Advantest posted record revenue, gross margin and operating margin; its limit is capacity, not orders. Agentic AI makes CPUs a third pillar of test demand and the CEO floated 20,000 units of capacity, but DDR inflation pulls 2H margin back to the low 50s.

Teradyne's Q3 revenue rose 70% YoY, but most of its AI upside isn't coming from GPU test. It's coming from HBM, power chips, hard drives and copper interconnect. The compute dual-sourcing ramp lands in 2H27, and the CPO test TAM is unchanged; the challenge is test time.

LandMark posted a record Q3 with EPS of NT$6.12, but the real story is the denominator: 250–300mW CW lasers for CPO yield 30–50% fewer dies per InP wafer. Capacity is locked by long-term contracts and set to grow 6x in two years, while management says gross margin will not keep climbing.

Corning's optical sales reached $2.38B, up 44% YoY and nearly half of company revenue, with segment net margin at 22.4%. The bigger shift: hyperscalers now prepay cash to fund Corning's fiber capacity, turning fiber into reserved supply.

Celestica grew revenue 83% and guided $40B for 2027, with awarded programs alone putting 2028 higher still. 1.6T overtakes 800G next year but stays mostly pluggable; the first rack-scale CPO program ramps in late 2027. The cost: net cash outflow in 2027.

NVIDIA used prepayments to lock up the CPO back end: SPIL optical-engine capacity 15x and a $1.1B KYEC US test site. No China vendors made the list. But CPO is still validation-scale in 2027, and its buyers are cash-tight neoclouds.

IP Networks grew 43% and AI & Cloud sales more than doubled: Nokia no longer needs to prove demand. It guided 25-30% Optical and IP growth for 2027, but cut its cash outlook to secure supply, and the next round will be decided at a 6-inch InP fab due to start production by year-end.

MaxLinear's pivot is now in the P&L: infrastructure revenue rose 210% and GAAP operating margin swung from -32.7% to +13.9%. At 1.6T, DSP pricing roughly doubles and a TIA rides along, but wafer prepayments pushed operating cash flow negative. The bottleneck is the fab, not demand.

Most thousand-accelerator scale-up domains are several single-hop islands. Google's Virgo backend has no OCS on the slides; OCS stays in TPU ICI and the front end, and between buildings it is still being weighed against patch panels. STT's view: switch radix and plane count set the scale.

At 448G, passive copper is squeezed to about 1 m and moves onto the package, SROI draws the line where optics become necessary beyond 3 m, and slow-and-wide optics bring three ways to cut fiber count. STT's view: 1 to 3 m goes to active copper first, and the key is the gearbox.

Part 1/6 asked whether the laser must sit outside the package; this part asks where it goes and who decides. At OCP, Lumentum, Nokia, Ayar Labs and Wiwynn all talked placement, loss and coolant conditions. STT's view: laser specs are now written by system builders.

Part 1/6 argued buyers want to switch suppliers; this piece asks whether the optical engine can actually be swapped. At OCP, Molex, Samtec and Dell all talked about board footprints, sockets and replaceable units.

The 448G copper bottleneck has moved: Luxshare-Tech's chip-side CPC is resonance-free to 110 GHz, close to PAM4's 112 GHz, but links with first-gen external twinax-to-twinax connectors stay clean only to 100 GHz. First-gen passive copper will likely start on PAM6, under 1 m.

TE's simulations show near-packaged copper loses twice as much as co-packaged copper at 448G, mostly from the first 100 mm next to the ASIC (PCB trace + board connector). The price of CPC is moving the connector into the ASIC package, where heat and skew spend part of the margin again.

Announced thousand-GPU scale-up domains are single-hop at only 48 to 144 GPUs. Eridu and Delos Data propose a 5,120-endpoint multi-rack single-hop design with 30 m of fiber, but it needs a petabit switch 10x larger than today's.

Ciena treats the whole fiber pair as the unit of scale-across bandwidth: multi-rail line systems cut a 20 Pb/s network from 22 amplifier huts to one, and a full-spectrum transponder could move coherent out of the router. Next up: a 1e-24 lossless physical layer.

Litra's OCS energy falls from 8 to 2 pJ/bit, but every point is the same ~3.2 W per port spread over faster lanes. The real news is splitting one fiber's 8 wavelengths across 8 destinations for sparse MoE traffic, if system loss holds near 3 dB.

Quintessent is not selling eight wavelengths but a laser line without InP: GaAs quantum dots bonded on silicon, tested at wafer level. The catch is 25 mW per wavelength, and the real test is the OCI source around OFC 2027.

The third path between copper and optics is really two architectures: Point2 upconverts with RF chips over plastic waveguide, while AttoTude moved to baseband with dielectrically enhanced twinax. At 448G twinax reaches under 1 m; filling 5–30 m favors the passive route.

Google proved OCS cuts cost 30% and power 40%, but 3D MEMS switches in milliseconds. Lumotive and iPronics list microsecond to sub-nanosecond options, and admit the real bottleneck is transceiver relock and the control plane.