As copper interconnects hit their physical limits, NPO is emerging as the practical lifeline for AI compute clusters. Tencent and Alibaba plan to start 3.2T NPO deployment by late 2026, while Huawei proposes a 7.2T spec aligned with GPU roadmaps.

Silicon photonics' original sin is that it can't emit light. Lasers hate heat, yet CPO asks them to sit next to a hot chip — making the light source the reliability ceiling, where one failed laser can cost an entire optical engine. External vs integrated sources and four light-source types each bet on a different future.

CPO isn't stalled because the components can't be built, but because packaging has switched to an entirely new method. Once the optical engine moves onto the ASIC substrate, photonic packaging must become semiconductor-grade. Whoever standardizes it for foundries and OSATs controls the CPO timeline.

Stacking optics into the chip in 3D wins electrically but pays across the board in thermal management. imec's paper quantifies 2.5D vs 3D CPO thermal crosstalk one number at a time: 8 K vs 60 K, a 13x spatial gradient, a 60x temporal gradient. All 18 figures explained.

Direct-detection optical receivers normally use only one terminal of the PD. NVIDIA recovers the wasted cathode current to build a differential TIA, improving sensitivity by roughly √2. The TIA doubles, but laser power across the whole link drops. A figure-by-figure walkthrough of VLSI 2026 C20.2.

In one day LandMark Optoelectronics locked in materials, orders and tools, and its stock hit limit-up at NT$3,255. But the five board resolutions are three facets of one story: materials, orders and machines, and companies lock these down by contract when they are lifelines, not when they are safe.

Five NVIDIA publications are not isolated research but one vertically integrated weapon, deliberately split into four layers. STT merges them into a single technology map: how far this optical interconnect stack is from volume production, and what shape it bends the supply chain into.

POET's technology direction fits 2026: between the EML shortage and CPO volume production, hybrid integration has real demand. But $2M in annualized revenue backing a $500M framework narrative is a big gap, and the Marvell cancellation is a warning. POET is a PIC springboard for Tier 2 module makers, not the next optical giant.

LandMark's 2Q26 gross margin hit 57% and EPS reached a record NT$4.62, with SiPh CW lasers now over 80% of revenue. The real signal: nearly NT$6 billion of CAPEX over three quarters, all aimed at 2.5x capacity by 2027, with multi-wavelength capability raising both the technical bar and ASPs.

True 3D means vertical stacking; 2.5D just places dies side by side — and that distinction decides who survives to 2030. OpenLight CEO Adam Carter names five gates that must converge (yield, alignment, thermal, testing, reliability) and flags scale-up as 3D's first volume window, 7–10x larger than scale-out.

At OFC 2026, Samsung announced that its 12-inch silicon photonics platform has entered volume production. With ultra-low-loss SiN waveguides and high-performance modulators, it is going after 1.6T optics and the next-generation AI CPO market.

As optics move face-to-face with the ASIC, CPO has become a packaging engineer's game. John Lau's review lays out every PIC/EIC integration approach, from 2D and bridges to 3D stacking and glass substrates, and shows why 51.2T can't be cleared in 2D.

The first bottleneck in silicon photonics isn’t silicon’s inability to emit light — it’s whether designers can get a box of foundry-guaranteed building blocks that will actually yield. That box is the PDK. Why is it so much harder than an electronic PDK, and why is it the foundry’s quietest moat?

Not one of the 14 talks debated whether CPO should happen, and that was the summit's most important signal. TSMC put 3.2 Tb/s on its 2026 roadmap, Advantest pegged test validation for 2027, and ficonTEC expects to ship half of its 25-year equipment total in the next 18 months. Once the debate ends, the winners shift from storytellers to whoever can deliver yield.