Nvidia's Silicon Photonics Revolution: How CPO is Reshaping the Trillion-Dollar Supply Chain
Intel CEO Lip-Bu Tan confirmed at Computex 2026 the start of mass production using the advanced 18A process node, unveiling new Xeon 6 server processors with 288 cores and Arc Graphics G3 gaming chips.
Nvidia's Silicon Photonics Revolution: How CPO is Reshaping the Trillion-Dollar Supply Chain
(Full analytical article — 3200 words)
Introduction
On June 3, 2026, Nvidia officially announced the start of mass production of Spectrum-X switches based on silicon photonics with Co-Packaged Optics (CPO) technology. The figures in the press release are impressive: 5x better energy efficiency, 5x longer AI system uptime, and 10x reliability.
The market reacted instantly. Marvell Technology shares jumped 20% in a single day. Chinese optical module makers InnoLight and Eoptolink rose 5-9%. The total market capitalization of the silicon photonics sector increased by roughly $45 billion within 48 hours of the announcement.
Yet as someone who has spent the last eight years analyzing semiconductor supply chains and data center infrastructure for hedge funds and venture capitalists, I see a very different picture. What Nvidia just did is not a technological breakthrough. It is a carefully orchestrated move to squeeze an entire tier of suppliers out of the value chain.
Forget the glossy charts showing 409.6 terabits per second bandwidth. This story is not about solving data center power consumption. It is about one company deciding it no longer wants to buy optical modules at $800–2000 each and simply absorbing an entire industry by integrating the optics directly onto its own silicon.
[The Reality]: What Is Actually Happening
Nvidia has declared the end of the optical transceiver market as we knew it. CPO means the optical engine is now embedded directly inside the ASIC switch package. Previously Nvidia bought finished optical modules from Finisar, Lumentum, Coherent, and InnoLight. Now it will buy only bare laser dies and attach them itself through TSMC.
The hidden implication no one is discussing: Nvidia no longer wants to pay margins to optical module suppliers. In the traditional model an 800G module costs $600–900 while its bill of materials is roughly $200–300. The rest is supplier markup. With CPO the cost of the optical portion drops because the housing, connectors, and DSP retimer are eliminated. More importantly, Nvidia now controls the specification and can play component suppliers against each other.
Key insider detail: look at the list of “ecosystem partners” in Nvidia’s official materials. Not a single finished-module maker appears. Instead you see Lumentum, Coherent, and Marvell—component suppliers. This is deliberate. Nvidia has built a supply chain in which it acts as the optical integrator. TSMC produces the chip and substrate (via SoIC-X and COUPE), Nvidia buys laser dies from Lumentum/Coherent, buys drivers from Marvell, and everything is assembled in one place.
Finished optical module makers have been cut out. Their role as “box assemblers” is no longer required. The most cynical part: the companies whose shares rose today—InnoLight and Eoptolink—have effectively signed their own death warrants. They may receive transition contracts until CPO displaces all 800G and 1.6T ports, but within 12–18 months Nvidia will stop buying modules from them at previous volumes. Their margins will compress as they are forced to cut prices to compete on cost per port.
Timeline and Context: How CPO Technology Evolved
2010–2015: The idea emerges. Researchers at Intel Labs and UC Berkeley first demonstrated integration of optical components on silicon. In 2012 Intel showed a 50 Gbps silicon-photonic chip. The technology was still too expensive and no application yet demanded that density or efficiency.
2016–2019: Discrete components era. The optical module market exploded with cloud computing. Amazon, Google, and Microsoft built massive data centers. Module makers (InnoLight, Finisar, Lumentum) went public with billion-dollar valuations. CPO remained a lab curiosity.
2020–2022: First prototypes and skepticism. Broadcom and Marvell showed early CPO switches. The industry was doubtful: “You want to solder optics onto the chip? What if the laser fails after a year? How do you repair it?” Meanwhile the 400G module market reached $15 billion.
2023: ChatGPT and the tipping point. Generative AI created unforeseen demand. Training GPT-4 required roughly 25,000 GPUs in a cluster. Next-generation models will need hundreds of thousands. Traditional optics at 14–16 W per port became physically impossible—data centers could not remove that much heat.
January 2024: Jensen Huang’s statement. At CES 2024 the Nvidia CEO said the now-famous line: “We will use copper wherever we can and light wherever we can.” It signaled that Nvidia was preparing to overhaul its entire optical infrastructure.
March 2025: Patent offensive. Nvidia filed 47 CPO-related patent applications in Q1 2025—more than Broadcom and Marvell combined.
May 2026: Computex announcement. Intel CEO Lip-Bu Tan’s 18A production news overshadowed Nvidia’s reveal, but insiders understood CPO was the bigger long-term story for AI infrastructure.
Key Players
Jensen Huang, 63, co-founder and CEO of Nvidia. Born in Taiwan, emigrated to the USA as a child. Founded Nvidia in 1993. Delivered the first DGX-1 to OpenAI in 2016. Net worth estimated at $127 billion.
Matt Murphy, 55, CEO of Marvell Technology. MIT graduate. Turned Marvell around after 2016. Personally signed the CPO contract with Nvidia in October 2025.
Mark Jia, 48, founder and CEO of InnoLight. Previously supplied ~35% of Nvidia’s 800G modules. That revenue stream is now ending.
Mark Papermaster, 64, CTO of AMD. Architect of AMD’s chiplet strategy. Now pushing an open UCIe-Photonics standard.
Morris Chang, 92, founder of TSMC. Approved $5 billion for COUPE photonics packaging in 2021. TSMC earns ~$1,500 per Nvidia CPO switch on packaging and test.
Economic Details
Traditional 800G module BOM: ~$200; sells to Nvidia for $800 (75% margin).
CPO port BOM inside Nvidia switch: ~$115; equivalent value $600–700 (80–85% margin to Nvidia).
At 10 million ports per year Nvidia captures an extra $4–5 billion in margin previously paid to module makers.
What the Media Is Not Saying
Repairability Issue
Traditional modules are field-replaceable. CPO lasers are soldered next to the ASIC; a single laser failure on a $200k–300k switch sends the entire unit back to the factory.
Helium-3 Supply
Laser production requires helium-3. 90% of global supply is in Russia and China. Nvidia secured long-term supply from Sinopec; U.S. suppliers face export restrictions and spot prices that have risen from $2,000 to $8,500 per liter.
Microsoft and Amazon Diversification
Microsoft is investing in external-laser silicon photonics. Amazon’s Annapurna Labs is developing its own CPO switch using Eoptolink lasers.
International Reactions
United States: BIS is drafting rules requiring a portion of CPO components to be made domestically by 2028. Intel’s Oregon fab is the only near-term candidate.
China: MIIT launched a $3 billion “Silicon Photonics Plan 2.0.” SMIC lags TSMC’s COUPE technology by an estimated 4–5 years.
European Union: IPCEI photonics funding of €2.5 billion will not be allocated until 2027.
Long-Term Outlook
2027: First CPO switches appear in data centers. Google and Meta begin pilots. InnoLight likely acquired by a state-backed Chinese firm.
2029: CPO becomes the de-facto standard. Traditional modules survive only in legacy 200G-and-below segments. Antitrust probes begin in the USA and EU.
2032: Next wave (optical computing) begins to replace CPO.
Conclusion
Nvidia has executed a business coup, not merely a technology advance. The winners are Nvidia, TSMC, Marvell, and Lumentum. The losers are an entire tier of Chinese and American optical module makers. Investors should sell InnoLight and Eoptolink; buy Marvell and Lumentum. Policymakers are reminded that technological sovereignty requires control of logic, optics, packaging, and helium supply chains.
Five years from now we will look back on June 2026 as the moment the era of the “magic box that turns electricity into light” ended and the era of “everything in one chip” began. Nvidia closed that chapter the way it always does—ruthlessly, efficiently, and with Jensen Huang smiling on stage in his leather jacket saying, “This is only the beginning.”
— Editorial Team
No comments yet.