Six Weeks On, Was the Samsung MOU Only A Foundry Story?

Key Highlights

• Samsung Electronics and Broadcom signed a memorandum of understanding on July 25, 2026, with the collaboration estimated by the companies at more than $200 billion across memory and foundry over five years through 2030, with no split disclosed between memory, foundry and packaging.
• Scope covers HBM supply for Broadcom's next-generation AI accelerators, Samsung 2nm and below process technology for Broadcom products including Wireless Broadband Communications solutions, and an expected extension into 2.3D and 2.5D advanced packaging built on the 2nm process.
• The MOU sits inside a broader package of Korea and US semiconductor cooperation announced at the same San Francisco summit and put at roughly $950 billion by the Korean presidential office, of which SK Group accounted for about $750 billion including an SK Hynix and NVIDIA memory and data center initiative valued at more than $500 billion.
• Broadcom's Q3 fiscal 2026 results, reported September 2, put AI semiconductor revenue at $16.7 billion, up 221% year over year and 56% of total revenue, against non-AI semiconductor revenue of $4.2 billion, with management guiding to roughly $115 billion of AI semiconductor revenue in fiscal 2027 on supply it describes as secured.
• Intel's Q2 2026 prepared remarks described industry-wide supply constraints across wafers, memory and substrates as the dominant challenge facing its own customers, naming the same silicon-side constraint set from the other end of the leading-edge market.
• Our read is that the memory and packaging elements carry more weight than the foundry headline suggested at the time, though nothing in the release allocates value between them.

The News

This July announcement came back to mind this week, revisited because the buyer has since told us more about its own constraints than the release did. Samsung Electronics and Broadcom announced on July 25, 2026 a memorandum of understanding to expand strategic collaboration across memory and foundry technologies. The companies estimated the value at more than $200 billion over five years through 2030, spanning HBM supply for Broadcom's next-generation AI accelerators, Samsung 2nm and below process technologies for Broadcom products including Wireless Broadband Communications solutions, and an expected extension into 2.3D and 2.5D advanced packaging. Read in July, the announcement invited a foundry-share interpretation. Read after Broadcom'sSeptember 2 earnings report, it invites a supply-security one, which is the subject of this note (Broadcom press release, July 25).

Analyst Take

We are deliberately late to this one, because the original MOU announcement seemed like only part of a story. Six weeks on, the Broadcom earnings print gives us a better read on what the July memorandum was, and we would rather err on the side of a useful interpretation. At announcement, consensus read it as a foundry story. We think the memory and packaging language carries more weight, and we hold that as a read rather than a finding, because neither company disclosed an allocation and no cash has to move under a memorandum of understanding. The bear case deserves a fair hearing first, and it is a serious one: this is a non-binding instrument, the value is a company-stated estimate, and it was announced inside a Korea and US cooperation package the presidential office put at roughly $950 billion, at a summit convened by a head of state. Announcements framed that way carry diplomatic weight alongside commercial intent. What has changed since July is not the document. It is what the buyer has since said about what actually limits it.

What was Announced

The agreement is structured across three layers, and the ordering in the release is worth noting. Memory comes first: Samsung and Broadcom plan to pursue strategic collaboration for supply of leading-edge memory including HBM, explicitly supporting Broadcom's next-generation AI accelerators. Foundry follows, focused on 2nm and below for Broadcom products including Wireless Broadband Communications solutions. Packaging is described as an expected extension, covering 2.3D and 2.5D integration built on the 2nm process.

We would infer that the packaging clause carries more weight than its conditional phrasing suggests. Wafer starts are the constraint the market talks about. Assembly capacity is harder to substitute, because a 2.5D slot cannot be resold or pulled forward the way a wafer allocation sometimes can. One distinction matters here and is easy to lose: the substrate capacity Broadcom discussed on its own call refers to a Singapore facility Broadcom is bringing up in fiscal 2027, which is a different asset from the 2.3D and 2.5D integration named in the Samsung release. Related, not identical. Both sit downstream of the wafer, which is the point.

The foundry entry point invites a reading rather than announcing one. Wireless Broadband Communications is the only Broadcom product family the release names, and it is not the AI franchise. We read that as a measured qualification path rather than immediate migration of flagship accelerator designs, though the release says "including" and not "starting with," so the sequencing is our inference and not a disclosure. Samsung's own framing positions memory, logic, foundry and advanced packaging as an end-to-end offering. Neither executive quote claims a design win. Both describe intent to collaborate, which is what an MOU is designed to carry.

Anyone who has run enterprise procurement at scale in the US, Europe and Asia recognizes the shape of this. The second supplier is rarely the cheaper one. It is the one you keep alive so that the first one has a reason to quote you a delivery date it intends to keep. Post-COVID that secondary source is more than leverage, it is due diligence.

Market Analysis

The constraint has moved, but it has not moved to one place. Broadcom's Q3 print suggests two clocks running at different speeds. The deployment clock is land, power and shell, which management described as dictating the specific timing of when capacity gets deployed, and which has construction lead times no supplier agreement can compress. The silicon clock is leading-edge wafers, substrates and memory. The Samsung MOU touches the second clock and does nothing for the first. That is a limit on how much this agreement can be made to explain, and it is the discipline a six-week lookback is supposed to impose.

Within the silicon clock, the memory question is genuinely two-sided. Management noted that HBM and the system memory around it are partly secured by customers rather than by Broadcom. Set against that, Broadcom has attributed consolidated gross margin dilution to rising memory content inside its XPUs, which places memory squarely in its own bill of materials. Both are true, and the MOU sits in the space between them. Broadcom appears to be securing a second berth in a harbor that is running out of them, rather than buying wafers.

The corroboration from Intel is worth something, with a caveat attached. Intel's Q2 2026 prepared remarks named wafers, memory and substrates as the dominant challenge facing its customers. Both companies are talking their own book, Broadcom to defend a supply-capped guide and Intel to defend a capex program, so this is not disinterested testimony. It is still notable that two books pointing in different directions produced the same three-item list.

The appetite for a second leading-edge source is not Samsung-specific, and Intel Foundry is the other live test on an immediate calendar. 18A is in volume ramp for internal products, 18A-P entered risk production, the 14A 0.9 process design kit is targeted for October, and management expects external 14A decisions from the second half of 2026 into the first half of 2027, with capacity spend following firm external commitments. Samsung brings a different configuration, holding memory, leading-edge logic and advanced packaging inside one company. Its Q2 commentary cited relatively high utilization below 8nm, 2nm project wins expected to more than double year over year, and a second Taylor fab breaking ground by year end. Counterpoint Research put Samsung at 7% of pure-play foundry against 73% for TSMC in Q2, and unconfirmed Korean and third-party reports have placed Samsung 2nm yields somewhere between the mid-50s and low-60s against a higher TSMC range, up sharply from reported levels a year earlier and none of it company-disclosed. Execution against those markers, not the headline value, is what settles this.

Looking Ahead

Going forward, we'll be tracking how the company converts language into disclosure, and the calendar is unusually cooperative. Broadcom presents at Goldman Sachs Communacopia on September 8 and reports fiscal Q4 on December 9. Neither event is obliged to mention Samsung. Silence is the default; Broadcom does not usually name foundry or memory counterparties. A named second source would be the surprise. That is why either outcome is informative: a management team that has just guided to roughly $115 billion of fiscal 2027 AI semiconductor revenue on supply it calls secured has a reason to describe second sources if they have become part of how that guide gets delivered, and no reason to describe them if the memorandum remains what it was in July.

Samsung's Tesla agreement is the useful template for what conversion looks like, because it is a named, filed, multi-year contract with a disclosed site and a production clock. This MOU has none of those attributes yet and names no manufacturing location at all.Three markers would change that: disclosed qualification of next-generation HBM (HBM4E or equivalent) for Broadcom accelerators; committed scope on the packaging clause; and a named site. Watch Intel's 14A commitment window on the same calendar. Two tests of one thesis. Capacity is the currency now.