ARM's Rally — The Next AI Bottleneck Is CPU Orchestration, Optical Links and Power Integrity

Fri, 22 May 2026 21:40:00 +0900

Related series: NVIDIA Q1 FY27 and Korean AI infrastructure / Samsung Electro-Mechanics’ KRW 1.5T silicon-capacitor contract / Understanding MLCCs and silicon capacitors / Samsung Electronics rerating thesis

ARM’s rally is not simply a statement that “CPUs matter again.” More precisely, AI servers are being redefined from GPU boxes into systems that combine CPUs, XPUs, HBM, optical links, power-stabilizing components, high-speed substrates and test infrastructure. ARM is the visible price reaction. But at the current valuation, a large part of the five-year success case is already embedded. For Korean investors, the more useful signal is the next bottleneck: Marvell’s custom silicon and optical links, Samsung Electro-Mechanics’ silicon capacitors, SK hynix and Samsung Electronics memory, and the high-speed substrate / test-socket layer.

Key Takeaways

ARM’s rally is a reclassification event: from mobile IP royalty company toward AI data-center CPU platform company. As AI agents run continuously, call tools, move data and coordinate GPUs, ASICs and memory, the CPU becomes the orchestration layer of the AI rack.

The external catalyst was NVIDIA’s Q1 FY27. NVIDIA reported $81.6B in revenue, $75.2B in Data Center revenue, and Q2 guidance of $91.0B ±2%. The market read that as acceleration, not peak-out, and began repricing the layers below GPUs: CPU, memory, interconnect, power and substrates. (NVIDIA)

ARM’s own story is strong. FY26 revenue was $4.92B, royalty revenue $2.61B, licensing revenue $2.31B, and non-GAAP EPS $1.77. The Arm AGI CPU is positioned as a data-center CPU platform with Meta as lead partner, and Arm says FY27-FY28 customer demand is already above $2B. (Arm)

The problem is price. Research OS local DB shows ARM closed at $298.23 on May 21, 2026. Against FY26 non-GAAP EPS of $1.77, that is roughly 168.5x earnings. Using a market capitalization around the low-$300B range and FY26 revenue of $4.92B, the stock trades around the 60x sales zone. The thesis is real; the stock is not cheap.

The practical alpha is therefore not ARM itself, but the bottlenecks ARM’s move points toward: Marvell’s custom silicon and optical fabric, Samsung Electro-Mechanics’ silicon capacitor, SK hynix / Samsung Electronics HBM and DRAM, high-speed substrates and test sockets.

The one-line conclusion: ARM is the signal. The alpha sits in the bottlenecks.

1. What ARM’s Rally Really Means

Calling this a “CPU renaissance” is only half right. The deeper message is that the constraint is moving.

In 2023-2024, the core AI infrastructure bottleneck was GPUs. In 2025, HBM and advanced packaging rose to the front. In 2026 and beyond, as AI racks become larger systems, CPU orchestration, memory movement, optical links, power integrity, substrates and testing all become limiting factors.

GPU scarcity
→ HBM / packaging scarcity
→ CPU orchestration / memory movement / optical interconnect bottlenecks
→ package-level power integrity / high-speed substrate / test bottlenecks

AI agents are not ordinary web requests. They call models, use tools, read documents, query databases and store intermediate outputs. GPUs handle compute, but CPUs coordinate task order, memory management, network I/O, security, error handling and system operation.

This makes the CPU the control plane of the AI rack. If the GPU is the engine, the CPU is the traffic controller. ARM’s rally reflects the market starting to price that role.

Arm’s FY26 results support the narrative. The company reported FY26 revenue of $4.92B and Q4 revenue of $1.49B, with record annual royalty and licensing revenue. Data-center royalties more than doubled year over year. (Arm)

But the right investment question is not “is ARM’s thesis right?” It is “does the current price still leave enough risk-reward?”

2. Arm AGI CPU: From IP Vendor to Silicon Platform

Arm has put the Arm AGI CPU at the center of the story. It describes the product as production silicon for agentic AI data centers, with Meta as lead partner. Arm also says commercial systems are available to order through ASRock, Lenovo, Quanta and Supermicro. (Arm)

The old Arm model was:

IP licensing
→ customers design and manufacture chips
→ Arm receives license fees and royalties

The AGI CPU extends that model:

IP + compute subsystems + production silicon
→ customers deploy the Arm platform faster at data-center scale
→ Arm gains more platform control

This is powerful, but it changes the risk profile. Arm is no longer only a neutral blueprint provider. As it sells its own silicon, customers and licensees may begin to see it as a potential competitor.

That is why regulatory risk matters. Bloomberg has reported that the FTC is looking into whether Arm could restrict or degrade license access while expanding its own chip ambitions. (Bloomberg)

The AGI CPU is a long-term call option, but it also makes Arm heavier, more complex and more exposed to channel conflict.

3. Valuation: Correct Thesis, Expensive Stock

The uncomfortable part of the ARM rally is the math. Directionally, the thesis makes sense. Valuation already discounts a long runway.

Research OS local DB shows ARM closed at $298.23 on May 21, 2026. Against FY26 non-GAAP EPS of $1.77:

FY26 non-GAAP P/E
= 298.23 / 1.77
= about 168.5x

Against FY26 revenue of $4.92B and a market cap in the low-$300B range:

FY26 P/S
= roughly low-$300B / $4.92B
= about mid-60x sales

This is not a price for “good current earnings.” It is a price for a 2030-2031 scenario: higher data-center CPU penetration, AGI CPU revenue, higher royalty rates and Armv9 / Neoverse expansion.

If Arm reaches something like $25B of revenue and $9+ of EPS by FY31, today’s price can be explained. But that is a five-year success case. Investors need to separate a correct story from a good entry.

Our view is Avoid at current price / Wait. The company is not the problem. The entry price is.

4. Better Alpha 1: Marvell and the Connectivity Bottleneck

If ARM represents CPU orchestration, Marvell represents custom compute plus optical connectivity plus switching. This may be the more actionable derivative idea.

Marvell reported FY26 revenue of $8.195B and non-GAAP EPS of $2.84. It expects FY27 revenue to approach $11B, with data center leading growth and interconnect revenue growing more than 50%. (Marvell)

Marvell is not just a networking chip company.

Layer	Meaning
Custom AI silicon	Design partner as hyperscalers build their own XPUs
Optical interconnect	Next bottleneck for intra-rack and inter-rack data movement
PCIe / CXL switching	Connects CPU, GPU, ASIC and memory pools
NVLink Fusion partnership	Strategic position inside heterogeneous NVIDIA-connected AI infrastructure

The Celestial AI acquisition adds a photonic scale-up platform. Marvell has said it expects the business to reach a $500M annualized run-rate by Q4 FY28 and a $1B annualized run-rate by Q4 FY29, if milestones are met. (Marvell Celestial)

The caveat: Marvell has already rallied. Research OS local DB shows MRVL closed at $190.69 on May 21, 2026. At this level, the May 27 earnings report needs to confirm the FY27 revenue path, data-center demand, custom silicon and optical interconnect momentum.

Our view: Wait / Buy on pullback.

5. Better Alpha 2: Samsung Electro-Mechanics and Power Integrity

In Korea, the clearest second-order bottleneck is Samsung Electro-Mechanics.

On May 20, 2026, Samsung Electro-Mechanics announced a silicon-capacitor supply contract worth approximately KRW 1.5T over two years, from January 1, 2027 to December 31, 2028. The company explained that silicon capacitors are installed inside high-performance semiconductor packages such as AI server GPUs and HBM to improve power stability. (Samsung Electro-Mechanics)

This is the same direction as ARM’s rally. As AI racks become systems rather than GPU boxes, CPU, GPU, HBM and ASIC packages become denser and draw power faster. Voltage noise and power fluctuations become a bottleneck.

Samsung Electro-Mechanics’ thesis is not “MLCC demand rises.” It is:

Old SEMCO:
MLCC + camera modules + FC-BGA

New SEMCO:
AI package power-integrity components
+ FC-BGA
+ AI-grade MLCC

A silicon capacitor does not replace all MLCCs. It is a high-performance complement used inside or very close to AI GPU / HBM packages. This can reclassify Samsung Electro-Mechanics from a passive-parts cycle name into an AI package power-integrity supplier.

The risks are clear: the customer is undisclosed, 2026 earnings impact is limited, margins are not disclosed, and the stock has already reacted. Our view remains Watchlist → Buy on pullback.

6. Better Alpha 3: SK hynix and Samsung Electronics Memory

ARM’s rally is positive for Korean memory as well. More CPU orchestration means more CPU-side memory, more data movement to HBM, more LPDDR / DDR / CXL memory pooling, and more server DRAM.

GPU-centric server
→ GPU + CPU + DPU + NIC + switch ASIC + HBM + DRAM
→ higher rack-level memory capacity and bandwidth

SK hynix is the cleanest HBM and server DRAM beneficiary, but it is also a crowded winner. The view is Hold / Pullback Buy.

Samsung Electronics is different. It is the catch-up option: HBM4, SOCAMM2, server DRAM, eSSD, HBM4 base die and foundry optionality. Since Arm AGI CPU manufacturing is tied to TSMC 3nm, one should not overstate direct Samsung Foundry exposure. But AI rack diversification still improves the value of Samsung’s memory and packaging ecosystem.

Company	Role	View
SK hynix	Proven HBM winner	Hold / Pullback Buy
Samsung Electronics	HBM catch-up + broad IDM option	Watchlist / Buy on confirmation

Samsung needs evidence: HBM4 qualification, volume, margins and foundry loss reduction.

7. Better Alpha 4: High-Speed Substrates and Test Sockets

As AI racks evolve from single GPU boards into dense systems, substrates and testing become second-order winners.

High-speed substrates and PCBs. More CPUs, GPUs, ASICs, HBM, NICs and switch ASICs mean more high-speed signaling and more complex board design. Isu Petasys, Daeduck Electronics, TLB, Korea Circuit and Simmtech are candidates. But direct orders, customer qualification, ASP and margin must be verified.

Test sockets. As chips become faster, hotter and more pin-dense, sockets and probe-card requirements rise. ISC, LEENO Industrial and TSE are relevant candidates. This layer can carry higher margins and less capex cyclicality than substrates, but product mix matters.

Layer	Candidates	What to Verify
High-speed PCB	Isu Petasys, Daeduck, TLB, Korea Circuit, Simmtech	AI-server revenue, layer count, ASP, OPM
Test sockets	ISC, LEENO, TSE	ASIC/HBM/CXL customers, product mix, new socket revenue
Package substrate	Samsung Electro-Mechanics, Daeduck, Korea Circuit	FC-BGA utilization, qualification, margins

The rule: do not buy the phrase “AI exposure.” Buy verified customers, orders and margins.

8. Strategy

Name / Domain	View	Reason
ARM	Avoid at current price / Wait	Correct thesis, but very rich valuation
Marvell	Wait / Buy on pullback	Custom silicon + optical + switching bottleneck
Samsung Electro-Mechanics	Watchlist → Buy on pullback	Silicon capacitor can reclassify it as AI package power-integrity supplier
Samsung Electronics	Watchlist	HBM4, SOCAMM, server DRAM, eSSD, foundry option
SK hynix	Hold / Pullback Buy	Cleanest HBM winner, but crowded
High-speed substrates	Watchlist	Beneficiary, but verify customers and margins
Test sockets	Watchlist	Chip-complexity consumables beta

Catalysts

Marvell earnings on May 27, 2026.
Arm AGI CPU supply expansion and conversion from demand to revenue.
NVIDIA Vera Rubin second-half ramp.
Samsung Electro-Mechanics follow-on silicon-capacitor customers or capacity announcements.
Samsung Electronics HBM4 qualification.
Korean substrate / test-socket companies separating AI revenue in quarterly results.

Invalidation

Arm AGI CPU demand fails to convert because of supply, margins or channel conflict.
FTC / regulatory risk damages Arm’s licensing model.
Marvell loses custom design wins to Broadcom or optical revenue ramps too slowly.
Samsung Electro-Mechanics’ silicon-capacitor contract carries weak margins or lacks follow-on customers.
Korean substrate / test names report AI revenue growth without OPM support.

Final Line

ARM’s rally is a correct revaluation signal. AI servers are no longer GPU boxes. They are CPU, XPU, HBM, optical, power-integrity, substrate and test systems.

But buying ARM after the move may confuse the signal with the asset. The better question is: which bottleneck has not yet been fully priced?

Our answer: ARM is the signal. The alpha sits in the bottlenecks.

Evidence Map

[Fact]

NVIDIA Q1 FY27 revenue was $81.6B, Data Center revenue was $75.2B, and Q2 revenue guidance was $91.0B ±2%. (NVIDIA)
Arm FY26 revenue was $4.92B, royalty revenue $2.61B and licensing revenue $2.31B. Arm said AGI CPU customer demand for FY27-FY28 was above $2B. (Arm)
Marvell FY26 revenue was $8.195B, non-GAAP EPS was $2.84, and the company expects FY27 revenue approaching $11B. (Marvell)
Samsung Electro-Mechanics announced a KRW 1.5T silicon-capacitor supply contract on May 20, 2026. (Samsung Electro-Mechanics)
Research OS local DB shows ARM closed at $298.23 and MRVL at $190.69 on May 21, 2026.

[Inference]

ARM’s rally signals a bottleneck shift from GPUs toward CPU orchestration, memory movement, optical interconnect and power integrity.
The better risk-reward may sit in Marvell, Samsung Electro-Mechanics, Korean memory, high-speed substrates and test sockets rather than ARM itself.

[Blocked]

Arm AGI CPU customer-level orders and margins.
Marvell customer-level custom silicon revenue.
Samsung Electro-Mechanics silicon-capacitor customer, ASP, margin and exact package position.
Direct ARM/NVIDIA/Marvell exposure for Korean substrate and test names.

This article is research commentary, not investment advice. US price data is based on Research OS local DB closing prices for May 21, 2026. Company data is based on official NVIDIA, Arm, Marvell and Samsung Electro-Mechanics materials. Scenario views are analyst frameworks and may be wrong. Data cut: May 22, 2026 KST.

ARM on Korea Invest Insights