Analyzing the Token Expenditure Index

Today on the AIDaily Brief, the shift from token maxing to token panic happened so quickly, I'm gonna explain why things are a lot different than a lot of the charts and analysis running around would make you think before that in the headlines, a preview of the upcoming SpaceX IPO AI Daily Brief is aaily podcast and video about the most important news and discussions in AI All right, friends, quick announcements before we dive in First of all, thank you to today's sponsors, KPMG, Section, Zencoder, and Out Systems To get an ad free version of the show, go to patreon dot com slash Ai daily brief or you can subscribe and novel podcasts If you want to learn more about sponsoring the show. Send us a note at sponsors at Ai dailybrief. ai, or you can check out the new Ai dailybrief. Ai slash sponsors By the way, one of the things that we have now on the new AI Daily Brief site, in addition to every episode having a whole page that organizes it into easy to share chunks is a sponsors page where you can go see all of the offers we've shared, like for example, getting a free month of Bolt Pro You can find all of that at aidailybrief.ai slash sponsors. And while you're there, check out the rest, send me ideas. We're going to be adding a lot here For now though, let's talk about the first big AI IPO of the year And we have a bit of an exciting Friday today After months of anticipation, SpaceX is conducting the largest IPO in history Now this has been one of the most hyped up events in markets for a very long time Investment banks have been battling it out for institutional sales, and the retail frenzy is already off the charts As of close of trading on Thursday, Bloomberg reports that retail investors submitted more than one hundred billion dollars in orders Yes, that is billion with a B Now, SpaceX was only selling seventy five billion worth of stock and reportedly reduced the retail allocation from thirty percent to twenty percent That means the retail allocation was almost seven X oversubscribed and would have been enough to fill the entire IPO by itself. The sale was priced atll one hundred thirty five dollars per share, a flat price set by SpaceX earlier in the process. Pricing implies evaluation just shy of one point eight billion Meaning the company will debut as the seventh largest company in the world ahead of Saudi A Ramco, Tesla, and Meta Some anticipate the flat pricing will increase day one volatility, as there was no price discovery mechanism in the IPO process. And much of the commentary has already declared this a retail bloodbath waiting to happen, and possibly an obvious market top for the AI bull run. A rare opinion piece for Reuters declared, There's a serious risk that investors piling into the world's largest IPO will get burned, especially the retail crowd. The analysis focused on the relative lack of revenue for a company of this size. Their twenty twenty five financials showed a five billion dollars loss on eighteen point seven billion dollars in revenue In contrast, Ma delivered two hundred billion in revenue last year, and even a Tesla that isn't at the top of its game managed ninety five billion Even after SpaceX signed megadataetter deals with Anthropic and Google over the past month, they're well short of revenue numbers that put them up alongside those other companies There also has been criticism around the way the company was marketed, with Goldman Sach simultaneously conducting the IPO and providing wildly bullish research analysis In a report last week, they forecast that SpaceX could hit four hundred seventy four billion in revenue by twenty thirty, with their AI division growing one hundredfold To some, this was less about plausibility and more about an analyst with a clear incentive forecasting a bajillion dollars in revenue There is also the sideidhow of Elon Musk on the verge of becoming the world's first trillionaire. Based on Bloomberg's net worth calculations, Elon was worth just shy of seven hundred billionars last month, with more than sixty percent of his wealth tied up in SpaceX The IPO pricing would bring his net worth to nine hundred seventy one billion any significant pop would push him over the line. Now many expect the IPO to be a bit of a circus, but one of the big questions is what will the implications be for the anthropic and open AI IPOs to come Some are saying this is the first chance the US market has to price an AI model company, meaning that if SpaceX does well, it could imply even greater valuations for the frontier labs Then again, there's also the potential with that line of thinking that SpaceX puts in the market top theoretically making it more difficult for open Aye and anthropic to get their IPO's out the door at a premium valuation I tend to disagree with this as the right way to look at things Two big reasons why First is the simple one. You can't really apply anything around Elon Musk to anyone else Love him or lohe them, he kind of operates in his own vortex. And I don't necessarily think that people are going to read this as a referendum on AI models as much as the pricing on the Elon Market Halo Secondly, though, I don't think that anyone is really looking at this as an AI model company I think that the eleventh hour shift to SpaceX as a neo clloud, totally changes the narrative equation First of all, it adds a whole bunch of billions to their revenue stream that look a lot more durable and interesting And second, it makes their whole push to get data centers in space look a lot more aligned and frankly plausible So yes, I do think that for some, the IPO will be AI related, but it's not going to be about models. It's going to be about the infrastructure buildout with a very heavy dose of Elon on top Now as I recorded, it is still early, so markets are yet to open in New York. I will, of course provide full coverage of whatever is important about the craziness to ensue on Monday's show I think though, if you're trying to look for a single take that avoids the hyperbole from either side, Economist Peter Atwater kind of nailed it when he said SpaceX has created an idiot moment for investors. Buy it and it goes down, you are an idiot. buy it and it goes up You were an idiot Speaking of very high net worth dudes, Jeff Bezos's AI startup Prometheus has closed their latest round of funding, valuing the company at a measly forty one billion dollars The round raised twelve billion dollars with participation from JP Morgan, Goldman Sachs, Black Rck, and Bezos himself. Prometheus aims to build what they're calling an artificial general engineer an AI system that can design and manufacture anything, including complex equipment such as jet engines. The company has already staffed up, hiring one hundred and fifty people across offices in San Francisco, London and Zurich In an interview, Bezos said the goal was to, quote, empower engineers to make an invention easier and faster So smaller teams can do much bigger things on much shorter time cycles Asks about fears of an AI jobs apocalpse, Bezos dismissed the premise He believes that AI will instead produce a labor shortage, because quote Even though you're shrinking the number of people needed by ten X, AI will create ten X more opportunities Bezos added There's going to be two earner income households where one earner drops out of the labor pool because there's going to be so much productivity Alongside their plans to produce AI that can accelerate the entire pipeline for physical manufacturing Prometheus is also looking at starting a fund for industrial buyouts Now there is no new news on this front, but in March, the Wall Street Journal reported on talks to raise one hundred billion dollars The fund would essentially take the private equity roll up model and apply it to the manufacturing sector using Prometheus's proprietary technology to improve productivity. All in all, Bezos dismissed the growing pessimism around AI, claiming that view is the quote opposite of reality All societal wealth is driven by invention, he said. Six thousand years ago, somebody invented the plow and we all got wealthier. Then much later, somebody invented the steam engine and we all got wealthier Prometheus seeks to do is to offer a set of tools that dramatically accelerates that invention loop Now for most people, what's interesting is the physical aspect of this As Chubby points out The problem is that the physical economy can't be scraped. There's no internet of manufacturing data to train on, which is exactly why the reported one hundred billion dollars vehicle to buy up legacy industrial companies is interesting You don't find that data, you acquire the factories that generate it Or as Dror Singularity put it, this is how the acceleration escapes the screen and enters atoms Next up, one brutal one, Meta has completed an operational split with Manus in compliance with orders from Chinese officials Bloomberg reports that Meta has firewild operations between the two companies. Manis staff are no longer able to access Metis's data systems, and Meta staff can no longer use Manis's tools for internal work. Now, by way of recap, Manus was one of Meta's flagship acquisitions as they reset their AI strategy coming into the year. They paid two billion dollars for the company just as twenty twenty five turned into twenty twenty six In March, however, the Chinese government opened an investigation into the deal and later barred Manis founders from leaving the country. Manis had attempted to circumvent Chinese tech export controls by first relocating operations to Singapore before courting the acquisition April Beijing ordered the deal to be unwound despite the workaround. This now leaves Manus in a difficult situation to say the least Sources said the company is attempting to raise a billion dollars to fund a buyback, but it's unclear if there are any takers. And while the product has seen updates since the separation was ordered, there is a heck of a lot less attention on Manis, since the rise of open source harnesses like openpenclaw and Hermes, and just in general, the aigantic push of the core harnesses like Clawud Code and Codex In China, the unwinding of the Misel has cast an absolutely chilling effect Manis's strategy of deccamping to Singapore before seeking foreign capital was a very common approach, which some even called the red chip corporate structure With Beijing cracking down on Manis, the Chinese tech industry has received the message that those times are over The Financial Times reports that numerous prominent startups are looking to unwind their foreign corporate structures to rein incorporate in China That Fund has completed the process in anticipation of a Hong Kong IPO, while Kimmi creator Moonshot, as well as Kling are considering doing the same Eugene Wang, an attorney at Shanghai based Wintell and Co said, Whether to dismantle the Rd chip structure is no longer in question. The key is how to complete the restructuring as cheaply and efficiently as possible Unsurprisingly, the AI industry is a particular focus Reports claim that Chinese officials are now seizing passports from key researchers and executives of private firms, which was previously beyond the pale Both capital and talent are facing a major crackdown as Beijing seeks to secure the strategically important industry Now over in Chipland, backlogs at TSMC are driving Google to consider Samsung for parts of their next generation chips The information reports that Google is evaluating Samsung's two nanometer process for some components of their tenth generation TPU's, codenamed Icefish Until now, Google has exclusively used TSMC for the full manufacturing process. However, the Taiwanese chipmaker has a years long waitlist and expects they won't have the capacity to meet demand for quite some time Customers then are beginning to look elsewhere Earlier this week, it was reported that Google had placed orders with Intel for their twenty twenty eight production run. Google will still be using TSMC to produce the actual processors, but Intel will provide advanced packaging services to mate the processor to networking circuits sources said that Google could turn to Samsung for the memory input output die, which marries the processor to memory chips Basically, what's emerging is a complex supply chain O TSMC just produces the processor, which requires the most advanced fabs Other companies including Samsung and Intel are increasingly producing less sensitive components Now, similar to the Intel newews from earlier in the week This doesn't appear to be a case of people being dissatisfied with TSMC's quality. It is simply the case that long wait times are forcing chipmakers to look elsewhere to keep up with demand Meanwhile, private equity companies continue to pile into data center investments as KKR and NNIidia announce a ten billion dollars construction company The new company is called Helix Digital infrastructure and will feature private equity giant KKR and Kuwait's sovereign wealth as capital partners. Nvidia will participate in the venture through the deployment of their chips and related infrastructure, and power company Vistra is attached to provide energy Helix said they have ten billionars in committed capital and have disclosed that they will be a wholly owned subsidiary of KKR. Adam Salipski, the former CEO of AWS is attached to Le the New venture. In a LinkedIn post, he commented, Data centers, power and connectivity have all too often been built on separate tracks That fragmentation has become an industry wide bottleneck This is slowing down the benefits of AI worldwide Now, this is one of several similar deals in recent months with Broadcom announcing a similar tie up with Apollo and Blackstone earlier this week. At the same time, real estate firm JLL, reports that almost half of data center projects around the country are being delayed So is it the case that increasingly bringing chipmakers, utilities and capital providers together in a single vehicle will reveal itself to be the best way to ensure that all of the components come together for a successful project? With Helix, we have another chance to see Finally today Goldman Sachs believes everyone is underestimating the AI infrastructure boom by a fairly wide margin Now, you wouldn't think most people would call the current forecast for AI CapExnd conservative, but that's exactly what Goldman strategists led by Ryan Hammond have done this week. The median Wall Street analyst believes the AI industry will deploy nine hundred twenty billion dollars to build AI data centers next year rising from around eight hundred billion for this year According to Hamm's team, those are rookie numbers In a research note, they wrote Consensus twenty twenty seven hyerscalar CapEx estimates are too conservative Their team now expects one point one trillion in AI spending for twenty twenty seven as a baseline scenario and one point four trillion do in a bullish scenario Now their key assumption is that AI demand is still in the opening innings They expect to see token consumption increase twenty four X through twenty thirty. driven by the widespread deployment of agents. Analysts wrote, highigher input costs also put upward pressure on the nominal dollars of CapEx required to support a given amount of token consumption In other words, excessive demand will keep the pressure on supply chains driving buildout costs even higher. Now you might be thinking that's a pretty bold claim in a week where many on Wall Street are focused on a corporate push to rein in token budgets, but Goldman believes that's just noise, with the signal being the expanding order books for the hypcalers Now as you will see very soon, I am firmly in the Goleman camp on this one, and there is in fact one specific new narrative on Wall Street that I would very much like to take on now One of the most important AI questions right now isn't who's using AI, It's who's using it well KPMG and the University of Texas at Austin just analyzed one point four million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers. they treat AI like a reasoning partner They frame problems, guide thinking, iterate, and push for better answers. And the good news, these behaviors are teachable at scale If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time Learn more at kpmg. com slash US slash sophisticated That's kpmg dot com slash US slash sophisticated. Here's a harsh truth Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized Half of companies have AI tools, but only twelve percent use them for business value Most employees are still using AI to summarize meeting notes If you're the one responsible for AI adoption at your company, you need seection Seection is a platform that helps you manage AI transformation across your entire organization It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value With the result, you go from rolling out tools to driving measurable AI value Your employees move from meeting summaries to solving actual business problems, and you can prove the ROI guessing if your AI investment is working Check out section at sectioni. com SEC Ti O n AI. com Coding agents are basically solved at this point. They're incredible at writing code Here's the thing nobody talks about Coding is may be a quarter of an engineer's actual day The rest is standups, stakeholder updates, meeting prep chasing context across six different tools And it's not just engineers. Sales spends more time assembling proposals than selling. Finance is manually chasing subscription requests. Marketing finds out what' shipped two weeks after it merged Zencoder just launched ZenflowWor It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools. Gira, Gmail, Google Docs Linear, calendar Notion. It runs goal driven workflows that actually finish. Your standup brief is written before you sit down. Review cycle coming up, it pulls six months of tickets and writes the prep doc Now you might be thinking, didnn't OpenClaw try to do this? It did, but it has come with a whole host of security and functional issues, which can take a huge amount of time to resolve Zencoder took a different approach. SoC two, type two certified, curated integrations, Titer security perameter. Enterprise grade from day one Model Agnostic and works from Slack or teelegram Try it at Zenflow. free This episode of the AI Daily Brief is brought to you by Out Systems, a leading agentic Sstems platform built for the enterprise. Organizations all over the world are building, orchestrating, and governing Aentic systems on the Out Systems platform and with good reason OutSystems open and unified platform allows teams to architect, deliver, and scale governed agentic systems with agility. Teams of any size and technical depth can use O Systems to build, deploy, and manage AI apps and agents quickly and cost effectively without compromising reliability and security Without systemystems, you can rapidly launch ideas from concept to completion. It's the leading andic Systems platform that is unified, agile, and enterprise proven, allowing you to accelerate growth, reduce operational friction, and deliver real enterprise impact with AI Out Systems. Build your agentic future Welcome back to the AI Daily Brief A, friends, it's my favorite time of year It's that time when some new set of numbers or in this case chart Inflame everyone on Wall Street to go into an absolute frenzy with their AI counter narratives, proving to themselves finally that this time they're right and the bubble is about to burst. Yes, the speed at which the investors have gone from token maxing to token panic is head spinning And yet The chart in question, this chart, the Silicon Data LLM token expenditure indndex, as shared by Citadel Securities Shockingly I just mean shockingly. It doesn't say what everyone on social media is saying it says In this episode, I'm going to explain why this chart which shows a big scary downward line on something called the token expenditure index has nothing to do with token demand, notothing to do with token volume and nothing to do with actual token expenditure which is not to say that there is an interesting signal there It's just not the signal that Wall Street is trying to look for And why this matters to you, even if you are not an investor, is that the story that it is telling is part of the shift from the token subsidy era to the token scarcity era that we've been tracking does have some interesting implications for how we all build All right So let's talk about where this chart started to come online It is obviously not coming into a vacuum. If you've been listening closely over the last couple of weeks, you've seen the professional investor class really start to take notice of headlines like this one about Walmart capping usage of their internal AI tool because there was too much demand or even more of Uber setting spending caps after it blew through its token budget in the first four months of the year This is the natural follow up to that. And in this context, Citadel just published a research note called tokenomics. The primary chart that it shares is the one that I just mentioned before, the Silicon Data LLM token Ependiture indndex, with that big scary downward line This of course, led to the perhaps expected onslaught of social media commentators, implying that this was somehow some very scary and big deal Failed crypto founder Mos Shake writ Citadel is one of the most significant hedge funds and they just dropped tokenomics. And it's not what you would have expected That scary sentiment, of course, went viral, getting over a half million views There were also endless AI slot posts like this one, from Thierary from RVy, who wrote Citadel securities just put institutional weight behind what the AI bulls won't say out loud When one of the most sophisticated trading firms on Earth starts writing about AI in the language of cost curves and rationing instead of limitless demand, the conversation is quietly changed The hype was about what AI could do. The reckoning is about what it costs. Now by the way, if you're asking me why I say that's an AI slot post, I would point you over to another Twitter post, this time from Nicholas Mugali, talking about a related topic, O AI's plan to cut token prices Weirdly ends with the same line The hype was what AI could do. The reckoning is what it costs And then of course, we have Zero Hedge, who have been nearly quivering with excitement over a new doom narrative to peddle Token priceices down six days in a row, longest streak since January Make that seven days. Tken price index slide back to mid January levels, fading much of the agentic frenzy of the past three months Of course, they made their point explicit with a blog post called Tkenomics equals Panic, and unfortunately, it's not just the zero hedges Real Vision's Andreas Stenner Larson writes, This is the chart that everyone should be watching Token pricing rolls over, everything from the memory trade to the broader hardware and data center trade is over for this cycle, in my humble opinion The whole setup depends on this Now, as you might be able to tell by now, I am absolutely allergic to this sort of pattern of discourse When I see the whole setup depends on this dot dot dot or someone loudly proclaiming something that is so obviously counter to the experience that everyone is having suggesting that somehow the agentic frenzy of the past three months has now returned to some pre agentic state. I immediately start to ask What's actually going on here? What is the chart actually trying to say did Citadel securities who have been held up as evidence of all of this actually even make any of the arguments that people are crediting to them So let's talk about this chart first As I think it's clear from those posts, the implication that people are trying to suggest is some combination of demand for tokens going down Volume of tokens going down Or, and this I guess is reasonable given that it's called the token expenditure index The total expenditure on tokens going down It turns out is not what this chart is actually trying to measure. And I can prove that to you by going to Silicon Data themselves, who took to Twitter to clarify They write LLM token Ependiture Index should really have been named the token expenditure prrice index. because it's an expenditure or usage weighted average token price index It tells you how much currently the entire market AI is paying for a million LLM tokens irrespective of models. The naming might have led to some misinterpretations, as some seem to have interpreted the index as either the total volume of tokens used or the average price of tokens In reality, The index captures something much more subtle than either interpretation It tells us the marginal willingness to pay for LLM models. Much credit to them for trying to clarify, even though the more dramatic assumptions would get their name out there more I even disagree with what they say their chart is saying, as I'll come to in a moment The specific and important note is that what this chart is measuring has nothing to do demand, nor total volume, nor total expenditure What this measures is the average amount The market is paying right now for a million tokens. So what this chart is actually saying and this line decrease. is that in mid June average price that the market was paying in practice for a million tokens had gone down from the peak at the beginning of June and was back at the level that it was paying around the beginning of May. Let me say that again average cost of a million tokens that buyers were paying In mid June had gone down from a peak of what they were paying for a million tokens at the beginning of June and was around what they were paying for a million tokens at the beginning of May. Now, as we'll discuss There is some interesting signal there. But what it's not is signal again on anything related to total demand for tokens, total volume of tokens consumed or total expenditure on tokens It's just about the average price paid for a million tokens Now here's where I disagree with their assessment They say it tells us the marginal willingness to pay for LLM models The idea being If you see the average price paid for a million tokens go down It means that some portion of the market is necessarily shifting their buying behavior From the most expensive basket of tokens two lower cost options And I think that that's partially true Or at least that could be one interpretation of what this data is saying It could also be saying, however, costs for tokens on offer from the frontier have gone down Now we know that that's not the case in this period But as we'll discuss, that might be something that happens coming up soon And more importantly Certainly for any listeners of the AI Daily brief It will come as no surprise that companies appear to be looking for lower cost token options As I have loudly said on this show and elsewhere I think every AI company is now in the token efficiency business The equation is really simple The shift from assisted to agentic use cases radically increases the amount of AI that companies use because of the real constraints of the physical world There are only so many tokens to go around. As demand starts to outpay supply, the prices of that people pay get high. And companies which never had to think about token efficiency or mixed basket models of some types of tokens for one use case and other types of tokens for other use cases, Now all of a sudden, do. That process is exactly what we've been tracking closely for the last couple of weeks And so in many ways, this chart just reflects exactly what we've been talking about The reason, however that I think it's even a little bit less impactful than they think evenven when interpreted correctly And why I don't believe that at least in full, they are right to say that it tells us the marginal willingness to pay for LLM models. has to do with the sources of their data or specifically the data they don't have The Silicon Data LLM token Ependiture indndex does not measure anything about the average prices that people are paying directly to the major labs themselves. They have no insight in other words. into the direct customer to open AI relationship or the direct customer to anthropic relationship. Now, obviously on a percentage basis The vast vast majority of token expenditure is going to be direct to those companies So what source of data could they possibly have? The Silicon Data indndex draws only from third party token routers Wait, you might be sitting there saying You mean the token routers that people explicitly go to use access to cheaper tokens Yes Token routers whose entire purpose in the market is to route different use cases more efficiently to lower cost and better models for their particular need, bringing costs down So this chart, which is neither about total demand nor about total volume nor about total expenditure But just the weighted average price of a million tokens based on data. companies whose entire purpose in the market is to provide lower cost alternatives My argument then is that this is going to greatly exaggerate actual shifts in behavior away from high cost frontier models and towards lower cost alternatives which again is not to say that there isn't signal I think that one could view this as a really good leading indicator of where advanced AI users are oriented In fact, I would argue that we're likely to see some follow on behavior over the course of the next six to twelve months that looks fairly similar, even from the companies that have their direct relationships with open AI and anthropic But what this certainly doesn't reflect is the average experience of buyers in the market right now It just simply doesn't And to be clear, the points that Citadel is even making in this are much less bombastic than those who are screenshotting it all over X In fact, I would bet that if you go read the note, you probably agree with a lot of it and a lot of it will remind you of what we've been talking about here The simplest version of the point that they're trying to make is that not all AI demand looks the same anymore and increasingly will be separated into different categories They put it as a bifurcation in frontier versus everyday AI usage, which sounds not dissimilar to the discussion here a couple of days ago about whether consumer and normal chat GBT usage should even be considered as the same thing as workk AI. At no point does Citadel argue that the implications are some cratering of demand for the most expensive tokens They write We do not think this implies that the frontier of inference intensive AI will be abandoned only that it is likely to be concentrated among a narrower set of firms with the balance sheets to absorb the compute costs the research depth to deploy it effectively And most important, the operating domain to scale the rewards from solving genuinely hard problems Another way to put that is that they're arguing that most of the most expensive AI is increasingly going to flow to the firms that can use it best But in a world of token scarcity, where there's already not enough of the best AI to go around Doesn't that just sound like the market efficiently allocating the most expensive AI to the people who can use it the most effectively That's not an AI bubble popping AI market rationalizing byy the way, despite the construction of that statement, that was absolutely not written by an LLM. It just came out of my mouth, but maybe I'm spending a little bit too much time with the LLMs if now I'm talking like that Now there's another really important part of this discussion. As people have been talking about all of these caps being set. One of the things that gets lost in the discourse is that those are the very most advanced firms who have already consumed sufficient AI to get to the level of agentic usage where they would have to start putting caps The vast majority of companies even close to that And here's one example. Finance company Ramp tracks how their customers spend money on AI They've been tracking for quite some time now. And they note that as the share of businesses using AI approaches one hundred percent The focus at their economics lab has started to shift to tracking the intensity of adoption. They're now tracking things like spend per employee And that spend per employee is a really important number Right now. T one percent of firms, those who are fully AI pilled, are spending about seven and a half thousand dollars per employee on AI That's a lot, right Those are the type of numbers where you're gonna see firms start to really ask what the ROI of that is or start to consider caps and more efficient options But again, that's just the top one percent When you move down to the top ten percent, That number comes all the way down to six hundred and ten dollars per month which you will note is less than half of Uber's fif five hundred dollars monthly per employee cap and the median firm in their index, right firmly in the middle of AI usage across all of RAM's customers who by the way, are going to be more tech savvy than the average business AI Snd sits at eleven thirty eight. one thousand one hundred and thirty eight dollars, eleven dollars and thirty eight cents. If we are looking at the market implications of a shift in the basket of token consumption towards lower cost options We have to contrast that against total growth in token demand and total growth in token volume In other words, actual token expenditure When the median company is still only spending eleven bucks a person on AI, the sheer amount of growth in total AI that will be consumed It is very hard for me to imagine a scenario, at least in anything in the short or medium term growth in the total amount of AI consumed, does not massively and I mean massively outweigh Any shift in the balance away from the most expensive tokens to less expensive tokens Put differently, if every firm followed Uber's example and set the cap at fif five hundred dollars per month The total increase in the market size for AI as firms go from eleven dollars thirty eight cents per employee per month to one five hundred dollars per month is going to dwarf any lost revenue on the other side because companies start to get more efficient. And what about these reports that openAI is considering drastic price cuts as a preemptive strike against anthropic who they think might also cut costs in a vicious price war for customers Will that tank revenue for the whole industry Maybe But then you have to ask about token margins Analyst Max Weinbach wrote, If openen AI does drop token pricing, this is likely because they've heard from customers they can't adopt AI at volume at the current pricing Margin is high now for served tokens. They could cut prices by like sixty percent and still be profitable in my opinion Now Max isn't coming out of nowhere Wellile, no one knows for sure the margins except the labs themselves, Weinback has done a lot of work to get to the unit economics of API tokens And his estimates are pretty similar to a lot of the other estimates that I've seen, which tend to guess something like seventy percent margins on API pricing for the most inference intensive tokens So summing up The argument is not that this token expenditure price index isn't a useful signal. It's telling a similar story to the one that we've been exploring here and that I think will shape the next period of especially entnterprise AI But at the end of the day All of these shifts look a lot more to me like markets doing what markets are meant to do and figuring out how to allocate scarce resources at the right price to different types of customers Look, if you take nothing else away from this If you ever see someone end a tweet with an ellipsis run in the other direction For now that's going to do it for today's AIaily brief. Areciate you listening or watching as always, and until next time Peace