| |
AI ARMS RACE |
THE CLOSE |
Knox Bennett, Lead Strategist, Strategic Compute · September 22, 2026 |
|
|
|
|
TODAY'S BEAT - MEMORY / VOLUME |
Samsung's HBM4 yield print turns the race into tons |
Around 80% yield, per Seoul Economic Daily, shifts Samsung from catch-up to volume. Allocation versus SK hynix and Micron is the fight. Samsung's 4nm base-die pull can crowd external foundry customers. |
Seoul Economic Daily, citing industry sources on Sept 20, says Samsung's HBM4 yield sat below 60% in early mass production and has climbed to around 80%. That is the golden-yield claim. HBM stacks many DRAM dies. One bad layer kills the stack. |
The next number is capacity. Average monthly wafer input is projected to rise from about 180,000 this year to about 250,000 next year, nearly 40%. HBM4 uses 1c DRAM and a 4nm base die coordinated inside Samsung. |
Yield was the catch-up story. Volume is the allocation story. SK hynix still holds the majority of Nvidia HBM. Micron is in the same line. If Samsung can ship tons, someone else's allocation gets tighter. |
|
|
|
|
What moved it: Tuesday sessionless seed held. Credit holds at 18. No new FedWatch primary vs Friday close / Monday seed. Valuation and Froth still estimated. Headline 72, flat. V13 C18 Cr18 F12 P11. |
|
|
| |
Why did Trump just ban America's most powerful AI? |
Banyan argues the ban is not what the headlines say - and names the San Francisco lab at the center of the story. |
|
|
Paid advertisement from Banyan. Not investment advice. Investing involves risk including loss of principal. |
|
|
|
|
This morning was Nvidia putting a vendor list on factory power. Tonight is memory. Who gets the physical tons. |
Do not file an 80% print as a shipping schedule. Seoul Economic Daily says third-quarter HBM4 revenue is projected to more than triple, and HBM4 is expected to take more than 60% of Samsung HBM revenue in the second half. I am repeating a Korean business daily. I have not audited a Samsung IR deck. |
Counterpoint, as cited there, has Samsung HBM share by revenue at 21% in the first quarter and 33% in the second. SK hynix moved from 58% to 50%. The gap narrowed. Majority is still SK. Nvidia Vera Rubin-class demand is why the tons matter. |
The foundry side is the sleeper. HBM4's base die is 4nm. Samsung can pull that capacity for its own stacks. External foundry customers on the same node can wait. |
Samsung memory benefits if 80% holds at volume. AI buyers who can take Samsung HBM4 leave a single-vendor line. Exposed: SK hynix and Micron if those tons ship, foundry customers who assumed 4nm was their lane, and AI programs still waiting on qualified stacks. |
The 80% figure can be a lot yield that does not survive customer qualification. SK hynix can hold Nvidia allocation even if Samsung has wafers. I would want a named customer taking Samsung HBM4 at volume, or a foundry customer complaining about 4nm pull, before I treat the mix as done. |
Watch who gets the incremental HBM4 tons first. Watch whether Samsung foundry 4nm looks like internal HBM support. |
| |
| |
80% yield, then 250,000 wafers |
Seoul Economic Daily puts HBM4 yield around 80% and monthly wafer input on a path from about 180,000 to about 250,000. Yield was catch-up. Volume is allocation. |
|
|
|
|
| |
| |
4nm base die can crowd foundry customers |
HBM4 pairs 1c DRAM with a 4nm base die inside Samsung. Internal pull for stacks can wait out external foundry work on the same node. |
|
|
|
|
| |
| |
ADVERTISEMENT |
|
|
The Companies Building AI's Future |
Every AI breakthrough requires more data centres, more power, more infrastructure. That buildout is creating opportunities well beyond the AI stocks everyone already knows. |
|
|
Paid advertisement from Stock Earnings. Not investment advice. |
|
|
|
|
| |
| |
THE LONG GAME |
An HBM line is a multi-year physical asset. Yield can print in a quarter. Qualification at the customer still takes lots and a board that holds. Commodity DRAM can tighten if more wafers go into HBM. |
The gauge's physical component stays sticky. Memory tons are still a plant and a customer who will requalify a stack. Volume is a sentence. Qualification is a year. |
|
|
|
|
MUST READS |
Wall Street Was Wrong About AI. Twice.
First they called it a bubble. Then they told you to pile into the Magnificent Seven. Green recommended Apple under $1, Amazon under $2, Netflix at $1.62 and Nvidia in 2004. |
One Utility Has Paid Every Quarter for 96 Straight Years
Also inside: a tech leader with a 15% payout ratio, leaving decades of room for dividend growth, a consumer staples name with twenty years of outperformance, and a healthcare leader with 61 consecutive years of payout increases. Seven in total. Free report. |
Follow the $50 Billion Buy-In
Wall Street just put billions behind a small group of names, and after a volatile first quarter the accumulation is building rather than fading. The report lays out the tickers, the rotation behind them, and why fresh capital is moving now. Free report. |
The gauge is 72 and flat. Credit held on a sessionless Tuesday seed. Valuation and Froth remain estimated. |
- Knox Bennett, Lead Strategist, Strategic Compute |
|
|
| |
| |
AI ARMS RACE
A Guardian Financial Publishing title |
PRIVACY & DATA HANDLING
We take your privacy seriously and we do not sell your personal information. Full details are in the Privacy Policy linked below. Paid advertisements may appear in this email. |
|
|
|
|
|
|
AI Arms Race is a publication of Guardian Financial Publishing (D/B/A) Stable Financial Publishing
Privacy Policy · Terms · Disclaimer |
Tidak ada komentar:
Posting Komentar