OCP APAC 2026: AI Compute Bottleneck Shifts to Networking, Copper vs. Optical Interconnect Debate Intensifies
Nashnova编辑部
At the OCP Asia-Pacific summit, industry leaders said the scaling bottleneck for cloud AI data centers has moved from compute to the network layer — with the copper-versus-optical debate and chip power-density constraints now reshaping the next wave of infrastructure investment.
Why has the bottleneck moved from compute to networking?
SemiAnalysis research director Dan Nishball noted that some end-customers are already down-speccing memory, redirecting the savings into scale-out. This means → networking will claim a larger share of AI server cost going forward.
In plain terms = chips can crunch numbers fast enough; moving data *between* chips is now the slower — and more expensive — part.
This reflects a phase shift: cloud AI is moving from "stack more compute" to "unclog the data pipes," making network transport the supply-chain segment drawing rising investment.
How much longer can copper hold? When does optical take over?
Astera Labs solutions architect Jeffrey Kung judges that 400G is copper's practical ceiling — beyond that, cable reach shrinks too short for data-center layouts. This means → past 400G, a direct switch to optical makes more sense.
Nishball frames the copper-vs-optics replacement debate as "essentially a false dichotomy," arguing copper stays mainstream and optics deploy only where necessary.
The split reflects differing views on transition pace — not *whether* to switch, but *how fast*.
What does the optical-interconnect timeline look like?
2026–2027: Near-package optics (NPO — placing optical modules right next to the chip package) expected to begin ramping.
2026–2029: Co-packaged optics (CPO — integrating optical modules *inside* the chip package) platforms mature and enter volume production.
Post-2030: On-chip photonics — transmitting data with light directly inside the chip — and all-optical architectures expected to advance further.
Why has chip power density become a hard constraint?
Applied Materials VP Subi Kengeri warned in his keynote: a modern full-reticle AI chip now draws roughly 3,600 watts — about 4 W per square millimeter on average, jumping to 8–10 W/mm² at peak.
Building a data center requires about $50 million per megawatt in upfront capital, depreciated over four to five years, and thermal management alone accounts for nearly 40% of operating expenses.
In plain terms = the stronger the chip, the costlier the cooling; cooling spend is catching up with the chip itself — so "energy-efficient compute" is a balance-sheet issue, not a slogan.
What is Applied Materials betting on with "system-technology co-optimization"?
Applied Materials advocates an STCO approach — system-technology co-optimization, improving the full stack from physics to software together. Key pillars: gate-all-around transistors + backside power delivery for advanced logic, 3D DRAM and high-stack HBM for near-memory compute, and advanced packaging.
A critical inflection in packaging: the shift from traditional micro-bumps to hybrid bonding — joining two chips directly, face-to-face — can raise interconnect density from thousands of I/Os per mm² to nearly one million.
Over the past two decades, packaging complexity — measured by die count, substrate size, silicon area, and transistor density — has grown 500,000 to 600,000 times.
How fast can the lab-to-fab cycle be compressed?
Applied Materials has committed $5 billion to its EPIC center (Equipment and Process Innovation Center), set to open this fall, bringing equipment makers, EDA developers, universities, and chipmakers under one roof.
The goal: compress a traditional 10-to-15-year lab-to-volume-production cycle. This means → if it works, the path from research paper to factory floor could shrink by more than half.
Whether the networking and packaging tracks can accelerate in sync within this co-optimization framework will be the key proof point for the next phase of AI infrastructure competition.
Content is for reference only, not financial advice.