The Terminator Distraction

Programming note: ARPU will return on October 1 to examine the evolving AI debate across capital markets.

When Cash Burn Meets Science Fiction

If you spent the past week reading the dispatches coming out of Silicon Valley's frontier AI labs, you could be forgiven for thinking you were reading the script of a high-concept sci-fi thriller.

First came the technical euphoria: OpenAI researchers were celebrating after their latest internal model reportedly cracked an impossible Millennium Prize math problem. Then came the panic: across the bay at Anthropic, a researcher quit in protest while colleagues publicly warned that the technology carried a 10% chance of killing all humans.

Finally, the drama reached the C-suite. In a rare truce between bitter rivals, Anthropic CEO Dario Amodei called to slow development, prompting Sam Altman to publicly agree that the industry needed to "pace the frontier"—before announcing that OpenAI would postpone its planned 2026 public offering because the risk of extinction made it an "ill-advised moment to go public."

It is a compelling narrative: the world's most brilliant minds, trapped in a tragic arms race, pleading for a slowdown to save humanity.

But when a company burning billions of dollars suddenly starts using science fiction to explain its corporate strategy, it usually pays to look at its income statement.

Sci-Fi Cover Stories and Price Wars

While the headlines were dominated by existential dread, a much more revealing detail slipped out at a Goldman Sachs technology conference. There, OpenAI's CFO quietly confirmed that the company had to slash prices on its flagship GPT-5.6 Luna model shortly after its release to stimulate usage.

For anyone looking at the numbers, that pricing detail is arguably far more telling than any philosophical debate over existential risk.

The economic reality facing frontier labs is that their core product—tokens—is caught in a brutal, deflationary squeeze. Open-source releases from Meta and cut-rate models from Chinese labs like DeepSeek have dragged down pricing across the industry.

But while the market price of an AI token is falling by 50% to 80% a year, the input costs—massive data centers packed with expensive Nvidia GPUs—remain stubbornly fixed. When your capital costs are rigid and your output prices are collapsing, your gross margins face severe structural compression.

In that light, pausing an IPO looks less like a moral awakening and more like basic corporate prudence. Telling Wall Street you are postponing a public listing because your models are "too dangerous for humanity" sounds heroic; admitting you are postponing it because you just started a margin-crushing price war is an entirely different conversation with an investment banker.

Anthropic, meanwhile, is attempting the opposite high-wire act. Rather than pause, it is still pushing toward a public listing this year, reportedly pitching investors on a staggering $30 trillion total addressable market—a figure equivalent to the entire GDP of the United States.

Marketing a $30 trillion TAM while your own researchers are warning of a double-digit chance of ending civilization might look like corporate cognitive dissonance. But the two narratives actually reinforce each other: if your software is dangerous enough to destroy the world, it must be valuable enough to justify an astronomical valuation.

The Concentraion Risk

The deeper analytical risk hanging over this market is not that an autonomous agent will escape a sandbox and take over a power grid. It is the extreme counterparty concentration propping up the entire infrastructure chain.

Consider Oracle's latest quarterly earnings. On the surface, the numbers looked solid, with Remaining Performance Obligations (RPO)—the company's contracted backlog—reaching a record $664 billion.

The catch? Roughly half of that backlog comes from a single customer: OpenAI.

That is a huge concentration of risk. A legacy enterprise tech giant with a market cap over $400 billion has tied half of its future cloud growth to an unprofitable AI lab that generates massive negative cash flow and has just pushed its public listing into an uncertain future.

Zooming out, this is the structural vulnerability sitting at the heart of the AI trade. The entire cloud buildout has become an insular loop: tech giants are pouring hundreds of billions into data centers largely to capture revenue from two cash-burning startups—OpenAI and Anthropic—that the tech ecosystem has to keep subsidizing with fresh equity. If either lab faces a funding crunch, hits a growth plateau, or renegotiates its compute commitments downward, that shockwave travels straight through Oracle, Microsoft, and the semiconductor supply chain.

For capital markets, this counterparty concentration matters far more than the cinematic imagery of Terminator. A financial cycle does not need autonomous rogue machines to stumble; it only requires the market to confront how little we still know about real-world enterprise return on investment. Four years into the generative AI boom, outside of software engineering, we still lack clear empirical evidence that corporate buyers are willing to spend hundreds of billions of dollars on recurring AI software licenses. Without that sustained end-user revenue, the multi-trillion-dollar infrastructure pyramid has no organic way to support itself.

Silicon Valley wants the world to brace for the apocalypse. Wall Street's focus is far more grounded: watching to see if real-world demand can ever catch up to all that AI capex.

Signal Stack

The operating reality beneath the headlines.

• Paying for Frontier AI Models Buys 4-Month Head Start at 5x the Cost (Ars Technica) – Mozilla finds the capability gap between US closed models and Chinese open-weights models has narrowed to 4.4 months, meaning the premium buyers pay for frontier access is now a time premium rather than a capability one.
• Is AI Impacting Global Labor Markets? (Goldman Sachs) – A 10% increase in occupational AI exposure is associated with just a 0.1 percentage point drag on annual headcount growth across France, Canada and the US, which puts the measurable labour effect three years into the buildout well below what the displacement discourse assumes.

📺 On Our Channel

Why China Thinks AI Can Reinvent Communism

Historically, communist central planning always foundered on one problem: information. Markets signal scarcity through prices. Planners, with no such mechanism, never had enough information to know what to produce or where to send it. Kevin Rudd—former Australian prime minister, ambassador to Washington, and a scholar who has spent decades reading the Chinese Communist Party's own ideological writings—argues in a recent lecture that Beijing may now see AI as the thing that closes that gap. We looked at what happens if the state starts treating algorithms as a substitute for the price mechanism.

Watch ARPU's deep dive on YouTube (13 Mins)

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.