RU

SK Hynix expands HBM4 output: betting on Cheongju P&T6 and 20K wafers a month

Published: 2026-11-15 · Author: AI Release · @ai_release1
SK Hynix expands HBM4 output: betting on Cheongju P&T6 and 20K wafers a month

⚡ The gist

🔍 What happened

On November 15, 2026, etnews reported, citing several equipment makers, that SK Hynix is negotiating with partners to purchase HBM4-related equipment. We're talking back-end here: bonding (die-to-die stacking), dicing, and inspection. Delivery and installation are slated for the first half of 2027, with some of the lines requested as early as Q1.

The bet is on Cheongju P&T6 — the same fab formerly known as M8, which was picked for HBM production back in 2025. The line there has been under construction since early 2026, and now the company plans to squeeze every last bit of free space out of the site. One of the outlet's sources estimates the gain at full utilization at "more than 20,000 wafers per month," counted in silicon wafers.

Why is an expansion even being discussed amid all the talk about AI market instability? The piece name-checks the "Jensen Huang scandal" and Nvidia's CEO change as sources of anxiety. Then come two counterarguments. First: at its Q3 earnings call, SK Hynix said its DRAM and HBM supply is essentially booked out through 2029. Second: Nvidia has lowered its HBM4 bandwidth targets for the Vera Rubin chip, which means the bar memory makers need to clear is now lower. SK Hynix itself is keeping its comments terse: "equipment orders are proceeding according to our plan, and we will respond to market growth without disruption."

🧠 Our take

Our read: this story isn't really about HBM4 per se — it's about the bottleneck moving from lithography to packaging. The orders are for bonding, dicing, and inspection — the steps where you can add capacity faster and cheaper than building a new cleanroom. The logic is simple: with demand booked out for years, it makes more sense to buy back-end gear than to wait for a new fab.

What this changes for the market. The lowered HBM4 bandwidth target for Vera Rubin is, in effect, permission to yield more good modules from the same line. If Nvidia really did lower the bar, SK Hynix gets extra yield headroom and can lock in capacity under the same contract. From experience, spec relaxations like this almost always trigger a surge in equipment orders — exactly what etnews is describing.

Who wins. The first tier is Korean and other back-end equipment suppliers — the outlet explicitly talks about a "rain effect" across the industry. The second is AMD and Broadcom: their HBM demand remains a hedge even if Nvidia runs into leadership trouble. The losers are those who were waiting for a pause in memory purchases to buy in cheaper. Looks like that pause isn't coming.

⚠️ What we don't know

The story's biggest weakness is its sourcing. All the numbers and timelines come from "several equipment makers" and one unnamed source in SK Hynix's supply chain. The company hasn't confirmed the investment amount, the exact number of tools being purchased, or even the final site selection. The framing is also worth a raised eyebrow: the "Jensen Huang scandal" and Nvidia's CEO change get a single line with no details, even though the whole narrative hinges on them. The things to watch: SK Hynix's upcoming quarterly reports and Nvidia's official press releases on Vera Rubin.

💬 Worth debating

  • Is the lowered HBM4 bandwidth target for Vera Rubin a forced concession from Nvidia, or a sober calculation that will make memory cheaper for everyone?
  • Expanding back-end capacity at the old Cheongju fab versus waiting for new sites: which offers more resilience if AI demand collapses?
  • Should you trust industry rumors about equipment purchases before an official capex plan drops?

Forecast: by the end of the first half of 2027, SK Hynix will either announce concrete parameters for the P&T6 expansion or start up the line without publishing any figures — if no official data appears by the end of Q2, it means some of the equipment orders got pushed back.

Disclaimer: not officially confirmed — the data comes from equipment suppliers, and the company has only issued a general comment.

Source: etnews

🤖 AI summary
  • Date: 2026-11-15
  • Source: www.etnews.com
  • Tags: SK, HBM4, Nvidia, полупроводники, AI, SKHynix, chips, neural networks
  • Key facts:
    • SK Hynix is in talks with suppliers to buy equipment for HBM4 : die bonding, dicing, and inspection.
    • The added capacity would go to Cheongju P&T6 (the former M8 fab), with the ramp targeted for the first half of 2027.
    • At full utilization, the fab would add more than 20,000 wafers per month; the company hasn't disclosed any investment figures.
    • On November 15, 2026, etnews reported , citing several equipment makers, that SK Hynix is negotiating with partners to purchase HBM4-related equipment.
    • We're talking back-end here: bonding (die-to-die stacking), dicing, and inspection.
    • Delivery and installation are slated for the first half of 2027, with some of the lines requested as early as Q1.
    • The bet is on Cheongju P&T6 — the same fab formerly known as M8, which was picked for HBM production back in 2025.
    • The line there has been under construction since early 2026, and now the company plans to squeeze every last bit of free space out of the site.
#SK#HBM4#Nvidia#полупроводники#AI#SKHynix#чипы#нейросети
← PreviousAOSP Android 17 built on a MacBook: userdebug, root, and ART tweaksNext →Tiro by The Playto: $3M ARR and a wiki built from meetings instead of transcripts

Source: www.etnews.com · post in Telegram