The Logic of the AI Stack

Financing Jensen's Cake

Nvidia's Jensen Huang has offered a clean, orderly way to think about the AI economy. At a recent conference, he described the industry as a "five-layer cake." At the bottom, you have Energy. Then comes Chips, Infrastructure, AI Models, and finally, Applications. It is a neat, logical stack—a structural framework for how a new economy is built.

Yet there is a slight complication with this framework: you have to pay for the entire cake upfront, but you don't find out how many guests are coming to the party until two years later.

This is the central dilemma of the AI revolution. While Huang sees an orderly stack, Dario Amodei, the CEO of Anthropic, sees a "cone of uncertainty."

In a candid interview with The New York Times, Amodei explained that while his revenue has grown 10x year-over-year, he faces a massive lag time. Data centers take years to build. He has to order hardware today for revenue that may or may not exist in 2027. If he under-buys, he loses the market. If he over-buys, he risks bankruptcy.

To bridge this gap, the industry heavily relies on a funding mechanism: vendor financing (often called "circular deals"). While critics call it a bubble, Amodei argued that the logic is actually sound—up to a point. He broke down the mechanics of why these deals happen:

• The Amortization Math: A gigawatt cluster requires roughly $50 billion in upfront capital. But amortized over a five-year useful life, that creates a "run rate" of $10 billion per year.
• The Shared Incentive: The cloud host (e.g. Microsoft) wants to move the hardware but knows the startup doesn't have $50 billion cash upfront.
• The Bridge: The vendor invests one year's worth of that cost (e.g., $10 billion) into the lab, usually in the form of equity. This effectively pre-funds the first year of the contract.
• The Growth Bet: For years two through five, the lab relies on its skyrocketing revenue growth to cover the remaining $40 billion.

"[The lab will] have the revenue at the right time" Amodei said, "So I don't think there's anything inappropriate about that in principle." It bridges the timing gap between building the infrastructure and earning the revenue.

The problem, as Amodei admits, is when you "start stacking these where they get to huge amounts of money." It's one thing to finance a $10 billion stream; it's another when your internal financial projection requires you to hit $200 billion in annual revenue just to break even. This is where the risk dial is turned too far.

But if you move down one layer in the stack—from the Model layer to the Infrastructure layer—the anxiety over future revenue disappears, replaced by a battle with physics.

In a conversation with a16z, Amin Vahdat, the VP of Google Cloud, described the view from the infrastructure layer. Because of the aggressive spending by frontier labs like Anthropic, and the record amount of capital flowing into AI startups, Google isn't seeing a demand cliff; they are seeing a supply wall.

"10x is an understatement," Vahdat said regarding the build-out compared to the internet era. "It's 100x what the internet was."

While Anthropic worries about the uncertainty of future revenue, Google is struggling to keep up with the current orders. Vahdat noted that even their 7- and 8-year-old TPUs (ancient history in AI terms) are running at 100% utilization. This is the data center equivalent of discovering that people are so desperate for transportation that they are paying top dollar to ride a 2012 Honda Civic. The demand for compute is so massive that Google literally cannot "cash the checks" their customers want to write because they are limited by power, land, and permitting.

This clarifies the state of the industry. The tension is one between the layers of the cake.

• Layer 4 (Models): Faces Financial Risk. They must spend billions today on the assumption that the market will pay for it tomorrow.
• Layer 3 (Infrastructure): Faces Physical Reality. They are receiving massive orders today but are constrained by the speed of pouring concrete and permitting power plants.

The industry isn't building "bridges to nowhere." It is attempting the fastest infrastructure sprint in human history. The danger is a duration mismatch: We are trying to finance a physical construction project (which moves at the speed of pouring concrete) with software venture capital (which moves at the speed of interest payments).

But this tension between finance and physics rests on one final, unspoken assumption about the top layer of the cake: that the guests will eventually show up to the party. Anthropic's financial risk only matters if the revenue model is fundamentally sound. Google's physical crunch is only a problem if the applications built on top of its servers create durable, profitable businesses.

Ultimately, the entire AI cake is being baked on credit. The question is not just whether the financial layer can outlast the infrastructure build-out, but whether the end-users of the applications are willing to pay for it all before the bill comes due. For now, the application layer has proven incredibly effective at generating usage, but far less so at generating profit. How that gap closes is, of course, a topic for another day.

More on AI Boom:

• Is A.I. Actually a Bubble? (The New Yorker)
• Is There Enough Data Center Capacity for AI? (Goldman Sachs)

On Our Radar

Our Intelligence Desk connects the dots across functions—from GTM to Operations—and delivers intelligence tailored for specific roles. Learn more about our bespoke streams.

Oracle's Buildout Bottleneck

• The Headline: Oracle has delayed the completion of some data centers for its massive OpenAI contract to 2028, citing labor and material shortages. (Bloomberg)
• ARPU's Take: Oracle's all-in, hyper-speed bet on building out AI infrastructure is now colliding with the messy realities of the physical world. The real-world shortage of skilled labor and construction materials is revealing a critical and potentially systemic vulnerability in its high-stakes strategy.
• The GTM Question: This delay gives competitors (Microsoft Azure, AWS) a tactical opening. If OpenAI cannot get the capacity it needs from Oracle on time, it may be forced to diversify workloads back to Azure or other providers. This weakens Oracle's "AI Supercluster" narrative, suggesting they over-promised on delivery timelines to win the marquee contract, potentially damaging credibility with other large AI customers who need capacity now, not in 2028.

Nvidia's China Tightrope

• The Headline: Following a US policy shift, Nvidia is now weighing an increase in H200 chip production to meet robust demand from Chinese tech giants, a decision complicated by its own supply constraints and pending approval from Beijing. (Reuters)
• ARPU's Take: The US government just opened a multi-billion dollar door for Nvidia, but the move was so unexpected that the company wasn't prepared to meet the flood of demand. Now, Nvidia is in a scramble to reconfigure its production plans, all while the ultimate gatekeeper is no longer Washington, but Beijing.
• The Operations Question: Ramping up H200 production creates a "capacity conflict" for Nvidia. Since the H200 shares TSMC's 4nm node with the H100 and other products, increasing its volume means potentially sacrificing capacity for newer Blackwell/Rubin chips or other lines. Nvidia will balance the lucrative but politically volatile Chinese demand against the strategic need to transition Western customers to next-gen architecture, risking inventory overhang if Beijing suddenly blocks the H200.

P.S. Tracking these kinds of complex, cross-functional signals is what we do. If you have a specific intelligence challenge that goes beyond the headlines, get in touch to design your custom intelligence.

You received this message because you are subscribed to ARPU newsletter. If a friend forwarded you this message, sign up here to get it in your inbox.