Broadcom’s $73B AI Backlog Validates the Photonics Investment Thesis

Now I have comprehensive data. Let me compose the report.

🔑 Key Points

• Broadcom's $73 billion AI backlog—with ~$20 billion in non-accelerator networking and optical content—provides the strongest quantitative validation yet that photonics interconnect has transitioned from niche technology to indispensable AI infrastructure, confirming a multi-year demand runway for optical networking suppliers.

• Nvidia's landmark $4 billion investment in Lumentum and Coherent (March 2, 2026) and Marvell's $3.25 billion Celestial AI acquisition signal that the semiconductor industry's most powerful players view photonics as a strategic chokepoint, making optical interconnect a first-order concern rather than an afterthought in AI data center design.

• Beyond Lumentum and Coherent, the greatest beneficiaries include Fabrinet (contract manufacturing), Ciena (data center interconnect systems), Marvell (custom DSPs + Photonic Fabric), Innolight/Eoptolink (800G/1.6T module volume), and Applied Optoelectronics (vertically integrated US-based manufacturing), each occupying distinct and defensible positions in the expanding optical supply chain.

---

2. Broadcom's AI Backlog: The Definitive Demand Signal

• Broadcom's $73 billion AI backlog encompasses far more than custom XPUs—it explicitly includes switches, DSPs, lasers, and PCIe components, validating the breadth of networking demand.
• Nearly $20 billion of the AI backlog is non-accelerator content, proving that optical networking is no longer a secondary revenue line but a material growth pillar.
• AI semiconductor revenue doubling year-over-year to $8.2 billion in Q1 FY2026 confirms hyperscaler spending acceleration has not plateaued.

2.1 The Numbers in Context

Broadcom reported $73 billion in AI-related backlog, covering custom XPUs, switches, DSPs, lasers, and PCIe components. The company expects these orders to be delivered over the next 18 months, representing nearly half of its total consolidated backlog. This is not simply a chip order book—it is a comprehensive infrastructure procurement commitment that spans the full stack of AI data center connectivity.

Consolidated revenue rose 24% year-over-year to $64 billion, while AI revenue climbed 65% to $20 billion for the year. Hock Tan stated: "We see the momentum continuing in Q1 and expect AI semiconductor revenue to double YoY to $8.2 billion, driven by custom AI accelerators and Ethernet AI switches."

Broadcom AI Semiconductor Revenue Trajectory ($B)

2.2 The $20 Billion Non-Accelerator Signal

The most important detail for the photonics investment thesis lies within the backlog composition. Since nearly $20 billion of the $73 billion backlog is nonaccelerator content, it is evident that networking and optical components are already becoming a material revenue source. During the Q4 earnings call, CEO Hock Tan described demand for optical components like lasers and pin diodes as "just going nuts."

Broadcom's AI switch backlog exceeded $10 billion at the end of fiscal 2025, driven by record bookings for its 102-terabit-per-second Tomahawk 6 switch. This means the remaining ~$10 billion of non-accelerator backlog consists of DSPs, optical lasers, and PCIe switches—components that flow directly through the photonics supply chain.

The implication is stark: for every custom XPU Broadcom ships, it is simultaneously driving billions in optical networking content. This creates a multiplier effect for the entire photonics ecosystem.

2.3 System-Level AI Delivery Amplifies Optical Content

The shift to certified system sales consolidates Broadcom content across accelerators, networking, optics, and interconnects, aligning with customer preferences for turnkey deployments. This system-level approach means optical content per rack is rising, as Broadcom now bundles switches, DSPs, lasers, and transceivers together rather than selling discrete components.

Broadcom CEO Hock Tan noted that networking costs in data centers are climbing, from 5% to 10% of capital expenditures today to an expected 15% to 20% by 2030. This structural shift means the networking TAM grows faster than overall data center capex—a critical tailwind for all optical suppliers.

---

3. Hyperscaler Capex: The $600B+ Spending Engine

• Hyperscaler capex is projected to exceed $600 billion in 2026, with approximately 75% directed toward AI infrastructure.
• Goldman Sachs research suggests consensus estimates have consistently underestimated spending by 30+ percentage points for two consecutive years.
• Power constraints and supply chain bottlenecks—not demand—are the binding limits on infrastructure buildout pace.

3.1 The Unprecedented Scale of Investment

CreditSights projects capex for the top 5 hyperscalers to increase from ~$256 billion in 2024 (+63% YoY) to ~$443 billion in 2025 (+73% YoY) and ~$602 billion in 2026 (+36% YoY). Approximately 75% of the aggregate hyperscaler capex in 2026 will be for AI infrastructure.

Goldman Sachs Research notes that consensus capex estimates have proven to be too low for two years running. At the start of both 2024 and 2025, consensus estimates implied capex growth of roughly 20% for the year. In reality, it exceeded 50% in both years.

Top 5 Hyperscaler Capex ($B)

3.2 Why Optical Networking Captures a Growing Share

As AI clusters scale beyond 100,000 compute nodes, the ratio of optical transceivers per GPU rises dramatically. The rapid expansion of global AI infrastructure is increasingly constrained not by raw compute but by bandwidth, latency, and reliability of the networking interconnects. These clusters require tens of thousands of optical modules to move data efficiently between GPUs, switches, and racks.

The datacom optical component market tells the story clearly. Cignal AI expects total datacom optical component revenue to grow at a 20%+ CAGR from 2024 through 2029, reaching nearly $29 billion by the end of the forecast period. The datacom optical component market exceeded $16 billion in revenue in 2025, representing growth of over 60% year-over-year.

---

4. The Photonics Investment Thesis: From Theoretical to Confirmed

• The copper-to-optical transition is now in full swing across data centers, with 800G deployments at scale and 1.6T ramp beginning in 2026.
• Co-packaged optics (CPO) represents the next inflection point, integrating optical transceivers directly with switch ASICs for radical power efficiency gains.
• Nvidia's $4 billion strategic investment in Lumentum and Coherent is the definitive endorsement that photonics is a strategic chokepoint in AI infrastructure.

4.1 The Copper Wall Is Here

The solution, long-promised but only recently perfected, is Silicon Photonics—the integration of laser-based data transmission directly into the silicon chips that power AI. This transition marks a fundamental shift in how AI clusters are built.

Over 80% of hyperscale data center links now use optical solutions. Scale-out networking has been fully optical for years, but the critical inflection now is scale-up networking—the GPU-to-GPU links within racks—beginning its transition from copper to optics.

Marvell stated: "As bandwidth and reach continue to increase, every connection point in the data center must move from copper to optical. For rack-to-rack scale-out and data-center to data-center connections this transition has already taken place. The next inflection point is within the rack, within the system, and even within a package."

4.2 Nvidia's $4 Billion Photonics Bet

On March 2, 2026, Nvidia made a landmark move that definitively validated the photonics thesis. Nvidia is investing a combined $4 billion in two companies developing photonics technologies. The U.S. chip giant announced it's investing $2 billion in Lumentum and the same amount in Coherent.

The tie-ups include multibillion-dollar purchase commitments from Nvidia and future capacity and access rights to advanced laser and optical networking products from both Lumentum and Coherent. Jensen Huang stated: "Together with Lumentum, NVIDIA is advancing the world's most sophisticated silicon photonics to build the next generation of gigawatt-scale AI factories."

This investment carries three profound implications. First, it confirms that laser supply is a strategic bottleneck. Second, it signals that CPO is transitioning from R&D to production scale. Third, it demonstrates that the two largest AI semiconductor companies (Nvidia and Broadcom) both view photonics as mission-critical.

4.3 The Technology Roadmap: 800G → 1.6T → CPO

For near-term deployments in 2025 and 2026, 800G transceivers in OSFP form factors represent the mainstream choice. The "main upswing" for 1.6T is projected for 2026. Co-Packaged Optics (CPO) is rapidly gaining adoption with 37% share as AI and high-performance computing workloads exceed the bandwidth and power limits of traditional optical and electrical links.

The technology progression creates recurring upgrade cycles that benefit optical suppliers multiple times over:

---

5. Optical Networking Suppliers Beyond Lumentum and Coherent

• The optical supply chain is a multi-layered ecosystem spanning module makers, component suppliers, contract manufacturers, and system integrators.
• Chinese manufacturers dominate 800G module volume, but Western suppliers control critical components and are reshoring production.
• The largest incremental growth opportunities exist in CPO-enabling technologies, coherent DCI systems, and vertically integrated US-based manufacturing.

5.1 Market Structure Overview

According to J.P. Morgan, the largest optical component suppliers are Coherent and Innolight (each with 20% market share), followed by Broadcom at 10%. Numerous smaller suppliers also are contributing to the expanding ecosystem.

Innolight, Coherent, and Eoptolink are the largest suppliers of Datacom modules, with Coherent, Broadcom, and Lumentum as key sources of critical optical components.

Optical Component Market Share (2025 Est.)

5.2 Marvell Technology (MRVL): The Scale-Up Optical Play

Marvell has made the boldest strategic bet among semiconductor companies on optical interconnects. Marvell announced that it has completed its previously announced acquisition of Celestial AI, a pioneer in optical interconnect technology for scale-up connectivity. Celestial AI brings its Photonic Fabric optical interconnect technology, designed to support high-bandwidth, low-latency connectivity across large-scale AI deployments. With this acquisition, Marvell further strengthens its leadership across critical interconnect technologies required for next-generation AI and cloud data center architectures.

Amazon Web Services Vice President Dave Brown said the acquisition will "help further accelerate optical scale-up innovation for next-generation AI deployments." Marvell expects initial revenue contributions from Celestial AI to begin in the second half of fiscal 2028, with revenue ramping meaningfully in the fourth quarter to a $500 million annualized run rate.

The Celestial AI deal positions Marvell uniquely: it can now offer DSPs, custom XPUs, Ethernet switches, and photonic interconnect fabric—the most complete optical-plus-semiconductor stack outside of Broadcom. This makes Marvell the single most important "beyond Lumentum and Coherent" photonics beneficiary from a platform perspective.

5.3 Ciena (CIEN): The Data Center Interconnect Winner

Ciena has emerged as a surprise beneficiary of AI-driven optical demand, particularly in the data center interconnect (DCI) segment. Ciena ended its fiscal year with an 18.8% increase in revenues to $4.77 billion. Management projects FY26 revenues to accelerate 23.7% to $5.9B, supported by a $5B backlog, meaningfully rising purchase obligations, and new hyperscaler wins.

Ciena has emerged from 2025 like the darling of optical networking. Ciena's stock is up by over 160 percent over the last six months. Of the three requirements, scaling across data centers is the biggest driver of growth, partly because Ciena's core business is wide area networking, but it has also started to generate business inside the data center.

Ciena's advantage lies in coherent optical systems for inter-data center connectivity. As hyperscalers build geographically distributed AI clusters, the demand for high-capacity DCI links grows faster than intra-data center networking. Ciena also acquired co-packaged optics startup Nubis Communications, positioning it for the next technology wave.

5.4 Fabrinet (FN): The Precision Manufacturing Backbone

Fabrinet has posted record-breaking sales of $1.13 billion for the final three months of 2025. Exceeding the billion-dollar mark for the first time, the total represents a 36 per cent increase on the same period in 2024.

Fabrinet has built an irreplaceable precision manufacturing moat in optical communications, positioning itself as the essential physical layer for AI infrastructure with 76% of revenue tied to datacenter interconnect and next-generation transceivers.

Fabrinet is the "TSMC of photonics"—it manufactures complex optical transceivers and subsystems for virtually every major player in the ecosystem, including Nvidia, Lumentum, and AWS. The Building 10 expansion—2 million square feet adding $2.4B in annual revenue capacity—represents the largest capacity increase in Fabrinet's history. Accelerating a portion for mid-2026 completion signals customer demand visibility extending 18 months forward.

5.5 Applied Optoelectronics (AAOI): The US Manufacturing Wildcard

AAOI represents the highest-risk, highest-reward play in the photonics supply chain. Total revenue rose 83% from 2024 to a record $456 million. Murry said that AOI expects to generate over $1 billion in revenue in 2026 with a non-GAAP operating profit of over $120 million, describing the revenue target as limited by production capacity and supply chain rather than demand.

The company's strategic value lies in its vertically integrated US-based manufacturing capability. It's stacking up design wins with hyperscale customers (Microsoft, Amazon, Oracle) for 400G and 800G transceivers. It signed a $4 billion, 10-year purchase agreement with Amazon. This domestic production capability has become increasingly important as geopolitical tensions and tariffs threaten Chinese optical supply chains.

5.6 Other Notable Beneficiaries

---

6. The Co-Packaged Optics Inflection

• CPO integrates optical transceivers directly with switch ASICs, reducing power consumption by 3.5x and eliminating pluggable module constraints.
• Nvidia and Broadcom have both committed to CPO architectures for next-generation switches, creating an inevitable demand ramp.
• The CPO market will create new winners, particularly laser source suppliers and advanced packaging specialists.

6.1 Why CPO Is Inevitable

Andrew Schmitt of Cignal AI stated: "AI data centre interconnects face growing bandwidth and power challenges. Co-packaged optics – integrating optics directly onto XPUs and switches – is the inevitable solution."

The technical breakthrough that solved this is Co-Packaged Optics (CPO). In early 2025, Nvidia unveiled its Quantum-X InfiniBand and Spectrum-X Ethernet platforms, which moved the optical conversion process inside the processor package using TSMC's COUPE technology. These systems support up to 144 ports of 800 Gb/s, delivering a staggering 115 Tbps of total throughput. By integrating the laser and optical modulators directly onto the chiplet, Nvidia reduced power consumption by 3.5x compared to traditional pluggable modules.

6.2 The CPO Supply Chain

At the VLSI conference, Nvidia highlighted several laser partners within its ecosystem: Lumentum for single high-power DFBs, Ayar Labs for DFB arrays, Innolume for quantum-dot mode-locked combs, and Xscape, Enlightra, and Iloomina for pumped nonlinear resonant combs.

CPO fundamentally shifts value in the optical supply chain. Pluggable transceiver module makers face disruption, while laser source suppliers and silicon photonics foundries gain leverage. This is precisely why Nvidia invested $4 billion to secure capacity from Lumentum and Coherent—they are the irreplaceable laser engine suppliers.

---

7. Risks and Counterarguments

• Customer concentration remains the largest systemic risk, with a handful of hyperscalers driving the majority of demand.
• Chinese optical suppliers present persistent competitive and geopolitical risk to Western manufacturers.
• Overbuilding cycles have historically punished optical companies with severe margin compression.

7.1 Concentration and Cyclicality

If hyperscalers decide to slow AI capex to digest the capacity they have already ordered, the $73 billion AI backlog will still exist but growth rates could compress quickly. Customer concentration is the second risk: a meaningful share of AI revenue depends on a small set of hyperscalers.

The optical networking industry has experienced brutal boom-bust cycles before. The dot-com crash devastated optical companies, and the 2023–2024 telecom capex pause similarly hurt suppliers. While these buildouts reinforce a bullish stance, it's important to recognize the cyclical downside. Once demand slows, these companies will lose significant operating leverage due to underutilized capacity.

7.2 Chinese Competition

Chinese manufacturers, led by Innolight and Eoptolink, have captured the majority of the merchant market share for 800G transceivers, securing approximately 60% of Nvidia's aggressive volume requirements. Western suppliers face a persistent cost disadvantage in commoditized transceiver modules, though they retain advantages in critical components, CPO technology, and US-based production.

The geopolitical angle cuts both ways. A 25% tariff on Chinese optical imports creates headwinds for Chinese suppliers but also raises costs for hyperscalers who depend on them. The period from 2025 marks a clear strategic shift toward onshore production. The launch of Mesh Optical Technologies in Los Angeles with an explicit "American-made" strategy is a direct response to tariff risks.

7.3 The DeepSeek Question

The January 2025 DeepSeek efficiency breakthrough briefly spooked investors into fearing hyperscalers would cut capex. Following the DeepSeek news, investor concerns were high that hyperscalers might reevaluate spending plans. However, those concerns have now been firmly put to rest, as every major hyperscaler not only maintained their aggressive investment plans but also issued capex guidance that exceeded Street estimates.

The counterintuitive reality: efficiency improvements in AI models actually increase total compute demand by enabling new use cases and expanding the addressable market. This is Jevons' Paradox applied to AI—lower cost per inference drives exponentially more inference.

---

8. Investment Framework: Positioning Across the Photonics Value Chain

• Investors should consider exposure across multiple layers of the photonics supply chain rather than concentrating in a single name.
• The highest conviction plays are companies that control laser source technology, the most constrained link in the optical chain.
• Valuation matters—many optical stocks have re-rated dramatically in 2025–2026, compressing the margin of safety.

8.1 Tiered Opportunity Assessment

8.2 The Underappreciated Opportunity

My strongest conviction "beyond Lumentum and Coherent" pick is Marvell Technology. Marvell is betting that the future of AI isn't just bigger chips, but better wiring. If they are right, they just bought the nervous system of the 2026 data center.

The combination of Marvell's existing DSP and switch leadership with Celestial AI's Photonic Fabric technology creates the only company capable of offering a complete optical-to-digital connectivity platform for scale-up AI clusters. Marvell expects meaningful revenue contributions from Celestial AI to begin in the second half of fiscal 2028, reaching a $500 million annualized run rate in the fourth quarter of fiscal 2028, doubling to a $1 billion dollar run rate by the fourth quarter of fiscal 2029. While the revenue timeline is distant, the strategic positioning is already being priced in as customers design next-generation architectures.

Fabrinet is the second most compelling name. It is the indispensable contract manufacturer for the optical ecosystem, and its capacity expansion directly correlates with industry growth. Sales are expected to increase 28.5% this year and 16.9% next year, while earnings are projected to grow 30.7% and 16.1% over the same periods.

---

9. Broadcom's March 4 Earnings: The Next Catalyst

• Broadcom reports Q1 FY2026 earnings on March 4, 2026, with the street expecting $19.1 billion in revenue and $8.2 billion in AI semiconductor revenue.
• Backlog commentary—whether the $73 billion figure grows, holds, or declines—will be the single most important data point for the photonics thesis.
• Networking mix and optical component demand color on the call will directly influence sentiment across the entire optical supply chain.

Broadcom's fiscal 2026 revenue estimate of $97.6 billion would be an improvement of 53% over last year. Broadcom's AI revenue is expected to grow over 150% YoY in FY2026, driven by $73 billion in AI backlog scheduled to be recognized over the next 18 months.

The key variables to watch:

1. Updated AI backlog figure — Does it grow beyond $73B as new orders offset shipments?
2. Non-accelerator mix — Does the $20B networking/optical backlog increase as a proportion?
3. Tomahawk 6 CPO commentary — Any acceleration in CPO deployment timelines?
4. Fifth XPU customer ramp — Does the new hyperscaler customer accelerate networking demand?

If Broadcom delivers on guidance and provides constructive backlog updates, it will serve as a catalyst not just for AVGO shares but for the entire photonics ecosystem—Lumentum, Coherent, Fabrinet, Ciena, AAOI, and Marvell alike.

---

10. Conclusion: The Photonics Super-Cycle Is Real

The evidence is now overwhelming. Broadcom's $73 billion AI backlog, Nvidia's $4 billion photonics investment, Marvell's $3.25 billion Celestial AI acquisition, and $600+ billion in hyperscaler capex collectively constitute the strongest validation of the photonics interconnect thesis in the industry's history. This is not a speculative bet on future technology—it is a confirmation that optical networking has become as essential to AI infrastructure as the GPUs themselves.

The investment opportunity extends far beyond the obvious names of Lumentum and Coherent. The optical supply chain is deep, multi-layered, and capacity-constrained at virtually every node. Marvell's scale-up photonic fabric, Fabrinet's precision manufacturing, Ciena's DCI systems, Applied Optoelectronics' US-based production, and even the fiber and connector infrastructure from Corning and Amphenol all represent distinct ways to participate in what is shaping up to be a multi-year photonics super-cycle.

The primary risk is not that demand disappoints—it is that supply cannot keep pace. That very constraint, paradoxically, is what makes laser source suppliers and manufacturing capacity holders so valuable in the current environment.

---

📚 Recommended Topics for Further Exploration

• Co-Packaged Optics (CPO) Technology Deep Dive: Understanding the specific silicon photonics architectures being deployed by Nvidia and Broadcom, and which component suppliers benefit most
• Marvell vs. Broadcom: The AI Networking Platform War: How Celestial AI's Photonic Fabric competes with Broadcom's Tomahawk 6 CPO architecture
• Geopolitics of Optical Manufacturing: The reshoring trend, Chinese supplier dominance in 800G modules, and tariff implications for the optical supply chain
• AI Data Center Power Constraints: How energy limitations are becoming the binding constraint on infrastructure buildout, and why power-efficient photonics technologies gain premium value
• 1.6T Transceiver Technology Competition: Silicon photonics vs. EML vs. VCSEL approaches, and which technology wins at scale for next-generation AI networks
• The Scale-Up Interconnect Transition: Why the copper-to-optics shift within GPU racks represents a potentially larger market than the already-optical scale-out networking segment