Nvidia Launches Mass Production of Spectrum-X Network Switches with Silicon Photonics
Nvidia has announced the full-scale release of Spectrum-X switches based on silicon photonics with CPO technology, delivering 5x better network energy efficiency than traditional solutions and accelerating AI factory deployments.
Nvidia's Silicon Photonics Trap: Why CPO Switches Are Not a Tech Revolution but a Supply Chain Massacre
You've seen the news. Nvidia announced mass production of Spectrum-X switches based on silicon photonics and CPO technology. The numbers are impressive: 5x higher energy efficiency, 5x longer AI system uptime, and 10x reliability. Investors are thrilled: Marvell shares jumped 20%, and Chinese optical module suppliers rose 5-9%.
But as someone who has tracked semiconductor supply chains and data center infrastructure for the past 8 years, I see a very different picture. What Nvidia just did is not a technological breakthrough. It is a carefully planned operation to squeeze suppliers out of the value chain.
Forget the flashy charts showing 409.6 terabits per second. Let's look at what is happening at the business model and supply chain level. This story is not about Nvidia solving data center power consumption. It is about how one company decided it no longer wanted to buy optical modules for $1000-2000 each and simply consumed an entire industry by moving the light onto its own chip.
I warned about this back in 2024 when Jensen Huang first mentioned "use copper wherever possible, use light wherever possible." Now it has happened. The consequences will be catastrophic for dozens of companies that are celebrating stock gains today.
[The Core]: What Is Really Happening
Nvidia has just declared the death of the optical transceiver market as we knew it. CPO (Co-Packaged Optics) means the optical engine is now built directly into the ASIC switch chip package. Previously Nvidia bought ready-made optical modules from Finisar, Lumentum, Coherent, and Innolight. Now it will buy only bare laser chips and solder them itself onto the substrate through TSMC.
The hidden meaning no one is discussing in the news: Nvidia no longer wants to pay margins to optical module suppliers. In the traditional model an 800G optical module costs $600-900 while its cost is around $200-300. The rest is the module maker's markup. With CPO switches the cost of the optical portion is lower because the housing, connectors, and DSP retimer chip are eliminated. But the key point is that Nvidia now controls the specification and can play component suppliers against each other.
The insider detail that changes everything: look at the list of Nvidia's "ecosystem partners" in the official materials. There is not a single finished optical module manufacturer. There are Lumentum, Coherent, and Marvell — component makers. This is no accident. Nvidia has built a supply chain where it acts as the optics integrator. The model is simple: TSMC produces the chip and substrate (via SoIC-X and COUPE technology), Nvidia buys laser chips from Lumentum/Coherent, buys drivers from Marvell, and everything is assembled in one place. Optical module makers have been cut out. Their role of "assembling boxes" is no longer needed.
And the most cynical part: those celebrating stock gains today — InnoLight (Chinese module maker), Eoptolink — have signed their own death warrants. Yes, they will receive transition contracts until CPO replaces all 800G and 1.6T ports. But in 12-18 months Nvidia will stop buying modules from them at previous volumes. Their margins will start shrinking as they are forced to cut prices to compete with CPO on a per-port cost basis.
Timeline and Context
Look at the timing to understand how tightly Nvidia planned this operation. In January 2026 at CES the company first announced that Spectrum-X would become the standard for the Vera Rubin NVL72 platform. In March at GTC 2026 Jensen Huang declared that CPO switches were entering "full mass production." On May 31, 2026 came official confirmation that mass production had begun. Three announcements in five months. This is not spontaneity. It is a planned campaign to seize the market.
Why now and not a year earlier or later? There are two reasons the press releases do not mention.
The first is that the data center power consumption crisis has reached the point of no return. An 800G port on traditional transceivers consumes 14-16 W. For a 100,000-GPU cluster that means megawatts just for optics. A CPO port consumes 5.2-5.6 W — a 60-68% reduction. The next generation, 1.6T, on traditional optics would consume 25-30 W per port — already impossible to cool. CPO is no longer the "better solution"; it is the only possible solution.
The second reason is limited production capacity at competitors. Broadcom is only promising its CPO version of Tomahawk 6 ("Davisson") in the second half of 2026. Marvell is also still catching up. Nvidia, however, already has signed contracts with TSMC for the COUPE platform and capacity is reserved. Taiwanese sources report that TSMC plans to ramp photonic integrated circuit production to 10,000 wafers per month by Q1 2027, with Nvidia as the primary consumer of that capacity.
Nvidia struck at the moment when competitors technically cannot respond in kind. By the time Broadcom releases its CPO switch, Nvidia will already have scale and cost-down advantages.
Winners and Losers
Let's rank everyone, because the "winners" in this story are not the ones shown on television.
Biggest winner — TSMC. Its COUPE (Cu-pillar bonding for photonics) and SoIC-X technologies have become the de-facto standard for CPO. Every CPO switch from Nvidia, Broadcom, Marvell, and AMD will be manufactured on TSMC lines. The Taiwanese giant is not just getting chip orders. It is gaining control over the entire "silicon photonics" segment analysts value at $100 billion by 2030. TSMC is the only player that can produce 3D hybrid bonding assembly of photonics with logic. Samsung and Intel lack this technology.
Next winner — Marvell. Its DSP chips (digital signal processors) for silicon photonics have become the standard for Nvidia's CPO solutions. Jensen Huang publicly called Marvell "the next trillion-dollar company." These are not empty words. Marvell has an exclusive contract for control chips across all Spectrum-X CPO ports. Volumes will be enormous — up to 100 million optical engines per year by 2028.
Loser #1 — traditional optical module makers. InnoLight (China), Eoptolink, Hisense Broadband. Their business depends 70-80% on selling modules to hyperscalers (Google, Amazon, Microsoft) and Nvidia. Now Nvidia is exiting the market as a module buyer. Hyperscalers will follow. Once Amazon sees Nvidia offering ready-made CPO switches with better efficiency and reliability, they will stop buying separate modules and start buying complete CPO racks. Demand for traditional modules will collapse.
Loser #2 — Broadcom and other switching-ASIC makers. They will also release CPO switches, but 6-9 months late. In that time Nvidia will capture the AI backend market (internal cluster networks) where bandwidth and power requirements are highest. Broadcom will be left with traditional enterprise networks where CPO is less critical. That is a smaller and less profitable market.
Chinese optical component makers — in the turbulence zone. On one hand, shares are rising on the news. On the other, their position is extremely vulnerable. Nvidia is not buying complete modules; it is buying only components (FAU arrays, laser chips, lenses). Chinese companies such as Source Photonics may retain share, but their margins will fall because they are no longer "assembling" finished products — they are selling "bricks" for Nvidia. Their negotiating position is weaker.
What the Media Is Not Saying
Journalists are now writing about a "breakthrough in photonic computing" and a "revolution in data centers." I will tell you what they are omitting, either because they do not understand it or do not want to alarm readers.
Insider #1: CPO does not work without liquid cooling.
Look carefully at Nvidia's switch specifications. Both SN6600 and SN6810 require liquid cooling. This is not optional. It is mandatory. The CPO engine generates less heat than traditional optics, but it sits directly next to the ASIC that reaches 80-90 °C. Laser chips (operating at 1310 nm) are extremely temperature-sensitive — a 1-degree deviation changes wavelength and degrades transmission.
What this means for the market: data centers that want to use Nvidia CPO switches must move to liquid-cooled racks. This requires complete reconstruction of the cooling infrastructure. Retrofitting a 100,000-GPU data center costs an extra $50-100 million. The media reports "60% lower power consumption" but forgets to mention these capital expenditures.
Insider #2: The maintainability problem no one talks about.
A traditional optical module is an FRU (Field Replaceable Unit). It fails — pull it out, plug in a new one, five minutes of work. The CPO engine is soldered onto the substrate next to the ASIC. It cannot be replaced in the field. If one laser chip out of 512 on the switch fails, the entire switch ($200,000-300,000) must be sent back to the factory for repair.
Nvidia promises "10x higher reliability." This is a marketing calculation based on lab tests with no vibration, temperature swings, or dust. In a real data center failures will occur. When they do, operators face a dilemma: keep spare switches worth $300,000 in inventory (expensive) or accept cluster downtime. Traditional modules allowed keeping $800 spares on the shelf and swapping them in minutes. That flexibility is now gone.
Insider #3: The real target is not hyperscalers but defense contracts.
Nvidia is selling CPO switches to Oracle, CoreWeave, and Lambda Labs. These are public customers. But there are customers no one mentions: the U.S. Department of Defense. Its "AI Next" program requires protected AI clusters with minimal power consumption for shipboard and mobile data center deployment. CPO is the only technology that can pack sufficient compute into a constrained power budget.
This explains why Nvidia is investing so aggressively in CPO despite the technological risks. Defense contracts carry 40-50% margins versus 20-25% on commercial deals. They are also protected from sanctions and trade wars. Part of Nvidia's CPO production capacity will be located in the United States (via TSMC Arizona) to comply with ITAR (International Traffic in Arms Regulations).
Forecast: Next 30 Days and 90 Days
Turn off the emotions. Forget the 20% stock jump you saw yesterday. Here is what will actually happen.
Next 30 days: Wave of analyst forecast revisions.
Within a month Morgan Stanley, Goldman Sachs, and other investment banks will issue reports downgrading optical module makers — InnoLight, Eoptolink, Hisense. Analysts will finally realize that the TAM for traditional modules is shrinking. Shares of these companies will correct 15-25% from current highs. Those who bought on CPO news will lose money.
At the same time ratings for Marvell and Lumentum will rise. Marvell will trade at a 35-40x P/E multiple instead of the current 25-30x. Analysts will begin modeling contracts for DSP chips supplying millions of CPO ports.
Next 90 days: First signs of trouble at hyperscalers.
Amazon and Google will announce delays in deploying their next-generation AI clusters. The reason: they planned to use traditional 1.6T optical modules but must now switch to CPO to stay within power budgets. The transition will take 6-9 months and require redesign of server racks. Their quarterly reports will show rising Capex without corresponding compute growth because of these delays.
Meta, which traditionally relies on Broadcom for networking gear, will find itself in a difficult position. Broadcom has no ready CPO switch until Q4 2026. Meta will have to either wait (and fall behind competitors) or buy Spectrum-X from Nvidia (and become dependent on its main GPU rival). The media will start writing about a "networking bottleneck" in the AI race, but no one will name the real cause — Nvidia's technological squeeze.
What about 12 months from now? CPO will become the standard for all new AI clusters. The optical transceiver market will shrink 30-40% from the previously forecast $200 billion. Dozens of mid-tier Chinese companies (Eoptolink, Hisense) will retreat to niches or be acquired. InnoLight, the largest player, will try to diversify into automotive LiDAR and consumer optics, but those markets carry 2-3x lower margins.
And Nvidia? Nvidia will gain yet another lever over customers. Now, to buy its GPUs you must buy its network switches. To buy the switches you must move to liquid cooling. To move to liquid cooling you must rebuild the data center. This is classic vendor lock-in, and it will work perfectly.
Nvidia engineers who developed CPO will receive $5-10 million bonuses. InnoLight engineers who assembled modules by hand will receive layoff notices. Technological progress always means someone gets fired. Today it happened to an entire industry. And no one in the news will tell you that.
— Editorial Team
No comments yet.