This website uses cookies

Read our Privacy policy and Terms of use for more information.

// AI Tangle

The Week AI Ate the Memory Supply

AI data centers are buying up the world's memory. Plus Anthropic's $518 billion compute bill, Bain's $6 trillion test, Samsung's bet on data centers, and a patent fight aimed at AI memory.

A single stick of RAM glowing orange on a workbench inside a vast blue-lit AI data center

A reader replied to an AI Tangle a couple of weeks back with a question I've been chewing on since. He'd noticed that building your own open source AI setup seems to be turning into a trend, and he has a spare desktop with a 4GB graphics card sitting around. He wanted to know if I'd done much of this. It took me straight back to the '90s, when every computer nobody wanted got Linux put on it and a second life.

I told him the truth. I run a 7B-parameter Qwen model on a Mac Mini with 32GB of RAM, and I wish I had a lot more RAM. Then I went and looked at what more RAM costs right now. That's this week's story. The same memory chips that let a small model run on your desk are the ones AI data centers are buying by the truckload, and the rest of us are paying the difference. (Tuesday's AI Lesson takes his actual question — what old hardware can really do — head on.)

// The Big AI Story

AI Data Centers Are Buying the RAM Your Next Laptop Needed

The memory squeeze tightened again last week. On September 30, market tracker TrendForce forecast that conventional DRAM contract prices will rise another 10–15% in the fourth quarter, with NAND flash up 15–20%, as suppliers point their most advanced production lines at server memory for AI. A day earlier, it raised its 2027 outlook for high-bandwidth memory, noting that HBM and ordinary DRAM are now fighting over the same limited wafer capacity. Those increases land on top of a year that already broke records — back in February, TrendForce projected that PC DRAM contract prices would more than double in a single quarter.

The cause is simple economics. HBM — the stacked memory that sits right next to the chips in AI accelerators — is the most profitable thing a memory fab can make right now, so capacity goes there first. In early September, a 36GB HBM3E part reportedly sold for about $2,100 on the spot market, four to five times its long-term contract price, and Samsung has reportedly committed about 70% of its memory capacity through 2031 to long-term deals. Laptops and phones get what's left. TrendForce says higher component costs are already weakening notebook sales and pushing PC makers to ship smaller SSDs to hold their price points. German laptop builder Schenker told Tom's Hardware that DDR5 laptop memory now costs six times what it did in July 2025, and it just raised prices on most of its gaming laptops by $113 to $228. It isn't only memory, either — the company says CPUs, graphics cards, and circuit boards are all getting pricier.

How long this lasts depends on who you ask. Acer's CEO expects average PC prices to climb 5% to 20% this quarter but says a shortage lasting until 2030 is "impossible" and sees prices easing from mid-2027 as more chip production comes online, especially in China. SK Group's chairman warned back in March that the wafer shortfall could run until 2030, with a gap of more than 20% and four to five years needed to build enough new capacity. One is a PC maker that needs prices to fall, the other a memory maker enjoying record margins — weigh both accordingly. Both camps agree on the next 12 months: any laptop refresh, workstation order, or GPU purchase in your 2027 budget costs more than last year's quote. And the spare machine sitting in a closet is suddenly worth more than it was.

// The Number

121%

That's how much TrendForce now expects blended HBM prices to rise in 2027 over 2026 — a forecast it revised upward last week, which tells you the forecasters don't think the shortage has peaked. It also reports that GPU and custom-chip makers are now weighing putting less HBM on each device to stretch supply.

Source: TrendForce

// 4 Quick Hits

1. Anthropic's IPO Filing Shows a $518 Billion Compute Bill

Anthropic's leaked prospectus puts numbers on the demand side of the memory crunch. According to TechCrunch's read of the filing, revenue grew twelvefold to nearly $4.6 billion in 2025 while the company posted an operating loss of more than $8 billion, and it plans $518 billion of spending on cloud, computing, and infrastructure in the coming years. Revenue hit $11.5 billion in the second quarter of 2026 alone, and the listing target is a valuation above $2 trillion. The filing also flags concentration risk, with nearly a quarter of last year's revenue coming from just two customers. For buyers, the takeaway is the scale of committed demand — commitments that size are what let memory makers sell out capacity years ahead.

2. Bain Says AI Needs $6 Trillion a Year to Pay for Its Data Centers

Bain's annual technology report says the AI industry needs about $6 trillion in yearly revenue by 2031 to justify the data-center buildout. Today's consumer and enterprise AI products might cover $1.2 trillion to $1.8 trillion of that, leaving roughly $4.2 trillion to come from products that mostly don't exist yet — Bain pencils in about $900 billion from physical AI like robots and digital twins alone — while annual infrastructure spending climbs toward $1.5 trillion. Local opposition has already blocked or delayed more than 75 projects worth $130 billion in early 2026 alone. For executives, this is the other half of the hardware story — the infrastructure bill is getting paid now, and the payback is a bet on applications still to come.

3. Samsung Puts $1 Billion Into an AI Data-Center Builder

Samsung Electronics and five affiliates are committing $1 billion to Helix Digital Infrastructure, the KKR-backed data-center company run by former AWS chief Adam Selipsky, bringing its pledged capital past $11 billion alongside Nvidia, Vistra, and the Kuwait Investment Authority. Samsung says the tie-up should boost sales of chips like high-bandwidth memory, and it plans to bring its cooling, battery, and construction businesses along too. When a memory maker buys a seat at its biggest customers' table, expect even more of its output to flow toward data centers.

4. A Patent Fight Takes Aim at the Memory Inside AI Servers

Memory-module maker Netlist filed a complaint with the U.S. International Trade Commission seeking to block imports of Micron memory chips it says infringe two of its high-bandwidth memory patents, along with Google, Nvidia, and Broadcom GPUs and servers that use them. Netlist won a $445 million verdict against Micron in 2024, filed a separate trade complaint in August, and settled its own memory patent fight with Samsung the same month after winning $421 million in verdicts. Micron declined to comment. Import bans are a long road, but with supply already this tight, legal risk now sits alongside price risk on anyone's AI hardware roadmap.

// 3 AI Tools
  1. NotebookLM · RESEARCH ASSISTANT — Google's notebook answers from the documents you load and cites them as it goes. The move most people miss: load the memory makers' last two earnings releases, TrendForce's quarterly price notes, and your own hardware refresh plan, then ask it when to buy and where your price risk sits. You get a procurement brief grounded in sources you can click through, not a chatbot's guess.

  2. Perplexity Comet · AGENTIC BROWSER — Perplexity's browser puts an assistant beside every tab. Before your next laptop purchase, open the same model at three retailers plus the manufacturer's spec sheet and have Comet build a side-by-side table that flags quiet downgrades in RAM or SSD size — exactly the cost-cutting move TrendForce says PC makers are using to hold prices.

  3. Helicone · LLM OBSERVABILITY — an open source gateway that logs every model request with its token count and cost. The novel use: before anyone pitches buying a GPU box, route a month of your team's AI usage through Helicone and export the totals. That one number tells you whether owning hardware or renting models actually pays — the question the rest of this week's editions are built around.

// The Extra Read

The best single explainer on why one gigabyte of HBM eats roughly three times the wafer capacity of DDR5, and why three companies controlling nearly all DRAM production turns a pricing problem into a geopolitical one. Read it for the sector-by-sector fallout the Big Story doesn't have room for, from PCs to cars.

So, back to that reader's spare desktop. In the '90s, the computer nobody wanted was the cheapest server you'd ever own, because the parts inside it were cheap and plentiful. This time the parts are the scarce thing. Before you send last year's laptop to the recycler, it might be worth a second look.

Mark R. Hinkle

Your AI Sherpa,

Mark R. Hinkle
Founding Publisher, The AIE Network
Follow me on LinkedIn


If you want to get in contact or give me feedback, reply to this email. I read every single one of them.