AI’s Big Guns Have a Serious Inflation Problem

Some of the most cash-generative tech firms and best-funded startups in history are frantically competing to secure enough hardware, fearing that otherwise they’ll be left behind in the race for superintelligence. And while these buyers are relatively price-insensitive, their chip suppliers tend to have dominant market positions with huge technical and financial barriers standing in the way of more competition. Controlling these bottlenecks is proving very lucrative.

Taiwan Semiconductor Manufacturing Co., the largest foundry for advanced chips, plans to invest a record $56 billion or so this year, but that still won’t provide enough cutting-edge products. Such constraints have prompted Elon Musk to consider building his own chip factory. That will cost at least $55 billion and possibly as much as $119 billion.

When demand exceeds supply, prices must go up. That’s why shares in AI hardware suppliers are outperforming those of most of their hyperscaler buyers, who are burning through mountains of cash to stay in the game. Alphabet Inc. is an exception to this, assisted by its in-house chip innovations.

“Nearly all the value has accrued to the chip layer, which is both unprecedented and unsustainable,” Goldman Sachs’ global head of equity research James Covello tells clients. “The chip companies are thriving at the expense of everyone above them in the chain.”

Hyperscalers’ latest earnings reports reveal the painful sting of inflation. Microsoft expects higher component prices to add $25 billion to its full-year capital spending bill, which now totals an eyewatering $190 billion. Meta hiked the midpoint of its forecast capex range by $10 billion, attributing it mostly to component costs, particularly memory chips.

Nvidia Corp.’s GPUs — the workhorses of AI costing tens of thousands of dollars each — have long been a money machine, generating 75% gross profit margins. While its latest models are incredibly expensive, they’re also much more efficient in terms of output and power usage. The company’s parallel processing and software innovations deserve reward. Yet its dominance of the market for AI accelerators, the hardware used to perform tasks, means its fat profits have been dubbed the “Nvidia tax.”

Tech companies are increasingly having to pay a “memory tax,” too, because datacenters are consuming so much of it. Part of the reason is that most advanced AI accelerators need a lot more high-bandwidth memory; and that’s provided by a very profitable and silicon-intensive form of dynamic random access memory, which offers fast, temporary storage for data and apps.

The big three DRAM suppliers — SK Hynix Inc., Samsung Electronics Co. and Micron Technology Inc. — have become stock-market darlings and are today worth more than $2.8 trillion collectively.

SK Hynix’s operating margins reached a record 72% in the latest financial quarter. Customers are “prioritizing securing volume over pricing,” the South Korean company says, unabashed. Samsung’s average selling prices for DRAM increased by more than 90% in the same three-month period compared with the quarter preceding it.

In aggregate, spending on various types of memory could account for 30% of hyperscalers’ capex in 2026, according to research firm SemiAnalysis. It was just 8% in 2024.

Hyperscalers and neoclouds (companies that rent GPUs to those needing computing capacity) remain confident that their spending will pay off, even if it’s often costing more than they thought. Supply constraints at least lower the risk that older, less efficient processors quickly lose their value. Rental prices are “increasing across the board,” says CoreWeave Inc., a prominent neocloud.

Still, the sheer amount of chip capex is giving hyperscalers a strong incentive to find ways to reduce their computing costs. One way is to tap alternative AI processors such as those from Advanced Micro Devices Inc. or to devise their own hardware — Alphabet’s tensor processing units, Amazon’s Trainium chips and Microsoft’s Maia 200, for example.

Amazon expects Trainium to save it tens of billions of dollars yearly on its own spending. Claude and ChatGPT’s owners, Anthropic PBC and OpenAI, have signed multibillion dollar contracts for the chips with Jeff Bezos’s firm, although in the near term most supplies are sold out or reserved.

Among other promising innovations, Google’s TurboQuant compression technique could help curb memory spending, and Arm Holdings Plc expects its new CPU to cut the cost of a gigawatt of datacenter capacity by $10 billion or so. Meantime, the tech elite’s largess is having unwelcome spillover effects. Importing AI-related hardware from Taiwan and elsewhere is worsening the US trade deficit — viewed by President Donald Trump as a barometer of economic failure.

And beyond the AI frenzy, manufacturers of smartphones, games consoles and PCs are struggling to secure memory chip supplies because their makers are prioritizing the more lucrative datacenter market and long-term contracts with hyperscalers. Consumer-tech product firms must either pass on price increases, lower device specifications or swallow a hit to their margins. Global smartphone sales are projected to decline around 13% this year, with budget handsets particularly affected. Nintendo Co. Ltd. has lifted the price of its Switch 2.

Semiconductor factories take years to build, so there’s no prospect of a swift supply response. It’s a notoriously cyclical industry and several companies suffered heavy losses not so long ago, making them wary of overexpanding.

When you factor in higher electricity prices caused by power-hungry data centers, AI will probably be quite inflationary for a while.

“Massive demand for semiconductors, memory capacity, and other components of the AI infrastructure buildout seems to be spilling over into consumer prices,” says Pimco economist Tiffany Wilding, pointing to the rise in personal-consumption inflation.

If the US Federal Reserve is unable to cut interest rates because of all this, the AI labs’ single-minded and astonishingly expensive pursuit of superintelligence won’t just look financially reckless. From a societal perspective it means we’ll all end up paying.


A message from Advisor Perspectives and VettaFi: Discover something new! Click here to register for our upcoming webcasts.

Bloomberg News provided this article. For more articles like this please visit bloomberg.com.

Read more articles by Chris Bryant