Character.ai: First Domino Falls

Perhaps the character.ai Chapter 11 bankruptcy, which allows it to continue operating while it conducts a sale of its assets, is the first step in a rationalization of the generative artificial intelligence market which virtually all observers would likely agree is inevitable. 

It appears the cost of legal battles are the immediate cause of the move. “Over the past year, Character.ai has taken meaningful steps to sustain and improve its business in the face of significant legal, regulatory, and operational headwinds,” said Colin Adams, company CEO.

The firm appears to have about $50 million to $100 million in liabilities and some $20 million to $25 million in assets such as cash. 

Character.ai has settled multiple lawsuits from families of children who died or attempted suicide after interacting with the firm’s chatbots. In October 2025, the company banned users under age 18 from having “open-ended chats” with its chatbot.

In January, Character.ai settled several lawsuits brought by families of teenagers who allegedly suffered harm. 

In May this year, the Pennsylvania government also sued the company alleging that its chatbot falsely presented itself as a licensed medical professional in the state and gave a “fake Pennsylvania license number” to a user.

Disney also sued for patent infringement, alleging the chatbot used its copyrighted characters without permission. The chatbot allegedly impersonated the likenesses of Frozen’s Elsa, Moana, Spider-Man’s Peter Parker and Star Wars character Darth Vader. 

Disney already has made a $20 million offer for some of Character.ai’s intellectual property, but other bidders could emerge, some observers suggest. 

Eventually, cash burn issues for many would-be contestants will cause many of the firms to follow character.ai. 

Even giant OpenAI has an annual cash burn rate between $14 billion and $17 billion, exceeding its revenues. 

And OpenAI does not expect to reach positive cash flow until about 2030. 

source: Reuters

But that same general argument could be made for all the leading and would-be contenders in the language model market. Generally speaking, investors might well believe that contestants with other existing businesses throwing off cash are better placed to win the battle. 

Conversely, upstarts gambling everything on the model business are arguably more exposed. Consider OpenAI, Anthropic and Perplexity, for example. Anthropic hopes to reach cash flow breakeven in 2028, while Perplexity hopes to do so between 2027 and 2028. 

That will matter greatly if capital availability becomes an issue. Recall the collapse of the dot-com bubble, when startups found they could no longer raise capital to fund their operating losses. 

Metric	OpenAI	Anthropic	Perplexity
Annualized Revenue (ARR)	$50 Billion	$65 Billion	$750 Million
Estimated Annual Burn	$14B to $17B	$5B to $6B	$100M to $150M
Burn % of Revenue	57%	33%	15% to 20%
Breakeven Target	2030	2028	2027 to 2028

Braking The AI Arms Race?

I was talking at lunch recently with some friends about why some leaders of artificial intelligence frontier models might actually favor slowing the rate of such model development, whether it can actually be done, and what constraints might exist.

Google’s development of Gemini 4 Argon illustrates one of the issues. Basically, frontier model development is open ended: it never stops. So that also means development work and investment never stops. It is an arms race.



source: Google

And the problem with arms races is that there is no obvious method of restraint. Competition does not necessarily converge on an efficient stopping point. There may be physical, financial, or technological limits, but there is often no natural point at which the participants independently decide that additional investment is no longer worthwhile.

Each participant's decision changes the payoff to everyone else. The security dilemma is that actions taken defensively by one participant can make the other participant feel less secure. So technological progress can continually move any stopping points.

“Parity” can exist at any level, from small to very large. And there is an asymmetry between the output gains and the input investment. Each additional increment of capability might not be worth all that much (it allows a firm to keep pace) but falling behind could be extremely costly (a firm loses the ability to lead the “winner takes most” market in share or profits).

It’s a positive feedback loop, very much similar to the cat-and-mouse game played by hackers and security professionals.

In other words, firms need to keep investing and keep running simply to stay in the same place, and, over time, customers might have trouble distinguishing between competitor offers.

But economic or financial constraints should eventually develop. Already, frontier model developers are facing resistance from financial interests that want to see revenue produced by all the investment. Political resistance also is growing.

So constraints seemingly will be imposed by external forces, and not market competition. At least so far, the marginal benefit of additional investment arguably outweighs the cost, as frontier models are striving for artificial general intelligence.

So contestants themselves are unlikely to “naturally” restrain investment levels. But external forces seemingly are gaining strength in the form of calls for lawmakers to “do something” about regulating AI. And even industry leader calls for some restraint while justified on ethical concerns, also strengthen the demand for some external regulation.

SpaceXSI, AT&T, Verizon, T-Mobile and Beauty Contests

AT&T, Verizon and T-Mobile have been under pressure since SpaceXSI suggested it would be competing in the mobile phone business, and has followed up by acquiring 800-MHz low-band spectrum ideally suited to inside-building connectivity, an important consideration since satellite-to-phone only works when there is line-of-sight. 

According to Elon Musk, the purchase is "the last critical piece of the spectrum puzzle needed for SpaceX to provide complete phone coverage in America".

On October 9, 2026, T-Mobile US shares plunged 13 percent. the company's worst single-day drop since 2013.

AT&T shares fell approximately almost 10 percent, the worst single-day drop since 2000. Meanwhile, Verizon Communications declined about nine percent,  Verizon's worst single-day loss since July 2002.

Skeptics will argue the impact is out of proportion to the nature of the threat, as SpaceXSI still has to acquire or build a terrestrial transmission infrastructure. 

The impact on SpaceXSI was the opposite. Based on SpaceXSI's current market capitalization of $2.14 trillion, 3.5 percent of its market value is approximately $74.94 billion. So it might be argued that the 3.7-percent increase in market valuation on Oct. 9, 2026, which was about $75 billion, reflects the $8 billion investment in spectrum. 

In other words, SpaceXSI valuation increased about 10 times more than the cost of spectrum acquisition. 

On Oct. 9, SpaceXSI gained $27 billion, even using the low figure of 1.24 percent increase for the day.

The mobile service providers lost about $50 billion. 

You might argue that is an example of the “Keynesian beauty contest” at work. The Keynesian beauty contest is an economic and behavioral finance concept introduced by John Maynard Keynes in 1936 to explain how speculative markets and asset prices fluctuate.

The basic idea is that investors buy some stocks not because of strong company fundamentals, but because they believe others will buy them and drive the price higher.

The analogy is a beauty contest in which judges are rewarded for selecting the most popular choices among all judges, rather than those they may personally find the most attractive. 

As it pertains to SpaceXSI, at least some investors arguably price the company not on its current cash flows or fundamentals, but on what they believe other market participants will pay. 

It is, in other words, a narrative-driven market. We saw this during the dot-com bubble. But it has happened before then, as well. We call these periods “manias” or “bubbles.” 

I’m not arguing we are in a similar bubble at the moment. Maybe we are, maybe we will be, but it is not certain we are at the moment. 

Instead, we do see a beauty contest logic at work, where valuation is less on fundamentals and more on expectations. It might be unfair, but SpaceXSI currently benefits from the beauty contest, while the big three mobile service providers suffer.

Era and Mania	The "Prettiest Face" (The Hype Asset)	The Real Underlying Catalyst	The Beauty Contest Driver	The Collapse Trigger
1636–1637 Tulip Mania	Rare, variegated tulip bulbs (e.g., Semper Augustus).	The introduction of exotic flowers to the Dutch merchant class and early futures contract innovations.	Buyers knew a single bulb couldn't generate income. They bought purely because the broader Dutch public was highly coordinated in bidding up bulb futures, ensuring a "greater fool" would buy it next.	In February 1637, buyers at a routine auction in Haarlem simply failed to show up, breaking the psychological chain of assumed demand.
1720 South Sea Bubble	Shares of the South Sea Company.	A British government debt-for-equity swap scheme wrapped in a theoretical monopoly over South American trade routes.	Investors recognized that actual trade was non-existent. However, because the King and high-ranking politicians were buying and promoting it, citizens bought shares assuming the royal endorsement would force the public to keep bidding it up.	The company's own insiders realized the stock was dramatically overvalued and quietly began liquidating their positions, sparking a mass panic.
1840s Railway Mania	Shares of newly proposed railway companies.	The industrial revolution and the British government relaxing rules on corporate formation, combined with cheap credit.	While railways were revolutionary, the market funded thousands of overlapping, impossible routes. Investors didn't look at track viability; they chased the "railway" tag because newspaper columns constantly validated it.	The Bank of England raised interest rates to combat inflation, draining the cheap credit that speculators relied on to buy shares on leverage.
1920s The Roaring Twenties	Industrial conglomerates, radio technology stocks, and investment trusts.	Widespread adoption of electricity, automobiles, mass media, and the creation of retail margin accounts.	Speculators freely admitted stocks were expensive. They bought because loose margin requirements (borrowing up to 90% of the stock price) meant a massive wave of new retail money was constantly entering the market.	Corporate earnings began cooling in mid-1929, leading to minor margin calls that snowballed into the catastrophic Black Thursday crash of October 1929.
1997–2001 Dot-com Bubble	Any company with a ".com" suffix or internet business plan.	The commercial birth of the World Wide Web and massive telecom network buildouts.	Traditional valuation metrics (P/E ratios) were abandoned. Investors bought companies with zero revenue because they knew institutional mutual funds were suffering from intense FOMO and felt forced to chase internet listings.	Major tech backbones (like Cisco and Intel) reported spending slowdowns, revealing that startup capital was burning out, which ended the assumption of infinite demand.
2024–Present AI Infrastructure Boom	Hardware providers (GPUs), liquid cooling firms, and next-generation data center networks.	The emergence of highly scalable Large Language Models (LLMs) requiring massive parallel computing power.	High-frequency trading algorithms and retail momentum traders target companies based on the size of their "AI capital expenditures." Traders buy not for the immediate dividend yield, but because tech giants are locked in an arms race and must keep buying hardware.	Highly variable; heavily tied to micro-catalysts, structural demand shifts (such as ultra-efficient compute open-source models), or data center budget fatigue.

Is AI Infra Customer Concentration a Bug or a Feature?

Common sense suggests that any supplier of goods or services is at risk when the customer base is highly concentrated. Indeed, we might argue that buyer diversification is a core strategic objective for any supplier of goods or services.

Unfortunately, high customer concentration is a basic feature of high-performance computing as a service, leading to high concentration on the supply side as well. And that leads to a high concentration of earnings growth in public markets as well. 

Perhaps few like this state of affairs, as it increases market risk for just about everyone in the investing value chain. But it really cannot be helped. 

Customer concentration, often defined as having a single customer represent 10 percent or more of revenue, or top three to five customers representing 30 percent  to 50 percent of sales, creates vulnerabilities:

• The "single point of failure" (revenue shock if a key buyer defects)

• Squeezed profit margins (buyers know they have leverage)

• Asymmetric counterparty and credit risk if a single buyer defaults

• Operational and capacity misallocation to the few large buyers

• Higher cost of capital, as lenders and equity investors view customer concentration as a risk. 

But the current artificial intelligence infrastructure market is skewed to a few customers supplying high-performance computing as a service (Google, Amazon Web Services, Azure, Oracle, SpaceXSI)

At the same time, the AI infrastructure boom is producing:

• A relatively small group of companies capturing a disproportionate share of incremental earnings growth in the public equity market

• A similarly small group of hyperscalers  responsible for a large share of demand for advanced AI computing infrastructure.

Micron alone accounts for 19 percent of all earnings growth in the S&P 500 in the third quarter of 2026 and Nvidia accounts for 15 percent, for example. That’s earnings growth, not the share of total market earnings. 

source: Seeking Alpha

Goldman analysts say the top 10 contributors account for approximately 68 percent of growth. So Goldman estimates that AI infrastructure companies (excluding hyperscalers) will account for 54 percent of third-quarter earnings growth.

Hyperscalers will contribute another 19 percent. Together, AI infra suppliers and hyperscalers account for 73 percent of earnings growth.

That will worry just about everyone. But such concentration reflects scale economics in the frontier AI model part of the value chain; the inference as a service value chain and the AI infra value chain. 

 

 

source: Seeking Alpha

There arguably are several reasons why the concentration exists:

• scale economies favor large buyers

• capital access matters

• utilization matters

• complementary assets matter.

Such characteristics are common in capital-intensive industries such as telecommunications, aerospace, semiconductor fabrication and electricity generation. 

source: J.P. Morgan 

In some sense, we might argue that concentration is a feature, not a bug. That is not to discount the risks. Everything hinges on eventual outcomes across the economy, especially the ability of the frontier model developers and computing as a service suppliers to demonstrate economic value for their costumers, leading to subsequent revenue growth for the hyperscalers. 

So concentration, in and of itself,  is not itself a reason to reject the AI infrastructure opportunity. Many other successful industries also are capital intensive and concentrated on the buyer and supplier sides. 

But such concentration also does magnify risk. A spending slowdown by the  hyperscalers will disrupt market valuations of firms in the industry. 

The wisdom of current investment pacing will be determined by the return on the capital being deployed, of course.

But buyer or supplier concentration, in and of itself, is not necessarily a bug. It is a feature of some industries, where scale economies are required.

General-Purpose Technologies Generally Do Come with Job Concerns

At least one study suggests that Millennials use artificial intelligence tools more than Generation Z. According to a survey conducted by Channel V Media, 47 percent of Millennials use generative AI, compared to 44 percent of Gen Z, 42 percent of Gen X and 32 percent of Baby Boomers.

According to that survey, Gen Z is the generation most anxious about AI and most afraid of losing its own job to it. 

source: Channel V Media 

But history might suggest such initial hesitations must simply be dealt with, as there is little, if any, precedent for people being able to opt out or avoid the impact of new general-purpose technologies. Non-adoption, like it or not, is rarely, if ever, widely possible. 

Technology	Initial worker reaction	Who tended to adopt fastest?	What happened to the "avoiders"?	Strategic lesson
Steam power	Fear of mechanization and displacement; Luddites famously attacked machinery in some industries	Firms and workers whose jobs could exploit mechanization	Avoiding machines did not prevent mechanization; occupations and industries changed	You couldn't opt out of the production system
Electricity	Initially viewed as an industrial substitute for existing power systems	Firms reorganizing factories around electric motors	Firms that merely substituted electric motors for steam got less benefit than firms that redesigned production	Adoption matters, but complementary organizational change matters more
Telephone	Concerns about disruption of existing communications and occupations	Businesses and professionals with high communication needs	Nonusers increasingly faced a disadvantage in a connected economy	Network technologies reward participation
Personal computer	Significant worker anxiety about clerical/administrative automation	Younger, more educated and occupationally exposed workers disproportionately	Computer nonusers increasingly concentrated in lower-wage occupations	Avoidance could protect a particular task, but not necessarily the career
Internet	Similar fears about retail, media, travel agencies, newspapers, etc.	Younger and more educated workers initially; adoption subsequently became nearly universal	Non-adoption increasingly imposed an economic and social penalty	The technology eventually became infrastructure rather than an optional skill
GenAI	Explicit fear that the technology will eliminate one's occupation	Younger workers and highly exposed occupations are among the earliest users	Avoidance may preserve some skills temporarily but risks losing complementary AI skills	Learning to use AI may be the modern equivalent of learning the PC

In all likelihood, the data simply reflects the fact that AI is arriving, and quite possibly eliminating some entry-level jobs, just as many Gen Zers are entering the work force.