Biphoo News

collapse
Home / Daily News Analysis / Sam Altman says it’s better for RAM to be in a data center than on your desk

Sam Altman says it’s better for RAM to be in a data center than on your desk

Oct 31, 2026  Twila Rosenbaum  31 views
Sam Altman says it’s better for RAM to be in a data center than on your desk

Sam Altman has a blunt, if unpopular, answer to one of the most pressing questions in consumer technology right now: the world's supply of memory chips is too scarce to be sitting on desks. Asked about supply chain constraints and the global RAM shortage, the OpenAI chief executive said flatly that it is “terrible” for DRAM to end up in personal computers instead of data centers.

The remark came during a wide-ranging podcast interview in which Altman also discussed artificial intelligence's effect on mathematics. But it was his side commentary on memory allocation that drew attention, because it lands directly on the wallets of anyone shopping for a laptop, a phone, or a desktop machine in the middle of a historic price run-up.

The argument, in his own words

Altman framed the issue as a collision between intuition and arithmetic. Many people, he suggested, assume that the spread of open-source models means inference will migrate onto local machines, relieving pressure on cloud infrastructure and reducing the need for enormous data center buildouts.

“A lot of people want local inference. And somehow, they assume that the problem is solved now because we are going to have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data center.”

He then turned to what he described as the global allocation question, arguing that distributed memory is fundamentally wasteful.

“But from a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time.”

His prescription was straightforward: concentration beats distribution when a resource is scarce.

“You want maximum utilization in the cloud if you're trying to deal with this very supply constrained environment. And somehow everyone's — many people's intuition is the opposite of that.”

The economics behind “maximum utilization”

The logic Altman is describing is a familiar one in computing, even if it is rarely applied to consumer hardware so bluntly. A data center operator can time-share a pool of memory across thousands of customers and workloads, pushing utilization rates far above what any single desk-bound machine can achieve. A workstation that renders video, plays games, or runs a local model in bursts may leave most of its memory idle for the majority of the day.

Multiply that idle capacity by hundreds of millions of PCs and the aggregate waste becomes enormous in a market where every wafer of DRAM output is already spoken for. From the perspective of a company buying memory by the ton, a stick of RAM in a home desktop is not just underused — it is capacity that cannot be scheduled, pooled, or resold to another tenant at three in the morning.

That framing also explains why hyperscalers have been willing to sign multi-year, multi-billion-dollar supply agreements. Locking in capacity years ahead is rational when the alternative is bidding on a spot market that is already tightening, and when the compute plan for a single AI campus can consume memory at a scale that dwarfs a mid-sized country's consumer electronics demand.

How memory became tech's scarcest commodity

The shortage did not appear overnight. It is the product of several forces converging at once. The first is the explosive growth of AI infrastructure, which consumes memory in two distinct ways: as conventional server DRAM for the tens of thousands of machines that surround accelerator clusters, and as high-bandwidth memory stacked directly onto AI chips.

High-bandwidth memory is especially demanding on manufacturing capacity. Each stack requires more silicon area and more advanced packaging per bit of storage than commodity DDR5, which means the same fab footprint produces fewer usable bits when the mix tilts toward AI accelerators. Memory manufacturers have responded by shifting wafer starts toward the higher-margin parts, tightening supply of the standard DRAM that laptops, phones, and desktops rely on.

Adding capacity is not a quick fix. A new memory fab involves billions of dollars of capital expenditure and takes years to move from groundbreaking to volume production, and the equipment queue for advanced lithography is itself constrained. The result is a classic inelastic supply curve meeting a demand curve that has gone nearly vertical, with long-term contracts absorbing output before consumer OEMs ever get a quote.

Why the price of a laptop is climbing

The knock-on effects are already visible at retail. Memory is one of the largest bill-of-materials line items in a mid-range laptop or phone, and when contract prices for DRAM and storage move sharply, device makers have limited room to absorb the difference. Some have trimmed configurations, shipping base models with less memory than they would have offered a year earlier. Others have simply raised prices, sometimes more than once in a single quarter.

Analysts expect the pressure to persist rather than fade quickly. Suppliers have publicly guided toward extended tightness: one major memory maker has said it expects the shortage to last until at least 2028, while another has projected constraints running through 2030. Those timelines reflect the reality that demand from AI buildouts is competing not just with consumer electronics but with every other industry that needs memory — automotive, networking, industrial, and telecommunications included.

The counterargument for local inference

Altman's position is not universally shared, and the case for putting memory on desks has real technical substance. Running models locally avoids per-token charges, keeps sensitive data off remote servers, works without a network connection, and eliminates the latency of a round trip to a distant data center.

Hardware vendors have leaned into that pitch. Unified memory architectures let a single pool of high-bandwidth memory serve both the CPU and the GPU, which is precisely the configuration that small and mid-sized models need. Quantization and distillation techniques have steadily reduced the memory footprint required to run capable models, and developer-focused desktop machines with very large memory configurations have become a niche category of their own.

The catch is scale. A locally hosted model serves one user; the same silicon in a cloud cluster serves thousands. Altman's point is essentially that in a supply-constrained world, the marginal bit of DRAM produces far more useful work in a shared facility than under a desk — regardless of how appealing the local experience may be for an individual developer.

The memory market has been here before

Cyclicality is a defining feature of the memory industry. The DRAM market boomed in 2017 and 2018 as mobile and server demand outran supply, driving prices to multi-year highs before collapsing in 2019 as new capacity came online and inventories swelled. A similar squeeze during the 2021–2022 period, driven by pandemic-era demand for PCs and cloud capacity, was followed by one of the sharpest downturns the sector has recorded, with manufacturers posting steep losses.

What makes the current cycle different is the concentration of demand. Previous upturns were broad-based, driven by millions of ordinary purchases that could cool quickly when prices rose. This one is anchored by a comparatively small number of very large buyers with multi-year capital plans and the balance sheets to pre-pay for capacity. That structure tends to lengthen shortages, because the buyers absorbing supply are the least price-sensitive participants in the market.

The allocation question nobody voted on

There is a political dimension to all of this that is only beginning to surface. Communities hosting large data center campuses have started pushing back on water use, electricity demand, and land consumption, and those fights are increasingly shaping local and national politics. The memory crunch adds a new argument to that debate: the same buildout driving up server demand is also making consumer devices more expensive.

For now, the allocation question is being settled not by consumers or regulators but by procurement departments. Multi-year supply agreements signed by cloud operators and AI infrastructure projects have effectively reserved a large share of global output, and consumer device makers are left to compete for what remains. Until new capacity reaches volume production, the price of a stick of RAM on a store shelf will keep reflecting a bidding war that individual buyers never entered.


Source: The Verge News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy