After the "Nvidia tax" comes the "memory tax"! AI infrastructure costs are spiraling out of control: Memory will account for up to 30% of capital expenditures.

```

The AI arms race is giving rise to a profound inflation crisis. After the high premiums for Nvidia GPUs were dubbed the "Nvidia tax," another "memory tax" is quietly taking shape.

According to estimates by market research firm SemiAnalysis, memory expenditures accounted for about 8% of total capital spending by hyperscale cloud providers from 2023 to 2024; it is estimated that this ratio will jump to 30% in 2026, and climb even higher in 2027 – nearly a fourfold change in four years.

This memory inflation has begun to make a tangible impact on financial statements of hyperscale cloud service providers. Microsoft expects higher hardware component prices to increase its annual capital spending by $25 billion, bringing the total to a staggering $190 billion. Meta has raised the midpoint of its capital expenditure forecast by $10 billion, and primarily attributes the increase to component costs, especially memory chips. The latest financial reports from these two companies epitomize the widespread predicament of "chip inflation" among hyperscale cloud providers. James Covello, Head of Global Equity Research at Goldman Sachs, bluntly stated:

Almost all value is accumulating at the chip layer, which is unprecedented — and unsustainable. The prosperity of chip companies comes at the expense of all other upstream companies in the supply chain.

From "Nvidia tax" to "memory tax": Cost pressure moves upstream

Nvidia GPUs have long been considered the "chief tax collector" in AI capital spending. With AI accelerator chip gross margins as high as 75%, and a monopoly in the AI accelerator market, Nvidia continues to levy implicit premiums on hyperscale cloud providers — a phenomenon dubbed the "Nvidia tax."

However, a new cost pressure is emerging. Bloomberg columnist Chris Bryant pointed out that technology companies are increasingly forced to pay a "memory tax" because the consumption of memory in data centers is staggering. The fundamental reason lies in the mainstream AI accelerators' need for large amounts of high-bandwidth memory (HBM), which is a lucrative and silicon-intensive kind of dynamic random-access memory (DRAM) that provides rapid, temporary storage for data and applications.

According to SemiAnalysis, key drivers of this trend include: DRAM prices are expected to more than double by 2026, with further double-digit percentage increases in 2027; since Q1 2025, LPDDR5 contract prices have tripled, and open market prices may exceed $10 per GB in Q1 2026; HBM continues to face structural supply shortages.

SemiAnalysis also highlighted a widely overlooked market dynamic: Nvidia enjoys "Very Very Preferred" (VVP) DRAM procurement pricing, which is far lower than that for hyperscale cloud providers or the broader market. This arrangement reduces Nvidia's own server cost exposure and lowers overall market price benchmarks, masking the true severity of supply shortages.

The big three reap windfalls: Memory industry profits hit record highs

Market power in memory is highly concentrated among three companies — SK Hynix and Samsung Electronics of Korea, and US firm Micron Technology. Their combined market capitalization now exceeds $2.8 trillion.

The suppliers' financial data is striking. SK Hynix’s latest quarterly operating margin soared to 72%, a record high; the company admitted that customers now "prioritize securing procurement volumes over price negotiations." Meanwhile, Samsung’s average DRAM price in the same period rose by over 90% quarter-over-quarter.

In market performance, the share prices of AI hardware suppliers have broadly outperformed the hyperscale cloud buyers. TSMC plans to invest about $56 billion this year — a record — yet still cannot meet booming demand for advanced products. Musk is even considering building his own chip factory, with estimated costs between $55 billion and $119 billion.

SemiAnalysis warns that the effects of memory inflation have partly been priced into 2026 capital expenditure guidance, but the 2027 repricing has yet to show up in Wall Street’s forecasts.

Tech giants seek breakthroughs: Can in-house chips break monopoly premiums?

Facing high procurement costs, hyperscale cloud providers are actively exploring in-house chip development to break free from reliance on Nvidia and memory suppliers.

Amazon’s self-developed Trainium chips are expected to save it tens of billions in spending annually. Notably, Anthropic and OpenAI have signed multi-billion-dollar chip procurement contracts with Amazon, though most short-term production capacity has been sold out or pre-booked. Google's Tensor Processing Unit (TPU), Amazon's Trainium, and Microsoft's Maia 200 all represent strategic layouts by hyperscale cloud providers to reduce external dependency.

On memory cost reduction, Google’s TurboQuant compression technology may play a role, and Arm Holdings states its next-generation CPUs can reduce data center construction costs by about $10 billion per gigawatt.

Yet, alternatives are currently limited by capacity constraints. Semiconductor factory construction takes years, making it difficult to respond to demand in the short term; coupled with the industry's marked cyclicality, many firms remain cautious about aggressive expansion.

Spillover effect: Dual costs for consumer markets and the macroeconomy

The AI hardware boom is spreading its costs throughout the entire economic system in various forms.

In consumer products, memory chip manufacturers prioritize capacity for the more profitable data center market and long-term hyperscale cloud provider orders, leaving makers of smartphones, game consoles, and PCs facing tight supply — forcing them to make tough choices between price hikes, lower specs, and squeezed profit margins. Global smartphone sales are projected to fall by about 13% this year, hitting low-end models especially hard. Nintendo has announced a price increase for the Switch 2.

At the macro level, some countries' AI hardware imports are widening the US trade deficit, directly contradicting the Trump administration's policy stance. Pimco economist Tiffany Wilding noted, "The huge demand for semiconductors, memory capacity, and other AI infrastructure components seems to be driving up consumer goods prices," citing a rise in personal consumption inflation data as evidence.

Bryan Bryant argued that if the Fed is unable to lower interest rates because of this, the single-minded and costly pursuit of superintelligence by AI labs will not only appear financially reckless — but from a societal perspective, everyone will end up paying the price.

Risk Warning and DisclaimerThe market has risks, and investments require caution. This article does not constitute personal investment advice, nor does it consider individual users’ specific investment goals, financial situation or needs. Users should consider whether any opinions, views, or conclusions in this article fit their circumstances. Investments based on this are at your own risk. ```