AIを形づくる5つの論争：収益とインフラ、利用層、主権、規制、データセンター

AI is, to put it mildly, a contentious field. On both a micro and a macro level, it is shaped by debates that will determine how it evolves. From what businesses want to buy, to what and how we should be focused on regulating, these are the most important debates shaping AI right now. The AI Daily Brief is a daily 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, Robots and Pencils, Harbor and Blitzy. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. And if you want to learn more about sponsoring the show, send us a note at sponsors at aidailybrief.ai. A little over a year ago, I released what would become my most popular episode ever. It was called Five Debates Shaping AI. And given how fast AI moves, it now functions almost like a time capsule.

The five debates I discussed on that show were the AI bubble discourse. Will entry-level jobs vanish? Does AI actually boost productivity? Is vibe coding overhyped? And should we accelerate or slow down? Today we're returning to that five debates format, and it's interesting to see in what ways the key questions have changed. The first debate shaping AI right now is a different version of that previous bubble conversation. The fall 2025 discourse about a bubble was completely exhausting. It was driven by a bunch of things, some of them legitimate, some of them a bit less so. For all the real concerns there were about circular financing, the nature of AI deals, the speed at which infrastructure investments were increasing, there was also generally Wall Street looking for something to be nervous about, and fairly dubious sourcing that followed from that, like the infamous MIT quote-unquote study that argued that 95% of generative AI pilots were failing. This year's version of the conversation has matured quite a bit. The first reason for that is that investors have a much better understanding of what we're actually

calculating when it comes to the demand and revenue side of this equation than we did back in September of last year. Giving the sincere AI bears the benefit of the doubt, the multiplication that they were doing was looking at the total number of available seats times 20 or 30 bucks ahead. And it was that math result that they couldn't square with the amount that was being spent on infrastructure. However, the first quarter of 2026 changed the way that most people think about this. It also did so by answering one of our other debate questions about whether VibeCoding was overhyped. The explosion of revenue this year, which ended up with Anthropic actually flipping and surging past OpenAI in terms of annualized revenue, was driven not by individual subscriptions, but by business spending through the API. It turns out that AI is not a seat game. It's a token game. And the upper bound of what a power user can use is not in the tens or hundreds of dollars per month, but in the thousands or even tens of thousands of dollars per month. As that story became clearer, the shape of the bubble discourse changed quite a bit.

As I record right now, the last reported numbers for Anthropics revenue had it at about a $65 billion annualized run rate, while recent reports from OpenAI suggested that their annual recurring revenue had neared $70 billion, although just before I started recording, another report had just come out suggesting that it was actually closer to $50 billion, with the discrepancy being between how OpenAI calculates revenue and Anthropic calculates revenue. Basically, when Anthropic gives their ARR numbers, they include CLAW tokens sold through partners, but do not remove the cut that goes to those third parties, and so some independent investors were trying to calculate OpenAI's revenue in the same way, which is what got them to the $70 billion number instead of the $50 billion that OpenAI has apparently shared with others more recently. In either case, the numbers are astronomical. As A16Z recently pointed out, labs have added more revenue in 2026 than all of public software combined. And yet at the same time, so have capital expenditures. After Q2 earnings calls, 2026 CapEx guidance stood at about $730 billion combined between

Alphabet, Amazon, Meta, and Microsoft, and jumped all the way to $825 billion if you included Oracle. Most of those companies had also increased guidance from where they thought at the beginning of the year. Moody's thinks that this big spender CapEx will approach a trillion dollars next year in 2027. And so the question animating markets now is, even with all that bonkers revenue growth, does it add up to enough fast enough to pay for the infrastructure build-out bill that's coming? Bain has argued that the AI industry needs about $6 trillion of revenue by 2031 to fund all this compute. They argue that today's consumers and enterprise use gets to $1.2 to $1.8 trillion, leaving a gap of about $4.2 trillion that has to come from new markets like AI search and ads, autonomous systems, robotics, drug discovery, and more. Bain found that based on its calculation, it is currently projecting about an $800 billion shortfall. And that's really the main money question that people are trying to figure out. Now, one thing that hasn't gone away fully from last year is the concern about circular funding, that a lot of the projections of revenue that the companies are making

are based on their commitments to each other. Basically, what happens is that every quarter, the hyperscalers show up to earnings and talk about both their current revenue, but also their revenue backlog, the revenue that they're projecting for the future that's justifying all that future spending on infrastructure. The problem is that that revenue is highly concentrated from just a couple of buyers, which are the AI model labs. Earlier this year, the information reported that about half of that $2 trillion backlog at Amazon, Microsoft, Google, and Oracle comes from just OpenAI and Anthropic. And so the bears say if things go badly, everyone is all bunched up together and the risk is much higher. Another type of concentration risk comes from where the revenue for OpenAI and Anthropic is coming. Ramp recently published that 80% of OpenAI and Anthropic's enterprise revenue comes from just 1% of their customers. It is important to note that the data that Ramp has access to is heavily skewed towards early adopter tech-forward companies and startups, meaning this may not represent things overall, but still that concentration risk is enough to give some people pause. I would argue that I think there are multiple interpretations.

The obvious one is that if those 1% of customers significantly change their behaviors, it could be catastrophic for OpenAI and Anthropics business However the other argument is that unless you believe that that 1 of businesses have use cases that are totally dissimilar from the 99 of other enterprises that it looks kind of like the total addressable market that still remains for OpenAI and Anthropics is absolutely enormous. Whatever the case, this debate is now being played out, not just in media outlets, but in the markets. As hyperscaler CapEx commitments have gone up, their ability to fund them off their balance sheet has ended. They have tapped into equity markets and are increasingly tapping into bond markets as well, looking at credit and debt as a way to foot the bill. That again creates a new type of risk that wasn't there when all these companies were just funding things from their own balance sheets. And we have seen some jitters in the bond market that suggest that maybe the days of easy capital have come and gone, and coming soon we'll have quite a bit of opportunity to see just how big the market's appetite for AI companies still is. The blockbuster event for the end of the year is of course Anthropics IPO, which will be followed sometime in 2027 by the OpenAI IPO, with Anthropics seeking a $2 trillion valuation.

So that is debate one, does the AI money math work? Debate two, is AI a mass market or a power user market? One thing that you'll notice in the difference between this year and last year's episode is that the debates that I think are most significant are a little bit more at a macro scale, as opposed to the micro of things like, is vibe coding overhyped? Does AI actually increase productivity? I think especially in the wake of agendic coding becoming such a powerful use case across so many different types of functions, a lot of those questions have shifted pretty dramatically. What a lot of those hype questions, though, have morphed into is about just how far the value extends. Is it possible, in other words, that this is the most incredible technology that's ever been created for high achievers, but not so much for everyone else? Certainly from a reach perspective, it doesn't seem like AI's viability is limited to just one ambitious type of person. OpenAI has recently shared that they are now at 1.2 billion weekly users, a jump from the billion number that they hit over the summer. More than two-thirds of Americans now use AI weekly,

and the percentage of American adults that report using AI every day has more than doubled in the past six months, from 8% in March to 19% in August. And yet, a finishingly small percent of households actually spend any money for AI. In fact, as of the most recent numbers we have, which admittedly are from April of this year, the share of U.S. households with a paid AI subscription is just 2.2%. 98% of households, in other words, are not paying for AI, meaning that those nearly two-thirds of American adults who are using AI every week are either getting enough value from it without paying that they don't feel the need to pay, or in their weekly usage, not getting enough value to consider paying. And, of course, it's not hard to see how this question starts to intersect with the previous debate about whether the money math matters. One slight counterpoint is that the amount that consumers are spending AI is growing significantly faster than the number of people who are paying for AI. Menlo Ventures recently estimated that global consumer spend was going to hit $40 billion in 2026, which is more than 3x what it was in 2025. And just like we saw when it came to enterprise

spend, consumer spending growth is also concentrated in a very slender part of the market. A16Z recently reported that the top 1% of AI spenders now outspend the bottom 50% combined. And that's talking about just the percentage of people who are paying for AI. The top 1% of paying AI users are spending an average of $903 a month, while the median customer spends just $25. Even in advanced technical use cases, there is still some of this power law distribution at play as well. At the beginning of September, Cursor reported that over the previous month, the top 10% of users accounted for nearly two-thirds of all tokens used. So, does this mean that AI is just for power users, or that the industry just hasn't been good at serving other types of users yet. There is certainly some increasing evidence for the latter. 2026, the year of agents, kicked off with the explosion that surrounded OpenClaw. OpenClaw brought the promise of agents to a wide cross-section of people for the first time, but it was also extremely technically complex. Throughout the year, there has been a lot of work done to try to bundle these types of features

in more clear, user-friendly packages, and that has really come to maturity in the last month or so. In mid-August, we got GrokBot, a team of always-on agents that was very similar in some ways to an OpenClaw team, except set up in a much easier user interface. But the big one is, of course, Muse. Muse is Meta's personal agent. And complete with its cute logo, it has become popular quite quickly. Muse, in fact, has sat for the last few weeks at the top of the Apple App Charts in the US. It had dethroned ChatGPT to do so, which is not something that's been easy for apps ever since ChatGPT launched. Muse is not strictly limited to personal use cases. There's a lot of work type things that you can do, but it kind of obliterates the work versus personal line by organizing your Muse solely around you. Muse is certainly not the only personal AI assistant that was getting traction. For example, there's been a ton of buzz around Instinct as well, but its success did make it clear that this was a form factor that every AI company was going to try on. OpenAI joined that party at their Dev Day event with the introduction of Dots.

and many early reports from inside these companies is that these types of features are changing in pretty fundamental ways the way they interact with AI. It is entirely possible that the entire AI industry has been getting away with fairly terrible product experiences for about four years now because the underlying intelligence that they give you access to has made up for the user experience deficits. However, should we continue to see adoption of these sorts of personal AI agents, I think many folks are going to have to update their priors about just how mass market AI can be. A new study from KPMG in the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes. Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs. These top performers, called AI amplifiers, weren't defined by what they knew alone,

but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at kpmg.com slash us slash AI amplifiers. At this point, it's no longer a question of whether companies are actively using AI. Using it well on the other hand is a whole different story Robots and Pencils though is a company that I can point to that is actually built for this time They an applied AI engineering firm working directly with clients on problems that matter to the business not experiments that live in a slide deck. Every engagement starts by working backwards from the outcome a client actually needs. If you're trying to tell real AI engineering apart from noise in this space, that's the difference maker. Head to robotsandpencils.com. Every episode, we cover the competition between OpenAI, Anthropics, SpaceX AI, Google, and Meta. Chances are you've already formed an opinion about who's leading. But every AI lab is taking a different approach, building different technologies, forging different partnerships, and developing a unique ecosystem. Harbor Capital Advisors' AI Lab Ecosystem ETF Suite gives investors a way to gain exposure to the AI ecosystem

they believe is best positioned for success. Search Harbor AI Lab Ecosystem ETFs wherever you invest or follow at Harbor Capital on X to learn more. Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information. Read and consider it carefully before investing. Risks include principal loss and artificial intelligence-related risks. Harbor UTSs are distributed by Foresight Fund Services, LLC. Harbor is not affiliated with AI Daily Brief, and the funds are not affiliated with, sponsored by, or endorsed by any AI lab. This is a paid advertisement and not personalized investment advice. Investing involves risk, including possible loss of principal. Every AI coding tool on the market does the same thing first. It starts writing code. Blitzy does the opposite. Before writing a single line, Blitzy spends days reverse-engineering your entire code base. Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade. The result is a dynamic knowledge graph that understands your software the way a principal engineer would after 30 years in the building. Other tools guess at context with grep searches and markdown files. Blitzy never guesses.

It builds true understanding first, then delivers over 80% of entire software epics autonomously. Validated, end-to-end tested, production-grade pull requests. That's why Fortune 500 engineering teams trust Blitzy with the code bases that matter most. See for yourself at Blitzy.com. That's B-L-I-T-Z-Y dot com. And yet still, at least from a business model perspective, it's clear that for the moment, business spend is the big driver of AI revenue, which is, of course, the key to justifying all of that infrastructure spend. And yet, over the course of this year, there have been some fairly big changes in how businesses think about their AI spend. The first phase was to be all excited about agentic coding, whether it was used by coders or by non-engineers, and to try to incentivize people to go experiment with figuring out how to use this new power for all sorts of valuable use cases. That's what got us the very short-lived token maxing era where you saw things like token leaderboards for who could use the most tokens, although companies very quickly found that that was quite an expensive proposition.

In fact, almost as soon as we started hearing stories about token maxing and token leaderboards, we then quickly thereafter started to hear about companies who were starting to impose token spend limits on their companies. The business models followed next, with companies like Anthropic and Microsoft starting to move away from the inherent subsidies inside their subscription accounts and push power users over to the API, where they actually had to pay for all the tokens that they were using. This has driven not only power users, but many enterprises to care about cost efficiency and performance per unit of cost as just as, if not more important than overall intelligence. Undoubtedly, China has been the early leader in that new efficiency push. Chinese models like DeepSeek V41 Flash and Kimi K3 and GLM53 all offered a different set of trade-offs that pushed many businesses to start thinking about experimenting with open-weight models. Open-weight models not only potentially represent a cost reduction, but because they can run on your own infrastructure, they also avoid growing concerns about whether the AI labs that are serving you the intelligence are going to use the traces of your usage of that intelligence

to ultimately compete with you. And these two themes, the business need for cost-efficient models and the business need for data sovereignty, have fairly dramatically shifted how AI companies are competing. First of all, from the leaders like OpenAI and Anthropic, we are seeing a growing focus on cheaper, more efficient models. We got that with GPT-6 Luna, a cheaper, faster version of GPT-6 Sol. And even more recently, Claude joined this party with Haiku 5.5, an extremely performant model for its cost. In fact, at this point, strictly from a cost perspective, these leading low-end American models are basically as cheap or at least fairly close to their Chinese competitors. And yet, that still doesn't answer the question of sovereignty. Are businesses really worried about the open AIs and anthropics of the world taking advantage of their data to ultimately compete with them? Or is that just the narrative that some have picked up? Certainly that is a story that Microsoft has decided to tell. Satya Nadella and AI CEO Mustafa Seliman have been beating the drum that companies shouldn't

have to pay for AI twice, once with their money and once with their data, and have even introduced new post-training products to address it. And this brings us back to this third debate. Does sovereign AI matter or just cheap AI? If just cheap AI matters, if that's really the primary consideration of businesses, they will likely just take advantage of the cheaper American models rather than deal with the complexity of running your own models on-prem or even post-training them. But if sovereign AI matters, that could be an entirely different situation. Certainly through all of this, the evidence points to the idea that companies are going to increasingly demand more options when it comes to models and are going to be less willing than it might have seemed previously to be locked into a single vendor's ecosystem. Maybe the best evidence of this yet came when Elon Musk recently announced that despite all of their emphasis on training their own models, when it came to their personal agent, GrokBot, SpaceX would be using the best backend model for any given task, even if it meant using not a Grok model. There is no doubt that companies are going to continue to compete

to have the smartest, most capable AI, but almost every other part of AI model training and delivery is going to, in some ways, be dictated by the answers to these questions about how much cheap AI versus sovereign AI and model lock-in actually matter to business buyers. The fourth debate shaping AI is about the right way to regulate it. We're past the point where there's a question of whether there needs to be some regulatory structure. The question now is in the details Very broadly speaking it a question of self versus government regulation although even that may be a temporary state of affairs So far the Trump White House has been focused on voluntary self style approaches In June, he signed an executive order committing AI companies to voluntarily give the government preemptive access to their models to review them before release. And then at the end of September, a slew of AI CEOs signed a voluntary safety pact, saying basically that it was their responsibility to pace themselves if their models presented on due risk. Certainly many within the AI industry have come to believe that we are at

or getting to a point where more deliberate attention is needed. The key phrase of 2026 on this front so far is the idea of pacing the frontier, of companies coming together to voluntarily slow down the speed at which things get released in order to avoid some of the worst potential consequences of breakaway agents in AI. And more and more, those concerns are less theoretical and more based on real-world events. The Hugging Face incident, in which agents from an unreleased open AI model broke containment and figured out how to get access to Hugging Face's servers in order to find results to a benchmark test that they were performing, has been for sure the clearest warning shot when it comes to the types of challenges we're going to face as agentic AI matures. And recently, it's not just been cybersecurity concern, but existential risk concerns that have made big headlines in mainstream media. Former Anthropic researcher Jacob Coxon has been doing a never-ending press tour at this point since he resigned, telling anyone who will listen and give him a platform about the high percentage chance that he ascribes to AI leading to human extinction.

As the Anthropic IPO gets clearer, investors are struggling to figure out how to put a price on this sort of risk, whether it's the rogue AI cybersecurity risk or even more dramatic risks. And in the meantime, we are seeing increasing legislative efforts for more dramatic regulatory action. Bernie Sanders, for example, introduced legislation to ban artificial superintelligence as well as temporarily pause advanced AI development, and we've seen almost endless discussion of kill switch bills designed to create a kill switch for AI systems that are causing undue harm. While right now, because of the White House's stance, the debate might be self-regulation versus government regulation, I think very soon society will decide that self-regulation is too limited, and instead the debate will shift to, which are the most important risks to regulate. Policies designed to improve cybersecurity considerations might look very different, in other words, from those trying to avoid AI disaster scenarios. I tend to think that almost wherever you land on the spectrum, you should welcome this shift as it moves us from having a very theoretical and frankly, fairly disempowering debate to one where what we think

could actually end up in the policy that gets presented. And the last debate shaping AI right now, given the upcoming midterm elections, is can data centers win over their neighbors? Data centers, of course, have become the physical manifestation of people's frustration with AI, as well as, frankly, I think, their frustration with the ability of moneyed interests in general to impose their will upon communities without those communities having a strong say. Opposition to data centers has grown significantly throughout the year. Putting it quite mildly, Pew at the end of September shared research showing how Americans' views of data centers have turned more negative. Since January, the percentage of people who had mostly bad views of data centers' impact on the environment home energy costs, and local quality of life all rose, and very few people want a data center anywhere near their home. This, by the way, is an extremely bipartisan position, with Republicans and Democrats disliking data centers in fairly equal numbers. Now, the AI industry has tried to combat some of the myths that have been driving this discourse. For example, the evidence suggesting

that data centers are driving up people's electricity bills is pretty limited. And yet, surprisingly, people being told that they're wrong about an issue that really matters to them hasn't really worked to change their attitudes. And as their attitudes have hardened, politicians have jumped right on board. Again, across party lines from Democrats to Republicans. And the playbook trajectory for the people building data centers at this point is pretty clear. They started this year with the Ratepayer Protection Pledge, committing once again voluntarily to make sure to buy or build the generation needed for their facilities so that it didn't increase the electricity costs for people in their communities. Turns out that's not enough. The next important thing is that companies have been disavowing NDAs. It used to be common practice that the negotiations that data center builders had with local officials were hidden in secret behind non-disclosure agreements, but that has been exhibit A for people who feel disempowered and not included in this process, and so companies have gotten the memo and pledged not to do that anymore. And finally, most recently, the data center builders have realized that it's not enough

to just make sure that electricity costs don't increase, and it's not enough to stop being opaque, but that they are going to have to invest a heck of a lot more money in the communities where they want to set up shop. That involves infrastructure investment, supporting local community initiatives, and yes, even direct payments to citizens. But will that be enough? It remains to be seen. And the reason that I think this is worth including as a debate shaping AI is that in many ways, data centers are, at least in part, also just a physical manifestation of the larger sentiment around AI, which remains, in a word, not good. And yet, despite the fact that Americans are worried about AI and report not liking AI, they are sure using a heck of a lot of it. The optimistic take there is that if usage of AI is important to them, for whatever reason, but that they have concerns about how AI exists right now, that creates an incentive for them to be involved in trying to make AI and the AI industry better. Democracy is a messy process.

But the fact that we are having all these conversations now, that this has become a political issue, that we're getting specific policy proposals to debate, all of these things I think are massive improvements from the previous state of the debate, which is just people screeching at each other on social media. So those are the five debates shaping AI right now. Certainly there are a lot more. And if you're interested, maybe we'll do a more technical or product and model focused version of this in the future. But for now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always. and until next time, peace.