AI Agent Business Models Diverge: Meta Burns Cash for Scale, Anthropic's Trillion-Dollar IPO Conceals a Compute Gamble
The global AI industry's competitive focus is shifting from chatbots to AI agents capable of autonomously executing tasks. At this critical inflection point, Meta and OpenAI launched their respective agent products almost simultaneously, but the two companies chose starkly different commercial paths. Meanwhile, Anthropic—widely regarded as OpenAI's strongest rival—is attempting to storm the capital markets at a $2 trillion target valuation. Its prospectus reveals massive compute commitments and a highly concentrated customer structure, adding a thick layer of risk to what would be the largest capital gamble in AI history.
On September 8, Meta officially launched its personal AI agent Muse in the United States. Three weeks later, on September 29, OpenAI unveiled its agent Dot at the DevDay developer conference. The two products share highly similar technical architectures: both run independent environments in the cloud, directly control browsers, and can continue completing tasks even when users close their screens. However, on the business model front, Muse's core features are free, with Meta planning to profit from transaction commissions; Dot, by contrast, is available only to premium subscribers of ChatGPT Pro and Business Plus—essentially an extension of the subscription service model.
This divergence is not merely a difference in pricing strategy. When AI agents begin selecting products and completing payments on behalf of users, whoever "pays the agent's salary" determines whose side it ultimately stands on.
Meta's Aggressive Experiment: Free for Scale, but at Staggering Cost
Muse's launch quickly ignited market sentiment, with Meta's stock surging 27% over 12 trading days and its forward P/E ratio jumping from 14x at the end of August to 21x. However, a deep-dive report published by Morgan Stanley in October revealed a sobering figure: Muse's per-user monthly service cost ranges from $3 to $130, with a base-case scenario of $37—translating to an annual cost of $445 per user. This already far exceeds Meta's global average revenue per user of approximately $70 per year.
Morgan Stanley analyst Brian Nowak stated bluntly in the report that Muse's subscription revenue will remain limited for a considerable period, as Meta's current strategic priority is expanding the user base and boosting engagement rather than monetization. He projects that true commercialization—an auction marketplace where businesses bid in real time for agent-completed transactions—will not materialize until late 2027 or 2028 at the earliest.
Even more concerning is the sharp deceleration in user growth. Data from alternative data provider TickerTrends shows that Muse's average daily net new users have plunged from a September peak of approximately 198,000 to roughly 74,000—a decline of 62.4%. As of October 6, Muse's daily active users stood at approximately 2.73 million, while Morgan Stanley's compute capacity analysis was built on a base of 1 billion daily active users—meaning Meta has already purchased CPU capacity roughly 366 times its current actual user count.
Muse faces headwinds far beyond slowing user growth. On September 20, Amazon blocked Muse's access to its platform, citing Muse's failure to clearly identify itself as an AI agent and its storage of user credentials without proper disclosure. Similar restriction reports emerged from Costco and Netflix. Telecom carriers began routing Muse-initiated AI customer service calls away from human agents, undermining its core "call on your behalf" functionality.
Anthropic's Capital Gamble: A Compute Black Hole Behind Explosive Revenue Growth
In contrast to Meta's aggressive expansion in the consumer market, Anthropic has chosen a B2B path centered on enterprise customers. According to a draft prospectus disclosed by Reuters, Anthropic recorded revenue of $4.59 billion in 2025, up roughly 12x from $386 million in 2024. More than 80% of that revenue came from enterprise and developer API usage fees, with subscription revenue accounting for just 17%.
Growth accelerated further into 2026: first-quarter revenue reached approximately $4.8 billion, already surpassing the entirety of 2025; second-quarter preliminary revenue was approximately $11.5 billion, bringing the first-half total to roughly $16.3 billion. By the end of July, annualized run-rate (ARR) had surpassed $65 billion—a more than 6x surge in just seven months.
Yet beneath the dazzling growth figures lies massive losses. The company posted a net loss of approximately $42 billion in 2025, of which $34 billion stemmed from accounting charges related to preferred stock valuation changes. Excluding those items, the operating loss was $8.06 billion, up 170% from 2024. The core driver of losses is the relentless expansion of upfront compute spending: of the $12.65 billion in total operating expenses in 2025, compute and infrastructure spending reached $7.33 billion—58% of the total, equivalent to 1.6x that year's revenue.
What truly unsettles Wall Street is the infrastructure spending commitments disclosed for the first time in the prospectus. Anthropic's total committed investment obligations for cloud, compute, and infrastructure over the coming years amount to $518 billion, of which approximately 80% (nearly $414.4 billion) constitutes irrevocable contractual commitments. Amortized over 10 years, the annual hard amortization burden reaches $41.44 billion—directly consuming 64% of the current $65 billion ARR.
These ultra-long-term, irrevocable compute commitments essentially constitute highly rigid long-term debt between Anthropic and cloud giants like Google and Amazon. If revenue growth falls short of expectations, the hefty fixed costs will rapidly drain the company's cash flow and trigger a liquidity crisis.
The Reefs of Circular Financing
The web of interests between Anthropic, Amazon, and Google is extraordinarily complex. Amazon and Google are both strategic investors in Anthropic—Amazon has committed up to an additional $25 billion in investment, while Google announced a $10 billion investment—and simultaneously serve as its compute suppliers and channel distributors. In 2025, distribution revenue through these two cloud platforms reached $2.16 billion, accounting for 47% of total revenue.
This symbiotic relationship has spawned a highly closed-loop circular financing chain: cloud providers commit investment, Anthropic purchases compute, the compute is distributed and monetized through the cloud providers—bolstering their revenue—the unicorn's valuation rises, and the cloud providers' investment returns multiply. Each link in the chain rests on the contract signed by the next link as its credit foundation—essentially a game of hot potato along the compute supply chain.
Vulnerability on the revenue side is equally concerning. In 2025, 24% of Anthropic's revenue came from its two largest customers, each accounting for roughly 12%, and many major customers have not signed long-term procurement contracts. The high concentration of revenue and the ultra-long-term rigid compute commitments create a severe mismatch in both duration and enforceability.
A Different Path for Chinese Competitors
Muse's launch has also drawn attention to comparable products in China. The core judgment of analysts at Bernstein and BofA Merrill Lynch is that the true prize in the agent competition is "contextual lock-in"—the deeper an agent understands a user's habits, history, and workflows, the harder it becomes to replace.
On this dimension, Tencent is seen as holding structural advantages. China's mobile internet is dominated by closed ecosystems, where external agents may never gain access to tasks within WeChat or Alipay—precisely the problem Muse encountered with Amazon, Costco, and Netflix, but in China this is the default state of the entire internet. WeChat agents require no separate download and are accessible with one tap within an app users open every day; mini-programs provide ready-made structured interfaces; and WeChat Pay processes trillions of yuan in transactions annually, allowing Tencent to take a direct cut from agent-completed transactions—the very model Morgan Stanley believes Meta cannot achieve until late 2027 or 2028 at the earliest.
The Chasm Between Valuation and Reality
Anthropic's valuation trajectory borders on the surreal: from $61.5 billion in early 2025, to $380 billion in the February 2026 Series G round, to $965 billion in the May Series H round, with this IPO targeting $2 trillion outright. Based on actual 2025 revenue of $4.59 billion, the static price-to-sales ratio reaches an astronomical 436x; even calculated against the $65 billion ARR, the ratio still exceeds 31x. The market's $2 trillion valuation effectively front-loads two years of anticipated explosive performance.
Notably, Anthropic is expected to post an operating profit of $559 million in the second quarter of 2026 (on an adjusted basis excluding stock-based compensation), marking the company as the first to knock on the door of profitability. But for a company saddled with over $400 billion in irrevocable compute commitments, whether a single quarter's profitability signal can be sustained remains an open question.
Morgan Stanley's analysis of Muse reveals another dimension of reality: Meta possesses compute capacity to serve 1 billion users, has fewer than 3 million actual users, and growth has fallen to one-third of its peak. Meta's valuation expansion rests entirely on a perfect growth narrative of 1 billion users, and the gap between narrative and reality is widening.
Meta's third-quarter earnings report on November 4 will be the first critical test of its AI strategy by the market. At that point, Meta will have to confront a core question: when user growth from a free strategy begins to slow, and per-user annual service costs far exceed average revenue contribution, how much longer can this cash-burning experiment for scale continue?
For the broader AI industry, the commercialization race for agents has only just begun. Meta has chosen a platform model centered on advertising and transaction commissions; OpenAI and Anthropic have chosen a paid model centered on subscriptions and APIs; and Chinese companies like Tencent have carved out a different path leveraging closed ecosystems and payment infrastructure. Whichever path ultimately prevails, one fundamental question persists: when AI agents begin making consumption decisions on behalf of humans, does their loyalty belong to the user—or to whoever pays their bills?