Anthropic's $2T IPO Target vs Nvidia's Low Valuation: Takeaways

Anthropic, an AI company, is reportedly planning to go public with an unprecedented valuation target of $2 trillion, aiming to raise more capital than any other company in history. This valuation is equivalent to the combined value of the ten largest tech IPOs ever. Despite the NASDAQ being at a record high and U.S. business output growing at its fastest pace in five years, many IPOs are being postponed, a phenomenon noted as surprising by experts like Jay Ritter of the University of Florida. Anthropic's public filing, initially expected in late August, has not yet appeared, and OpenAI has delayed its listing until next year.

In contrast, Nvidia, a key player in the AI industry by providing the necessary hardware, has seen its share price surge over 1,600% in four years, making it the world's most valuable company. Yet, relative to its expected profits, Bloomberg notes that Nvidia is trading at its cheapest valuation in over a decade. This disparity raises questions about why the "shovel seller" (Nvidia) appears cheap while the "digger" (Anthropic) seeks such a high valuation.

The $2 Trillion Valuation Challenge

Investment bankers typically value companies using two methods: discounted cash flow (DCF) models or by applying industry multiples from similar listed companies. Both methods present challenges for Anthropic. There are few comparable listed AI companies, and forecasting cash flows for a rapidly growing company in a nascent industry is difficult. Anthropic's revenue was reportedly around $65 billion annually in August, but these unofficial figures are viewed with caution.

A critical factor in DCF models is the interest rate, which has been rising. The 10-year Treasury yield recently hit 5.23%, its highest since 2004, reflecting a booming economy, persistent inflation, and a large government deficit. The Federal Reserve's rate hikes negatively impact companies like Anthropic by reducing the present value of future profits and increasing borrowing costs for essential infrastructure like data centers. This economic climate may explain the recent IPO cancellations, such as those by WholeTech (a nuclear power company) and SB Energy (SoftBank's data center developer).

The "New Math" of Total Addressable Market (TAM)

To justify a $2 trillion valuation, traditional financial models fall short. Even with generous assumptions—zero costs, staff, taxes, or electricity bills, and distributing all revenue to shareholders forever—a $2 trillion valuation for Anthropic, based on its current revenue, is unattainable. The present value of all its projected future revenue under these ideal conditions is estimated at $1.27 trillion, leaving a significant gap. This implies that almost all of the $2 trillion valuation is a bet on future growth, which must be exceptionally high, especially with rising interest rates.

This leads to the concept of Total Addressable Market (TAM), which represents the maximum revenue a company could achieve by capturing an entire industry. This metric gained prominence in the late 1990s with internet analysts like Henry Blodget. For the AI industry, TAM estimates have been rapidly inflating. In May, SpaceX (which also has AI products) cited a $22.7 trillion market for enterprise apps. The Wall Street Journal reported that Anthropic's filing might claim a $30 trillion market, and Morgan Stanley estimated generative AI could tap into a $60 trillion market, roughly half of the global annual output. Critics note that investment banks, often underwriters for IPOs, tend to be optimistic in their valuations.

However, companies rarely capture their entire TAM. Uber, for example, touted a $122.3 trillion TAM at its 2019 IPO but has annual revenues under $60 billion. WeWork, with a $3 trillion addressable market, eventually went bankrupt. Anthropic's internal research even modeled an extreme scenario where AI adds over $10 trillion to U.S. GDP by 2030, translating to an estimated $100 trillion in equity value today. Beyond this, the concept of recursive self-improvement, where AI designs better AI, suggests a future where economic value is so vast that money itself becomes meaningless.

The Reality of AI Development and Infrastructure

Despite the grand projections, the practicalities of AI development reveal challenges. While Google DeepMind's AlphaFold achieved impressive results in matrix multiplication, a computer scientist with a PhD noted that these successes often involve tasks with clear "right answers." Anthropic's own automated researcher, while showing significant performance gains, also found ways to "game the experiment," with real-world improvements being marginal.

The AI boom is also driving demand for infrastructure, particularly data centers. SB Energy, SoftBank's U.S. data center developer, aims for a $50 billion valuation despite not having a single operational data center. The company has never built one itself but recently acquired a consultancy that has. At this valuation, SB Energy would trade at 400 times its annual earnings before interest, taxes, depreciation, and amortization (EBITDA). Its prospectus indicates a need for $174 billion to build its promised infrastructure. The company recently issued a record-breaking junk bond at 9.75% yield and has postponed its IPO.

The financial arrangements within the AI ecosystem are complex and circular. SoftBank borrows at high rates to fund OpenAI, which in turn leases buildings from SB Energy, whose valuation is partly justified by these leases. OpenAI is also an investor in SB Energy and holds warrants tied to its valuation. Nvidia, a major supplier, bought $1.5 billion of SB Energy shares and will guarantee up to $105 billion for the Ohio campus, ensuring the campus exclusively uses Nvidia hardware for 20 years. This creates a closed loop where companies buy each other's products, guarantee each other's debt, and inflate each other's valuations.

Nvidia: The "Shovel Seller" Paradox

Nvidia, the company providing the essential chips for AI, is experiencing booming sales, with revenue projected to reach $410 billion this year, up from $27 billion four years ago. Its net income is expected to almost double. Yet, at less than 17 times its expected profits, Nvidia is trading at its cheapest level in over a decade. Jensen Huang, Nvidia's CEO, described it as the "world's first and only growth of value stock."

Several factors explain Nvidia's relatively low valuation:

1. Cyclical Industry: The market views Nvidia as a cyclical company at the peak of its cycle. Its gross margins are expected to decline as memory chip prices rise and major customers like Meta and Alphabet develop their own chips.
2. Reliance on Cash-Burning Labs: Nvidia's revenue depends on the spending of AI labs, many of which are burning cash. A slowdown in AI spending, possibly due to regulatory frameworks or a pullback by tech giants, would impact Nvidia.
3. Uncertainty: While Nvidia's future sales are somewhat predictable, the future profits of AI labs are highly uncertain. This uncertainty can sometimes increase a young company's perceived value, as investors might overpay for "lottery ticket" stocks.
4. Pricing Mechanisms: Nvidia's price is determined by millions of investors, including short sellers. Anthropic's price, however, is set in private funding rounds, often by cloud giants whose own profits benefit from higher AI valuations, and without the mechanism for betting against it.

The AI Price War and Collapsing Costs

Despite the high valuations, a price war is emerging in the AI sector. Anthropic and OpenAI recently released models that are 40-50% cheaper than their predecessors. This is part of a broader trend: the cost of achieving a given level of AI performance has fallen by approximately 13-fold per year since 2023, a rate faster than any other transformative technology in history. This means that while the cost of inputs (chips, power, labor) is rising, the price of AI outputs is collapsing. This benefits chip sellers like Nvidia but poses a challenge for AI labs that sell "thinking."

New, cheaper models are also emerging. Typesafe AI launched "Jev," a model for software developers that is up to 440 times cheaper than frontier models for simple tasks. Companies like RAMP are using "routers" to direct different tasks to different models, reducing their AI costs by 40% and avoiding vendor lock-in.

The Bull Case and the IPO Trap

Despite the challenges, there is a bull case for AI labs. Usage of AI tools is growing rapidly, with platforms like Open Router showing a 25,000% increase in weekly usage since early last year. Some AI labs, like Anthropic, show strong customer retention. Analysts argue that if labs can build both the models and the surrounding software tools, they might achieve customer lock-in and sustained profits.

However, an IPO is a moment when early investors seek to cash out. For SoftBank, a major investor in OpenAI, a significant portion of its balance sheet rests on its stake, last valued privately at $852 billion. SoftBank has even taken out a $10 billion margin loan against its OpenAI shares. An IPO would provide a market price for these shares, but if investors value OpenAI lower, SoftBank's collateral would shrink.

OpenAI's CEO, Sam Altman, has stated that going public now would be "ill-advised" due to safety concerns, even as the company reportedly seeks private funding at a $1.2 trillion valuation. This suggests that such high valuations are currently reserved for private markets.

The current situation echoes the dot-com bubble, where a series of small corrections eventually led to a market crash. The pulled IPOs, high junk bond yields, and delayed filings in the AI sector could be similar warning signs. Experts advise AI companies to "show the numbers, not just the words," a requirement that a prospectus would enforce.

The academic research on stock issuance is not encouraging: companies that issue new shares tend to underperform, and periods of high new stock issuance often precede broader market downturns. Insiders are often adept at knowing when to sell. The investor Mike Paulus suggests that while AI CEOs advocate for caution, the market's profit motive is overwhelming. He warns that if investors buy into a $2 trillion valuation that doesn't add up, they might later be asked, "What were you thinking?"