Goldman Sachs is saying the AI networking market is set to explode from $15 billion to $154 billion, and Nvidia just splashed $6 billion in a month on optical firms—what does "copper is history, optics are here" really mean?
At the end of April, Goldman Sachs dropped a major report: (Optical Networking: The Next Mega Trend in AI Infrastructure).
Key takeaway: The TAM (Total Addressable Market) for AI network interconnects is projected to skyrocket from $15 billion to $154 billion, a nearly 9x increase. The CPO (Co-Packaged Optics) segment is set to contribute $91 billion, accounting for 59% of the total.
Almost simultaneously, Nvidia splashed $6 billion on three optical companies within a month: $2 billion into Lumentum, $2 billion into Coherent, and $2 billion into Marvell.
When the world's largest AI chip company starts frantically acquiring optical assets, and when Wall Street's most influential investment bank defines optical networks as a 'super trend'—this signal cannot be ignored.
Today, let's break down this track from start to finish.
One, why can’t copper wires keep up?
Let’s start with an intuition: AI training GPU clusters are getting larger. Nvidia's currently mass-produced GB300 NVL72 squeezes 72 GPUs into one rack, while the next-generation Rubin Ultra NVL576 aims to connect 576 GPUs across 8 racks into a supernode.
Data must be transmitted between GPUs. You can think of GPUs as workers in a factory; after each worker finishes their part, they need to pass the results to the next worker to compile before continuing to the next step. The more workers there are, the greater the demand for data transmission between them.
Currently, data transfer between GPUs mainly relies on copper wires (electrical signals). However, copper wires have three physical limits:
First, the bandwidth ceiling. Copper wires have significant attenuation when transmitting high-frequency signals; the longer the distance, the slower the speed. It's alright for short distances within racks, but it struggles across racks.
Second, power consumption is exploding. The power consumption of copper wire transmission increases dramatically with speed and distance. If a NVL576 supernode used all copper interconnects, the electricity costs for the optical network portion would be astronomical.
Third, density limitations. Copper cables are thick and heavy; when you need to pull thousands of connections between 8 racks, physical space simply cannot accommodate it.
The solution is to replace copper with light—light signal transmission is faster, consumes less power (Nvidia claims a 5x efficiency improvement), can transmit over longer distances, and the cables are thinner and lighter.
That's why Goldman says, 'copper is a thing of the past; optics are here.'
Two, how is the $154 billion market calculated?
Goldman Sachs provided a very specific set of data:
From GB300 NVL72 (currently in mass production) to Rubin Ultra NVL576 (shipping in 2027-2028), the network interconnect value of each computing unit skyrockets from $315,000 to $9.4 billion—a 29-fold increase.
A Rubin Ultra NVL576 supernode requires: 324 optical engines, 162 external laser sources, 5184 fibers, and MPO connectors. The material cost for the optical scale-up (interconnect within nodes) alone reaches $800 million.
Breaking down the $154 billion TAM: 69% (around $106 billion) comes from scale-up (high-speed interconnect within nodes), and 31% comes from scale-out (interconnect between nodes/data centers). CPO accounts for $91 billion, assuming a 29% penetration rate—if penetration is higher, the numbers will be even larger.
Another trend accelerating this market is the projected penetration of Silicon Photonics technology, which is expected to surge from 6% in Q1 2024 to 46% by Q4 2028. This means traditional EML (Electrically Modulated Laser) solutions are being replaced, and the entire supply chain is undergoing a major technological shift.
Three, distribution of profits in the supply chain: who eats the meat, who drinks the soup?
This is the most critical issue.
The optical interconnect supply chain can be roughly divided into four layers:
First layer: passive materials (fiber, connectors).
Representative company: Corning (GLW). Fiber is the foundational infrastructure; regardless of which technology route (CPO, pluggable optical modules, OCS) is taken, fiber is needed. Many people's first reaction is to buy Corning—'whoever wins, it wins.'
But in reality, fiber is a low-margin passive material with low technological barriers. Moreover, Corning is a diversified company; optical communications are just one of its business segments. It also has display glass (for mobile/TV screens), automotive glass, telecom carrier equipment, and a bunch of traditional businesses. Even if AI optical communication demand surges, growth will be diluted by other low-growth businesses.
Corning is a stable allocation—unlikely to drop much but limited upside. It isn't a target for seeking excess returns.
Second layer: light sources/lasers.
Representative companies: Lumentum (LITE), Coherent (COHR), Sivers Semiconductors (SIVE).
Lasers are the core component of CPO—the optical engine needs a laser source to generate light signals. Goldman clearly pointed out in its report that the supply of CW lasers and EMLs will be 'very tight' from 2025 to 2026, and it may not balance until the second half of 2028.
Tight supply means pricing power lies with the sellers. This is similar to the logic of SK Hynix's HBM memory chips—when supply is tight, margins skyrocket.
Nvidia invested $2 billion in both Lumentum and Coherent to secure laser supply. Sivers Semiconductors is a small Swedish company specializing in InP (Indium Phosphide) laser arrays specifically for CPO. They just reached a deal with Jabil for 1.6T optical transceivers and are collaborating with POET Technologies on the optical engine ELS module. On May 29, it was added to the MSCI Sweden Small Cap Index, and passive funds are about to buy in.
But be aware of the risks with Sivers: a market cap of 16.5 billion Swedish Krona (about $1.6 billion), with projected revenue of only 304 million Swedish Krona (about $32 million) in 2025 and a net loss of 223 million Swedish Krona. Its price-to-sales ratio is as high as 46, far exceeding the average of 5 for European tech stocks. The stock price skyrocketed by 172% in a month—this has already priced in quite a bit of optimism.
Third layer: active chips (DSPs, switch chips, optical engines).
Representative companies: Marvell ($MRVL), Broadcom ($AVGO).
This layer is where the profits are thickest in the supply chain, and it has the highest technological barriers.
DSP chips (digital signal processors) are responsible for converting electrical signals into a format that optical signals can understand, effectively acting as the 'translators' of optical communications. Marvell is close to a monopoly in this field. Nvidia invested $2 billion in Marvell, not out of friendship, but because without Marvell's DSP, Nvidia's GPU clusters cannot utilize optical interconnects.
Switch chips are responsible for data routing in optical networks, determining where data goes. Broadcom is the absolute leader in this field. They showcased a 102.4T CPO switch at the OFC 2026 conference, having already delivered CPO products to customers in October 2025, and will start mass production in 2026.
Why are active chips the most profitable? Because anyone can lay fiber, but only a handful of companies worldwide can design high-performance DSPs and switch chips. Technological barrier = pricing power = high margins. Broadcom's semiconductor business has maintained a gross margin of over 70% for a long time.
Fourth layer: system integration and equipment.
Representative companies: Zhongji Xuchuang ($300308), Fabrinet ($FN).
Optical module manufacturers are responsible for packaging lasers, DSPs, fibers, and other components into usable products. They are the 'assembly factories' of the supply chain, with relatively low margins but benefiting from volume growth. Goldman has significantly raised the target price for Zhongji Xuchuang (from 791 yuan to 1187 yuan), indicating that module manufacturers can also benefit from growth dividends, though their profit thickness is less than that of upstream chips.
Four, an investment framework: "Just because you can't avoid it doesn't mean you'll make big money."
This is the key understanding to grasp the entire supply chain.
Many people's investment logic is 'find the unavoidable companies'—fiber is unavoidable, so buy Corning; GPUs are unavoidable, so buy Nvidia. But 'unavoidable' is just a necessary condition, not a sufficient one.
What really determines profit thickness is replaceability:
Fiber is unavoidable, but many companies make fiber; the technological barrier is low, and margins are thin. DSP chips are also unavoidable, but only a few companies like Marvell produce them; the technological barrier is extremely high, and margins are thick.
The same logic has been validated once in the semiconductor storage sector: HBM memory chips are unavoidable; only SK Hynix and Samsung can manufacture them, with margins exceeding 70%. HBM's packaging substrates are also unavoidable, but many manufacturers produce them, with margins just above 20%.
What investors should look for is not the 'indispensable' segments, but the 'irreplaceable' ones. The two may sound similar, but the profit gap is vast.
Applying to the CPO supply chain:
Corning → indispensable but replaceable → stable allocation Marvell/Broadcom → indispensable and irreplaceable → high elasticity targets Sivers/Lumentum/Coherent → indispensable and temporarily irreplaceable (supply tight) → maximum elasticity but also maximum volatility.
Five, signals from Nvidia's $6 billion shopping spree.
Nvidia's $6 billion investment in optical companies within a month isn't just financial investment; it's strategic hoarding.
Nvidia's investment in Lumentum and Coherent is to secure laser supply—Goldman’s report states that laser supply will be tight until 2028. Their investment in Marvell is to secure DSP chips—without DSP, you can’t build a CPO switch. Nvidia is also working on CPO and plans to start mass production of CPO switches in early 2026.
When Nvidia starts to personally lay out optical interconnects, it means one thing: the bottleneck of AI infrastructure has shifted from 'computation power' to 'connection.' The GPUs are sufficient (at least the roadmap is clear), but the data highways between GPUs still need to be built.
This follows the same script where Nvidia crazily hoarded HBM chips two years ago—first, it was insufficient computation power (buy GPUs), then insufficient memory bandwidth (buy HBM), and now it's insufficient network bandwidth (buy optical interconnects). Each bottleneck shift spawns a new batch of infrastructure winners.
Six, where are the risks?
First, the penetration pace of CPO may not meet expectations. Goldman assumes a 29% penetration rate for CPO by 2028, but CPO has structural drawbacks: once the optical engine fails, the entire module may be scrapped (unlike pluggable optical modules that can be replaced individually). High maintenance costs mean data centers may adopt it more conservatively, and penetration may be lower than expected.
Second, valuations are not cheap. Sivers has a price-to-sales ratio of 46, while Coherent and Lumentum have also surged significantly in the past six months. After Goldman released this report, many targets have moved from the 'independent insight' phase to the 'Wall Street consensus' phase—most of the easy money has already been made.
Third, the technological route is uncertain. CPO is not the only option; pluggable optical modules are also upgrading (1.6T → 3.2T), and OCS (optical circuit switching) is another direction (Google's TPU v7 is already using it). Which route will ultimately become mainstream is still undecided. Goldman also mentioned that CPO and pluggable optical modules will coexist in the long term.
Fourth, the export control risks of InP substrates. CPO lasers rely on InP materials, and Goldman’s report mentions that if InP export controls tighten further, the supply chain could face issues.
Seven, summary.
The investment mainline for AI infrastructure is shifting from 'computation power' to 'connection.' GPU → HBM storage → optical interconnect; each bottleneck is transmitted layer by layer. Each layer of transmission spawns a new batch of winners.
The investment logic in the optical interconnect space is exactly the same as that in the previous HBM memory chip: find the segments with the highest technological barriers, tightest supply, and thickest profits. In this supply chain, active chips (DSPs, switch chips) are the most profitable layer, lasers are the tightest supply layer, while fibers are the largest volume but thinnest profit layer.
In every technological wave, the first to be noticed are the front-line stars (GPUs, large models), while the infrastructure behind them (storage, networking, cooling) is only discovered later. Optical interconnects are transitioning from being 'last to be discovered' to 'Wall Street consensus'—Goldman's report is the turning point.
'Copper is a thing of the past; optics are here'—this is not just a slogan; it's backed by $154 billion in cold hard cash.
That's my analysis; feel free to discuss.