Last month, the mathematical community warned that AI solving mathematical problems is a threat to intellectual work. The statement is well-intentioned, but it ignores the fundamental nature and evolution of human existence. It treats our current methods of intellectual toil as sacred ends rather than stepping stones to a more fulfilling future. This is at odds with history. Progress throughout human history has meant graduating from solving mere survival challenges to tackling higher-order problems.

Hunter-gatherers focused on daily survival

If you go back 30,000 years, what was the daily problem you had to solve?

You woke up, left your dwelling, and risked your life hunting wild animals or gathering berries, fruits, and nuts for your family. Meanwhile, your extended family hopefully stood guard to protect your children and dependents from predators.

You had to gamble daily on whether an unfamiliar plant or mushroom was nutritious or lethal. The entire day was consumed by immediate survival risks.

Agricultural advances led to permanent settlements

Around 10,000 years ago, that changed with the domestication of wheat.

Domesticating wheat marked the beginning of the end for the hunter-gatherer lifestyle. We no longer had to chase down dangerous animals or experiment with poisonous plants just to survive. We could reliably reproduce crops, build granaries, and establish settled villages.

That shift had profound impacts. It freed up time to spend with families and raise kids in relative safety. A secure food supply allowed for life to prosper. The population exploded.

Political stability unlocked human creativity

Once we had a roof over our heads and a secure food supply, we asked a bigger question. What if we built more complex societies?

We built governance and institutions. We organized political structures, established laws, and coordinated across larger populations.

Political stability unlocked art. Because people were not hunting all day or defending immediate borders, they could spend months chipping away at a block of marble to produce a magnificent sculpture, or spend years painting intricate murals.

Technological progress mechanized art. Today, computers and machines can render intricate art, carve stone, or reproduce images in a flash. While handcrafted art remains deeply valuable, there is not enough patronage to sustain it as a primary economic engine.

Technological progress unlocked large-scale knowledge work

Knowledge work—coordination, communication, data analysis, academic research, engineering, and more—is far less physically taxing than spending a lifetime chipping stone. It is a better problem to solve.

Just as living in a village was a better problem than dodging wild animals, doing knowledge work has been a better problem than backbreaking physical crafts.

Now we are building systems that can automate significant parts of that knowledge work, including mathematical solutions and scientific reasoning.

If history is any guide, that is not a bad thing.

Automation of knowledge work frees us to tackle higher-order problems

Every time humans solved a foundational problem, we moved to a higher-order problem.

Automating cognitive execution does not destroy intellectual work—or the purpose of human existence for that matter. It frees up collective human brainpower for tackling the next tier of challenges. Instead of grinding through the mechanics of the work, we get to frame the challenges that matter. We get to focus our collective energy on asking the right questions.

There will be better problems to solve. Look around, there are so many already!

Last week, Anthropic launched a legal plugin for its Cowork assistant. Shortly thereafter, stock market chaos ensued, wiping $300B off software and data stocks. The technology-heavy Nasdaq Composite dropped for three straight days starting from February 3. Many investors are worried that AI could dampen the prospects of SaaS companies like Adobe, SAP, Salesforce, ServiceNow, Thomson Reuters and Workday. However, SaaS products do more than just accomplish tasks. Let us understand what else people buy when they buy B2B SaaS products.

1. Reliability and security: You pay SaaS companies to get your job done—every time. They ensure high availability and durability, allowing you to focus on your core competency. These applications typically come built-in with industry standard security protocols that often win over your IT department. When something breaks, you can generally expect prompt, high-touch service to get things up and running again.

2. Integrations: You pay for the integrations that facilitate your business processes. Imagine a sales rep closing a deal in Salesforce that then automatically creates a sales order in SAP. Or think about all the data connectors in Tableau that power your business intelligence. You pay because you would rather not deal with third-party integration negotiations and software maintenance.

3. Best-in-class: You pay for the most up-to-date way of doing things. SaaS companies exist to serve you the best possible version of their products. Their product managers speak to their customers every day to learn, validate and solve pain points.

AI enables you to build many SaaS alternatives quickly. However, if these alternatives are to replace SaaS, they also need to address the what else that people buy SaaS products for. Some tech-savvy companies would try to build their own alternatives and perhaps succeed to varying degrees. Even so, most companies would probably prefer to not do this undifferentiated heavy-lifting. They would prefer to focus on their core competency instead.

This does not mean AI has no future in SaaS—quite the contrary. AI will turbocharge many SaaS products—by running inside applications. The likely winners would be companies which can create the most value using AI—particularly the ones that can do away with expensive workforce training needed to use many SaaS products. New SaaS or SaaS-like companies will pop up but existing ones with strong distribution have the odds stacked in their favor—as long as they do not fumble.

AI has made the competitive moat of SaaS narrower. This puts pressure on the pricing models of established players. To adapt, SaaS is already moving away from seat-based to consumption-based pricing. This pay-as-you-use pricing is more aligned with clients’ business outcomes. It also means the value to clients is increasingly in line with the cost to SaaS providers. This commoditizes task execution. In a world where task execution is a commodity, AI-powered alternatives would replace B2B SaaS when they can address the what else in addition to accomplishing tasks.

There is a proposed ballot initiative in California that aims to impose a one-time 5% wealth tax on billionaires who were California residents on January 1, 2026. The initiative needs 875K signatures to qualify for the November ballot—given it does not get blocked by courts. If Polymarket bets from January 4 are any guide, there is a 73% chance that it will make it to the ballot, and a 56% chance that it will become the law. If it does become the law, how would it impact California’s tax revenues?

The Legislative Analyst’s Office thinks it could lead to a temporary bump in tax revenues at the expense of longer run income tax revenues. That is probably a good guess. A state cannot stop people from leaving for another state with better tax benefits. Rich tax filers are often more mobile than the average person and can choose to exercise that mobility if they do not like their state’s tax policy. Or at least that is how the theory goes.

Let us look at a few examples as to what has happened in the recent past when states have tried to tax the rich. This is not an exhaustive list of examples. Write to me if you have more.

Maryland

In 2008, Maryland started taxing income in excess of $1M at 6.25%—up from 4.75%. In the first year that the new tax rate was in effect, Maryland saw a 29% drop in the number of tax returns with tax liability that reported income of more than $1M. Income tax revenue from $1M-plus tax bracket declined by $222M between 2007 and 2008 tax years. During this period, Maryland’s total income tax revenue dropped by $318M—a 5% decrease. The surtax expired at the end of 2010. The state’s income tax revenue started recovering in 2011.

There is a caveat here though. The Maryland surtax overlapped with the Great Recession, which lasted from December 2007 to June 2009. You could say that higher income filers were more impacted during the recession because a bigger portion of their incomes came from stocks. The S&P 500 index returned -37% in 2008 but went up every year from 2009 through 2017. However, Maryland’s tax revenue bottomed out in 2009 and as we will see below, Oregon’s in 2010—both bottomed out midway through their surtax periods.

Oregon

Starting in 2009, Oregon increased the tax rates on joint-filer income between $250K and $500K to 10.8%, and to 11% for income above $500K. The previous top marginal rate was 9%. Oregon’s total income tax revenues dipped by $360M between 2008 and 2009 tax years—a 7% drop. The surtax expired at the end of 2011 and settled at a lower rate of 9.9% for joint-filer income above $250K. In 2012, Oregon’s income tax revenue recovered and then some.

You might note that both Maryland and Oregon had a surtax period of three years. Both states had enacted temporary measures to shore up short-term revenues and/or budget gaps. It is fair to say that they were not wildly successful. Let us see if we can say the same thing about Washington state, which has implemented a new tax that is not temporary.

Washington state

In 2023, Washington state enacted a 7% tax on capital gains in excess of $250K. The first year’s capital gains tax haul was $848M while people awaited a verdict on the constitutionality of the tax—Washington’s constitution prohibits graduated taxes on income. This tax haul was more than three times what the state had expected. In 2024, capital gains tax revenue declined 58% to $356M even as the S&P 500 climbed 23%. For now, the lesson is—set low expectations and you might succeed!

Tax filers are not sitting ducks

Tax filers, by and large, respond to changes in tax policies. Some filers are more able and willing to exercise their mobility. Freedom of movement across states is the right of every American in a way that movement across countries is not. As a result, states have less leverage than the federal government when it comes to taxation. States can still capture a portion of the utility (benefit) that residents get by living in that state. When states demand more, able and willing people reevaluate their utility and sometimes decide to move. That is why states have a mixed record, at best, of taxing the rich.

It is difficult to choose top books for the year—not the least because this is my first time publishing such a list. However, as promised last week, here is a list of books that are subjectively worth sharing.

1929 by Andrew Ross Sorkin

This book is about the stock market crash of 1929 and touches on the Great Depression that followed. The author has reconstructed a day-by-day account of the dealmaking, brinkmanship and fragility of the Wall Street bigwigs of the age. It is also a timely read given the fears about an AI bubble, which we talked about last month.

Chip War by Chris Miller

This book takes you on a captivating ride from the beginnings of the semiconductor industry to the present day. Semiconductors had somewhat faded from public consciousness until the 2020-2023 global chip shortage. However, the industry is as competitive and fast-moving as ever, and increasingly intertwined with geopolitics.

Founder vs Investor by Elizabeth Zalman and Jerry Neumann

This book about venture capital pits the viewpoints of a founder against those of an investor. Founders and investors often do not see eye-to-eye on topics such as money, valuation, control, exits, etc. The authors use a dialogue format to bring out their agreements, disagreements and everything in between.

Golden State by Michael Hiltzik

California rose from a Western frontier to the epicenter of American innovation in a rather short time. Its history is replete with societal upheavals, frontier spirit and boom-and-bust cycles. This book examines California’s history, and brings the reader to the challenges and opportunities of today.

The Thinking Machine by Stephen Witt

Nvidia’s hardware was behind AlexNet—a groundbreaking image classifier released in 2012—that kicked off a renaissance in deep learning. Nvidia and its longtime CEO, Jensen Huang, have not caught a breath since. This book presents a gripping tale about Huang, Nvidia and the rise of AI.

Back in January 2020, I published the following article on my website. I am re-sharing the original article—verbatim and with the same feature image—for new readers. I will publish my list of top reads from 2025 next week.

Until a few years ago, I stayed largely clear of books. I would still spend hours reading every day—just not books. The content I used to consume mostly spanned news and news analysis. After I reached a certain saturation level, I realized that this was not really helping me grow. So, I stopped using my Pocket app.

Suddenly, I found myself with valuable extra time at hand. This was when I started reading books. First couple of months seemed very mechanical but eventually, I got used to it. I always had a few books waiting in line and the allure of the next book was enough to keep me reading.

Fast forward to today, I cannot imagine a life without books. I read more than 100 books in 2019. Most of the books I read are non-fiction, and I am never short on recommendations.

Where do I get my recommendations?

Books. Authors refer to other books very frequently in their own books, and I do a quick search every time to see whether I would like to read them. A smaller share of recommendations come from blogs, word-of-mouth, etc.

Once I locate a book on Goodreads and decide on it, I quietly add it to my queue on Libby—an app that gives me access to books from local libraries. There is usually a wait time for most books but I always have something to read since my queue is long enough that a new book turns up every few days.

How do I choose what I read?

I read on a wide range of topics. There is really no catch-all category. More importantly, I actively read viewpoints that I disagree with. This usually leads to one of three scenarios—tempering of disagreement, agreement or, no change in the status quo.

How do I consume books?

I find e-books and audiobooks most appealing. However, I also buy hard copies of certain books that I would like to read again.

How do I keep track of the books I read?

I maintain a simple sheet which contains what I have read so far and what I would like to read. Sometimes, I would jot down a few lines which summarize what I learnt from a given book. I also maintain a top 100 list (in no particular order) on this website that I update periodically.1

What if you find it hard to get started?

Might I suggest that you try making a commitment contract on stickK.com. Posting your pledge to read on social media or simply committing in front of friends can also help.

How does the math work out for 100 books a year?

On an average, you need to read two books every week to make it to 100. Now, let us say that a typical book takes eight hours to read. This means that you need to read 16 hours a week to finish two books per week, which amounts to reading a little over two hours every day.

So, there is not much of a secret to reading 100 books a year. It is mostly about willingness to learn and adjusting your lifestyle around it.

The prevailing wisdom in Silicon Valley, so far, has been that the world needs a lot of AI compute. Data center spending is expected to hit $425B in 2025 (S&P Global report 1). OpenAI alone has $1T in compute deals lined up (S&P Global report 2). By 2030, total AI data center spending could amount to $5T (McKinsey report).

How much of this spending is driven by AI demand outstripping supply in the near-term? Probably some non-trivial amount. How much of this spending is driven by anticipation of future AI demand in a fixed timeframe? A lot. Sounds a bit irrational, eh? In a BBC interview this month, Alphabet CEO Sundar Pichai called out ‘elements of irrationality’ in the current AI boom (BBC report).

Let us use two frameworks to dissect some of the irrationality.

Framework #1: Keynesian beauty contest

In his 1936 book—The General Theory of Employment, Interest and Money—John Maynard Keynes wrote the following about behavioral dynamics among investors:

…professional investment may be likened to those newspaper competitions in which the competitors have to pick out the six prettiest faces from a hundred photographs, the prize being awarded to the competitor whose choice most nearly corresponds to the average preferences of the competitors as a whole; so that each competitor has to pick, not those faces which he himself finds prettiest, but those which he thinks likeliest to catch the fancy of the other competitors, all of whom are looking at the problem from the same point of view. It is not a case of choosing those which, to the best of one’s judgment, are really the prettiest, nor even those which average opinion genuinely thinks the prettiest. We have reached the third degree where we devote our intelligences to anticipating what average opinion expects the average opinion to be. And there are some, I believe, who practise the fourth, fifth and higher degrees.

If we were to generalize Keynes’ analogy, we would say that some market participants engage in first degree thinking—they pick the options they like the most. There are participants that act based on second degree thinking—they pick the options that they think participants with first degree thinking might choose. Then there are participants with third degree thinking—they pick the options that they think participants with second degree thinking might choose. And so on.

In a sense, you can say that participants with first degree thinking are the most naive, or perhaps the most irrational. It follows that second degree participants are acting based on their understanding of the judgment of the most irrational participants. Third degree participants are acting based on second degree participants. You could say, at some level, that irrational participants are driving the game.

Framework #2: Prisoner’s dilemma

This framework came about in the form of a puzzle in 1950. Here is a version of the puzzle (UMich’s website).

Two bank robbers happen to meet. They decide to pull a job together.

The cops nab them, but without enough evidence to convict. They need a confession. And they know both robbers are unlikely to talk, since if neither implicates the other, the cops can keep them in jail for only 30 days.

So they put the two in separate cells. They go to the first prisoner and say:

“If you rat on your partner and he stays mum, we’ll let you go and he’ll do ten years.”

“If you both rat on each other, you’ll both do eight years.”

Then they go to the second prisoner and say the same thing.

If the two robbers could cooperate, they would likely agree to stay silent. This gets both of them out of jail in 30 days. However, because they cannot cooperate, each robber is likely to pursue their self-interest and rat on the other. This gets both of them eight years in jail. Here, pursuing self-interest gives you a worse outcome than acting as if you were cooperating i.e., acting based on second degree thinking.

Getting to trillions on compute

Our framework #1, Keynesian beauty contest, tells us that irrational market participants could breed more irrationality in the market. Because there are very few companies vying to hyperscale compute, irrational investments from just one player could spur investments from others. In the piece about everyone wanting their own AI chips, we talked how nobody wants to be left behind in the AI race. When you are in a race to secure massive amounts of compute, data center spending can become your interim scorecard.

Our framework #2, Prisoner’s dilemma, tells us that the collective outcome of a group suffers in the absence of cooperation. Obviously, corporate cooperation, aka collusion, is a no-no. Although nobody is stopping you from acting based on second or higher degree thinking. However, there is little trust among market participants. They are working with different information and nobody wants to share the spoils. As a result, each participant ends up maximizing their self-interest. This is how we end up spending trillions of dollars on compute.

Some caveats

1. Nobody knows how much compute would be needed and in what timeframe.

2. Irrationality could be in the eye of the beholder.

   1. A high-risk, high-reward strategy that appears irrational to most of the market could be a rational decision for you depending on your situation.

   2. Some market participants could be working with superior information and/or structural advantages. Their seemingly irrational decisions could be more rational than we think.

Silicon Valley is awash with AI startups. A few like OpenAI and Anthropic are building AI models and using them to power chatbots, coding agents, video generation tools, etc. Many startups are applying those AI models to offer tools that address industry-specific use cases like customer service, drafting legal briefs, etc. Others are using the AI models to run their own chatbots and coding agents. However, a large number of AI startups do not have sufficient ownership of their value chain. This increases their exposure to disintermediation, copycats, price wars and feature races.

Let us understand this using an example. Say, Company X builds an AI model, and makes it available to the public via a chatbot. The chatbot answers all your internet curiosities. It can even do things for you, like manipulate Microsoft Excel files. The chatbot quickly gains hundreds of millions of users. You spot an opportunity—you think people would like to be able to manipulate their Excel files without ever leaving Excel. If you were to build an Excel add-in, you could make a lot of money. So you build the add-in using the AI model’s API. People love it! You raise money to scale your startup.

To understand what is likely to happen next, let us ask a few questions.

What is the operating environment of Company X? It costs $100M or more to train an advanced AI model (Techcrunch report). Company X would likely keep racking up bills to train new models to stay ahead of the competition. They better make a lot of money from their models. Their investors are watching. The time is ticking.

What does product success look like for Company X? Mass adoption. Consumers should use their AI models to chat, generate images and videos, shop, etc. Businesses should use their APIs to build applications and workflows.

How would Company X price its products? Their prices would probably be very close to or below their unit costs. Company X is probably thinking their unit costs will come down over time (Moore’s law and/or software optimizations). They might even give away products for free—Company X would want to remove any friction that could slow their adoption curve.

How many users does Excel have? Close to a billion.

What is Company X thinking about launching its own Excel add-in? An AI-powered add-in for Excel has massive market potential. With hundreds of millions of users of its own, Company X has a warm start to enter and penetrate the market for productivity applications.

What is Microsoft thinking about launching its own Excel add-in? Most users would want this functionality. This might even expand the market for Excel users. Why should a third-party use their platform for free (Microsoft's terms)? Even if Microsoft gets a commission from each sale, it would be more accretive to book the full revenue themselves. They would want to launch a native Excel feature, potentially using the same API your startup is using. Microsoft would not mind pricing at or below their unit costs to drive adoption, though they probably have more pricing power and could also monetize it through bundling (The Verge report).

How competitive is your startup’s Excel add-in? Not very competitive. If Microsoft launches an Excel native feature without any big fumbles, their distribution gives them a big user base instantly. This probably kills your startup. If Company X launches an Excel add-in, they also get a sizable user base—because they have a warm start. Your startup hits on tough times but if you have a great product, you stand a chance. The problem, however, is that your product is easily copiable because someone else’s AI model is a big part of your core offering. If Company X copies your product, you likely cannot compete on price—because their price is your cost.

Your initial market position does not predestine your startup to doom. You can still create value with a great product but if the market is lucrative enough, a deadly footrace with Company X and Microsoft would ensue. It would be detrimental to be a price taker in a market where you cannot compete on price. To improve your odds of success, you would want to own more of your startup’s value chain that gives you a sustained competitive edge.

If you are an accredited investor (as per the SEC) and venture capital firms are willing to take your money, you can bet on startups to generate returns. You would agree to lock up your capital in VC funds for 10 years or more, while they put your money to work by investing in unproven business models. This is a fair deal if you are compensated for the risk you are taking. At the very minimum, VC returns should beat the returns from public markets. If your VC fund returns 50% more than what you would have otherwise got from, say the S&P 500, over the same period, you can say your public market equivalent ratio or PME ratio is 1.5. But something is changing that puts VC returns at risk.

Dot-com era returns

All VC firms are not created equal. People chase some VC firms to give them money, because they have consistently good returns (WSJ on Sequoia Capital’s $7.2B fundraise). Good returns bring more capital and better deal flow. Three Sequoia funds in the 2000s had returns between 8x and 11x (Business Insider report). However, average returns of the U.S. VC industry are less impressive.

VC funds that started investing in 1999—also called 1999 vintages—at the peak of the dot-com bubble, had a PME ratio of 0.89, meaning their weighted average returns were 11% lower than what you would have got from indexing the S&P 500 (data from this paper). Funds from 2000 had a PME ratio of 0.78. However, funds that started investing early in the boom and were probably investing at more reasonable valuations, had high PME ratios. 3.14 for 1995 funds, 4.34 for 1996 and 2.68 for 1997. Funds from 1998 also did well with a PME ratio of 1.74.

To invest, or not to invest

If you were a venture capitalist in 1999, you saw startups going public at astronomical valuations. Here is a snippet of a CNET report from 1998 about TheGlobe.com, an early social media startup:

TheGlobe.com took Wall Street by surprise with its initial public offering this morning, blasting out of the gate with its first trade of the day at 87 a share--a ninefold increase from its target price.

Venture capitalists at the time probably knew the boom would not continue forever and that it would take a lot to justify the high valuations. But when everyone is riding the wave, it is difficult to sit out. Sitting out could be bad in other ways. Here is FT on the change of command at Sequoia:

Sequoia Capital’s Roelof Botha was ousted by top lieutenants who lost confidence in his ability to keep Silicon Valley’s most powerful venture capital firm ahead of rivals.

Under Botha, Sequoia has taken a more cautious approach to AI investment than some rivals. The firm invested a little more than $20mn in OpenAI in 2021, when the ChatGPT maker was valued at about $20bn, and has boosted that stake in subsequent rounds, said people with knowledge of the deals. When OpenAI raised funds at a $260bn valuation this year, Sequoia offered to invest $1bn, but ultimately was given a stake of a fraction of that, according to people with knowledge of that deal.

As a venture capitalist at the peak of a boom, you do not seem to have great options. Sitting out could cost you, but getting in could depress your returns. In 1999, VC firms chose to get in, perhaps thinking they could get out before the bubble burst. U.S. VC investments totaled $8B in 1995, inching up to $21B in 1998 (source). They hit $54B in 1999. $105B in 2000. VC firms also started taking bigger swings with average deal size increasing 125% from $4.4M to $9.8M between 1995 and 1999. Average deal size was $5.8M in 1998—the last year of the dot-com bubble in which funds returned more than public markets.

2025 looks like 1999

It is becoming difficult to ignore the similarities between 1999 and 2025. In the first three quarters of 2025, U.S. VC investments blew past the 2024 total investment of $215B (NVCA report). Their total investment in the first three quarters of 2025 was in the neighborhood of $250B (KPMG report 1 and report 2). At this rate, venture capital investments as a share of U.S. GDP would be double the 1999 level by the end of 2025. Between 2022 and 2024, average deal size was in the range of $12M to $15M. In the first three quarters of 2025, it stood at $26M.

Caveat investor

As a VC firm, all of this might still be okay if valuations are reasonable, you have sufficient downside protections (Business Insider report), your portfolio companies have a path to profitability in the right timeframe (WSJ report) and your investors are not begging for liquidity (Carta report). You get the idea.

The Nasdaq Composite, which is heavily weighted toward technology stocks, is up +123% in the last three years. We would not know for sure it is a bubble unless it pops. Economists say a bubble is an ‘upward price movement over an extended period of fifteen to forty months that then implodes’ (Manias, Panics and Crashes). Regardless of when the implosion happens, we are pretty sure that the upward price movement criterion is satisfied. Cyclically adjusted price-to-earnings ratio, also called Shiller P/E ratio, recently crossed 40 (WSJ has a chart). It has done that only once before and that was during the dot-com bubble. The long-run average is close to 17.

Stock prices could keep going up for a while or they could crash today. We have no way of knowing. However, there are structural factors at play that can give us an insight into how this will end.

Sudden decline scenarios

Big technology companies are investing many tens of billions a year. Some are more creative in financing than others. Meta funded its Hyperion data center in Louisiana using off-balance sheet financing. The company is leasing the data center in four-year chunks along with a ‘residual value guarantee for the first 16 years of operations’ (Meta’s statement). It was creative enough that they had to check with S&P that it would not be considered debt, thereby impacting their credit ratings. S&P released a report saying it is okay, for now.

Meta has enough cash flow to make the payments. But when you pile up many of these deals, it does make the investors jittery. A company’s stock price is, in theory, the present value of its expected future cash flows. If you start eating into those cash flows without showing evidence of commensurate returns, investors will sour at some point and typically when you announce your earnings (Business Insider report).

Credit agencies like S&P could also change their minds when larger portions of these companies’ cash flows start going towards mandatory lease payments, servicing bonds, etc. Big technology companies have a long way to fall in terms of credit ratings (Microsoft’s rating is better than of the US government) but that does not mean that a credit rating downgrade would not spark a selloff.

Slow decline scenarios

If everybody keeps investing in the hopes of not being left behind and the market does not grow sufficiently, we could soon be looking at overcapacity. And overcapacity is not good for prices. This could be a recipe for profits to get competed away. It might become clear over months or even years.

Another factor is having continued market access to monetize the investments. Chinese open AI models have become the standard, capturing 63% of open model users (MIT Technology Review). In a fragmented market, the profits of American AI companies are not guaranteed. While American cloud hyperscalers might still run the compute for Chinese models, many customers will switch to more cost-effective Chinese providers. A general perception that China is winning the AI race is probably enough to start deflating the bubble. Any visible market share loss will add to the perception.

Plateau scenarios

Stocks plateauing after an upward price movement cannot be called a bubble. This is a desirable outcome for most people (unless you are Michael Burry). It is also the least likely because it involves a tightrope walk balancing investments and returns. Companies do not necessarily have to show that they can make their entire money back in the next few years but their revenue growth should affirm sound unit economics to pacify investors. There would be a divergence as some companies are able to make good on their money and others are not. If their net effects cancel out, we would get a plateau.