Original Title: The Economics of AI Usage and What's Next For SaaS | Benedict Evans on a16z

Original Translation: Yanlin Hang, Z Finance

In June 2026, Silicon Valley. Anthropic's annualized revenue has more than quintupled in the past 12 months, reaching $47 billion. The four major tech giants' AI capital expenditure guidance for this year exceeds $700 billion, nearly double the total investment of the global telecommunications industry.

As the entire industry convinces itself with “the risk of underinvestment is greater than the risk of overinvestment,” Benedict Evans discussed the mobile data crisis of 2008 in the podcast studio of a16z.

Those were chaotic years following the release of the iPhone. AT&T launched unlimited data plans, users went crazy watching YouTube, and the network collapsed in an instant, with operators spending hundreds of billions to expand capacity. Ultimately, all the cool applications were built by others, and operators only earned "pipeline fees."

Evans releases a talk each year titled AI Eats the World, regarded in Silicon Valley as an important reference point for observing technology cycles. Unlike most optimists, he tends to look for bad news in history. In his view, the gap between today's $20/month ChatGPT subscription and the underlying token costs of over $10,000 is based on the same illusion that generated those $500 billion data bills back then. Pricing is severely disconnected from costs, and everyone pretends not to see it.

“All bets are still open.” He admits he cannot predict the outcome. But when the efficiency of models improves by 100 to 200 times each year, and nearly one trillion dollars of capital floods into this sector, he believes at least one thing is certain: today’s luxury of “ROI-based pricing” will not last long.

The following are Evans's six judgments on the core contradictions of AI economics:

1. The foundational models are not products, and value will eventually shift upstream. Model companies are likely to become mere sellers of water, repeating the mistakes of chip manufacturers, ISPs, and mobile operators. They build amazing infrastructure but fail to capture the most profit.

2. Programming is currently the only field that has really found PMF. Agentic Coding has transitioned from “somewhat useful” to “game-changing,” but aside from that, most scenarios still linger on the fringe of “opening ChatGPT once a week to try it out.”

3. The pricing system is collapsing. As model efficiency improves by 100-200 times each year, and CapEx floods in at a trillion scale, today’s luxury of ROI-based pricing cannot continue. Tokens will inevitably fall into a commodity price war like mobile data.

4. AI will not end SaaS but will redefine the boundaries of software. Where should probabilistic LLMs be placed in the tech stack, on the top or bottom? Enterprise software will enter a new round of chaotic games between “Excel vs. dedicated software,” with more software, more competition, and more uncertain profit margins.

5. History can only explain, not predict. Analogies from the mobile internet, cloud computing, and the PC era are useful, but none can tell you whether OpenAI will become the next Windows or the next Netscape.

6. The real questions are moving out of the tech circles. What does AI mean for law firms, investment banks, consulting firms, and Hollywood? The answers are not in San Francisco but in the hands of industry insiders who know “what junior employees are really doing.”

01 OpenAI and Anthropic strategic differentiation

Erik Torenberg：Benedict, welcome back to the a16z podcast. The last time you were here, we discussed the first version of your talk, AI Eats the World. Now that you have been working on it for nearly a year and a half, you always start your talks with "what are the big problems". This time, I would like to ask first: what have we learned since you initially gave that talk? Which predictions came true? Let's review.

Benedict Evans：Let’s talk about what happened over the past year. I believe we have seen a clearer differentiation in product strategy, and we have observed competitive tension — this competition is no longer just about “making models bigger, faster, investing more computing power”.

OpenAI's strategy has gone through several shifts — from “putting all bets in every direction at once” to “maybe we should double down on the programming area.” Clearly, Agentic Coding has started to really pay off. Consequently, all focus in the tech world is highly concentrated on this area, which has achieved absolute product-market fit, with customer demand so strong it’s nearly overwhelming. Of course, this has also brought about the issue of supply shortages, around capacity, pricing, supply-demand imbalances, and capital expenditure pricing — which is the current situation we see. This is the point we are in now — previously, we thought it was interesting and exciting, but we were not completely certain about what it could be used for. Now, it can indeed be used for programming. Whether it can be used in other areas, the answer is almost certainly yes, but programming is where it really works right now.

Now the focus has narrowed significantly. In addition, data is continuously rising: models are getting larger, CapEx is constantly increasing, usage is on the rise, and people are using it more and more. But the fundamental questions you raised two or three years ago still largely remain unanswered.For instance, we do not know if there will be a single absolute winner in the model field, we do not know whether they can gain value at the top of the value chain, we do not know where the boundaries of model capabilities lie, and we do not know when consumers will transition from weekly to daily usage with current technology. So, all of these questions remain unresolved.

Erik Torenberg：Speaking of programming, did we foresee that it would become the first truly exploding application scenario?

Benedict Evans：If you look at it from a deterministic angle, you could argue: Who is most interested in tinkering with these things? Software developers. And what do software developers want to do with these things? Of course, it is software development itself. So from this very naive perspective, software development is the highest priority. I often compare this moment to the internet in1997, 1998, or the personal computer era at the end of the 1970s and the beginning of the 1980s — everything was very exciting, but it was still unclear what exactly it was for, because it hadn’t truly matured. In the early days, the primary thing people did with personal computers was to create more computers, and now the first thing people do with LLMs (and larger LLMs) is also to create more computing power. So this is not surprising.

However, a noticeable shift occurred at the beginning of this year:Agentic Coding transitioned from “somewhat useful” to “truly transformative.” I am not sure if anyone could have accurately predicted when this would happen, nor if it would be the first application to emerge. Some people will say in hindsight that they saw it coming, but I do not think anyone could predict with certainty when all of this would occur or that it would emerge first in the form of programming.

Erik Torenberg：So on an organizational level, what have we learned? What does it mean for junior engineers, senior engineers, the job landscape, and team organizational forms?

Benedict Evans：I think we are not really talking about anything yet. Six months ago, this thing was hardly usable.Now everyone is scrambling to figure out what it actually means. If you get too caught up in the noise and details, seizing on something someone said at an event and thinking the sky is falling, you will be thrown into chaos. It will take at least another two or three years for everything to stabilize, not to mention there is a significant supply-demand contradiction regarding pricing, leading to unexpected consequences.So, we have no idea what future teams will look like.

I think people are beginning to raise some new questions, the most obvious of which is: are you still hiring junior employees? If so, what do they do? Why did you use to hire junior employees? Do you hire them to do what they themselves can do, or to do something else? What happens if you automate an entire class of work that people used to do? This question has become more real in the realm of software development because you are indeed automating a lot of what people used to do. So these questions have shifted from a theoretical level to reality. But I don’t think anyone can claim to know what the market structure will be like in the next three to five years — if you think you know, then you’re insane.

Erik Torenberg：Let’s talk about OpenAI. What has most surprised you? How do you interpret their strategic evolution and the challenges they face in the future?

Benedict Evans：This has always been a place filled with dramatic conflicts. It’s obvious that their CEO is on medical leave, which has altered the situation.

In the second half of last year and into the fourth quarter, the question posed from the outside was: “The model itself is quite good, but besides that, what are you doing to get people to use these things for other purposes?” It was almost like saying,“Go ask ChatGPT for 15 ideas on how to create value based on the infrastructure, and just do all of them,” and that’s pretty much how OpenAI operated. Meanwhile,Anthropic, with relatively weaker financial strength, said: "We are focused on coding.” And then they actually made coding work. As for whether that was intentional or a happy accident, it's for others to judge. But it's clear this tactic has worked.

But the problems remain: what is actually working now are software development and parts of other areas. There are many people who are just excitedly dipping their toes on the fringes, using it a bit. The internal differentiation within Silicon Valley is also quite obvious — on one side are those who have bought piles of Mac Studio and are running open-source models around the clock, while on the other side are the roughly forty to fifty percent of people who think it is somewhat useful but only used it once last week. The question is, how do you bridge that gap? I do not believe there is a simple answer to that. Software development has indeed crossed that gap, but many people from other areas are still scratching their heads and only sampling it.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。