> ## Content Index
> Fetch the complete content index at: https://espresso.cafecito.tech/llms.txt
> Use this file to discover other available public pages before exploring further.

# DRAM makers lock 2027-28 output as AI inference debate continues
- URL: https://espresso.cafecito.tech/dram-makers-lock-2027-28-output-local-ai-push/
- Published: 2026-11-02T13:11:06.000Z
- Updated: 2026-11-02T13:11:06.000Z
- Description: Samsung and SK Hynix report record Q3 2026 profits as shortages extend into 2028-2030. But the debate over local vs cloud AI inference depends on whether locked output holds through 2027-28.
- Author: Barista @ Cafecito
- Tags: HPC & Semiconductors, Technology, Data Centers, Semiconductors, High Performance Computing, #cortado-4208b759-81e1-5e2c-b99e-5f976128b429

Memory chipmakers are pre-empting future supply. Samsung reports Q3 2026 operating profit of 107.4 trillion won (\~$80B), a ninefold year-over-year jump, with DRAM margins above 80% and its 2027 HBM supply sold out. SK Hynix, in the same stretch, reports record Q3 2026 revenue of 104.62 trillion won (\~$78.1B). Both firms say shortages extend into 2028-2030.

The analytical finding is that the scarcity is being priced in by manufacturers even as OpenAI argues local inference wastes DRAM. Whether the two trends reinforce or cancel depends on how much of the locked output converts into 2027-28 fab throughput — a number not yet announced.

### Scarcity: numbers and who's buying

Samsung's Q3 2026 profit of 107.4 trillion won against 12.17 trillion a year earlier is the anchor. Long-term agreements reportedly cover 60-70% of its production capacity; HBM4 sales are cited as supporting AI data-center demand. Its Pyeongtaek P5 facility is slated online in 2028\. SK Hynix separately reports 14 key customers holding LTAs over 50% of capacity, with HBM4 scaling by end of 2026 and shortages projected to 2030.

DDR5 prices in Japan rose up to 761% over the past year; a 32GB DDR5-7600 kit bought for $110 in 2024 is now listed above $1,500, one report says. Micron exited its Crucial consumer memory business in December 2025, citing AI demand.

### Altman's argument and the allocation counter

Sam Altman told Stripe co-founder John Collison in a Cheeky Pint conversation around November 1 that 32GB local rigs sit at roughly 5% memory utilization, while cloud inference runs near continuous load. The same framing appears in an earlier October 30 piece: local AI inference wastes \~95% of DRAM. The estimate carries no supporting data, one account notes. OpenAI's own GPT-OSS 20B supports 16GB local devices, which softens the universal-waste claim.

OpenAI reportedly signed LOIs with Samsung and SK Hynix in October 2025 for up to 900,000 DRAM wafers per month — roughly 40% of global DRAM output per one coverage summary. One framing calls that "constraining memory supply"; the figure is attributed, not independently verified.

- Altman's claim: 32GB local rig uses \~5% of memory; cloud runs near continuous load
- Counter: GPT-OSS 20B supports 16GB local devices
- OpenAI LOI: up to 900,000 DRAM wafers/month with Samsung and SK Hynix, reported October 2025
- Samsung LTA coverage: 60-70% of production capacity

### Consumer knock-ons and the unresolved test

NVIDIA's RTX 5000-series GPUs are pricing up with release delays; PC and smartphone shipments are being cut. Sony's PS5 Pro sits at $1,399 after a March increase. Apple and Qualcomm are promoting on-device AI models, a competing architecture against centralized memory.

The deciding marker is whether the 900,000-wafer LOI volumes and the >50% (SK Hynix) / 60-70% (Samsung) LTA shares hold through 2027-2028\. If they do, local-AI and offline-access benefits keep losing ground to utilization economics; if allocations loosen, the debate returns to user choice rather than scarcity.

[![Beans](https://beans.cafecito.tech/favicon.ico)Explore more coverage on BeansSearch this story across publishers→](https://beans.cafecito.tech/search?q=DRAM%20makers%20lock%202027-28%20output%20as%20AI%20inference%20debate%20continues&utm%5Fsource=espresso.cafecito.tech&utm%5Fmedium=referral&utm%5Fcampaign=cafecito&utm%5Fcontent=dram-makers-lock-2027-28-output-local-ai-push)