RAM prices: Sam Altman, AI and when they will fall

Q: What did Sam Altman say about RAM?

On the Cheeky Pint podcast he argued that keeping the world's scarce DRAM in personal computers, used about 5% of the time, is an inefficient allocation, and that the cloud uses it better. After the criticism he clarified on X that the problem is not consumers but memory production that is too low.

Sam Altman appeared on a podcast and one part of the conversation went viral. The OpenAI CEO argues that running AI models on your own computer, so-called local inference, is a waste: the memory in a PC sits idle most of the time, while in a data center it is almost always busy. Said at a time when consumer RAM prices have more than tripled in a year, it was never going to go unnoticed.

We talked about it in an episode on our channel.

▶ Watch the video on YouTube (in Italian)

Below we lay out the facts: what Altman said, why RAM is so expensive, who makes it and when we can expect prices to stop rising.

In 30 seconds

What Altman said. On the Cheeky Pint podcast he called it "terrible" that part of the world's already scarce DRAM sits in people's computers, used about 5% of the time. The cloud, in his view, uses it far better. He later clarified that the problem is not consumers but memory production, which is too low.
Why RAM is so expensive. The chips that train and serve AI models consume huge amounts of memory. A 32 GB kit that cost 100-150 euros in early 2025 now easily reaches 500. Production is in the hands of three companies, Samsung, SK hynix and Micron, which together hold almost 90% of the DRAM market.
When prices will fall. Not before 2028. The first new fabs come online between 2027 and 2028, and even then a halt to increases followed by a gradual decline is more likely than a price crash.

What Sam Altman said about local inference

Inference is when an already trained AI model answers a request. It can run in the cloud, on the servers of OpenAI, Anthropic or Google, or locally, on your own computer, with a downloaded open source model.

Talking with Stripe's John Collison, Altman compared the two from a resource point of view:

locally, memory is poorly allocated: a model on your laptop works for a small fraction of the day, and the rest of the time that RAM sits idle;
in the cloud, the same memory serves thousands of users in turn, and is almost always in use.

The reasoning itself is not wrong. But it applies to any computing resource: your computer's processor also spends most of its time waiting. So why is everyone talking about it now?

Why Altman's words made everyone angry

Because for more than a year memory has been a scarce good. The same companies that make RAM for PCs and smartphones also make the high bandwidth memory (HBM) that goes into AI accelerators, and AI companies, backed by huge investments, can pay far more for it than any consumer. Production capacity has moved to where margins are highest, and prices have gone up for everyone.

You can see it on the shelf: a 32 GB RAM kit that cost 100-150 euros in early 2025 now easily reaches 500. The criticism of Altman ran along two lines:

ownership. Many people want to run models locally precisely so they don't depend on a company: for privacy, to avoid usage limits set by someone else, to avoid a subscription. Being told it is wasteful, by the person selling the alternative, came across as a conflict of interest;
prices. AI companies are accused of driving up, through their demand, the cost of a component everyone needs.

Altman replied on X that the clip lacked context: the problem, he wrote, is not consumers but a memory supply far lower than needed. On the podcast itself he had said he hoped for an abundance of cheap compute for anyone. Looking at the numbers, he has a point on the diagnosis.

Who makes RAM: a market in the hands of three companies

Global DRAM production is concentrated in three companies: South Korea's Samsung and SK hynix and the US-based Micron. According to TrendForce, in the second quarter of 2026 they split the market like this:

Manufacturer	Country	DRAM revenue Q2 2026	Share
Samsung	South Korea	61.0 billion dollars	39.4%
SK hynix	South Korea	38.6 billion dollars	24.9%
Micron	United States	36.0 billion dollars	23.3%

Together they hold almost 90% of the market. In the same quarter revenue for the whole DRAM industry grew 59.5% over the previous three months, driven almost entirely by higher prices: shipments grew only slightly and supplier inventories are at historic lows. SK hynix closed the third quarter with an 80% operating margin.

The companies do not publish official production volumes, but the picture is clear: supply has grown a little, demand has gone vertical.

Why they don't increase production

Because it cannot be done quickly, even if they wanted to. Memory fabs are among the most advanced industrial plants in the world, and the existing ones are already running at full capacity. Producing more requires new fabs, and that takes a long time:

building a fab takes years: Micron started building its first new fab in Idaho in October 2023, and its first DRAM wafer is expected in mid-2027;
even once a fab starts, output ramps up gradually and only becomes significant in volume after a few quarters;
realistically, tripling a manufacturer's capacity is a 5-10 year project.

In the short term, meeting current demand is simply impossible.

Why they didn't build them earlier

AI is not new this year, and it is fair to ask why manufacturers didn't move sooner. There are two answers.

The critics' version: prices kept high on purpose. Some argue that manufacturers kept supply low to push prices up. At the moment there is no evidence of this. Those who make the claim, however, point to a real precedent: between 2001 and 2002 Samsung, Hynix (now SK hynix), Infineon and Elpida secretly agreed on the prices of DRAM sold to Dell, HP, Apple, IBM and others. The US Department of Justice fined them more than 730 million dollars in total, and 18 executives were charged, many receiving prison sentences.

The official version: the memory market is cyclical. This is the analysts' explanation, and the more solid one. Memory lives in cycles: years of shortage and high prices, followed by years of oversupply and collapsing prices. A fab costs tens of billions and has very high fixed costs: if demand falls just as it comes online, margins collapse. In the early years of AI nobody could tell whether demand would stay this high for long, and investing to chase a spike could have been catastrophic. As late as November 2025 TrendForce described the industry's 2026 capital spending as "cautious": +14% for DRAM, with limited impact on supply growth.

What is being done and when RAM prices will fall

In the short term, not much. In 2026 and 2027 capacity gains will come mainly from more efficient processes in existing fabs and some reclaimed cleanroom space, not from new plants. In its third quarter 2026 results, Samsung expects AI-driven demand in 2027 to far exceed industry supply, even more than this year.

From 2027-2028 things start to move. Between 2025 and 2026 the three manufacturers started building several new fabs:

Micron: its first Idaho fab will produce its first wafers in mid-2027, the second by late 2028, the first New York fab in 2030;
SK hynix: the first phase of its Yongin cluster in Korea reaches full operation in 2027;
Samsung: is speeding up construction of the new P5 lines in Pyeongtaek, about six months ahead of schedule.

The choice, today, is rational: even in the worst case AI will keep being used in some form, so a certain level of investment is justified. It remains a cautious investment relative to demand, even if it is huge in absolute terms.

What to expect. Prices will stay high, or rise further, at least through 2027. From 2028, with the first new fabs up to speed, it is reasonable to expect increases to stop and then a gradual decline. A quick return to early 2025 prices, on the other hand, is unlikely.

What it means for software teams

For companies this is not just news for hardware enthusiasts. Servers, workstations and laptops cost more, and the comparison between cloud and owned infrastructure needs to be redone at today's prices. At the same time, choosing an AI model is not only about cost: where the data lives, who can change the terms of use and how much you depend on a single vendor are questions that weigh as much as the bill.

Our advice is the same as always: start from the problem you need to solve, not from the technology. We wrote about it in AI in business: where to start. And if the new PC can wait until 2028, it might really be worth waiting.

Frequently asked questions

Why is RAM so expensive in 2026?

Because demand for memory for AI has exploded and supply cannot keep up. Manufacturers have shifted capacity towards more profitable data center memory, and existing fabs are already running at full capacity. A 32 GB kit that cost 100-150 euros in early 2025 now easily reaches 500.

What did Sam Altman say about RAM?

On the Cheeky Pint podcast he argued that keeping the world's scarce DRAM in personal computers, used about 5% of the time, is an inefficient allocation, and that the cloud uses it better. After the criticism he clarified on X that the problem is not consumers but memory production that is too low.

Who makes RAM?

Almost all the world's DRAM is made by three companies: Samsung and SK hynix, from South Korea, and Micron, from the United States. In the second quarter of 2026 they held 39.4%, 24.9% and 23.3% of the market respectively, almost 90% in total.

When will RAM prices go down?

Not before 2028. The first new fabs start producing between 2027 and 2028, and volumes grow gradually. For 2027 Samsung expects demand to far exceed supply. From 2028 a halt to increases and a gradual decline are likely, not a crash.

Are RAM makers keeping prices high on purpose?

There is no evidence they are doing so today. There is a precedent, though: between 2001 and 2002 Samsung, Hynix, Infineon and Elpida fixed DRAM prices and were fined more than 730 million dollars in the United States. The most widely accepted explanation for the current shortage is the cyclical nature of the market, which makes investing in new fabs risky.

Why aren't new memory fabs being built right away?

Because it takes years: Micron's first new fab in Idaho, started in October 2023, will produce its first wafers in mid-2027. And a fab costs tens of billions: if demand fell just as it was ready, the manufacturer's margins would collapse.

Should I buy RAM now or wait?

If you can wait, forecasts point to high prices at least through 2027 and a gradual decline from 2028. If you need it now, buy only what you need and consider second-hand parts or a configuration you can expand later.