Optical Interconnect Advances on Multiple Fronts as Korean Equipment Makers Capture HBM4 Dividends
Miles Bennett
AI compute expansion is reshaping two sectors at once: optical interconnects have splintered into five coexisting technology tracks, while Korean semiconductor equipment makers are riding a real order wave from accelerating HBM4 investment — in both cases, the beneficiary pool is wider than the market expected, with no single winner.
Why isn't one optical-interconnect technology winning outright?
Once single-node compute crosses the exascale threshold, traditional copper-based electrical interconnects hit bandwidth, latency, and power walls simultaneously — optical interconnects become unavoidable.
But optical interconnect is not a monolith. Distance, power, and cost requirements vary so widely across use cases that no single approach covers them all.
This means → the industry logic is "use case dictates technology choice" — multiple tracks coexist and divide labor long-term, rather than one displacing the rest.
LPO, LRO, FRO — how do the three pluggable routes compare?
LPO — linear-drive optics that strip out the DSP chip entirely — delivers the lowest power: Macom data shows 800G multimode module power dropping from over 13 W to under 4 W, but it only works within 500-meter short-reach links.
LRO — a half-retiming compromise that keeps partial signal repair — is the more pragmatic middle ground: an IEEE March 2026 report shows LRO holding power below 20 W at 1.6 Tbps, versus over 30 W for full-DSP, while still allowing air cooling.
In plain terms = LPO saves power but can't go far; LRO costs a bit more power but covers more ground — which is why nearly every company demoing 1.6T LPO at OFC 2025 also showed an LRO product as a complement.
Silicon photonics, CPO, NPO, TFLN — where do the longer-range bets stand?
Silicon photonics — a platform technology that moves data on chips using light instead of electricity — keeps gaining share: LightCounting projects 2026 as the first year silicon-photonic transceiver revenue crosses 50 % of the $4 bn+ total market; Yole forecasts the segment growing from $278 M in 2024 to roughly $2.7 bn by 2030, a 46 % CAGR.
CPO — co-packaged optics, embedding the optical engine directly inside the switch chip — is viewed as the endgame. Broadcom has shipped the first 51.2 Tbps CPO Ethernet switch; Nvidia plans to deliver a CPO InfiniBand system in H1 2026. LightCounting sees the CPO market reaching $10 bn by 2030; Coherent revised that up to $15 bn. But a single optical engine costs $35,000–40,000, and the non-pluggable design creates real maintenance challenges.
NPO — near-package optics, a bridge between pluggable and CPO — is the primary path chosen by Chinese GPU chip makers; global market valued at $3.8 bn in 2025, with a 19.3 % CAGR through 2034. TFLN — thin-film lithium niobate, a new electro-optic modulator material — carves out its own lane in high-speed modules: a 1.6T-DR8 transceiver built on TFLN draws just 20 W, 20 % less than conventional designs. Zheshang Securities sees TFLN not as a disruptive replacement but as a critical complement filling gaps in legacy materials.
Why are Korean equipment makers suddenly landing big contracts?
The driver is straightforward: Samsung and SK Hynix are accelerating HBM4 and advanced-DRAM buildouts in H1 2026, pulling equipment procurement forward across the board.
The beneficiary pool spans front-end process tools, advanced packaging equipment, and test gear — not one segment eating alone, but the entire chain benefiting in parallel.
This means → Korea's equipment supply chain is pivoting from dependence on the traditional DRAM cycle to deep alignment with HBM4, advanced packaging, and AI infrastructure capex.
How strong are the orders at three bellwether equipment firms?
TES (front-end process tools): Q1 revenue hit KRW 97.2 bn (~$64 M), operating profit KRW 22.2 bn — up 73.4 % QoQ; signed roughly KRW 42.7 bn in equipment supply contracts with SK Hynix in May–June. FnGuide forecasts Q2 revenue rising to KRW 109.9 bn.
Hanmi Semiconductor (thermocompression bonders for chip stacking): Q1 was pressured by scheduling delays, but Hanmi then signed a KRW 44.2 bn supply contract with SK Hynix, with deliveries extending into early September. FnGuide projects Q2 revenue jumping to KRW 227.6 bn, operating profit KRW 110.3 bn.
Techwing (test equipment): Q1 revenue KRW 52.4 bn, up 51.4 % YoY; its Cube Prober is purpose-built for testing HBM stacked structures and is considered a key yield-improvement tool. Techwing signed a KRW 9.7 bn contract with Samsung in May and has already been shipping to Micron's Malaysia fab.
How long can this equipment order wave last?
SEMI forecasts AI demand will sustain global wafer-fab equipment sales through 2026–2027.
The key variable lands in H2: whether Samsung, SK Hynix, and Micron further expand HBM4 capex will determine the depth and durability of this order cycle.
In plain terms = the H1 orders are already booked; it is the H2 expansion decisions that will determine whether equipment makers enjoy a one-time wave or a sustained upcycle.
Content is for reference only, not financial advice.