Skip to content
Primores AI wiki

AI Compute Costs: Why Memory and GPU Prices Rose in 2026

AI Compute Costs

TL;DR: AI data centers buy the same memory chips and factory capacity that PCs, phones and ordinary servers need, and in 2026 they outbid everyone else. TrendForce’s series for conventional DRAM contract prices runs +45–50% in the last quarter of 2025, then forecasts of +90–95%, +58–63%, +13–18% and +10–15% for the four quarters of 2026. SK hynix says long-term agreements already cover more than half of its 2027–28 production and expects demand to exceed supply through 2030. GPU rental turned too: prices fell in 2025 and rose in 2026. The first sign of strain on the demand side came on 11 November 2026, when TrendForce reported notebook shipments down 15.8% in a quarter and some PC makers adjusting their memory orders. For a business this explains dearer laptops, servers and GPU time. It has not shown up in AI token prices, which fell through 2026.

⚠️ Read 2026-11-11. TrendForce’s figures after 4Q25 are forecasts made at the start of each quarter, not measured outcomes. Statements about supply come from the companies that sell memory. Nothing we read ties GPU rental prices or AI API prices to memory costs. Items marked “carried” come from our 9 November research and were not re-read for this page.

The mechanism

Three companies make most of the world’s DRAM: Samsung, SK hynix and Micron. Each wafer they process becomes one kind of memory or another, and three things pulled wafers toward AI.

AI memory uses more wafer. HBM (high-bandwidth memory) is the stacked, fast memory that sits beside the processor in AI accelerators. Micron has said HBM3E takes about three times the wafer supply of DDR5 for the same amount of memory. That figure reaches us through press reports of Micron’s statements, not a Micron document (carried).

Server memory pays better, so it goes first. TrendForce in March 2026: “suppliers continue prioritizing server DRAM due to its superior profitability” and “have simultaneously reduced shipments to PC OEMs and module makers.” In September: “DRAM suppliers continue to prioritize advanced-process capacity for high-performance server products”.

The biggest buyers reserved their share years ahead. SK hynix, in its results for the third quarter of 2026: “Long-term agreements secured for more than 50% of 2027~2028 production capacity.” Everyone without such an agreement, which is most businesses and every consumer, buys from what is left.

Our read of what follows: the price spike lands hardest on buyers with no contract. TrendForce noted in July that server price gains would “moderate as a portion of procurement is governed by long-term supply agreements”, and in September that increases for some suppliers would “lag the market average” because of the pricing terms in those agreements.

What prices did

TrendForce’s conventional DRAM contract prices, quarter on quarter:

QuarterChangeReported or forecastTrendForce release
4Q25+45–50%Reported26 Feb 2026
1Q26+90–95% (PC DRAM “over 100%”, server DRAM “around 90%“)Forecast, raised from 55–60%2 Feb 2026
2Q26+58–63%Forecast31 Mar 2026
3Q26+13–18%Forecast3 Jul 2026
4Q26+10–15% (PC DRAM +13–18%)Forecast30 Sep and 11 Nov 2026

Two readings of that table are both true. The rises are slowing sharply: by July TrendForce said consumer buyers were “reaching their affordability limit”. And prices are still going up every quarter, from a much higher base. On 6 November TrendForce said it expects DRAM price increases “to gain further momentum in 1Q27”.

The makers’ results show where the money went. SK hynix reported an operating margin of 80% on revenue of 104.6 trillion won for the quarter, its first above 100 trillion.

GPU rental turned at the same time

A price index kept by Uvik, a software company, from providers’ public price lists:

  • “GPU rental prices fell in 2025 and rose in 2026.” AWS cut its on-demand H100 price by 44% in June 2025. Then a SemiAnalysis index of one-year H100 contracts rose from $1.70 to $2.35 per GPU-hour between October 2025 and March 2026, up 38%, and Nebius raised its on-demand H100 price from $3.85 to $4.50 on 1 October 2026.
  • In October 2026 an H100 rented for a median $4.20 per GPU-hour on demand across 16 providers, with list prices from $3.20 to $12.29.
  • The big clouds charge 1.7 to 2.7 times what GPU-specialist clouds do for the same chip.
  • Three-year commitments at AWS, Google Cloud and Azure cut H100 prices by 56% to 57%.

Prices we read directly on 11 November 2026: Nebius lists on-demand H100 at $4.50, H200 at $5.40, B200 at $8.50 and B300 at $9.50 per GPU-hour. AWS’s reservable Capacity Blocks list an eight-H100 instance at $47.757 an hour ($5.970 per GPU) and an eight-B300 instance at $129.168 ($16.146 per GPU). AWS says those prices “are updated regularly based on trends in supply and demand” and are next due to change in January 2027.

Neither AWS nor Nebius gives memory as a reason for its prices. The two markets tightened in the same year; a causal link between them is not in anything we read.

Who pays for it

  • Laptops. TrendForce: for a mainstream $900 notebook, the CPU, DRAM and SSD made up “approximately 68% of the bill of materials” in the third quarter of 2026. Notebook shipments fell 15.8% that quarter, and it projects a low-single-digit decline for 2027. PC makers are planning more private-label models with “lower-specification memory chips and CPUs with fewer cores”.
  • Phones. Brands are turning to lower-cost reballed memory, meaning reclaimed chips prepared for reuse, in budget models. TrendForce puts it at 3–4% of smartphones in 2026 and 7–9% in 2027, with a warning about reliability and warranty costs.
  • Small computers. Raspberry Pi raised prices in February 2026 because “the cost of some parts has more than doubled over the last quarter”, blaming “competition for memory fab capacity from the AI infrastructure roll-out.”
  • Anyone renting GPUs or hosting their own model. Higher rental prices move the break-even for self-hosting. See glossary/open-weight-models.

What it hasn’t done: raise token prices

No vendor we read links its API or subscription prices to memory or GPU costs, and list prices for AI tokens went down in 2026, not up. Anthropic’s current models list below the ones they replaced (Opus 5.5 at $4 / $20 per million tokens against $5 / $25 for Opus 5; Fable 5.2 at $8 / $40 against $10 / $50 for Fable 5.1). Menlo Ventures’ 2026 report estimates the blended price of a closed-model token fell 24% in a year and the average open-source token fell 3.3 times.

We can’t tell from public sources how long that holds. Two things to watch on tools/ai-plans-and-pricing: introductory API prices that double on a set date, and usage limits on seat plans, which vendors can tighten without changing a price.

When does it ease?

The makers say not soon. SK hynix “expects the structure in which demand exceeds supply to persist through 2030.” Samsung’s third-quarter deck expects 2027 shortages to tighten further compared with 2026 (carried). Makers are the sellers, and none of them names a peak.

The case that it turns:

  • Supply is coming. Analyst Paul Kedrosky (17 August 2026, opinion): Samsung and SK hynix “spent KRW 43.2T on semiconductor facilities in the first half of 2026, up 35 percent”, and “the capital already committed makes a severe 2028–29 downcycle increasingly plausible.” His point about timing: “Memory prices are set by marginal availability, so even a shrinking supply deficit can sharply reduce scarcity premiums well before the market reaches outright oversupply.”
  • Demand is bending. TrendForce on 11 November 2026: steep price rises “prompted some PC brands to place orders beyond their actual requirements”, finished-goods inventories are rising, and “some manufacturers are planning to adjust memory procurement for 4Q26—and even 1Q27”. Six weeks earlier it had said PC makers “continue to procure aggressively”.
  • It has crashed before. Memory prices fell hard in 2019 and 2023 after inventories built up (carried, from trade-press reports of TrendForce forecasts at the time).

Our read: the direction has not turned, the first inventory signal has appeared, and nobody credible is offering a date.

What a business can do

  • Check the memory in any 2027 laptop quote. TrendForce reports that some PC makers plan lower-specification memory to hold prices. A model at the same price may ship with less RAM than this year’s.
  • Buy what you need, and don’t stockpile. Prices are still rising, which argues against waiting. The inventory build-up and the supply response argue against buying years ahead.
  • Treat a GPU commitment as a bet on price. A three-year commitment cuts the hourly rate by more than half and locks it in. That is a good trade if prices keep rising and a bad one if 2028 looks like Kedrosky expects. AWS charges a Capacity Block “at the prevailing rate at the time of purchase”, so the date you reserve matters more than the date you use it.
  • For AI work, prefer what you can switch off. Hosted APIs passed none of the 2026 hardware rises through to token prices. Owned or committed hardware carries the price risk yourself. Data-control requirements can outweigh that; cost alone usually doesn’t. This is our inference from the sections above.

FAQ

Why did RAM get so expensive in 2026?

Memory makers moved factory capacity to the memory AI servers need, which uses more wafer and pays better, and the largest buyers reserved future production with multi-year contracts. Ordinary memory for PCs and phones got what was left.

When will RAM prices go down?

Nobody credible gives a date. The makers expect shortage through 2027 and beyond. The counter-signs are a large wave of new factory spending and, as of November 2026, PC makers starting to trim memory orders as their unsold stock builds.

Are GPU rental prices rising?

They rose in 2026 after falling in 2025. A one-year H100 contract index rose 38% between October 2025 and March 2026, and Nebius raised on-demand H100 prices 17% in October 2026.

Will AI API prices go up because of this?

It hasn’t happened in public price lists: token prices fell through 2026. No vendor we read ties its prices to hardware costs, so we can’t say how long that lasts.

Should we buy computers now or wait?

Buy what you need for the next refresh cycle, check the memory specification in the quote, and don’t buy years ahead.

Do not cite

ClaimWhy not
”SK hynix is sold out through 2029”Not in SK hynix’s release, which says agreements cover more than 50% of 2027–28 capacity; the 2029 line comes from Korean press reports
”DRAM prices rose five- or six-fold in 2026”That compounds quarterly forecasts as if they were outcomes
”PC makers are cutting 2027 memory orders” as a general factTrendForce says some are adjusting plans (11 Nov 2026), six weeks after saying they were buying aggressively
”GPU rental got dearer because of RAM”Neither AWS nor Nebius says so
A specific retail price for a RAM kitThe figures in circulation come from press headlines and snippets we could not read in full
”Demand exceeds supply through 2030” as a forecastIt is a seller’s expectation

Key Takeaways

  • AI demand raised memory prices for everyone by taking wafers and by reserving future supply under long-term contracts.
  • TrendForce’s contract-price rises slowed from +90–95% to +10–15% a quarter during 2026. Prices are still rising.
  • GPU rental prices rose in 2026 after falling in 2025; committing cuts the rate and locks it in.
  • The first demand-side strain showed on 11 November 2026: notebook shipments down 15.8% in a quarter and some PC makers trimming memory orders.
  • AI token prices fell over the same period. Hardware cost has not reached the API price list.
  • Buy what you need, check the specification, and be careful with multi-year commitments.

The article built on this research: Why is RAM so expensive right now, and will it go down?

Sources

By Andrej Ruckij