Jonathan Tower 🦄🦄🦄🦄🦄🦄

10/10 Winners earn returns above their cost of capital across hardware replacement cycles. Losers may have great technology and packed facilities. Wall Street eventually insists on being paid. My latest: The AI Gold Rush Is Colliding with the Cost of Capital: tinyurl.com/4es7r4yv

9/10 Today's hyperscalers have stronger balance sheets, and inference may absorb enormous compute capacity. But the question remains: who captures the economic surplus, and who merely finances its creation? A full data center proves little.

8/10 The telecom boom already ran this experiment. Global Crossing and WorldCom built for an explosion in internet traffic. They got the demand forecast right, then drowned in debt and excess capacity. The internet won. Plenty of its financiers didn't.

7/10 CoreWeave illustrates the tension: extraordinary demand, expensive infrastructure, customer concentration, and refinancing exposure. J.P. Morgan cited credit-default-swap spreads near 1,200 basis points. Revenue growth doesn't repeal credit risk.

6/10 Utilization is another trap. A cluster with $100 million in annual revenue potential earns $90 million at 90% utilization, but only $60 million at 60%. Debt service barely notices. Discount prices to fill capacity, and margins take another beating.

5/10 Then comes the hardware treadmill. A $600 million GPU cluster depreciated over six years carries $100 million in annual expense. At three years, it's $200 million. Depreciation is noncash, but replacing obsolete chips is painfully real.

4/10 Assume that project earns $100 million annually before interest, taxes, and hardware replacement. Raising borrowing costs from 5% to 9% cuts cash remaining after interest from $70 million to $46 million. That's a 34% haircut without losing a customer.

3/10 Consider a $1 billion AI infrastructure project funded with 60% debt. At 5% interest, annual borrowing costs are $30 million. At 9%, they're $54 million. Same facility, same customers, same GPUs. An extra $24 million goes to lenders.

2/10 S&P Global Ratings projects more than $1.3 trillion in cumulative capex among six hyperscalers by 2027, with negative free operating cash flow expected through that year. Even the richest tech companies cannot finance infinity for free.

1/10 You can be absolutely right about the future of AI and still lose a fortune financing its arrival. Silicon Valley is betting that insatiable demand for compute will justify almost any investment. The cost of capital has other ideas. 🧵

10/10 Before hunting for bargains, trace where independent customer cash enters, which obligations survive slower growth and whether backstops withstand the same shock. My latest: “The Tech Selloff Has a Circular Financing Problem.” tinyurl.com/ywmedhjx

9/10 An LP can own venture stakes in model developers, supplier stocks, infrastructure loans and data-center real estate. Four allocation categories can depend on the same buyers. Private marks adjusting more slowly than public shares do not make that exposure disappear.

8/10 Those decisions travel through the network. Infrastructure providers face weaker utilization, suppliers receive fewer orders, and lenders reassess collateral. A supplier holding customer equity can suffer twice: weaker sales and a declining investment.

7/10 A falling share price does not cancel purchase orders. Cash-rich companies can keep investing. Pressure builds where expansion requires fresh capital. Lower valuations make equity more dilutive; tougher lending terms can force spending cuts or capacity delays.

6/10 CoreWeave illustrates the overlap: Nvidia is an investor, equipment supplier and buyer of computing capacity. Those relationships can strengthen the business. Their value as protection still depends on how each participant withstands the same industry slowdown.

5/10 Investors can mistake an equity investment, a purchase commitment and rising revenue for three independent endorsements of value. All three may depend on the same forecast of eventual customer demand. That is one assumption wearing three suits.

4/10 A supplier can deliver real equipment and recognize legitimate revenue while financing the customer’s purchase. The valuation question is whether that customer can sustain spending once the supplier’s capital runs out.

3/10 Chipmakers invest in companies that buy their equipment. Cloud providers fund developers that purchase their computing capacity. These arrangements can build useful infrastructure while making revenue harder to underwrite. Both can be true.