Custom XPU on Korea Invest Insights

Broadcom Reconfirmed a $100B+ AI Semiconductor Run-Rate for 2027: How Korea Should Read the Value Chain

Fri, 05 Jun 2026 09:30:00 +0900

📚 Connected context This is a follow-up to the Marvell/Broadcom Korea bottleneck preview, the Jensen Huang-Marvell read-through, the laser-source follow-up, and the Samsung-Hynix-Micron parity note. Relevant hubs are the Korea semiconductor value chain hub, AI HBM hub, and AI PCB/substrate hub.

TL;DR

Broadcom’s Q2 FY2026 result is not negative for Korean semiconductors. The better read is that AI custom silicon demand continues to flow into HBM, package substrates, networking, and power-integrity components.

The stock fell because expectations were even higher, not because the underlying AI demand collapsed.

Core View

Broadcom reconfirmed the FY2027 AI semiconductor revenue frame of more than $100B. The market wanted an upward revision, not just confirmation. Korean investors should treat the selloff as an expectation reset and re-check the durability of HBM, FC-BGA, AI networking, silicon capacitors and test/packaging bottlenecks.

The Korean translation is:

Priority	Layer	Korean names to watch	View
1	HBM / server DRAM / eSSD	Samsung Electronics, SK hynix	Core layer. Custom ASIC broadens the HBM customer base beyond NVIDIA GPUs.
2	FC-BGA / Si-Cap / MLCC	Samsung Electro-Mechanics	Most direct non-memory read-through, but valuation discipline matters.
3	AI networking PCB / high-speed boards	Isu Petasys, Korea Circuit and peers	Second-order beta to AI Ethernet and switch ASIC demand.
4	TC bonders / sockets / inspection	Hanmi Semiconductor, ISC, Leeno and peers	Follow-through depends on real orders.

One-line conclusion: Broadcom’s fall is not a bearish signal for Korean AI infrastructure. It is a reset in over-heated expectations. The first name to revisit is Samsung Electronics, followed by SK hynix and Samsung Electro-Mechanics.

What Broadcom Actually Confirmed

[Fact] Broadcom reported Q2 FY2026 revenue of $22.187B, up 48% year over year. GAAP net income was $9.310B, non-GAAP net income was $12.074B, and free cash flow was $10.262B. (Broadcom Q2 FY2026)

[Fact] AI semiconductor revenue was the key. Q2 AI semiconductor revenue was $10.8B, up 143% year over year, and Q3 AI semiconductor revenue is guided to $16.0B, implying more than 200% year-over-year growth. Q3 total revenue guidance is $29.4B. (Broadcom Q2 FY2026)

Q2 AI semiconductor revenue mix = $10.8B / $22.187B = about 48.7%
Q3 AI semiconductor guidance mix = $16.0B / $29.4B = about 54.4%

Broadcom is no longer just a diversified semiconductor/software company with VMware attached. AI custom silicon is approaching half of revenue.

[Fact] Earnings-call summaries indicate Broadcom still frames FY2026 AI semiconductor revenue at about $56B and FY2027 at more than $100B. It also cited Q2 AI semiconductor bookings of more than $30B. (MarketBeat)

The important word is “still.” Broadcom did not introduce a new, higher 2027 frame. It maintained the existing one. That is where the market was disappointed.

Why the Stock Fell

The absolute numbers were strong. The problem was expectation.

First, investors expected more than the Q3 AI semiconductor guide of $16B. Some reports put market expectations around $16.3B to $17.2B. (Reuters)

Second, the FY2027 $100B+ frame was a confirmation, not an upgrade.

Third, margin mix became a debate. The AI custom silicon mix is rising, and investors questioned whether that could pressure gross margins even as operating leverage remains strong.

So the message is not “AI semiconductor demand is rolling over.” It is: AI ASIC leaders are now priced for very large upside, and anything short of another step-up can trigger a reset.

Why It Matters for Korea

Broadcom’s AI revenue is not only custom XPU. It also includes AI networking. More custom AI chips mean more HBM, high-performance packages, power-integrity parts, high-speed boards, test sockets and optical/networking hardware.

[Inference] The custom AI compute stack is widening from NVIDIA GPUs into Google TPU, Broadcom XPU, OpenAI accelerators, Meta MTIA, Anthropic compute and other hyperscaler designs. Korea participates mostly through the physical bottlenecks: memory, substrates, passive components, test and high-speed interconnect.

Korean Stock Read-Through

Samsung Electronics is the balanced read-through. It has HBM catch-up, server DRAM, eSSD/NAND and a long-term foundry/base-die option. The key check is HBM4E/HBM4 customer approval and whether DS earnings estimates keep rising.

SK hynix remains the cleanest HBM exposure. But the investment question is no longer simply “HBM is good.” It is whether the stock is attractive versus Micron and Samsung Electronics after the recent move. That relative-value angle is covered in the Samsung-Hynix-Micron parity note.

Samsung Electro-Mechanics is the strongest non-memory translation. FC-BGA, MLCC and silicon capacitors sit directly under AI ASIC and AI networking growth. The thesis is right, but the stock already reflects a large part of the re-rating, so repeated orders and margin proof matter.

Hanmi Semiconductor, ISC, Leeno and testing/socket names benefit only when the package/test capex turns into orders. Here the right posture is watchlist, not automatic buying.

Execution

The event supports Korean AI infrastructure, but it does not say “buy everything.”

Layer	What to verify
Samsung Electronics	HBM4E/HBM4 approval, DS earnings revision, foreign flow stabilization
SK hynix	Relative PER vs Micron, HBM price/mix, foreign selling pressure
Samsung Electro-Mechanics	Si-Cap repeat orders, package-solution margin, AI customer diversification
High-speed PCB	AI networking orders, ASP, capacity allocation
Packaging/test	Real orders, not only TAM expansion narratives

Final View

Broadcom reconfirmed a large AI ASIC market. The market sold the stock because it wanted an even larger upgrade. For Korea, that distinction matters. The demand signal remains positive, but the investable alpha should be sorted by physical bottleneck and valuation.

The cleanest path is Samsung Electronics as the core, SK hynix on relative-value pullbacks, and Samsung Electro-Mechanics only with valuation discipline and repeated order evidence.

Sources

Marvell's Trillion-Dollar Story and the Broadcom Read-Through: What Does It Mean for Korea?

Tue, 02 Jun 2026 20:45:00 +0900

Context This is a follow-up to the pre-earnings Marvell/Broadcom Korea bottleneck note, Marvell Q1 FY2027 and Korean semiconductors, and the AI infrastructure multiple map. Related hubs: AI PCB and Substrate Hub, Korea Semiconductor Value Chain Hub, and AI HBM Hub.

TL;DR

Jensen Huang calling Marvell a potential next “trillion-dollar company” does not mean Marvell replaces NVIDIA GPUs. The better interpretation is that custom XPUs, AI networking and optical interconnects have become important enough to be pulled into the NVIDIA AI factory architecture.

For Broadcom, this is a 70/100 positive read-through. It is positive because Broadcom’s AI business sits exactly in custom XPU, Ethernet switching, optics, SerDes, CPO and 3.5D packaging. It is not 100/100 because Marvell is the cleaner NVIDIA NVLink Fusion partner, while Broadcom is more tied to hyperscaler custom silicon and open Ethernet / UALink-style ecosystems.

For Korea, the first-order read-through is not HBM alone. The direct language is FC-BGA, AI network PCB, silicon capacitors, high-speed substrates and power integrity.

Priority	Korean exposure	Why it matters	Stance
1	Samsung Electro-Mechanics	FC-BGA + MLCC + silicon capacitor bundle	Cleanest structural proxy, but price discipline matters
2	Korea Circuit	Broadcom AI data-center ASIC FC-BGA qualification option	Higher beta, needs public customer / production proof
3	Isu Petasys	AI Ethernet / switch / networking PCB beta	Thematic beta, needs backlog and ASP proof
4	Samsung Electronics / SK Hynix	HBM / eSSD second-order beneficiaries	Core AI memory exposure, but not the first-order event
5	Hanmi Semiconductor	Advanced-packaging equipment derivative	Needs actual orders and customer diversification

One-line conclusion:

Marvell’s trillion story and Broadcom’s earnings checklist point to the same place: after GPU/HBM, the next AI infrastructure bottleneck is connectivity, substrate and power integrity. In Korea, Samsung Electro-Mechanics is the cleanest structural proxy; Korea Circuit and Isu Petasys are higher-beta translations.

1. What Happened

[Fact] Reuters reported on June 2, 2026 that Jensen Huang joined Marvell CEO Matt Murphy on stage during Computex week in Taipei and called Marvell the next “trillion-dollar company.” Marvell shares rose more than 24% in U.S. premarket trading. Reuters also noted that Marvell’s market cap was still below $192bn and that NVIDIA had earlier invested $2bn in Marvell. (Reuters via Investing.com)

[Fact] Marvell and NVIDIA announced an NVLink Fusion partnership on March 31, 2026. Marvell will provide custom XPUs and NVLink Fusion-compatible scale-up networking, while NVIDIA will provide Vera CPU, ConnectX NICs, BlueField DPUs, NVLink interconnect, Spectrum-X switches and rack-scale AI compute. (Marvell 8-K / EX-99.1)

[Fact] NVIDIA describes NVLink Fusion as a rack-scale AI infrastructure platform that enables hyperscalers and custom ASIC designers to integrate custom CPUs and XPUs with NVLink scale-up interconnect and OCP MGX rack-scale architecture. The modular portfolio includes NVIDIA GPUs, Vera CPUs, NVLink networking, CPO switches, ConnectX SuperNICs, BlueField DPUs and Mission Control software. (NVIDIA)

So the point is not simply “Marvell is good.” The larger point is:

NVIDIA is not leaving custom silicon outside the AI factory. It is trying to absorb it into its rack-scale architecture.

2. What Marvell’s Trillion Story Really Means

Marvell’s message is straightforward: AI scaling depends on connectivity. Its Computex 2026 keynote was titled “The Future of AI Scaling Depends on Connectivity.” The company argued that AI infrastructure is expanding beyond a single rack and even beyond a single data center, making data movement across accelerators, servers, racks, campuses and geographically distributed data centers a key bottleneck. (Marvell)

Investment translation:

Shift	Meaning	Investment read
GPU-only to heterogeneous compute	GPUs, custom XPUs, CPUs and DPUs coexist in the rack	Not NVIDIA versus ASIC, but ASICs entering NVIDIA fabric
Compute bottleneck to connectivity bottleneck	Connections between chips and racks matter more	Ethernet, NVLink, optics, SerDes and switch ASICs re-rate
Board-level to package-level bottleneck	Power, signal integrity and packaging drive performance	FC-BGA, high-speed PCB, Si-Cap and MLCC demand expands

[Fact] Marvell reported FY2027 Q1 revenue of $2.418bn, up 28% year over year, and guided FY2027 Q2 revenue to a $2.7bn midpoint. Data-center revenue was $1.833bn, or 76% of total revenue. (Marvell Q1 FY2027)

[Inference] Marvell is being re-rated not because of a few cents of EPS, but because it has unusually pure exposure to the AI data-center connectivity bottleneck. That still does not mean it replaces GPUs. It means custom XPUs and networking have become core pieces of the AI factory system.

3. Why the Broadcom Read-Through Is Positive, but Not Perfect

Broadcom is not the direct protagonist of the Marvell event. But the read-through is strong because Broadcom’s AI franchise sits on the same layers:

custom AI accelerator / XPU
AI networking
Ethernet switch
optical DSP / CPO
SerDes / retimer / PCIe Gen6
3.5D packaging

[Fact] Broadcom reported FY2026 Q1 revenue of $19.311bn, up 29% year over year. Q1 AI revenue was $8.4bn, up 106% year over year, and the company guided Q2 AI semiconductor revenue to $10.7bn. Q2 total revenue guidance was $22.0bn, with adjusted EBITDA margin guidance of about 68%. (Broadcom Q1 FY2026)

The simple math:

Broadcom Q2 AI semiconductor QoQ growth
= 10.7 / 8.4 - 1
= +27.4%

Q2 AI semiconductor share of total revenue guide
= 10.7 / 22.0
= 48.6%

Q1 AI revenue share of total revenue
= 8.4 / 19.311
= 43.5%

[Fact] Broadcom’s Q2 FY2026 earnings call is scheduled for June 3, 2026 at 5:00 p.m. ET, or June 4, 2026 at 6:00 a.m. KST. This note is therefore a pre-call checklist, not a post-earnings verdict. (Broadcom Q2 Event)

What to Listen For

Checkpoint	Strong signal	Korea read-through
Q2 AI semiconductor	Above $10.7bn	Confirms custom XPU / AI ASIC demand
Q3 AI guide	QoQ growth	Supports 2H26 substrate / package demand
AI networking	Ethernet switch, optics, SerDes strength	Isu Petasys, Korea Circuit, Samsung Electro-Mechanics
Tomahawk 6 / CPO	Production volume, lead-time or supply constraint	High-speed PCB, FC-BGA, optical ecosystem
EBITDA margin	Holds near 68%	Custom silicon is not low-margin outsourcing

[Fact] At OFC 2026, Broadcom showcased a gigawatt-scale AI cluster portfolio including 3.5D XPU, 102.4T Ethernet switch with CPO, 400G/lane optical DSP, 200G/lane Ethernet retimers/AEC and PCIe Gen6 switches/retimers. (Broadcom OFC)

[Fact] Broadcom announced that its Tomahawk 6 family switch series is shipping in production volume, and that it enables 128K-XPU networks using only two switch tiers. (Broadcom Tomahawk 6)

The nuance:

The Marvell call-out is category-positive for Broadcom, but not company-specific proof. Marvell is the cleaner NVIDIA NVLink Fusion partner; Broadcom must prove the same connectivity thesis through its own AI revenue, networking and margin numbers.

4. The Common Message: After GPU/HBM Comes Connectivity

Separate Marvell and Broadcom, and you see competition. Put them together, and you see the bottleneck migration.

AI infrastructure layer	Old market focus	New bottleneck emphasized here
Compute	NVIDIA GPU	GPU + custom XPU hybrid architecture
Memory	HBM	HBM + eSSD + memory attach + KV/cache hierarchy
Interconnect	NVLink / Ethernet	Scale-up, scale-out and scale-across connectivity
Package	CoWoS / HBM stack	2.5D/3.5D, FC-BGA, high-layer substrates
Board	Server PCB	AI networking board, low-loss materials, high-speed MLB
Power integrity	Rack power	Package-level Si-Cap, MLCC and PDN components

[Inference] Reading this only as “HBM demand is stronger” misses half the point. HBM is still core. But Marvell and Broadcom are pointing to the components that move data, stabilize power and package chips together.

5. Korea Priority #1: Samsung Electro-Mechanics

Samsung Electro-Mechanics is the cleanest Korean large-cap translation of this read-through.

FC-BGA: custom XPUs, switch ASICs, server CPUs and network chips require large, high-layer, high-reliability package substrates.
MLCC / silicon capacitor: higher AI package power density makes die-near power-integrity components more important.
Customer qualification: package-internal components carry stronger certification barriers than consumer components.

[Fact] Samsung Electro-Mechanics reported Q1 2026 revenue of KRW 3.2091tn and operating profit of KRW 280.6bn. Component Solution revenue was KRW 1.4085tn, up 16% YoY. Package Solution revenue was KRW 725.0bn, up 45% YoY. The company cited MLCCs for AI servers / ADAS and FCBGAs for AI accelerators / server CPUs as growth drivers. (Samsung Electro-Mechanics Q1)

[Fact] Samsung Electro-Mechanics signed a silicon capacitor supply contract worth about KRW 1.5tn with a global large-scale company. The contract runs from January 1, 2027 to December 31, 2028. The company says silicon capacitors are installed inside high-performance packages such as AI-server GPUs and HBM to improve power-supply stability. (Samsung Electro-Mechanics Si-Cap)

So the company frame changes:

Old classification	Better classification after this event
MLCC cyclical component maker	AI server / network power-integrity supplier
Package substrate vendor	FC-BGA platform for custom XPU / AI network ASIC
Smartphone component company	AI package-inside component supplier

But this is not a “buy at any price” conclusion. Samsung Electro-Mechanics has already re-rated sharply. The evidence that matters now is Package Solution margin, AI substrate mix, Si-Cap yield, customer diversification and repeat contract visibility.

6. Higher-Beta Option: Korea Circuit

Korea Circuit has lower scale and lower certainty than Samsung Electro-Mechanics, but it may have higher beta to a Broadcom AI networking surprise.

[Inference] Based on user-provided SK Securities research, Korea Circuit has prior Broadcom communications FC-BGA supply experience, is reportedly undergoing AI data-center ASIC FC-BGA qualification in 1H26, may begin small supply in 2H26, and could see meaningful impact in 2027. P3 FC-BGA plant utilization is a key variable.

The boundary:

Item	Current confidence
Broadcom-related communications FC-BGA experience	Research-based inference
AI data-center ASIC qualification	Not public official fact
2H26 small supply	Speculative until confirmed
2027 earnings impact	Needs evidence

Therefore Korea Circuit is not a confirmed beneficiary. It is a Broadcom AI DC FC-BGA option. If Broadcom’s call shows strong custom XPU and AI networking demand and Korea Circuit later confirms qualification, production or P3 utilization, the re-rating case strengthens.

7. AI Networking PCB Beta: Isu Petasys

Isu Petasys could react quickly because the event language is filled with AI Ethernet, switch ASIC, optical interconnect, 800G/1.6T and CPO/NPO.

But its investment logic is different from Samsung Electro-Mechanics:

Item	Samsung Electro-Mechanics	Isu Petasys
Core exposure	FC-BGA + MLCC + Si-Cap	AI networking PCB / MLB
Confidence	Higher	Needs customer and mix proof
Stock character	Structural proxy + valuation discipline	Theme beta + earnings proof
Key metrics	Package revenue, OPM, Si-Cap	Backlog, high-layer MLB ASP, customer mix

[Inference] Strong Broadcom commentary on Tomahawk 6, AI Ethernet switches and optical DSPs would be positive for high-speed networking PCB names. But “Broadcom is strong, so every PCB is strong” is too broad. Low-loss materials, layer count, yield, customer certification and ASP must show up.

8. Why HBM and Hanmi Are Second-Order Here

This event is not negative for HBM. If custom XPUs grow, HBM demand can diversify beyond NVIDIA GPUs into hyperscaler custom accelerators.

However, the first-order signal from Marvell/Broadcom is connectivity, not HBM.

Layer	Directness to this event
custom XPU / AI networking	Very high
FC-BGA / high-speed substrate	High
Si-Cap / MLCC / power integrity	High
HBM / eSSD	Medium, second-order
TC bonder / advanced package equipment	Medium, order confirmation needed

Hanmi Semiconductor is a valid derivative option, especially because its recent IR notice widened the TC-bonder TAM from HBM into 2.5D package, AI system semiconductors, OSAT and HBF. But equipment names need orders, customer names, delivery timing and margin proof. See the Hanmi IR follow-up.

9. Post-Broadcom Scenarios

Bull Case

Conditions:

Q2 AI semiconductor revenue at or above $10.7bn
Q3 guide shows QoQ growth
Strong AI networking, Tomahawk 6, optical DSP and CPO commentary
Better custom XPU customer or 2027 visibility
Adjusted EBITDA margin holds near 68%

Read:

Samsung Electro-Mechanics is structurally reaffirmed
Korea Circuit can be reframed as a Broadcom FC-BGA option
Isu Petasys can react as AI networking PCB beta
HBM and Hanmi follow as second-order extensions

Base Case

Conditions:

Q2 AI semiconductor revenue near $10.7bn
Q3 grows but does not accelerate
Networking commentary positive but not supply-constrained
Margin stable

Read:

Keep quality exposure centered on Samsung Electro-Mechanics
Avoid chasing Korea Circuit / Isu Petasys without earnings proof
Maintain the existing HBM large-cap thesis

Bear Case

Conditions:

Q2 AI semiconductor below $10.7bn
Q3 guide slows
custom XPU customer ramp delays
AI mix hurts margin
Inventory or customer timing dominates the commentary

Read:

High-beta Korean substrate names may correct first
Samsung Electro-Mechanics remains the structural name, but valuation discipline becomes stricter
Pause the Broadcom/Marvell read-through trade

10. Final View

Do not reduce this to “Marvell went up.” Jensen Huang highlighted Marvell because custom XPU and connectivity are becoming core pieces inside the NVIDIA AI factory.

For Broadcom, this is positive but still needs proof. The June 3 U.S. post-market Q2 call has to confirm AI semiconductor revenue, Q3 growth, AI networking / optics / CPO and margin defense.

For Korea, the message is clearer:

The first-order read-through is not just HBM. It is FC-BGA, high-speed PCB, silicon capacitors, MLCC and power integrity.

Practical priority:

Stance	Korean exposure	Reason
Core watch	Samsung Electro-Mechanics	Cleanest AI package power-integrity proxy
High-beta watch	Korea Circuit	Broadcom AI DC FC-BGA option
Theme beta watch	Isu Petasys	AI networking PCB beta
Secondary	Samsung Electronics / SK Hynix	HBM/eSSD second-order exposure
Derivative	Hanmi Semiconductor	Advanced-packaging equipment orders still need proof

One-line summary:

Marvell’s trillion story should be read in Korea not as “HBM goes higher” alone, but as “AI infrastructure’s connectivity, substrate and power-integrity bottlenecks get more valuable.”

Evidence Classification

[Fact]

Reuters reported Jensen Huang’s Marvell “trillion-dollar company” remark and Marvell’s premarket surge. (Reuters via Investing.com)
Marvell and NVIDIA announced an NVLink Fusion partnership on March 31, 2026. (Marvell 8-K / EX-99.1)
NVIDIA describes NVLink Fusion as a rack-scale AI infrastructure platform for custom CPUs/XPUs and NVLink scale-up architecture. (NVIDIA)
Marvell FY2027 Q1 revenue was $2.418bn, data center was 76% of revenue, and Q2 revenue midpoint guidance was $2.7bn. (Marvell Q1 FY2027)
Broadcom FY2026 Q1 AI revenue was $8.4bn, Q2 AI semiconductor guidance was $10.7bn, Q2 total revenue guidance was $22.0bn, and adjusted EBITDA margin guidance was about 68%. (Broadcom Q1 FY2026)
Broadcom Q2 FY2026 call is scheduled for June 3, 2026 at 5:00 p.m. ET. (Broadcom Q2 Event)
Samsung Electro-Mechanics Q1 2026 Component Solution revenue was KRW 1.4085tn and Package Solution revenue was KRW 725.0bn. (Samsung Electro-Mechanics Q1)
Samsung Electro-Mechanics signed a KRW 1.5tn silicon capacitor supply contract for 2027-2028. (Samsung Electro-Mechanics Si-Cap)

[Inference]

Jensen Huang’s Marvell call-out signals NVIDIA is absorbing custom XPU into the NVLink Fusion ecosystem rather than treating it only as an external threat.
The read-through for Broadcom is category-positive, but Marvell has the cleaner NVIDIA ecosystem angle.
Korea’s first-order read-through is FC-BGA, high-speed PCB, silicon capacitors, MLCC and power integrity.

[Speculation]

Korea Circuit’s AI data-center ASIC FC-BGA qualification and 2H26 supply are research-based assumptions, not public official facts.
Whether Broadcom’s call immediately re-rates Korean substrate names depends on Q3 guide, customer ramp and AI networking commentary.
Samsung Electro-Mechanics’ Si-Cap and FC-BGA mix must still prove sustained margins and repeat orders.

[Blocked]

Broadcom Q2 FY2026 actual results had not been released as of this note’s timestamp: June 2, 2026, 20:45 KST.
Korea Circuit customer names, ASPs and AI DC FC-BGA revenue timing are not public.
Isu Petasys’ direct Broadcom / Marvell customer exposure and AI networking revenue mix are not public.
Samsung Electro-Mechanics’ Si-Cap customer name, product margin, yield and post-2028 repeat contract status remain undisclosed.