding everything because AI isJust not economically viable." No, no,no, no, no, no, no, no, no, no. Theywill find a way. Sure, it may be in acouple of years, but you know what? They'regoing to find a way to make this thingviable, but for now, we're starting tosee the cracks. Things just can't be asfree as they once were, and the amountof usage you're going to be getting isclearly and obviously going down. Idon't want to be the person's like,"Oh, Mr. anti AI for all reasons. Likethat's why I make this video to show youhow ridiculous their marketing is. Thisis why they do such hyped out marketing. This is why Daario is constantly tellingyou, "Hey, you're out of a job here veryvery soon. You, you, you. Yeah, we're goingto take your job." I feel super badabout it. Oh, I, I, I want to talk a bitabout the point he just made beforeabout the like amount of compute we getgoing down. I really don't see that. Ifyou measure it purely by dollars, sure. If you measure it by tokens to the bestFrontier models maybe, but the cost ofintelligence is consistently going down. A good example of this is the artificialanalysis intelligence index. GBT55 is amore expensive model than prior. They 2xthe token costs, but they also made itmore efficient. And if you look at thehighest end X high versions of thehighest end models, 55 is smarter than54. 55 got 60 points and 54 got 57. Itis smarter. Where this gets much moreinteresting is when you turn on 5.5medium. When you put 5.5 medium in thegraphs, you'll notice 5.5 mediumperformed roughly as well as 5.4x highdid. Same score there. But what's muchmore interesting is the actual cost torun. This isn't the cost per token.People get too crazy about thosenumbers. To do the grocery store exampleagain, if somebody was buying grapes andyou wanted to charge them, you couldcharge them by the number of grapes orby the weight of the grapes. And we arelooking at the cost per grape. Effectively, when we look at the tokencounts, but the actual cost is not justhow expensive it is per token. It's howmany tokens were needed to solve yourproblem. And that shows very interestinginformation here, where GPT-54 and GPT-55xare high. Despite, remember GPT-55 is 2x moreexpensive per token, it used fewertokens that it only cost like 20%more, from 2,850 bucks to 3,350bucks for the exact same data drop run,everything. And Sonnet, which is cheaperthan GPT-55 per token, ended up beingeven more expensive because it used waymore tokens. And obviously, Opus isbasically double the price at $5,300. But here's where things get cool, wherewe don't have to worry about cost goingup a ton. If you were happy with howsmart 54 was, you're in for a treatbecause 55 medium is just as smart andcosts less than half as much. 55 mediumcosts $1,200, where 54x high costs $2,800.That is what's exciting here. While thefrontier continues to get more expensive,As we're pouring more compute into it,at any given level of intelligence,prices are dropping fast. 54 happened a few months ago, like two months ago, not even.And 55, a month and a half or so later, took that level of intelligence,cut the price in half,and then offered more intelligence for 20% more money.And if you're willing to go a little bit lower and want to try out 55 low,which is honestly quickly becoming my default,it only cost $500 to run this. That is a sixth the price of 54,and the intelligence level isn't that far off.Like, yeah, it's dumber,but it's still 51 points.That still puts it up there with Claude Sonnet 46,which was way more expensive.And again, if we compare the number of tokens,things are very entertaining because Sonnet 46 max did 200 million tokens to run this benchmark.Opus was more efficient at 110 million,but those tokens are very expensive. 54x high did 120 million tokens,so slightly more than Opus.When we look down here,55 was down to 75 million. Massive tokenreduction. 55 medium was only 22 millionand 55 low was the lowest amount oftokens used of any of the things we haveselected at 7 million tokens to scorehigher than deepseek 4. The models onthe furthest left and right in thischart got the same score. The one thatused the most tokens, Deepseek V4, andthe one that used the least tokens, 55low, got the same score. That's what's socool. The labs that are putting theireffort in the right places are making things more efficient. There is effortto make these things cheaper. You mightnot see this because you look at howexpensive the subscriptions are, howmany messages you get, or how expensivethe tokens are, but when you look at howexpensive real work is to complete, thecost is going down rapidly as long asyou're not using anthropic models. Soagain, just to emphasize how silly and[ \_\_ ] all the numbers here are andthat like the way things are priced toYou is not based on the amount ofelectricity that it costs, but it'sbased on the availability of compute. We've now established that 55, despitebeing 2x more expensive by tokens, endsup being a lot cheaper per run,especially if you use it on medium orlow. It ends up being way cheaper than54. To go back to our favorite examplefrom earlier, Microsoft, why is it that55, which worst case is 2x moreexpensive, but generally speaking isactually quite a bit cheaper than 54,why is it a 7.5xmessages per request when 54 is 1x? Ifthis was about the cost to run, thesenumbers make literally no sense at all. The reason these numbers are here inthis way and that Opus is 15x is becausethis actually comes down to how muchprovisioning is available. How many GPUsare running 54? How many are running 55? How much demand is there for thosedifferent clusters? How muchavailability is there in those clusters? 55 costs more opportunity for Microsoftto serve because they're trying to sellthe 55 inference to companies andenterprises right now. These numbers arenot meant to actually make more orless money. They are meant to get you asan individual paying 40 bucks a monthout of the way of random Fortune 500companies that are paying way more pertoken. That's all this is. And if youthink of this any differently than that,you are falling for conspiracy theory. That isn't actually how any ofit works. And by the way, we're now wayover 2 hours and this one request isstill running and my co-pilot usage isstill at 0.4%. So if you think that thischange that GitHub Copilot is making isthem being evil and greedy, you justdon't understand how these costs work.You just don't. And that's fine,but don't complain about things youdon't understand. Go understand it. So Iwanted to jump on that because it's justa very, very common misconception thatI've seen around a lot. I just feel sobad about me taking all of your money. I'm so sorry. It's so dangerous. Butthis is the reason why they're doing it.They need to raise money. This is whyGoogle doesn't do it. They don't need toraise the same kind of capital that uhDaario and Sam need to raise. >> Yeah. Again, I just this is wrong.Google was subsidizing really hard.Google and Microsoft's problem isn'tthey have more money so they couldsubsidize longer, and Google chose notto. Microsoft chose to. It's just thatenterprise, these things take a longertime to do and do right. Microsoftdoesn't want to take hits to theirreputation. So they extended co-pilot'sgenerosity for longer than they shouldhave. Google doesn't really know whatreputation means, so they made changesway quicker that made their already badproduct basically unusable. That's theother piece here is that co-pilot beingpart of VS Code makes it solid. Generally speaking, the C-pilot developer ecosystem is okay. It's not myfavorite thing. There are otherharnesses I like much more, but theCopilot CLI is totally fine. This is notthe case for Google solutions. Google'sCLI is not great, and Google's IDE isabsolute garbage. So, Google didn't havemuch to lose there. But I feel like allof these nuances were missed becausethere's so much nuance here. There is somany things you have to be tapped in onto understand what's going on. Andagain, I do not fault Prime for notunderstanding these things and notcommunicating these things because hiscore point still stands. Things arechanging and we're getting used tothings working one way. It is gettingdifferent. But I also think they'regetting better in a lot of ways, too. The fact that you can pay a hundredbucks a month and have enough compute tobuild like three full startups as anend user is unbelievable. The fact thatThese models are getting so much moreefficient and smarter is incredible. Myprevious video where I said that we werehitting the ceiling for what models werecapable of is so stupid in hindsight, andretrospect, that's actually kind offunny. So, I have made much dumber flubsand much less informed videos andopinions than anything I've been talking about, but I ended up going way deeper inthe weeds and learning all of thesethings. So, that's why I wanted to jump on this. I still think you should gowatch the original video. It's linked inthe description. While I overall agreewith Prime's analysis here that theeconomics of all of this are shifting,and changing, I think it's a lot lessabout how much money they can squeezeout of you as an end user, and it's muchmore about the amount of compute theyhave available. That's all I have to sayon this one. I hope this breakdown washelpful, and until next time, save yourGPUs., Prime Is (Mostly) Right About AI, YouTube, 01kqw01nbpm9sfdpxenbk96cjr.