Token factories are expanding while token prices are falling

Hyperscalers are spending heavily on GPUs, HBM, networking, power, and data centers. Some are issuing debt or locking in long-term power and leasing commitments. At the same time, AI labs and model platforms are cutting subscription and API prices to win users.

That makes the setup uncomfortable for investors. Semiconductor and memory suppliers receive the order first. Data-center owners absorb depreciation, power cost, and build-out risk first. If AI service prices fall before customers absorb the cost, the market stops asking whether AI matters and starts asking when the spending comes back as cash flow.

The infrastructure stayed. Many railroad companies did not.

The British Railway Mania of the 1840s is a useful comparison. Railroads were real industrial technology. They lowered logistics costs and connected cities, factories, and markets. But rising railway shares pulled in too much capital too quickly. Public records describe 263 railway acts authorized in 1846, with a large portion of authorized lines never completed.

The lesson is simple. Infrastructure can be essential for society without guaranteeing good returns for the shareholders who financed it at the wrong price. AI data centers are similar. Usable compute may become a strategic asset, but not every data-center project will earn a strong return on capital.

The internet was right. Fiber was needed. Telecom equities still broke.

The late-1990s telecom and fiber bubble is even closer to today’s AI debate. The thesis was sound: internet traffic would surge, so backbone networks, submarine cables, and fiber capacity would be scarce. The problem was timing and capital intensity. Too much capacity arrived before demand could pay for it, bandwidth prices fell, and leveraged network operators failed.

Global Crossing built a global network but never posted a profitable year before filing for bankruptcy in 2002. WorldCom also collapsed after an accounting scandal. Yet the installed infrastructure did not disappear. Later, Google, Amazon, Netflix, and cloud businesses grew on top of cheaper network capacity.

AI faces the same question. Is the final winner the company that built the data center, or the company that turns cheaper tokens and compute into customer workflow and recurring cash flow?

Even essential products can destroy capital

DRAM and memory have always been necessary for computing. But the industry repeatedly moved through demand surges, price spikes, capex expansion, oversupply, price collapse, and producer exits. Elpida Memory filed for bankruptcy in 2012 and was later acquired by Micron.

Today, higher HBM and memory prices help suppliers. They also raise token cost for hyperscalers and AI labs. HBM strength is therefore a two-sided signal: good for memory margins, but potentially painful for AI service ROI.

The direction can be right while the cost curve defeats the company

Solar was also directionally right. Installations grew massively over time. But some manufacturers could not survive Chinese supply growth, faster cost declines, and price competition. Solyndra received a large U.S. loan guarantee and still failed.

AI has the same risk. “AI demand will grow” is not the same statement as “this data-center or model operator will earn a high return on capital.” When prices fall quickly, the operator with the largest fixed cost feels the pressure first.

AWS and Tesla show the other side of the cycle

There are successful examples. AWS was not just server rental. Amazon connected internal infrastructure skill, developer APIs, payments, operating scale, and customer lock-in. Cloud prices fell, but AWS defended the business with scale, service bundles, and operating efficiency.

Tesla also came close to failure in 2008 before strategic financing helped it survive. The point is not that every risky infrastructure bet fails. The point is that in major industrial transitions, being directionally right is not enough. Financing survival matters.

The capex cycle has several possible endings

The AI growth thesis remains. The liquidity bill is larger.

The current setup is growth-positive but ROIC-sensitive. The market no longer pays a premium for the word AI alone. It is starting to separate companies that convert capex into revenue, margin, and free cash flow from those that merely spend.

Companies that turn expensive infrastructure into customer cash flow

The strongest candidates combine distribution, internal demand, software lock-in, balance-sheet strength, custom silicon, and data-center execution. They can defend economics even if token prices fall.

Suppliers with real benefit but heavy expectations

GPU, HBM, power, cooling, and networking suppliers remain major beneficiaries. After large moves, however, investors should wait for earnings revisions, utilization, support levels, and capex durability rather than chase every spike.

The numbers that separate productive capital formation from a capex chicken game