What did Sam Altman say about laptop RAM?
On the Cheeky Pint podcast, OpenAI chief executive Sam Altman was asked whether open-weight models would let people run AI at home. His answer turned the question into an argument about global memory allocation.
According to The Verge's transcript of the episode, Altman said: “it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer.” He described that memory as “being used five percent of the time.” He added: “You want maximum utilization in the cloud if you're trying to deal with this very supply-constrained environment.”
The Verge published the exchange on 30 October, framing it around OpenAI's own buying spree. Tom's Hardware reported the same remarks and placed them about 29 minutes into the episode.
The full interview is titled “Why Sam Altman is against a single superintelligence.” Cheeky Pint is the podcast Altman recorded with Stripe co-founder John Collison. Nothing in the quoted passages is a product announcement. It is a chief executive explaining why he wants more of the world's memory.
Read narrowly, the comment is not about whether a laptop needs RAM. Browsers, spreadsheets and video calls all need memory at the same time. The claim is about where the next unit of DRAM capacity should go.
Read broadly, it is a statement of priorities from a company that has reserved a large share of future memory output. That has direct consequences for anyone shopping for a laptop.
Why is RAM so expensive in 2026?
The price of memory has been climbing all year. TrendForce projected in late September that conventional DRAM contract prices would rise 10-15% in the fourth quarter, with NAND flash up 15-20%. NAND flash is what laptop SSDs use for storage.
A TrendForce summary published on 20 October, citing Commercial Times, put the expected fourth-quarter DRAM increase at about 20%. It also noted that prices had risen roughly 40% in the third quarter and about 80% in the second.
The suppliers are reporting record results. Samsung's third-quarter revenue reached KRW 195.3 trillion, up 14% from the previous quarter, and operating profit hit KRW 107.4 trillion. Its memory business set records for both revenue and profit.
Samsung said memory demand in 2027 will “far exceed the industry supply.” The Verge reported that memory executive Jaejune Kim told analysts tight conditions would continue into 2028. It also reported that Samsung's 2027 HBM output is already sold out.
SK hynix painted a longer picture. Its third-quarter revenue passed KRW 100 trillion for the first time at KRW 104.6 trillion, with an operating margin of 80%. The company said it expects demand to exceed supply through 2030.
SK hynix has signed long-term agreements covering more than half of its 2027 and 2028 capacity with 14 customers, including cloud providers. Those contracts typically run five years.
OpenAI is one reason. In October 2025 it announced partnerships with Samsung and SK hynix under its Stargate plan, targeting 900,000 DRAM wafer starts per month. That capacity is committed years before the laptops competing for it are designed.
Memory makers are also shifting their product mix toward AI. SK hynix expects HBM4 to become its mainstream product in the fourth quarter, overtaking HBM3E in shipments and revenue. High-bandwidth memory is sold to data centers, not laptop makers.
Supply cannot adjust quickly. SK hynix said leading-edge processes are technically difficult and new production facilities take a long time to build. It expects demand to exceed supply through 2030 for that reason.
The companies are investing, but not in a way that helps a $500 laptop soon. Samsung spent KRW 29 trillion on capital expenditure in the third quarter. It expects about KRW 84.4 trillion for the full year, mostly in its chip division.
SK hynix plans to ramp its M15X fab for HBM4 and bring its Yongin Phase 1 site into full operation next year. Both efforts are aimed primarily at AI memory rather than the laptop market.
China's CXMT is working on a denser 4F² DRAM design that could cut cell area by about a third. TrendForce reports that full commercialisation is generally expected after 2028, and that yields and data retention remain unresolved.
The result is a market where the buyers with the longest contracts and the deepest pockets set the price. Laptop makers buy conventional DRAM and NAND in a market where AI server demand is the priority.
What does the memory crunch mean for laptop buyers?
Two research firms now agree that prices are cutting demand. IDC expects global PC shipments to fall 11.3% in 2026, with a 20% year-on-year drop in the fourth quarter. It sees no meaningful relief before the end of 2027.
IDC vice president Jean Philippe Bouchard said: “We're not seeing any relief to the memory shortage situation before the end of 2027.” The firm also forecasts average selling prices up 17% this year.
Its research manager Jitesh Ubrani said: “the overall trajectory for average selling prices is firmly upward.” He added that pricing is unlikely to return to 2025 levels even as memory capacity expands.
Counterpoint Research measured a 21% year-on-year fall in third-quarter PC shipments. It expects shipments to plunge 27% in the fourth quarter and stay broadly flat in 2027.
Counterpoint senior analyst Minsoo Kang said vendors are choosing to “prioritize configurations that can better absorb component inflation.” In practice, that means fewer cheap models and higher default specifications. Some vendors have revived DDR4 memory and 4GB graphics cards to hold prices down.
Apple offers a visible example. The MacBook Neo launched in March with an education price of $499. It rose to $599 in June and now starts at $649 on the Education Store, $150 above launch. The entry model still has 8GB of RAM and 256GB of storage.
The chip side is feeling it too. Intel reported third-quarter revenue of $16.5 billion, up 21% year on year, but guided the fourth quarter to $15-16 billion. Analysts had expected $17.1 billion.
Chief financial officer David Zinsner said: “We see the PC slowdown continuing at least through the first half of next year.” Intel shares fell 29% on Friday, their worst day since the 1970s, according to CNBC.
The pull-forward effect makes the numbers look worse before they look better. IDC says first-quarter 2026 shipments grew 3% because buyers accelerated purchases ahead of expected price rises. That borrowed demand is now gone.
Commercial buyers face a different trade-off. Large organisations often refresh on fixed schedules and sign multi-year leases, so they have less room to wait for prices to fall. Consumers can simply delay.
For buyers, the practical effect is narrower choice. The cheapest configurations disappear first, because a $500 laptop has less room to absorb a memory price increase than a $2,000 one.
Students and small businesses are the most exposed. They tend to buy at the low end and keep machines for years. A higher purchase price persists for the life of the device.
Can a laptop still run AI locally?
Yes, but the models are smaller than the ones Altman sells. OpenAI's own open-weight release shows the split. gpt-oss-20b can run on devices with 16GB of memory. gpt-oss-120b needs a single 80GB GPU.
Microsoft has built a specification around that middle ground. A Copilot+ PC needs an NPU rated at 40 TOPS or more, 16GB of RAM and a 256GB SSD. Those machines run features such as Live Captions and Recall on the device.
Microsoft has also pushed local models beyond that tier. Its Phi Silica model now runs on standard Windows PCs without a Copilot+ NPU, using the GPU instead. Microsoft says prompts and responses stay on the device.
Qualcomm chief executive Cristiano Amon has argued the opposite case to Altman. “Whoever has presence on the edge is going to win. The edge is where the humans are,” he told TIME in January. He said those devices are not useful without models running locally, with the latency and privacy that implies.
The gap is capability, not possibility. A 20-billion-parameter model can draft text, summarise and answer questions. OpenAI positions gpt-oss-20b against o3-mini, not against its frontier systems.
Memory capacity also decides which features a laptop can offer. Recall, Live Captions and local image generation need RAM reserved for models. That comes on top of the memory the operating system and applications already use.
That is one reason 8GB machines are becoming a harder sell. The MacBook Neo's entry model still ships with 8GB, and it is the configuration that has absorbed the steepest price increases.
So Altman's point survives in a narrow form. If the goal is to serve the largest models to the most people, cloud utilization matters. If the goal is privacy, offline use or low latency, local memory is doing real work.
Who gains from sending RAM to the data center?
Memory makers gain most directly. Samsung's device solutions division posted KRW 147.1 trillion in revenue and KRW 106.5 trillion in operating profit in the third quarter. SK hynix plans about KRW 100 trillion in shareholder returns this year.
OpenAI gains if cloud inference stays the default. Its business is built on models that run in its own data centers and are sold by subscription or token. Tom's Hardware noted that Altman was, in effect, defending that model.
PC makers absorb the squeeze. They buy memory on contracts, pass costs to buyers, and lose volume when prices rise. IDC's shipment decline and Counterpoint's fourth-quarter forecast describe that trade-off.
There is also a conflict of interest worth naming. The person arguing that memory should sit in data centers runs a company that rents data-center capacity. That does not make the argument wrong. It does mean the claim should be tested against evidence.
Where does Altman's argument break down?
The first weakness is the number. Neither published account attributes the 5% utilization figure to a study or a measurement. It is an estimate offered in conversation, not a published benchmark.
The second is fungibility. Memory soldered into a laptop cannot be shipped to a data center later. The choice was made when the device was built, often years before the shortage became a headline.
The third is that utilization is not the only measure of value. A fire extinguisher is idle almost all the time. Local memory provides privacy, offline access and predictable latency, which cloud capacity cannot match for every task.
The fourth is demand destruction. PC shipments fell 21% in the third quarter, according to Counterpoint. Buyers who delay or downgrade reduce demand for the memory that goes into laptops.
The fifth is that the cloud is not free of waste. Altman's comparison assumes high data-center utilization, but the published transcripts offer no figure for it. The claim is an assertion about a private fleet.
There is also a product-mix question behind the allocation debate. Memory makers are prioritising HBM and server DDR5 over the LPDDR5 shipped in thin laptops. Capacity moves toward those products even without a policy decision.
None of this means the shortage is fake. It means the allocation argument is a policy choice presented as arithmetic. The question is who gets to decide, and who bears the cost of the answer.
What should laptop buyers watch next?
Watch the fourth-quarter contract settlements. TrendForce's 10-15% projection will be replaced by actual prices, and those set the cost of laptops shipping in early 2027.
Watch the low end of the market. The MacBook Neo's education price has risen twice this year, from $499 to $649. Counterpoint says vendors are prioritising configurations that can absorb higher component costs.
Watch long-term contracts. SK hynix has committed more than half of its 2027-28 capacity to 14 customers. TrendForce notes that price ceilings under such agreements are starting to take effect, which may slow increases without reversing them.
Watch product decisions. The Verge reports rumours of another MacBook Neo with more RAM early next year. If Apple raises the memory floor, other vendors may follow.
Watch what other AI executives say. If Google, Microsoft or Anthropic make the same allocation argument, it becomes an industry position rather than one chief executive's aside.
For now, the safest reading is the boring one. Memory is scarce, prices are rising, and the cheapest laptops are bearing the brunt. Altman's comment explains the direction of travel, not the end point.