Marvell Technology is expanding its effort to reduce the energy required to move data inside artificial-intelligence facilities, unveiling what it describes as the industry’s first optical networking chips manufactured on a 2-nanometer process.
The company presented the technology at ECOC 2026, an optical communications conference in Málaga, Spain, according to a Marvell press release cited by TheStreet’s report published by the Miami Herald.
Marvell is also demonstrating a 102.4-terabit co-packaged optics platform, positioning the company to supply components for multiple approaches to high-speed data-center networking.
The announcements reflect an emerging challenge for AI infrastructure: delivering more computing performance without allowing networking equipment to consume an unsustainable share of a data center’s limited power supply.
New chips target optical transceivers
Marvell’s new 2nm products include digital signal processors designed for optical transceivers. These modules convert electrical signals into light—and then back into electrical signals—so large amounts of information can move quickly between servers and other data-center equipment.
Reducing the manufacturing process to 2nm is intended to improve energy efficiency. The company’s industry-first claim relates specifically to the manufacturing node rather than the transmission speed.
Broadcom introduced a 3nm signal processor supporting 400 gigabits per lane in March, according to a company announcement referenced in the source report. Marvell is therefore not claiming to have reached that transmission rate first; its distinction rests on moving the technology to a smaller process.
That difference could become important as AI operators seek to control electricity consumption while expanding the number and performance of accelerators inside their facilities.
Marvell supports competing optical designs
The networking industry has not settled on a single way to connect increasingly powerful AI systems.
Traditional pluggable transceivers can be inserted into or removed from networking equipment. Co-packaged optics, by comparison, place optical components close to the switch chip, potentially lowering the power needed to transmit data.
Research presented by Meta at the previous ECOC conference found that a co-packaged optical design used approximately 65% less power than a comparable pluggable module, according to SemiAnalysis data cited in the report.
Such findings could threaten demand for some components used in pluggable modules. Marvell’s strategy, however, spans both sides of the technological shift.
The company is working to lower the power consumption of pluggable signal processors through its 2nm design while also developing the 102.4-terabit co-packaged optics platform.
This approach could allow Marvell to benefit whether data-center operators continue relying on removable modules or move more aggressively toward integrated optical systems.
A large transceiver market is taking shape
The commercial opportunity extends beyond a single product demonstration.
LightCounting expects more than 100 million transceivers in the 1.6-terabit and 3.2-terabit categories to ship during a five-year period, according to figures reported by SDxCentral and cited in the source article. Nearly half are expected to use 400G optical technology.
Marvell and Broadcom are competing for a share of that projected demand.
Winning early designs can have lasting value because transceiver manufacturers must test and approve the chips used in their products. Once a component completes that qualification process, replacing it can require additional time and expense.
Marvell is betting that the energy savings associated with its 2nm technology will help it secure those long-term design commitments.
GlobalFoundries agreement strengthens U.S. production
Marvell’s product announcement followed the expansion of a multiyear manufacturing agreement with GlobalFoundries.
The companies announced additional silicon-germanium production capacity at GlobalFoundries’ facility in Burlington, Vermont, on Sept. 17, Reuters reported.
Silicon-germanium technology is used to make amplifiers and drivers for optical connections. GlobalFoundries said the expanded production would support pluggable, near-packaged and co-packaged optical systems.
That breadth provides Marvell with another way to participate in the market regardless of which networking configuration becomes dominant.
During an Aug. 27 earnings call, Marvell executives said the amplifier-and-driver operation was meeting or exceeding its progress toward a $1 billion annual revenue run rate, according to a transcript cited by TheStreet.
Data-center revenue drives record quarterly sales
Marvell develops semiconductors used to move, store and protect data for cloud providers, telecommunications companies and other enterprise customers.
Its data-center revenue increased 46% during the most recently reported quarter, helping total quarterly revenue reach a company record of $2.74 billion.
That performance has tied Marvell’s outlook closely to continued spending on AI infrastructure.
The company’s shares gained 2.4% in premarket trading on Sept. 21, according to Seeking Alpha. However, Benzinga attributed part of the increase to a broader technology-sector rally, making it difficult to connect the movement solely to Marvell’s optical announcements.
The stock opened at $251.40 after closing at $244.25 on Sept. 18 and was trading near $250.66 at 10:13 a.m. Eastern time, according to StockAnalysis figures cited in the report. At that level, it remained approximately 24% below its 52-week high of $329.88.
Short-term market prices can change quickly and should not be treated as evidence that a product strategy has succeeded.
Analysts await Marvell’s October investor presentation
Wall Street’s attention is now turning to Marvell’s investor event scheduled for Oct. 6.
Morgan Stanley increased its price target for the company from $246 to $268 on Sept. 21 but maintained an Equal Weight rating, TheFly reported. The firm suggested that Marvell could use the event to discuss a path toward more than $40 billion in revenue by fiscal 2030.
That figure would be more than three times the approximately $12 billion in revenue expected for the current fiscal year, based on StockAnalysis estimates cited by the source. It remains an analyst projection rather than company guidance.
StockAnalysis listed 40 of 45 tracked analysts as assigning Marvell either a buy or strong-buy rating, with none recommending a sale. Their average target was $289.04. The same data valued the shares at approximately 46 times projected earnings, indicating that substantial future growth was already reflected in the price.
Energy efficiency becomes an AI priority
Technology companies are expected to continue investing heavily in AI infrastructure. Estimates from Evercore and Bank of America cited by CNBC suggest combined capital expenditure by major technology companies could exceed $1 trillion in 2027.
Yet every facility has limits on how much electricity it can access and distribute. Power consumed by networking equipment is unavailable to the processors performing AI training and inference.
That constraint is increasing the value of components capable of moving data while using less energy.
Marvell previously projected approximately $300 million in revenue from scale-up optical products in fiscal 2028. Management said in August that it now views the opportunity as meaningfully larger, according to the earnings-call transcript.
What happens next
Marvell’s Oct. 6 Investor Day may provide updated financial expectations for its scale-up optical business and clarify how heavily its long-term outlook depends on near-packaged and co-packaged technology.
Before then, the company’s ECOC demonstrations establish the technical direction of its strategy: improve the efficiency of conventional pluggable components while developing products for more integrated optical architectures.
The next stage of AI infrastructure competition may therefore be determined by more than processing speed. As data centers confront strict power limitations, the companies that move information most efficiently could gain an increasingly important advantage.
