Guangdong Juyao Intelligent Technology Co., Ltd. · Laser Software & Process R&D
23°07′N 113°15′E · GUANGZHOU · PURE STATIC MONOGRAPH
← Back to news
Industry2026.10.21· 3 min read

Industry: NVIDIA lines up six supply-chain partners — CPO optical-engine capacity targeted at 15x by mid-2027

The bottleneck for AI chips has long stopped being “can’t design it” — it is “can’t build it, can’t test it fast enough”. NVIDIA laid six supply-chain partners on the table in Berlin at once, and what it quoted was not orders but capacity multiples: 15x for CPO optical engines, 2.5x for substrates, a doubling for testing — with advance payments written straight into the release.

JUYAO editorial concept: a co-packaged optics (CPO) module — an optical-engine chip with glowing fibre ribbons connecting to the main chip package, dark-blue industrial style
COVERJUYAO editorial visual (conceptual illustration, not a vendor photograph).
EXECUTIVE SUMMARY

On 21 October, NVIDIA announced at GTC Berlin an expansion plan with six supply-chain partners (Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron) to meet AI-infrastructure demand. Key figures: with Siliconware Precision Industries (SPIL), optical-engine capacity for co-packaged optics (CPO) is targeted at 15x by mid-2027 and to double again by end-2028; substrate agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron will in aggregate drive a 2.5x capacity increase over three years; Amkor’s interposer wafer capacity rises more than 300% by end-2028; testing capacity doubles from 2026 to 2027, including a $1.1 billion capacity-reservation advance to King Yuan Electronics (KYEC) for new U.S. testing facilities targeting high-volume production in April 2027. “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint,” said Jensen Huang in the release.

01 / 15x optical engines: CPO capacity named in an NVIDIA release for the first time

The line laser people should read first: NVIDIA and SPIL plan to take Taiwan’s optical-engine manufacturing capacity — dedicated to co-packaged optics (CPO) — to 15x by mid-2027, doubling again by end-2028. CPO puts optical transceivers right next to the chip package and is the standard answer where electrical interconnects can’t keep up inside AI clusters; this is the first time it has appeared in an official NVIDIA expansion list as a “capacity multiple”, which says CPO has moved from roadmap to a volume item worth reserving real capacity for.

Mind the wording: 15x is a “target”, not existing capacity; the release discloses neither the baseline absolute output nor per-product allocations — third-party reads (e.g. xenospectrum) also warn against multiplying the figures together to back out GPU shipments. Fine for the trend, discount the absolutes.

02 / Six partners, four fronts: substrates, packaging, testing — optics is one of them

The expansion runs on four fronts. Substrates: agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron total 2.5x over three years. IBIDEN gets a “multibillion-dollar” capacity-reservation advance (agreement signed earlier this year) for its Ono Plant in Japan, targeting 3x with ramp from Q1 2028; Samsung Electro-Mechanics brings its Sejong and Busan upgrades, with NVIDIA-allocated output projected at a “significant portion” of its annual substrate capacity by 2028; Unimicron’s KF substrate line starts adding capacity in Q2 2027 and reaches full capacity by mid-2028.

Advanced packaging and testing: Amkor expands interposer wafer capacity in South Korea by more than 300% by end-2028 (i.e. 4x+); testing capacity doubles from 2026 to 2027 — U.S. (KYEC’s new facilities, $1.1B advance, high-volume April 2027), Singapore (high-volume February 2027) and Taiwan in parallel. The geography is deliberate: substrates in Japan/Korea/Taiwan, packaging in Korea, testing in the U.S./Singapore/Taiwan, optical engines in Taiwan.

03 / “Capacity is the constraint”: the era of buying capacity with advance payments

Huang’s line deserves quoting in full: “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint.” It explains why a release can carry a figure like a $1.1 billion advance payment: when capacity caps growth, buying capacity matters as much as buying chips — the advances lock in shipment certainty for 2027–2028.

The signal for the optical-communications chain is direct: the 15x target for CPO optical engines is the highest multiple on the list. The AI cluster interconnect bottleneck is shifting from “are there optical modules” to “can optical-engine capacity keep up” — and this release is the first to hand the “keeping up” job explicitly to a supply-chain expansion plan. Per NVIDIA’s timeline, mid-2027 is the first checkpoint.

PRIMARY REFERENCES & LINKS

  • NVIDIA official press release: NVIDIA and Supply Chain Partners Expand Capacity to Accelerate AI Infrastructure Buildout
    GlobeNewswire (via Morningstar mirror), published 2026-10-21; vendor press release

    Vendor-stated plan targets; figures not independently verified; contains forward-looking statements

    Open source link ↗