The Scott McNealy Test of AI Valuations

Programming note: ARPU will return next Thursday to look at the economics of AI deployment.

31x Revenue

In 2002, two years after the dot-com bubble burst, Scott McNealy sat down with BusinessWeek and gave what might be the only interview in Wall Street history where a CEO explained in detail why anyone who bought his stock was an idiot.

His company, Sun Microsystems, had been one of the great winners of the early internet. Its servers ran the web, its revenue was real, and at the peak of the frenzy its stock traded at $64 a share, or roughly 10 times revenue.

Looking back at that valuation, McNealy walked investors through the math of what paying 10 times revenue actually implied:

At 10x revenues, to give you a 10-year payback, I have to pay you 100% of the revenues for 10 straight years in dividends. That assumes zero cost of goods sold. That assumes zero expenses, which is really hard with 39,000 employees. That assumes I pay no taxes, which is very hard. And that assumes you pay no taxes on your dividends, which is kind of illegal. And that assumes with zero R&D for the next 10 years, I can maintain the current revenue run rate.

Now, having done that, would any of you like to buy my stock at $64? Do you realize how ridiculous those basic assumptions are? What were you thinking?

Twenty-four years later, Wall Street is preparing to run the same arithmetic with a much larger number.

According to a confidential IPO prospectus leaked to Reuters, Anthropic is preparing to go public at a valuation of around $2 trillion. With Anthropic recently crossing $65 billion [1] in annualized revenue, that works out to about 31 times revenue.

Sun Microsystems, of course, was never growing at 1,000% a year. Public markets are forward-looking discounting mechanisms, and paying a 31x multiple is standard practice if you believe an exponential revenue curve can continue. But for Anthropic to justify that premium over the long run, investors need to believe at least three things simultaneously:

• The market for AI becomes enormous.
• Anthropic captures a disproportionate share of that market's economic value.
• The cost of staying at the frontier does not consume most of the profit.

The first assumption is the easiest one to sell. The other two are where the math becomes much more complicated.

The Narrative of Scale

It helps to justify a rich valuation if you have a compelling narrative about future market size.

That is where the Total Addressable Market (TAM) comes in handy.

When near-term multiples look stretched, the time-tested corporate playbook is to zoom out until the numbers look small by comparison. The run-up to Anthropic's public debut has produced some truly cosmic benchmarks:

• In May, SpaceX told investors that the addressable market for enterprise software and AI was $23 trillion—roughly three-quarters of the entire United States economy.
• Anthropic's leaked filing raises the bar, citing a potential market of $30 trillion.
• Morgan Stanley went further still, publishing an estimate that generative AI could eventually touch $60 trillion—or roughly half of the annual economic output of the entire planet.

If you assume the TAM for machine intelligence is $30 trillion, paying $2 trillion for the leading frontier lab starts to look like a bargain.

The historical catch, however, is that creating economic value has never been the same thing as capturing it. The railroads transformed global commerce, but the builders went broke while steel mills captured the surplus; the dot-com boom laid the internet's fiber, but the telecom companies rarely kept the profits.

In the AI stack today, that same structural leak is already visible: the strongest pricing power sits in the compute layer, not the models. Companies providing the underlying silicon—from TSMC to Nvidia—can dictate terms because you cannot run AI algorithms without chips.

The model layer, meanwhile, looks like a capital pass-through. Anthropic spent $7.3 billion on compute last year just to generate $4.6 billion in revenue, and has signed up for an eye-watering $518 billion in future compute commitments (80% of which are non-cancellable). The money comes in the front door from software budgets, and goes out the back to fund someone else's data centers and GPU clusters.

The Four-Month Moat

The deeper problem for frontier labs is that their pricing power requires a durable technological lead—and right now, that lead is short.

Research from Epoch AI suggests that top open-weight models trail the proprietary frontier by an average of just four months.

That is an uncomfortably narrow advantage for an asset that costs billions of dollars to train.

Once comparable intelligence becomes available more cheaply, pricing power starts to erode. Anthropic and OpenAI have already responded with aggressive cuts to mid-tier model pricing, while open-weight alternatives continue to narrow the performance gap.

This puts frontier labs on an expensive treadmill. To command premium pricing, you need to stay at the frontier. Staying at the frontier requires another multi-billion-dollar training run. And a few months later, competitors catch up and the cycle begins again.

The shorter the moat, the more expensive it becomes to maintain—and the more money flows back down the stack to the hardware suppliers already holding the strongest bottlenecks.

That makes Dario Amodei's recent calls for government-mandated "safety pacing" economically interesting. In September, Amodei explicitly urged frontier labs to slow the advancement of model capabilities. Anthropic's own policy framework goes further, arguing that sufficiently advanced AI may eventually require government oversight closer to nuclear energy or financial regulation than ordinary software.

A regulated frontier would effectively freeze the competitive treadmill in place. The longer each model generation remains scarce, the longer frontier labs can charge frontier prices before the market catches up.

And that leaves prospective investors with a rather unusual proposition.

On one hand, Wall Street is pitching a $2 trillion valuation on the promise of capturing a $30 trillion slice of the global economy. On the other hand, Anthropic's own leadership is spending its time warning that the technology is dangerous, and the government ought to step in and slow everyone down.

It is rare to see a company seek the largest valuation in human history while actively asking regulators to put a speed limit on its industry. But if you are trading at 31 times revenue with a moat that expires three times a year, you need all the time you can get.

Footnote:
[1] An important accounting caveat: per the leaked S-1, Anthropic's actual revenue for full-year 2025 was $4.6 billion. The reported $65 billion figure is an "annualized run-rate"—an extrapolated, non-GAAP metric with no standard definition. According to Reuters, Anthropic calculates this using a hybrid formula: taking the last 28 days of consumption-based sales multiplied by 13, adding monthly subscriptions multiplied by 12, and recognizing the gross value of sales through cloud partners like AWS before partner revenue shares. Because roughly 80% of its revenue is usage-based, that 28-day window is a volatile snapshot highly sensitive to short-term compute spikes. Taking $65 billion at face value gives Anthropic the benefit of its best-case velocity (compressing what would otherwise be a 435x multiple on trailing GAAP sales down to 31x), but it is an extrapolated artifact rather than locked-in recurring revenue.

Signal Stack

The operating reality beneath the headlines.

• Tech's Trillion-Dollar Internal Inconsistency (Apollo) – Apollo's chief economist notes that analysts covering tech expect the sector's operating cash flow to more than double to roughly $2.4 trillion by 2028, while analysts covering the rest of the S&P 500—tech's customers—forecast far more modest cash flow growth, meaning one of the two sets of estimates has to be wrong.
• What Is the AI Capex Breakeven Rate? (Financial Times) – Goldman estimates hyperscalers need roughly $300 billion of annual AI revenue just to break even on 2026-27 capex, and about $636 billion to earn the 30% return on capital they generated before the buildout — roughly ten times the current run rate.

You received this message because you are subscribed to ARPU newsletter - you may unsubscribe anytime at the bottom of this email.

If a friend forwarded you this message, sign up here to get it in your inbox.