TSMC Is Mass-Producing Light Instead of Copper to Wire Together AI's Newest Chips. Nvidia Already Bet $4 Billion That the Rest of the Industry Can Keep Up.

Terry Tsao, who oversees SEMI's global operations, opened this week's SEMICON Taiwan conference with a warning that had nothing to do with transistor counts. Data movement inside an AI cluster, he said, can now consume more energy than the computation itself. After a decade of chipmakers competing on how fast a processor can think, the industry's newest constraint is how fast it can talk to itself.

The number behind that shift is straightforward. Copper wiring, the material every server rack has used to connect chips for decades, starts losing signal integrity above roughly 100 terabits per second per node, right around the speed AI accelerator clusters are now pushing. Electrical signals need power-hungry equalization circuits to survive that distance over copper. Light doesn't.

[[IMG1]]

What Changed

Co-packaged optics, the practice of building the light-based interconnect directly into a chip's package rather than bolting it on later, moved from lab demonstrations to commercial production lines this year. TSMC's COUPE platform is now in mass production, and the company is targeting a substrate-level version, one that integrates optics even deeper into the chip package, for the second half of 2026. Foxconn expects to start shipping CPO-based switches in the third quarter. Taiwan's ShunSin Technology has reached mass production on modules running at 51.2 and 102.4 terabits per second. UMC produced its first mass volume of silicon photonics wafers at a facility in Singapore, and GlobalWafers has moved select silicon photonics products into volume production as well.

Powertech Technology, meanwhile, is partnering with Broadcom on low-volume CPO production this year, with full integration planned for 2027. Broadcom, whose cost of insuring its own debt hit a record in late August as it races the same AI buildout, is positioning itself alongside Nvidia as one of the two companies most likely to define whose optical architecture becomes the industry standard.

Nvidia isn't waiting to see how that contest plays out. The company has committed $4 billion, split evenly between Lumentum and Coherent, for multi-year procurement of the lasers and optical components CPO systems depend on. That's a direct hedge against the kind of supply squeeze Nvidia has already lived through with CoWoS advanced packaging and high-bandwidth memory, both of which constrained GPU shipments earlier in this AI cycle.

Why It Matters

Two separate bottlenecks are forming underneath the optics transition, and only one of them is getting attention. The advanced packaging capacity TSMC and its peers are racing to build for CPO is the visible one. Less visible is a shortage brewing in traditional wire bonding, the older packaging method used for the CPUs, networking chips, power management ICs, and other peripheral silicon that surround every AI accelerator. Trade publication Digitimes reports that outsourced assembly and test (OSAT) capacity has tightened steadily since late 2025, with new capacity added in 2026 filled almost as soon as it comes online and order visibility for some IC design customers now pushed beyond 2027, driven by the sheer volume of AI servers being built rather than by any single flashy component.

Advanced substrates carry a similar warning. AI GPUs and ASICs require five to ten times more advanced ABF substrate material than a traditional CPU does, because the packages are larger and carry more layers. Analysts at TrendForce have flagged a potential shortage of high-end ABF substrates in 2027, the same year the wire bonding gap is expected to widen. If either constraint proves real, it would cap how fast AI accelerators can actually ship, regardless of how much compute capacity TSMC and its foundry peers build.

The investment case is already ahead of the production numbers. SEMI projects global 300mm fab equipment spending will climb from $133 billion this year to $151 billion in 2027, the first year that figure has crossed $150 billion, and to $172 billion by 2029. Total semiconductor equipment billings hit $135.1 billion in 2025, up 15% from the year before, the base the current spending forecast builds on. Lumentum shares have gained roughly 140% in 2026, a run that a Barclays note extended again in late August even without a fresh company-specific catalyst. Coherent's run has been choppier, up more than 80% for the year in mid-August before giving back much of that gain to sit closer to 43% higher now, evidence that conviction varies stock by stock even inside the same trade. Holders chasing Lumentum's momentum are betting the production ramp described above actually lands on schedule.

What to Watch Next

TSMC's substrate-level COUPE rollout in the second half of this year is the nearest checkpoint. If it slips, expect the optics trade to cool the way it did briefly in early August, when Coherent and Lumentum both pulled back sharply ahead of earnings before resuming their climb. Foxconn's third-quarter CPO switch shipments are the next concrete data point after that, followed by whatever capacity commitments emerge from the wire bonding and ABF substrate squeezes analysts are now flagging for 2027.

The Pulse24 Take

The AI trade has spent two years fixated on how much compute Nvidia can ship and how fast hyperscalers can build data centers to house it. This shift argues that the more binding constraint sits somewhere less glamorous: the wiring, the packaging, the substrate. None of those show up in a keynote slide about model performance, and all of them can slow a buildout that looks unstoppable on a chip roadmap.

Nvidia's willingness to pre-pay $4 billion for laser supply is itself worth reading as a signal. Companies don't lock in multi-year procurement at that scale for a component they expect to stay abundant. The optics stocks pricing in years of demand growth may be right about the direction, but names that have already climbed anywhere from roughly 40% to well over 100% on a production ramp still in its first year carry real room for disappointment if wire bonding, substrates, or the pace of TSMC's own rollout run behind schedule.