Nvidia and SPIL Plan 15x Optical-Engine Capacity by Mid-2027

Nvidia confirmed plans with Siliconware Precision Industries (SPIL) to expand manufacturing capacity 15 times for optical engines used in co-packaged optics by mid-2027, SDxCentral reported from Nvidia’s GTC event in Berlin on 21 October 2026. Nvidia said it aims “for another doubling” by the end of 2028, according to the same report.

The capacity plan arrived alongside research on the next step. At SEMICON West 2026, Nvidia research scientist Nikola Nedovic presented a prototype called Optics-on-Interposer (OOI), which moves the optical engine onto the silicon interposer that carries the chip, TechNews reported on 20 October 2026. A post by Daniel Romero (@HyperTechInvest) on X on 25 October 2026 drew the two together.

Business Pill · LEAD TIME

A short explainer of lead time: why the wait is part of the price. It teaches the general idea only and says nothing about any company in this story.

What Nvidia Said in Berlin

SDxCentral quoted Ian Buck, Nvidia’s VP and GM of hyperscale and high-performance computing, in a press briefing: “Our new Spectrum-6 CPO switch is the building block for scale-out networking for Vera Rubin DSX AI factories. We’re very excited about bringing co-packaged optics to the AI data center and working with SPIL and others to expand our capacity.”

The same report quoted Nvidia CEO Jensen Huang: “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint.” SDxCentral said the GTC Berlin plans also named Amkor, Samsung Electro-Mechanics and Unimicron, and that Nvidia signed supply deals in 2026 with optics firms including Corning, Lumentum and Coherent.

SDxCentral said Nvidia describes the optical version of its Spectrum-6 switch as packing 409.6 Tb/s of bandwidth across 512 ports at 800 Gb/s each.

What Was Posted on X

@HyperTechInvest posted on 25 Oct 2026 that Nvidia unveiled Optics-on-Interposer (OOI), demonstrated DWDM at 256 Gb/s across eight wavelengths over one fiber, and that SPIL plans to increase optical-engine capacity 15-fold by mid-2027. The post says Nvidia has not confirmed whether Feynman will use OOI.

The Step Past Co-Packaged Optics

Co-packaged optics puts the optical engine in the same package as the switch or GPU. OOI goes one step closer: according to TechNews, it moves the optical engine from the organic substrate onto the silicon interposer, keeping electrical signals inside the interposer’s shorter, denser wiring. Nedovic proposed the idea in a 2023 paper, “Beyond CPO: A Motivation and Approach for Bringing Optics onto the Silicon Interposer,” which is listed on Nvidia Research’s site.

TechNews said an Nvidia comparison of a 200T ASIC package put the optical components across 100 mm on an organic substrate and inside 8 mm on the interposer. It reported a research-stage dense wavelength-division multiplexing (DWDM) demonstration of 8 wavelengths at 32 Gb/s each, or 256 Gb/s on one fiber path. Nedovic said Nvidia is proposing and evaluating new technologies for future products while working through the obstacles to commercialising them, TechNews reported.

The slides attached to Romero’s post come from two earlier conference talks. One is an OFC 2026 paper by Nvidia researchers, “A 256 Gb/s DWDM Optical I/O in a 3D-stacked EIC/PIC Silicon Photonics Platform,” dated 16 March 2026, whose slides say the platform is “Based on TSMC COUPE.” The other is Nedovic’s ISSCC 2026 forum talk, with a slide stating “Interposer integration removes bandwidth bottleneck between ASIC and OE.”

Sources: as listed in each row. OOI and the DWDM result are research-stage.

Why Optics Capacity Is the Bottleneck

TrendForce said in July 2026 that Nvidia had begun shipping Spectrum-X CPO switches to select partners, and that the supply of optical engines, silicon photonics chips and advanced packaging is expected to be the main factor setting the pace of expansion.

SemiAnalysis’s Dylan Patel put optics third on his list of chip-supply bottlenecks, after substrates and memory, in a panel transcribed by Fireside Alpha on 24 October 2026: “even if everyone wanted to do CPO in 2028, or even one major company wanted to do CPO in 2028, they could not satisfy their volumes in terms of lasers, in terms of fiber attach, in terms of all these different components that go into it.”

Where It Meets Our Coverage

Our Lumentum piece reported CEO Michael Hurlston telling Bloomberg Television on 9 October 2026 that Lumentum is “almost sold out for 2029.” Our Patel piece covered his estimate of what a gigawatt of AI capacity costs. Our power piece tracked the other constraint: grid capacity for new data centres.

The Structural Read

As we read it, the Berlin plan is about supply: SDxCentral reported 15 times the optical-engine capacity with SPIL by mid-2027 and “another doubling” by the end of 2028.

As we read it, the SEMICON West talk is about the step after that: TechNews reported optics moving from the organic substrate onto the silicon interposer, from 100 mm to 8 mm beside a 200T ASIC.

As we read it, the constraint is named by others too: TrendForce said optical engines, silicon photonics chips and advanced packaging would set the pace, and Patel put optics third on his bottleneck list.

Three Implications

FOR PACKAGING SUPPLIERS SDxCentral said Nvidia’s GTC Berlin plans named SPIL, Amkor, Samsung Electro-Mechanics and Unimicron.

FOR OPTICS MAKERS SDxCentral said Nvidia signed supply deals in 2026 with Corning, Lumentum and Coherent; Lumentum’s CEO said in October the company is “almost sold out for 2029.”

FOR FUTURE CHIP PACKAGES TechNews reported that Nvidia presented OOI as research and that Nedovic said Nvidia is evaluating new technologies for future products.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Nine Layers of AI.

The framework puts it this way: the AI economy is “a layered industrial stack” and “at any given moment, the binding constraint sits at a different layer than it did six months ago.”

As we read it, Nvidia’s optics plan and its interposer research target optics and packaging, which TrendForce and Patel list among the current constraints. The framework was published in May 2026 and does not discuss Nvidia’s plan.

What Is Not Established

The 15 times figure and the end-2028 doubling are from SDxCentral’s report of Nvidia’s GTC Berlin briefing; we read them there, not in an Nvidia or SPIL release, and SDxCentral does not give the starting capacity. OOI and the 256 Gb/s DWDM result are research results; Romero’s post says Nvidia has not confirmed whether its Feynman architecture will use OOI. The SEMICON West figures come from TechNews’s report, not from slides we read. We did not contact Nvidia, SPIL or TSMC.

The Bottom Line

Nvidia told SDxCentral it plans 15 times more optical-engine capacity with SPIL by mid-2027 and another doubling by the end of 2028. At the same time, its researchers are showing optics that sit on the silicon interposer itself, in 8 mm rather than 100 mm, according to TechNews.