Samsung and TSMC Reportedly Join Forces on Nvidia Custom HBM

Samsung is working with TSMC to supply custom high-bandwidth memory for Nvidia, according to a report from Verdict. The arrangement, if confirmed, would pair the world's two largest contract chipmakers on a single product line rather than in their usual roles as rivals.

The deal centers on high-bandwidth memory, the stacked DRAM that sits beside AI accelerators and, increasingly, defines their performance limits. Custom HBM variants go further than standard parts: the memory is co-designed and co-packaged with a specific customer's logic, which ties the memory supplier to the foundry that builds that logic.

In Nvidia's case, the logic comes from TSMC. Samsung supplies memory. The reported collaboration would therefore place Samsung's HBM inside an TSMC-fabricated, Nvidia-branded package — a configuration that requires the two manufacturers to align on interface specs, packaging flows and qualification schedules.

Why would Samsung and TSMC cooperate?

The commercial logic is straightforward. AI memory demand has outrun supply, and Nvidia buys HBM at a scale no single vendor comfortably satisfies. SK Hynix has led the HBM race in recent generations; Samsung has chased it. A direct line into Nvidia's custom programs would give Samsung volume it has struggled to win outright.

For TSMC, the calculus is equally plain. Its largest AI customers want qualified HBM attached to their logic with minimal friction. Cooperating with Samsung widens the pool of approved memory without TSMC making memory itself — a business it has avoided.

For Nvidia, a second qualified HBM source strengthens negotiating position and hedges supply risk. Three parties, three incentives, one package. Rarely does the industry arrange its rivalries so neatly.

What remains unconfirmed?

The report does not specify volumes, timelines or which HBM generation the custom parts would target. Samsung, TSMC and Nvidia have not, as of this writing, issued matching public confirmations of the arrangement.

That leaves open questions that matter commercially:

• Which HBM generation — current or next — the custom chips would use
• Whether Samsung's HBM would ship in TSMC's advanced packaging or a hybrid flow
• When volume production would begin, and at what scale
• How SK Hynix, the incumbent leader, responds in Nvidia's supplier mix

Analyst estimates of HBM market share and pricing shifts should be treated as projections until any of the three companies discloses shipment figures.

The packaging angle

Custom HBM tightens the link between memory and logic manufacturing because the memory stack must be qualified against a specific foundry's process and packaging. Coordinated HBM4-class designs push memory makers toward co-optimization with the logic supplier rather than selling catalog parts.

This is the practical reason a Samsung-TSMC collaboration is plausible even amid fierce foundry competition. Advanced packaging has become the meeting point where rivals must interoperate because customers demand it. TSMC builds the compute; Samsung builds the stack; Nvidia owns the specification.

What comes next

Confirmation, when it arrives, will likely take the form of supply agreements or qualification announcements rather than a joint press conference. Watch Samsung's HBM shipment disclosures and Nvidia's supplier commentary over coming quarters. If the reported collaboration holds, the AI memory market shifts from a two-horse race toward a structured three-way dependency — and Samsung gains the Nvidia design win it has pursued for years.