A $200 billion handshake

Here's the deal: on July 24, at an "AI Summit" held at The Midway in San Francisco, Samsung Electronics and Broadcom signed a strategic MOU worth roughly $200 billion over five years, running through 2030.

That number is hard to feel, so here's a reference point. Samsung Electronics posted annual revenue in the ~$220 billion range in 2025. This agreement therefore contemplates doing business with a single counterparty at a scale approaching a full year of Samsung's total revenue, spread over five years. An MOU is a framework, not booked backlog, and we'll come back to that caveat.

But the number isn't the interesting part. The scope is. What Samsung is selling Broadcom isn't just memory. It's leading-edge memory including HBM, sub-2nm foundry process, and advanced packaging — three layers inside one agreement. Samsung has spent years saying "we're the only company that does memory, foundry and packaging under one roof." This is the first time that claim has been cashed as a single large-customer contract.

The two companies at the table

Start with Broadcom. Consumers rarely encounter the name, but in AI silicon it sits directly behind Nvidia. It does two big things. First, custom AI accelerator (ASIC) design — Broadcom co-designed Google's TPU, and it's the design partner behind Meta's MTIA line. When a hyperscaler decides "let's build our own chip instead of buying Nvidia," Broadcom is effectively the address they go to. Second, networking silicon — lashing tens of thousands of GPUs into one cluster takes switches and Ethernet chips, and Broadcom owns that market too.

Put differently, Broadcom holds most of the AI data center that Nvidia doesn't. And it's fabless: it designs, someone else manufactures. Until now that someone was overwhelmingly TSMC.

Samsung had the mirror-image problem. It could build; it couldn't fill the lines. Its foundry has trailed TSMC badly, and it lost the HBM lead to SK hynix — Q1 2026 HBM share is reported around 58% for SK hynix with Samsung and Micron near 21% each. Samsung has been pushing hard to close that: in February 2026 it became the first in the industry to mass-produce and ship HBM4 built on 1c DRAM with a 4nm base die, and in May it sampled HBM4E to customers.

So this deal is two companies with opposite gaps filling each other's. Broadcom can design but needs capacity and memory. Samsung has capacity and memory but needed an anchor customer.

What's actually in it

Item Detail
Signed July 24, 2026 (local time)
Venue The Midway, San Francisco — "AI Summit"
Form Strategic memorandum of understanding
Scale ~$200 billion over five years, through 2030
Memory HBM and other leading-edge memory solutions
Foundry Sub-2nm process applied to Broadcom's networking silicon and other key lines
Packaging Advanced packaging as part of an integrated solution
Samsung HBM4 World-first mass production shipment, February 2026 (1c DRAM + 4nm base die)
Samsung HBM4E Customer samples, May 2026

The foundry line deserves unpacking. What was specified is that Samsung's sub-2nm process will be applied to Broadcom's next-generation high-speed data communication chips. As AI clusters scale, the bottleneck stops being compute and becomes the data moving between chips. Bolt on all the GPUs you like — if the switches connecting them can't keep pace, aggregate performance stalls. That's why networking silicon now needs leading-edge nodes too, and routing that volume into Samsung's fab is the substance of this announcement.

Why packaging is the third leg is worth a beat as well. A modern AI chip isn't one logic die. It's a compute die with multiple HBM stacks alongside it, all bound into a single package. That assembly step became a hard constraint on AI chip supply over the past few years. When design (Broadcom), memory (Samsung), logic fabrication (Samsung) and packaging (Samsung) sit inside one agreement, there's real room to cut coordination overhead and lead time between steps. That's what "integrated solution" actually means here.

One thing to be clear about: this is an MOU. It's an agreed direction and scale, not confirmed backlog, and actual orders get placed product by product and year by year. The $200 billion figure behaves more like a ceiling on five years of cooperation than a revenue forecast, and realized volume will move with market conditions.

What each side gets

Samsung gets three things.

First, a platform to claw back HBM share. The HBM market has been decided largely by who captures more of one giant buyer — Nvidia — and SK hynix won that round. Broadcom represents a different demand axis: hyperscaler custom silicon. Winning that axis lets Samsung grow share while reducing dependence on Nvidia. This is the most valuable piece for Samsung.

Second, an anchor customer for the foundry. Samsung's foundry problem was never purely technical; it was utilization. Install a 2nm line with nothing to run on it and yields don't climb and costs don't fall. Foundry is a business where volume produces yield and yield produces cost. A customer of Broadcom's size feeding steady volume starts that flywheel. That's why Korean coverage described the foundry business as having "gained wings."

Third, proof for the bundle. Samsung has pitched memory-plus-foundry-plus-packaging for years without many marquee wins to show for it. TSMC is strong in packaging but has no memory; SK hynix is strong in memory but has no logic foundry. If this runs well, Samsung finally owns a reference case for one-stop positioning.

Broadcom gets supply chain diversification. Today it depends on TSMC for most leading-edge logic. The trouble is that Apple, Nvidia, AMD and Qualcomm are all queued for the same advanced capacity — even a very large customer can't escape the queue. Layer geopolitics on top and Taiwan concentration becomes a board-level item, not a procurement one. A second axis at Samsung improves both negotiating leverage and risk posture.

Broadcom also doesn't make memory. Selling custom AI accelerators means sourcing HBM from somewhere, and locking volume during a tight HBM cycle is competitive advantage by itself.

The weak link is execution on Samsung's side. Volume only flows once sub-2nm yields and HBM4 quality qualification actually clear. An MOU is paper; semiconductor contracts get decided in qual testing. Samsung has previously announced large customer wins that produced less volume than expected because of yield issues, and that history deserves a clear-eyed look.

How deals like this have gone before

Mega-scale, long-horizon semiconductor pacts have both success and failure precedent.

The success template is Apple and TSMC. In the early 2010s Apple had Samsung fabricating its A-series chips, then moved to TSMC as the smartphone fight intensified. Apple's volume gave TSMC the cash flow and yield-learning opportunity to invest ahead of every subsequent node — and that compounding produced today's gap. The lesson is that one anchor customer can bend a foundry's entire technology trajectory. That's precisely the effect Samsung wants from Broadcom.

The failure pattern also exists. Plenty of chip companies have announced splashy MOUs that never converted into revenue, generally for the same reason: at MOU stage, price and volume aren't fixed. The customer is really buying an option and will place actual orders based on competitive conditions at the time. In foundry specifically, if you don't pass qual, having a contract doesn't move wafers.

One recent precedent matters here. Two days before this article, on July 27, Samsung and SK were reported to have unveiled AI cooperation packages with the US totaling roughly $950 billion. The Broadcom deal sits in that same current. This isn't a one-off — it's part of a continuous repositioning of Korean semiconductors inside the AI infrastructure supply chain. Any single announcement is just an MOU. Several of this magnitude landing within weeks is a signal.

How competitors counter

TSMC takes the most direct hit, though it isn't wobbling. Its leading-edge share and yields remain dominant, and not all Broadcom volume is moving. TSMC's counter is predictable: lean on Arizona and Japan fabs to defuse the geopolitics argument, and expand CoWoS packaging capacity to relieve the bottleneck itself. What stings TSMC most isn't the lost volume — it's that a customer now holds one more bargaining card.

SK hynix will read this announcement most carefully. Its strategy for defending the HBM crown has been deep integration with Nvidia; a growing Broadcom axis disperses the market's center of gravity. SK hynix is also moving toward custom HBM with client-specific base dies, and that's likely where the next round gets decided. Its structural constraint is that without a logic foundry, it can't make Samsung's bundle offer.

Micron is expanding HBM capacity fast on the back of US domestic-production support. If a Samsung–Broadcom axis hardens, expect Micron to push harder into other custom-silicon customers.

Nvidia's calculus is different. The custom accelerators Broadcom designs are direct substitutes for Nvidia GPUs; the more hyperscalers shift volume to their own chips, the weaker Nvidia's grip. But Nvidia is also a Samsung HBM customer, which leaves Samsung selling to both sides. For a memory company, it doesn't much matter who wins the chip war — that's an old truth of this industry.

And Intel. For a company trying to rebuild a foundry business, this is bad news. If Samsung claims the "credible alternative to TSMC" position with results, the opening Intel was aiming at gets narrower.

So what changes

If you're in semiconductors, the metric to watch isn't the dollar figure — it's qual test outcomes and 2nm yield. Volume typically takes several quarters to flow after an MOU. Samsung foundry utilization and HBM shipment numbers from H2 2026 into early 2027 are the real report card.

If you invest, Samsung Electronics now warrants a separate read on the foundry segment. The stock has effectively traded on the memory cycle; an anchor foundry customer opens room for the valuation logic to shift. With the same caveat: MOU, and execution decides.

If you build AI infrastructure, the implication is that the custom silicon path just got more practical. With Broadcom designing and Samsung supplying memory, logic and packaging, the road to your own chip is shorter. Expect more attempts to exit the Nvidia queue.

If you follow the Korean economy, bundle this with the other announcements of the past several days. Korean semiconductors are trying, on multiple fronts at once, to shift from component supplier to integrated solution provider in the AI infrastructure chain. Succeed and the industrial structure changes; fail and all that's left is the capex.

If you're just reading the news, one sentence: the next round of the AI chip fight isn't Nvidia versus everyone — it's "who can build you the whole chip." What Samsung sold here wasn't HBM. It was that integration.

🥄 Three Things You're Probably Wondering

— Does the whole $200 billion become revenue? No. An MOU is a framework; actual orders are placed product by product. The figure is closer to the total scope of cooperation the two will discuss over five years, and realized value depends on market conditions and qual results. Dropping the headline number into a revenue model is a mistake.

— Is Samsung catching TSMC now? Too early to say that. The gap in leading-edge share and yield is still large, and this covers sub-2nm for part of Broadcom's lineup. The direction is clear, though: what Samsung needed wasn't another technology announcement, it was a customer to fill the lines, and it landed one.

— Can Samsung overtake SK hynix in HBM? Not easily. 58% versus 21% is a wide gap and SK hynix's position in Nvidia-bound volume is solid. But as custom silicon demand grows the market disperses, and that's where Samsung can compete. The pivot is who wins large-customer qualification first at the HBM4/HBM4E generation.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!