Samsung wins a $200 billion deal with Broadcom, expanding cooperation on memory, chips, and advanced packaging

Samsung lands a $200 billion, five-year order from Broadcom, covering 2nm foundry manufacturing and HBM4 memory supply. Samsung’s 2nm mass-production yield is still stuck around the 60% mark—does this order mark a technological turnaround, or is it Broadcom’s second choice amid worries about TSMC’s capacity?

(Background: Trump calls for chips to return to the U.S.! TSMC’s gross margin will need to be eroded by 4 percentage points, with a price hike again in 2027?)
(Extra context: Nvidia and the SK Group toss in a $500 billion plan: 2GW AI data centers to start in 2027, locking in HBM4 supply)

Table of contents

Toggle

• What does a single memo buy?
• Who is anxious
• Before the yield passes

$200 billion is the five-year foundry order amount Samsung Electronics secured from Broadcom this time. A foundry shop said to have only a 50–60% mass-production yield for 2nm—still hovering on the edge of the stable production threshold—got the check. The question may not be whether Samsung’s technology has suddenly leapt forward, but whether Broadcom doesn’t want to be left with only one option—TSMC.

What does a single memo buy?

On July 25, Samsung Electronics announced that it has signed a five-year, deal running through 2030, cooperation memorandum with Broadcom worth over $200 billion. The MOU was signed on July 24 during the AI Summit in San Francisco. It covers two parts: memory supply and wafer foundry manufacturing.

On the memory side, Samsung will supply Broadcom’s next-generation AI accelerators with HBM4 and HBM4E high-bandwidth memory; on the foundry side, Samsung will use 2nm and below process nodes to manufacture Broadcom-designed ASICs. The cooperation also extends to advanced packaging, including 2.3D and 2.5D integration technologies.

This order didn’t appear out of thin air. At the time, South Korean President Yoon Jae-in was in Silicon Valley hosting an AI summit. During the itinerary, he met Nvidia CEO Jensen Huang, OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, and Broadcom CEO Hsiao Fu-yang. Samsung Electronics Chairman Lee Jae-yong, SK Group Chairman Choi Tae-won, Hyundai Motor Group Chairman Chung Ui-sun, and Naver founder Lee Hae-jin were all present, and after the event they released the “San Francisco AI Declaration.”

Nvidia’s cooperation projects with Naver, and Nvidia’s cooperation with SK Hynix, were also announced during the same trip.

Who is anxious

Over the past half-year, Samsung and SK Hynix have basically thrown all their effort into pushing AI chip orders for a very practical reason: TSMC’s advantage in advanced foundry is widening—not shrinking.

In 2025, TSMC captured nearly 70% global wafer foundry market share, with revenue of $122.5 billion; Samsung had 7.2% and revenue of $12.6 billion. In other words, Broadcom’s $200 billion, five-year order exceeds by more than fifteen times Samsung’s full-year foundry revenue in 2025.

This isn’t the result of Samsung’s technology suddenly overtaking. It’s more likely Broadcom’s own calculations. In recent years, Broadcom has grown its AI ASIC business by designing chips for cloud giants like Google and Meta, with order sizes becoming ever larger. If foundry capacity were completely locked to TSMC, it would effectively hand its own capacity and pricing power to TSMC. Then when TSMC raises prices or prioritizes capacity schedules, Broadcom would have nothing else to say.

Choosing Samsung as a “second supplier” spreads risk—and also helps rein in TSMC’s negotiating leverage. This is a common chip-industry logic of diversifying risk. It’s just that the amount this time is exceptionally large, and it also lays bare the gap that already existed between Samsung and TSMC.

Another driving force comes from the South Korean government. In this Silicon Valley trip, Lee Jae-myung—pushing diplomacy resources entirely into semiconductor and AI infrastructure—bundled the overall scale of cooperation between South Korea and the U.S. to about $950 billion. SK Hynix’s and Nvidia’s investment plans are also included. Anthropic CEO Amodei even stated more directly at the summit: it has already signed supply agreements with Samsung Electronics and SK Hynix. This company, which raised funding in May this year with a post-money valuation of $96.5 billion, named Samsung, SK Hynix, and Micron as key future supply partners for its memory and chip needs.

Only when geopolitics, capacity anxiety, and capital align do these three forces come together to form this order.

Before the yield passes

Samsung’s 2nm process mass-production yield is rumored to be about 50–60%, still right at the edge of the stable mass-production threshold; TSMC’s 2nm mass-production yield for the same generation is already above 70%.

In other words, the numbers written in the MOU and the numbers that can actually ship are two different things. A memo is only an intention and a framework. What truly determines whether Samsung can absorb the $200 billion is whether, over the next few years, it can raise its 2nm yield from 50–60% to TSMC-level. But improving yield has never been an engineering problem that can be accelerated by political will alone.

For Samsung and SK Hynix, this order is more like an admission ticket than a guarantee. Broadcom is willing to put eggs in two baskets because it worries about TSMC’s dominance; but if Samsung’s yield never catches up, Broadcom could shift the order’s weight back to TSMC at any time.

The South Korean government wants to double memory capacity within five years and pull out “world-class manufacturing capability.” It sounds ambitious, but ultimately it has to be delivered through wafer-by-wafer yield numbers—not through a photo at a summit.