Samsung to build Nvidia's NVHBM base dies on TSMC 3nm
Nvidia's custom memory program is reportedly moving onto TSMC's process lines — with Samsung coming along for the first time. And in Philadelphia, the jobs-vs-delivery-robots argument has produced its first arrest.
Nvidia wants all three of its memory suppliers building its custom NVHBM base dies on one TSMC process — and Samsung has reportedly agreed.
The Seoul Economic Daily reported Sunday that Nvidia asked Samsung, SK Hynix and Micron to manufacture the base dies for its custom NVHBM memory on TSMC's 3nm process, using sixth-generation 1c DRAM for the core dies above them, in a story Reuters carried. Samsung accepted after an internal review, which would make it the first time the company uses TSMC at the HBM manufacturing stage — the one step of memory production Samsung has never let outside its own fabs. According to the report, Samsung would keep building standard HBM and other customers' custom products in-house, and is weighing its own foundry as a fallback if TSMC's capacity runs short. None of Nvidia, Samsung, SK Hynix, Micron or TSMC commented.
For readers who haven't tracked the acronym: NVHBM is Nvidia's custom-memory design, announced in August, that moves the memory controller off the compute die and down into the base die — the silicon floor every HBM stack is assembled on. Nvidia says the swap delivers 30% more memory bandwidth, 15% less HBM power, and up to 25% more compute-die area for cores, measured against standard HBM4E, from a single specification that any of the three memory makers can build. AWS's Annapurna Labs was the first outside customer to adopt it, for its Trainium4 chips. What Sunday's report adds is the manufacturing answer nobody had: which foundry makes that floor, for all three suppliers at once.
Memory has been the tightest input in the AI buildout for two years, and Nvidia's custom program is how the biggest buyer in the market shapes what its suppliers build. Whoever owns the base-die process decides how fast a new stack can be qualified and how much of the design is common across suppliers — which is the control question Sunday's report puts on the table.
That single detail reorganizes who controls the most valuable layer in the memory stack. Base dies are the part SK Hynix, Samsung and Micron have always owned end to end — the layer where their process know-how and margins live. Routing them through one external process gives Nvidia one qualification path instead of three vendor-specific ones, and gives TSMC a foothold in a manufacturing stage it has never touched. For Samsung, the arrangement undercuts the turnkey pitch it has used to sell memory — design, fab and stack under one roof — which may be exactly why it hedged the deal with a fallback to its own foundry. The counter-argument comes from Micron, whose COO Manish Bhatia told TrendForce that moving base dies to an external foundry should lift selling prices and margins: the suppliers trade a layer of process control for price.
The specifics rest on a single Korean newspaper — Reuters carried the Seoul Economic Daily's reporting rather than sourcing it independently, and no company has confirmed anything. The direction, though, was reported independently by TrendForce on October 9: Micron had said NVHBM and HBM4E base dies would come from an external foundry, and Korean trade press reported SK Hynix already runs HBM4 base dies on TSMC's 12nm process with 3nm prioritized for HBM4E and later custom parts. No volumes, prices or timelines have surfaced for any of it. Still, the pattern is consistent enough to read: memory's most profitable layer is being standardized around one Nvidia specification and, increasingly, one Taiwanese process.
A 78-year-old professor emerita was arrested in Philadelphia for standing in front of an Uber Eats delivery robot.
Police arrested Carolyn Marvin on October 14 on a Center City sidewalk after she repeatedly stood in the path of the robot and refused officers' orders to move, then charged her with disorderly conduct, per ABC13's report. "I know they're taking the jobs of human delivery workers and so, I have been protesting them whenever I saw them by standing in front of them and preventing them from going forward," Marvin told the station. Marvin is a 78-year-old Penn professor emerita who taught freedom of expression, and she plans to fight the charge on First Amendment grounds, according to PhillyVoice — she has been blocking the robots for months.
The Avride-built Uber Eats bots have become a regular presence in Philadelphia streets this year, and the backlash has been building with them: one was kicked over in Center City in April. Marvin's arrest was captured on video and has drawn national pickup over the weekend. Nothing in the reporting identifies a city rule specifically about obstructing a robot, which is why police reached for the generic disorderly conduct charge — Marvin isn't accused of damaging anything, only of refusing to get out of the way. That gap is the story. Cities have spent two years licensing autonomous delivery onto their sidewalks and writing rules for how the machines should behave; nobody has written the rule for a human who declines to move. A Penn professor of free expression testing that gap in court is about the best fact pattern the issue could hope for.
What to watch: whether Samsung, TSMC or Nvidia puts an on-record confirmation — or a denial — behind the NVHBM foundry report, and whether Samsung's dual-track hedge survives first contact with real order volumes; on the sidewalk, the docket date for Marvin's First Amendment challenge, and whether other cities see copycat blockades once the charge is public.
Is standing in front of a delivery robot protected protest or obstruction? Write to us via our contact page.
