Broadcom Just Signed a $200 Billion Pact With the Foundry That Has 6.5% of the Market. The Target Is the One With 72%.

TL;DR

On July 24, Samsung and Broadcom signed a memorandum of understanding the two companies estimate at more than $200 billion over five years, through 2030. The bundle: HBM4 and HBM4E memory for Broadcom's AI accelerators, 2-nanometer-class foundry work, and 2.3D/2.5D advanced packaging, pitched as one turnkey stack. Broadcom is the company that builds custom AI silicon for Google and OpenAI, which makes this the largest bet anyone has placed on a second source for AI chip manufacturing besides TSMC.

What is actually in the bundle

Per Bloomberg and Samsung's own release, the agreement spans three layers that normally get bought separately: high-bandwidth memory for Broadcom's AI accelerators, chips built on Samsung's 2nm-and-below process technologies, and advanced packaging including 2.3D and 2.5D integration. Seoul Economic Daily reports the memory side covers HBM4, built on Samsung's sixth-generation 1c DRAM with a 4nm base die, plus the follow-on HBM4E.

The word Korean coverage keeps using is turnkey, and it is the actual product here. A modern AI accelerator is not one chip; it is a logic die and stacks of HBM sitting on an interposer, fused together by advanced packaging. Today those pieces usually come from different vendors on different continents, with everyone fighting over allocation. Samsung is pitching the whole thing under one roof: one vendor, one invoice, one throat to choke.

One detail from Samsung's release that the headlines flattened: the named 2nm foundry work covers Broadcom's wireless broadband connectivity products, while the HBM supply targets its AI accelerators and the packaging ties both together. Vice Chairman Young Hyun Jun's framing: "AI is driving unprecedented demand for tightly integrated semiconductor technologies spanning memory, logic and advanced packaging."

The name on the other side of the table

The $200 billion figure will get the headlines, but the counterparty is the story. Broadcom is the quiet giant of custom AI silicon: it has co-designed Google's TPUs for roughly a decade, and last October it signed a 10-gigawatt pact to build OpenAI-designed accelerators, with racks targeted to start deploying in the second half of 2026. When a hyperscaler wants off Nvidia's price list, Broadcom is who they call.

Until now, essentially all of that silicon has been manufactured by TSMC. That is the context that makes this MOU more than a supply agreement: the biggest buyer of custom AI chip capacity just publicly committed to qualifying a second manufacturer.

Per TrendForce data reported by TelecomLead, the top-10 foundries booked a record $47.95 billion in Q1 2026, and TSMC's share grew from 70.4% to 72.3% quarter over quarter. The AI boom has been making the foundry market more concentrated, not less. Every AI builder pays for that concentration eventually: it is embedded in GPU prices, in HBM lead times, and in why one earthquake off Taiwan shows up in your inference bill.

Samsung needed exactly this

For years, Samsung's leading-edge pitch deck had a blank space where the anchor customer should be. That started to change in July 2025, when Tesla signed a $16.5 billion deal, running through 2033, to build its AI6 chips on Samsung's 2nm-class process at the Taylor, Texas fab. Elon Musk called that number "just the bare minimum." Broadcom is the second marquee name in a year, and it is an order of magnitude larger on paper.

The memory half may matter even more than the foundry half. SK Hynix has owned the HBM relationship with Nvidia through the entire boom while Samsung fought through qualification delays. Landing Broadcom as a named anchor customer for HBM4 and HBM4E, in the middle of a global memory crunch, is the win Samsung's memory division has been chasing for three years.

The Seoul subplot

The signing happened at The Midway in San Francisco with South Korean President Lee Jae-myung in the room, mid-swing through Silicon Valley, where he also met Sam Altman, Jensen Huang, and Dario Amodei, per the Korea JoongAng Daily. ANI reports Korean and US tech giants agreed to pursue roughly $950 billion in semiconductor and AI partnerships over five years, unveiled at the same summit. The Samsung-Broadcom MOU is the flagship of that package, which means it is partly industrial policy wearing a term sheet.

Caveats

• This is a memorandum of understanding, not a purchase order. Bloomberg's headline calls it a contract; Samsung's own release says MOU with an estimated value. The estimate is the two companies' projection of the collaboration's worth through 2030, not booked revenue.
• There is no disclosed breakdown between memory and foundry, no binding volumes, and no pricing.
• The 2nm foundry work named in the release covers Broadcom's wireless broadband connectivity products. Neither company has said the compute dies for Broadcom's AI accelerators, Google's or OpenAI's included, are moving off TSMC.
• Samsung still has to execute: 2nm yields at scale and HBM4 qualification at Broadcom are prerequisites, not givens. Marquee pacts flex in both directions.

Key Takeaways

• Samsung and Broadcom signed an MOU estimated at more than $200 billion over five years through 2030, spanning HBM4/HBM4E memory, 2nm-class foundry work, and 2.3D/2.5D packaging.
• Broadcom is the custom-silicon partner behind Google's TPUs and OpenAI's 10-gigawatt accelerator program, so this is a public commitment to a second source beyond TSMC.
• TSMC held 72.3% of foundry revenue in Q1 2026 to Samsung's 6.5%, and the AI boom has been widening that gap, per TrendForce.
• It is Samsung's second marquee 2nm win in a year after Tesla's $16.5 billion AI6 deal, and its first named anchor customer for HBM4 during a global memory crunch.
• The fine print matters: it is an MOU with an estimated value, the named 2nm work is connectivity silicon, and nobody has said AI compute dies leave TSMC.

Sources: Samsung Newsroom, Bloomberg, Fortune, Seoul Economic Daily, TelecomLead (TrendForce data), Broadcom Investor Relations, CNN, ANI