Marvell's MRVL Surpasses $290, Driven by AI Infrastructure and Connectivity

On June 3, 2026, Marvell Technology's stock price, $MRVL, reached $290, a new all-time high.

Increased by 254% over the past 12 months, less than 40 three years ago.

If we start counting from when Matt Murphy became CEO in 2016—when the stock price was under $10 and the market cap was less than $2 billion—that’s a 30x increase.

But the increase itself is certainly not what we want to discuss in this article.

What this article truly seeks to understand is: What price is the market actually assigning to Marvell? And beneath this price, are people still using an outdated cognitive framework to interpret this company?

Many people on the street call Marvell “the mini Broadcom”—the second-largest player in custom AI chips, picking up the scraps left by hyperscalers after Broadcom. This characterization isn’t entirely wrong, but it has a critical blind spot: it assumes Marvell is merely a smaller version of Broadcom. Yet, Marvell’s true value lies precisely in the fact that it is nothing like Broadcom.

Marvell occupies a unique niche. AI infrastructure is shifting from “stacking GPUs” to “building systems,” and this position will become increasingly valuable.

This article will try to clarify this matter.

The full text is approximately 15,000 words; please read at your own pace, teachers~

I. What exactly is Marvell selling?

To understand Marvell, the first step is to throw away the label of “chip company.” It doesn’t make GPUs, doesn’t build CPUs, and doesn’t sell memory. It sells “connection”—enabling data to flow at the speed of light between AI chips, servers, and data centers.

Looking at it separately, three businesses:

Marvell is the undisputed leader in high-speed optical DSPs. Approximately 70% of DSP chips used in global data center optical modules above 400G are manufactured by Marvell.

Every time you see news about a surge in shipments of optical modules for AI data centers, Marvell is quietly collecting the profits.

Why is the 70% market share so hard to challenge? High-speed optical DSP is not an ordinary chip. It must simultaneously handle signal modulation, demodulation, error correction, and clock recovery—under rates like 800G and 1.6T, managing signal attenuation and noise at the physical layer becomes extraordinarily complex.

After spending approximately $10 billion to acquire Inphi in 2021, Marvell has accumulated over five years of mass production experience in this field, progressing through iterations from 5nm to 3nm. Broadcom is also catching up, but this first-mover advantage cannot be easily erased by spending money alone.

In March 2026, Marvell launched four new 1.6T DSPs—Ara T, Ara X, Petra, and Aquila M—covering everything from short-reach to long-reach, and Ethernet to InfiniBand. During the FY2027 Q1 earnings call, Murphy stated plainly: the growth outlook for the optical interconnect business has been raised from 50% to over 70% for FY2027.

It’s not that the market changed; it’s that they underestimated how strong the demand was.

This is the area most closely watched by the market. The logic is simple: Amazon doesn’t want to pay NVIDIA for every GPU, so it designed its own AI training chip called Trainium. But Amazon doesn’t manufacture chips—it needed someone to help with design and mass production. That someone is Marvell.

Marvell currently has 18 custom XPU design projects underway, serving three hyperscalers—Amazon, Microsoft, and Google—with a lifetime revenue funnel reaching $75 billion. Custom chip revenue for FY2026 is approximately $1.5 billion, and is expected to more than double by FY2028.

However, there’s an uncomfortable aspect to this business: the gross margin is lower than that of standard products. For Q1 FY2027, the non-GAAP gross margin was 58.9%, compared to Broadcom’s 77.5%. The reason is straightforward—you’re working for Amazon, not selling your own standard products, which means higher R&D investment and stronger customer bargaining power.

We'll go into more detail later.

The Ethernet switching chip is expected to exceed $600 million in FY2027 (doubling year-over-year), driven by the rigid demand for high-speed switching as AI clusters scale from hundreds to over 100,000 GPUs. Enterprise SSD and HDD controllers, the traditional core business, generate stable cash flow but are seeing their share gradually shrink due to pressure from AI-related business growth.

Put the three pieces together, and the picture becomes clear: Marvell is not a “do-everything” chip company, but one that has built end-to-end connectivity capabilities centered around AI data movement—from SerDes inside chips, to PCIe/CXL switching between chips, to optical DSPs between racks, to coherent optical modules between data centers—each step has been actively addressed.

Understanding this makes it clear why NVIDIA invested $2 billion in it. It also helps explain why calling Marvell a "mini Broadcom" is a misinterpretation.

II. In the AI Era, "Connection" Has Become the Main Character

Over the past two years, all the spotlight has been on GPUs. Computing power is muscle—the thicker, the better. But when AI clusters scale from thousands of GPUs to 100,000 or even 500,000, a fundamental physical limitation emerges: copper cables can only transmit signals up to 3 meters; beyond that, signal degradation renders them unusable.

A GPU can be the most powerful brain in the world, but if the signal transmission between neurons can't keep up, even the highest intelligence is useless. In a cluster of 100,000 GPUs, each GPU may spend 30% to 50% of its total runtime waiting for data.

This is why optical interconnects have become the centerpiece. Light can transmit hundreds of meters, even several kilometers, with almost no attenuation. The larger the cluster and the more GPUs involved, the higher the proportion of optical connections—not growing linearly, but superlinearly.

Barclays estimates that optical port shipments will double in 2026 and double again in 2027. Marvell’s optical interconnect business is expected to grow by approximately 90% annually this year and next—wait, the estimate has now been revised to over 70%, but based on actual performance, 90% may still be an underestimate.

The key point is that this trend is not something that will last just one or two years. As long as AI model parameters continue to grow and training and inference clusters continue to expand, the demand curve for optical interconnects will not flatten.

This is not an issue of a market cycle; it is a long-term structural trend determined by the laws of physics.

For example: AI infrastructure is a city undergoing rapid expansion—GPUs are the buildings themselves, while Marvell sells the plumbing, wiring, and highways. You can switch architects to build different houses, but once the infrastructure is laid, replacing it is far harder than constructing the buildings.

Three: From $10 to $290: An Undervalued CEO

In 2016, Marvell was a stock abandoned by the market.

Founders Sehat Sutardja and Weili Dai were forced to step down due to an accounting investigation and governance crisis, prompting SEC intervention. The company had expanded into many areas—mobile communications, printers, consumer electronics—but none reached the top three in their respective industries. With the stock price under $10, customers began to question whether the company could survive.

That year, the activist hedge fund Starboard Value intervened, orchestrating a textbook management overhaul, bringing in Matt Murphy as CEO from Maxim Integrated.

However, I think Murphy deserves a bit more mention: he spent 22 years at Maxim, rising from front-line sales to Executive Vice President, overseeing the company’s entire product development, sales, and profitability. He’s not the typical “tech-genius” type of semiconductor CEO—he’s an extremely pragmatic and highly focused businessman.

He once said something that left a deep impression on me: “My dad worked on Apple’s first sales team, so from a young age, I understood—no matter how great the technology, if you can’t sell it, it’s worth nothing.”

After Murphy took office, he did three things that looked simple but were extremely difficult to execute:

First, cut.

Mobile communications discontinued. Printer chip business discontinued. Consumer electronics discontinued. Wi-Fi/Bluetooth business sold to NXP for $1.76 billion (2019). Automotive Ethernet business sold to Infineon for $2.5 billion (2025).

All resources are focused in one direction: data center infrastructure.

Second, buy.

In 2018, acquired Cavium for $600 million—specializing in ARM server CPUs and DPUs. In 2021, acquired Inphi for $10 billion—specializing in optical DSPs—a transaction that transformed Marvell’s trajectory. By end of 2025, acquired Celestial AI for $3.25 billion—focused on silicon photonics and photonic interconnects. In early 2026, acquired XConn for $540 million—specializing in PCIe/CXL switching chips.

Four acquisitions, each filling in a piece of the "AI connectivity" puzzle.

Third, bind.

Murphy pursues something called "long-term visibility"—years of predictable revenue certainty. By signing a multi-year agreement with AWS covering custom AI chips, optical DSPs, AEC DSPs, PCIe retimers, DCI optical modules, and Ethernet switching chips, it’s not a one-time transaction, but a comprehensive system-level partnership.

The 10-K discloses that certain capacity reservation agreements have terms ranging from 4 to 10 years.

Result? Upon acquisition, FY2016 revenue was approximately $2.65 billion (reported as $2.32 billion in the 2017 annual report), with slim profits. FY2026 revenue reached $8.2 billion (+42% YoY), with non-GAAP EPS of $2.84 (+81% YoY).

Ten years later, a second-tier chip manufacturer that emerged from a governance crisis has become a core supplier of AI infrastructure.

I’ve repeatedly verified this rule in investing: “CEO change → strategic focus → major acquisitions → securing key clients.” When this entire chain is successfully executed, and each step is supported by clear financial evidence, the company is worth your serious attention.

Four: NVIDIA’s $2 Billion: Endorsement or Acquisition?

On March 31, 2026, NVIDIA announced a $2 billion strategic investment in Marvell, purchasing 2 million shares of convertible preferred stock with an initial conversion price of approximately $91.84, representing about 2.4% equity upon full conversion.

On that day, Marvell rose 13%. However, what excited the market at the time differed from what I later reflected on repeatedly.

The market sees this as NVIDIA putting its real money behind a stamp of approval: “This is a partner I endorse.” The logic is sound. $2 billion is not a PR expense—it’s a strategic investment. In 2026, NVIDIA made a series of heavy investments in optical interconnect companies—Coherent ($2 billion), Lumentum ($2 billion), Marvell ($2 billion)—pouring $6 billion into the same sector. This signal is louder than any analyst report.

But what we should focus on more is the collaboration framework—NVLink Fusion.

NVLink Fusion is NVIDIA’s “semi-custom AI infrastructure” platform. Third-party vendors, such as Marvell, can provide customized XPU accelerators that directly integrate into NVIDIA’s high-speed interconnect network. NVIDIA provides the Vera CPU, ConnectX network cards, BlueField DPU, NVLink interconnects, and Spectrum-X switches.

You hyperscalers want to replace GPUs with your own custom chips? Sure thing—I’ll use NVLink Fusion to connect your chips into my ecosystem. You can have someone else make the chips, but the connection layer is still mine.

It’s nothing short of brilliant. Turning the “enemy” into a “customer”—the more hyperscalers try to escape NVIDIA’s GPUs, the more they need NVIDIA’s networking. And Marvell happens to be the company that builds custom chips for hyperscalers, while also helping NVIDIA build its interconnection ecosystem.

When the clam and the sandpiper fight, the fisherman benefits.

The Next Web published an analysis with a sharp but accurate headline: "NVIDIA's $2 billion investment in Marvell is not an investment—it's a toll booth."

NVIDIA has positioned itself at every entry point of the ecosystem through this investment. But from another perspective—Marvell itself is part of the toll booth: one hand builds chips for cloud providers, while the other helps NVIDIA build its network.

Both sides rely on it, and both sides are sending money to it.

Of course, this also means a perpetual tension: NVIDIA $NVDA is both a partner and a competitor. It develops its own networking chips and is investing in silicon photonics, making the boundaries between collaboration and competition consistently unclear.

However, my personal assessment is that, at this stage of building AI infrastructure, NVIDIA needs Marvell more than Marvell needs NVIDIA. This is because the hyperscalers’ demand for customized chips is structural and irreversible; by refusing collaboration, NVIDIA would essentially hand the entire market over to Broadcom.

Five, what exactly is the difference compared to Broadcom?

After viewing this table, many people quickly conclude: “Broadcom is better—it’s ten times larger in scale, has 20 percentage points higher gross margin, and isn’t significantly more expensive in valuation.”

You can't say it's wrong, but you've missed two of the most critical things.

First, the gross margin gap stems from structural reasons, not because Marvell is unable to generate profits.

Broadcom's 77.5% is not a pure semiconductor figure: it includes VMware's software revenue, and its 67% EBITDA margin significantly inflates the consolidated gross margin.

Looking solely at the semiconductor segment, it’s approximately 60%-65%. Marvell’s 58.9% is indeed lower, but the gap isn’t as significant as it appears. Moreover, as the scale of Custom ASIC production increases and R&D costs are spread out, there is a clear path to improved gross margins—the company targets a medium-term non-GAAP operating margin of 38%.

Second, Marvell is not the "number two" in optical DSP—it's the "number one."

With a 70% market share, Broadcom is the one chasing in this field. Optical DSPs are precisely the segment most benefited by the shift in AI infrastructure from being primarily training-focused to being primarily inference-focused— inference clusters are far more distributed than training clusters and demand a higher density of optical interconnects.

I’ve always found Duan Yongping’s classification framework very useful: A “Class B business” is one where you do better than others, but others are also doing it; a “Class A business” is one where others simply can’t do what you do, or if they try, they can’t catch up. Marvell’s optical DSP is closer to Class A; custom ASICs are closer to Class B, but the depth of customer lock-in and switching costs are extremely high, making the Class B characteristics quite strong.

The market views Marvell through the "small Broadcom" framework, naturally concluding it's not worth $180 billion. But when viewed through a three-dimensional framework of "leader in optical interconnects, second-largest custom chipmaker, and NVIDIA ecosystem partner," the valuation logic changes.

Six: Numbers don't lie

Financial data is the "only standard" for testing any narrative—take a look at Marvell's latest results and guidance:

FY2027 Q1 key figures:

Quarterly revenue of $2.418 billion, up 28% year-over-year and 9% quarter-over-quarter, reaching a new all-time high. Exceeded the midpoint of guidance by $18 million.

Data center revenue amounted to RMB 1.833 billion, accounting for 76%, up 27% year-over-year and 11% quarter-over-quarter.

· Non-GAAP EPS of $0.80, in line with expectations. Operating cash flow of $639 million—also a record high.

· Q2 Guidance: Revenue of approximately RMB 2.7 billion (midpoint), up 12% quarter-over-quarter / 35% year-over-year. Data center revenue expected to grow in the high single to low double digits quarter-over-quarter.

Several trends are particularly noteworthy:

The growth rate is accelerating.

FY2026 full-year growth of +42%, FY2027 guidance of +40%, FY2028 target of +45%. Accelerating growth from a $8 billion base indicates this is not merely inventory restocking or cyclical rebound—it’s structural demand rising.

Operating leverage has been unleashed.

EPS growth (81%) significantly outpaced revenue growth (42%), driven by the combined effects of economies of scale from custom chips, improved production yields for optical DSPs, and strict cost control by Murphy.

Custom chips have a "hidden goldmine"—attach. At the 2025 Custom AI Investor Event, the company disclosed an easily overlooked statistic: by 2028, the TAM for custom XPU is approximately $40.8 billion, while the TAM for attach components around XPU—such as network cards, scale-up fabrics, security coprocessors, and memory pooling chips—is about $14.6 billion, with a compound annual growth rate of 90%. Many focus solely on “who designed the most expensive AI chip,” but the real profits lie in the supporting roles.

PEG rough estimate: Forward PE around 23-24x, revenue growth rate approximately 40%, PEG around 0.6.

Comparing Broadcom’s forward P/E of approximately 30–41 times and revenue growth of around 20%, its PEG is roughly 1.5–2.0. Looking through the simplest lens of PEG, Marvell’s current valuation doesn’t appear expensive. Of course, PEG is only a rough starting point—the real variables are whether growth can be sustained, gross margins can improve, and the competitive landscape won’t deteriorate.

Seven: The Story of Light: Celestial AI and the Next Journey

Light DSP is Marvell’s present; Celestial AI is its future.

In December 2025, Marvell announced the acquisition of Celestial AI, a startup specializing in photonic fabric technology, for $3.25 billion upfront. If Celestial AI achieves cumulative revenue of $2 billion by FY2029, the total consideration could reach $5.5 billion.

The price isn't cheap. Why is Murphy willing to pay?

Celestial AI addresses the next physical bottleneck in AI chip interconnects: copper cables have reached their limit.

Inside current AI servers, GPUs are connected via NVLink—offering high speed but short distances. With eight GPUs per accelerator card, four cards per rack, and hundreds of racks per cluster, the physical limits of copper cabling have become a bottleneck for the entire system.

Celestial AI’s Photonic Fabric replaces electricity with light to enable direct optical interconnects between chips—16 Tbps bandwidth per chiplet, halved power consumption, and nanosecond-level latency.

In simple terms: DSP alone is like expanding the highway inside a data center from two lanes to eight, while Celestial AI is installing teleporters directly between buildings.

Another detail about this acquisition: Amazon supported the transaction.

Marvell even issued a warrant to Amazon, allowing Amazon to purchase up to $90 million worth of Marvell stock based on its procurement volume of Photonic Fabric products. Amazon does not lightly endorse suppliers—its endorsement reflects its genuine need for this technology.

Marvell expects Celestial AI to begin contributing meaningful revenue in the second half of FY2028, with an annualized revenue of $500 million in Q4 FY2028 and $1 billion in Q4 FY2029. If the roadmap is executed successfully, the optical interconnect business will transform from a "first business" into a "super business"—a full-stack optical connectivity platform spanning DSP, silicon photonics, and photonic fabric.

Including the 540 million XConn acquisition in early 2026, Marvell now holds a complete “electrical + optical” interconnect puzzle: chip-level SerDes → inter-chip PCIe/CXL switching → intra-rack optical interconnect → inter-rack optical DSP → coherent optical modules between data centers.

On the AI interconnection front, no other company has an equally comprehensive layout.

Eight: After a 254% increase, don't forget there are still risks.

When researching investments, what matters most isn't finding out "why it will go up"—there are plenty of reasons during a bull market. What matters is identifying "what could cause it to fall," and then deciding whether you're willing to accept that risk. I think @aleabitoreddit's approach to risk management is something everyone should take the time to understand.

Marvell recently lost the primary design rights for Amazon’s next-generation Trainium3 to Taiwan’s Alchip. The company emphasized that Trainium2.5 will continue to be designed by Marvell, “ensuring no revenue gap.”

But the market sees another side: the largest custom chip customer did not choose its longtime partner for the next-generation product. This is not a good sign.

In FY2026, the top ten customers contributed 82% of revenue, with two accounting for more than 10% each. The 10-K candidly states: “The current level of capital expenditure on AI infrastructure may not be sustainable in the long term.”

If either Amazon or Microsoft scales back its custom chip initiatives, Marvell’s revenue will take a direct hit.

Non-GAAP gross margin of 58.9%, nearly 20 percentage points lower than Broadcom's. This is not temporary—it's structural. Custom ASICs are inherently a services business—designing dedicated chips for clients, who own the final product, inherently limiting your pricing power.

Economies of scale can improve, but cannot fundamentally solve, the issue.

If future revenue growth is primarily driven by custom ASICs (with low gross margins) rather than DSPs (with high gross margins), revenue growth and margin expansion will trade off against each other. The market’s valuation multiples may not be as generous as bull investors expect.

NVIDIA invested $2 billion but is also building its own network chip team. Spectrum-X switches, BlueField DPUs, and NVLink interconnects directly or potentially compete with Marvell’s switching chips and custom ASICs. A 2.4% stake is not control—it’s an ecosystem tie.

If NVIDIA decides to internalize more value chains of NVLink Fusion—such as manufacturing more optical interconnect chips—Marvell’s position could become precarious.

Since 2026, CEO Murphy has cumulatively sold approximately $5.3 million (in three sales, at prices ranging from $98.70 to $177.26), CFO Willem Meintjes has sold approximately $4.7 million, COO Chris Koopmans has sold approximately $2.73 million, and CDO Sandeep Bharathi has sold approximately $13.14 million.

There were no internal增持 transactions.

The absolute amount is not significant relative to the ownership stake (Murphy still holds approximately $131 million in stock after the sale), and all transactions were executed through a pre-arranged 10b5-1 plan.

But the signal is clear: the stock price has reached an all-time high, and the group most familiar with the company is selling, with no one buying.

At the very least, ask yourself again: Did I buy this stock because I understood its value, or because I saw it rose by 254%?

10-K disclosures require locking in capacity 26 to 52 weeks in advance, with some agreements lasting 4 to 10 years.

TSMC’s 5nm/3nm production capacity is in fierce competition with GPU manufacturers (such as NVIDIA and AMD), and the delivery lead time for DSPs has already extended to six months.

Marvell misjudged demand—either committing to too much capacity only to see demand decline, or facing insufficient capacity when demand exceeded expectations—the penalties will directly hit the financial statements.

This supercycle of AI infrastructure rewards not only those with the right technology, but also those whose supply chains never falter.

After discussing all these risks, what is my conclusion?

These risks are real.

Losing Trainium3 is no small matter, and the structural issues with gross margins cannot be resolved overnight; insider selling warrants caution. However, I am not siding with Marvell’s bears for three reasons:

First, the bad news about Trainium3 has been overshadowed by the guidance doubling custom chip revenue for FY2028.

Even after losing its largest customer, the next-generation product still provided doubling guidance, indicating that the pipeline for other customers—Microsoft Maia, Google Axion, and the undisclosed new hyperscaler—is stronger than the market realizes.

Second, the moat created by optical interconnects is real—and it's widening.

70% DSP share + Celestial AI silicon photonics + XConn’s PCIe/CXL switching = a full-stack capability others cannot replicate in the short term.

Competitors may win a few custom chip orders, but no one can catch up to Marvell’s years of accumulation in optical interconnects within the next three to five years.

Third, a PEG of 0.6 provides a certain margin of safety.

A 40% revenue growth rate corresponds to a 23x forward P/E—this valuation isn't "the market treating it as the next Broadcom," but rather "the market still hesitating whether it deserves to be cheaper than Broadcom."

Nine: Some Thoughts on the Era

Peter Thiel presents a provocative argument in "Zero to One" that makes many entrepreneurs uncomfortable: competition is for losers; truly great companies create monopolies.

All failed companies are the same—they failed to escape competition.

When it comes to investing, this framework forces you to ask a sharp question: Is the company you're researching struggling in a highly competitive market, or does it hold a monopolistic position in a market it has defined itself?

What's interesting about Marvell is that it's doing two things at once.

In the custom AI chip market, it is a "participant" and a challenger to Broadcom.

In the fields of optical interconnects and high-speed DSP, it is the absolute market leader—the monopolist.

NVIDIA's $2 billion investment essentially confirms Marvell's monopolistic value in the "connectivity" dimension with real capital.

Thiel says one characteristic of monopolies is that the market appears smaller than it is—"Monopolies often conceal their dominant position to avoid attracting regulatory attention."

Marvell, on the contrary: its monopoly status has been overlooked by the market, as everyone is focused on its gap with Broadcom in custom chips.

Conduct a thought experiment: Marvell currently has a market capitalization of approximately $250 billion, corresponding to about $11.5 billion in revenue for FY2027, resulting in a price-to-sales ratio of roughly 21.7x. However, within the $11.5 billion, the data center segment accounts for approximately $9.2 billion and is growing at a rate of over 50%.

Valued separately at Broadcom’s valuation multiple (approximately 25-30 times revenue), this segment alone is worth $230-276 billion.

The market's valuation of $250 billion implies a discount on the data center business and the other businesses being given away for free.

Of course, this "segment valuation" is overly simplistic—Marvell’s data center business would not truly receive Broadcom’s multiple, given differences in gross margin structure, higher customer concentration, and a weaker position in custom chips compared to Broadcom.

But it at least provides a starting point for reflection: the market’s valuation of Marvell is likely still anchored in the old narrative—“that’s the company that lost Trainium3”—rather than the new reality—that it’s the only company globally generating scaled revenue across optical DSP, silicon photonics, and custom AI chips.

My judgment might also be wrong. The competition in custom chips could be more intense than I thought, the growth rate of demand for optical interconnects might not be as optimistic as the model predicts, and Celestial AI’s $1 billion annualized revenue target may not be achievable.

But now, I’m willing to bet on the direction of “AI connectivity.” It’s not because Marvell is the best company, but because it’s in the right position.

Ten: Epilogue — Light and Civilization

At this point, I’d like to step outside the framework of investing and say a few broader things.

Every leap in human civilization, when looked back upon, has not been due to a single-point breakthrough, but rather to an upgrade in connectivity.

Words allow thoughts to transcend time, printing enables knowledge to flow across social classes, the telegraph delivers information across oceans, and the internet connects the entire human brain into a single network for the first time.

Each time, what truly changes the world is not the content itself, but the speed and scope of its flow.

The AI era is replaying the same story.

We’ve focused too much on the “brain”—larger models, greater computing power, smarter reasoning. But the brain has never existed in isolation.

No matter how intelligent a person is, if they cannot communicate with others, they are merely an island. The same is true for a cluster of 100,000 GPUs—if data cannot flow freely between them, even the most powerful computing capacity is just silent silicon.

Light is the messenger of this era.

From a physics perspective, light is the universe’s ultimate speed limit for information transfer. It took us thousands of years to master it—from signal fires to fiber optics, from Morse code to 1.6T DSP signal processing. Now, as humanity attempts for the first time to build a true “silicon-based brain,” we return once again to that same ancient question: How do we make thoughts—whether carbon-based or silicon-based—flow at the speed of light?

Marvell’s story appears on the surface to be a decade-long turnaround for a chip company. But upon deeper reflection, it touches on a more fundamental truth: in any complex system, the value of “connections” ultimately surpasses the value of “nodes.”

In the internet era, the combined value of routers and fiber optics eventually surpassed that of any single server.

In the era of social networks, the value of the platform exceeds that of any individual content creator.

In the AI era, the same logic is playing out again—while everyone competes for the crown of the "smartest mind," the real winner may be the one quietly building neural networks.

We stand at a remarkable and delicate historical juncture. For the first time, humanity has the ability to build something smarter than itself—and whether this thing can truly become “intelligent” depends on our ability to solve what seems like a mundane engineering problem: enabling light to travel freely between chips.

This, in itself, carries a certain poetry.