Silicon Photonics and the Coming Obsolescence of the AI Infrastructure We Just Built

The trillion-dollar AI buildout has a physics problem. The solution is already in production — and the investment window is open right now.

There is a quiet irony embedded in the current artificial intelligence gold rush. The hyperscalers — Microsoft, Google, Meta, Amazon — have committed north of $320 billion through 2027 to build the infrastructure that will run the next generation of AI workloads. NVIDIA's market capitalization briefly touched $4 trillion. New data centers are being planned at a pace that the Department of Energy projects will double or triple electricity demand within four years. The numbers are staggering, the ambition is real, and the fundamental engineering problem at the center of all of it is this: the wire is the bottleneck.

Copper, the material that has carried data between computing components for the better part of a century, is hitting its physical ceiling. In modern AI data centers, up to fifty percent of total energy consumed is spent not on computation but on moving data around and cooling the heat that movement generates. The chips have outpaced the wires. NVIDIA's latest GPU clusters can process data at rates that the electrical interconnects linking them simply cannot sustain without catastrophic energy cost and signal degradation. The CPU made the 1990s. The GPU made the 2010s. The wire is about to make or break the 2030s — and the industry knows it.

The technology that resolves this problem is silicon photonics: moving data not as electrical current through copper but as light through glass. The concept is not new. Fiber optic cables have been moving data between continents since the 1980s. What is new — and what makes this one of the most significant investment themes of the current decade — is the scaling of photonic technology down to the chip level, integrating it directly into the packaging of the processors that run AI workloads. The result is data transfer that is faster, cooler, vastly more energy efficient, and capable of the bandwidth densities that AI at scale actually requires.

Anastasiia Nosova, the semiconductor engineer turned technology analyst behind the Anastasi In Tech platform, has been covering this transition with the advantage of someone who spent years inside the chip design industry before pivoting to explain it to the world. Her framework for understanding the AI hardware challenge is usefully simple: compute, memory, and interconnect are the three pillars, and improving any one of them while neglecting the others just moves the bottleneck. The industry spent the last decade improving compute. It is now paying the price on interconnect. Nosova's assessment of where the next breakthrough comes from is unambiguous: silicon photonics, and specifically the tight integration of photonics with electronics at the chip packaging level. Not as a future possibility — as an engineering trajectory already in motion.

What Silicon Photonics Actually Does

The analogy that clarifies this fastest is the highway. Electrical interconnects — copper traces on circuit boards, pluggable transceiver cables — are the existing road network. They were designed for a certain volume of traffic, they are congested, and they cannot be widened beyond a physical limit without the resistance and heat generation that makes the entire system progressively more inefficient at scale.

Silicon photonics is the new highway. It transmits data as pulses of light through silicon waveguides and optical fiber, moving information at speeds that approach the fundamental limit of physics rather than the engineering limits of a conductor. The bandwidth ceiling is orders of magnitude higher. The power required to move a given amount of data is a fraction of the electrical equivalent. The heat generated — the nemesis of every data center operator staring at an electricity bill — drops accordingly.

The immediate commercial application of this technology is in optical transceivers: the modules that convert electrical signals from a processor into light for transmission and back again at the other end. The data center industry has been upgrading through successive generations of transceiver speed — from 100 gigabits per second a decade ago to 400G, then 800G, and now the ramp toward 1.6 terabits per second, with 3.2 Tbps already in development. TrendForce projects that global shipments of 800G-and-above transceivers will leap from 24 million units in 2025 to nearly 63 million in 2026 — a 2.6-times jump in a single year. LightCounting, the leading optical industry research firm, estimates the AI cluster optics market alone reached $16.5 billion in 2025 and will hit $26 billion in 2026. Sixty percent growth in twelve months is not a trend. It is a structural shift.

But the pluggable transceiver is only the first phase of this transition, and arguably not the most important one. The more significant development — the one that represents both the larger technical challenge and the larger long-term opportunity — is co-packaged optics.

Co-Packaged Optics: Where the Real Money Is

Co-packaged optics, or CPO, takes the principle of optical data transmission and moves it from the cable plugged into a switch to inside the processor package itself. Rather than converting data from electrical to optical at the edge of a board and back again — a conversion that consumes power and introduces latency at every step — CPO integrates the photonic engine directly alongside the switch ASIC or processor die, on the same substrate, in the same package.

The power savings are dramatic. Broadcom's own modeling suggests CPO can reduce system-level power consumption by up to fifty percent compared to pluggable optics. For a data center running a 100,000-GPU cluster — the kind of installation that hyperscalers are building right now — that translates to a reduction in annual interconnect electricity costs of between $8 million and $12 million per year. The total-cost-of-ownership case is not marginal. It is the kind of number that makes CPO adoption not a question of whether but when.

Broadcom and Marvell are both actively developing CPO reference designs for 51.2 and 102.4 terabit-per-second switch ASIC generations, with volume production targeting 2027 deployment by leading hyperscalers. NVIDIA announced in March 2026 a $4 billion investment split between Lumentum and Coherent — the two dominant players in the laser and photonic integrated circuit supply chain — structured as multi-year strategic agreements designed explicitly to secure the optical component supply that NVIDIA's next-generation Rubin GPU platform will require. NVIDIA's Rubin Ultra architecture is being designed from the ground up around silicon photonics networking. In parallel, NVIDIA's Spectrum-X Photonics platform for Ethernet is targeting deployment in the second half of 2026.

These are not exploratory research investments. These are supply chain securing moves by a company that can see its own product roadmap and has concluded that the optical transition is not optional.

TSMC is providing the advanced packaging backbone through its COUPE platform, which unfolds in three stages — from optical engines for current connector form factors through co-packaged integration using CoWoS packaging, toward full native optical integration. GlobalFoundries, which describes itself as the largest dedicated silicon photonics foundry, reported that its silicon photonics revenue doubled in 2025 and is expected to nearly double again in 2026, with what management calls a "clear line of sight" to a $1 billion-plus run rate by the end of 2028. The company's optical networking SAM is projected to grow at roughly 40 percent CAGR through the end of the decade. Tower Semiconductor is investing $650 million to triple its silicon photonics capacity. GlobalFoundries also acquired Singapore's Advanced Micro Foundry in late 2025, positioning its "GF Fotonix" platform as the dominant China-free supply chain option for U.S. hyperscalers — a selling point that geopolitical conditions may make mandatory rather than merely attractive.

The M&A activity tells the story as clearly as any market report. Marvell's $5.5 billion acquisition of Celestial AI. Nokia's acquisition of Infinera. AMD quietly acquiring Enosemi for optical I/O capability. These are multi-billion-dollar bets by companies with direct visibility into their customers' multi-year roadmaps, and they are all pointing in the same direction.

The Supply Chain Chokepoint Nobody Is Talking About

Buried inside the bullish market projections is a materials problem that deserves more attention than it receives in mainstream coverage of this sector. Silicon photonics, despite its name, is not entirely a silicon story. The lasers that generate the light in photonic systems are primarily built on indium phosphide, a semiconductor compound with optical properties that silicon cannot replicate. And the global supply of indium phosphide is severely constrained.

Global demand for InP devices reached two million units in 2025. Production capacity was approximately 600,000 units — a 70 percent supply gap. Orders at major InP suppliers are reportedly fully booked through 2026, with buyers stating publicly that price is not the constraint; securing the quantity is. Only two or three companies in the world manufacture InP substrates at meaningful scale — AXT and Sumitomo control roughly 75 percent of supply between them. New production facilities take years to build. Equipment lead times run 18 to 24 months. And China controls a significant portion of existing production, meaning export permit requirements add a geopolitical dimension to what is already a capacity problem.

This chokepoint has two investment implications that run in opposite directions. For companies in the InP supply chain — the substrate manufacturers, the epitaxial wafer growers, the laser chip producers — constrained supply against surging demand means pricing power and margin expansion in the near term. For the broader buildout of photonic infrastructure, it represents a genuine pacing constraint that makes the most optimistic adoption timelines vulnerable to supply chain reality.

The Incumbent Risk: What Happens to Everything Already Built

Here is where the investment thesis gets uncomfortable for anyone with exposure to the current AI infrastructure stack.

The hyperscaler buildout — the $320 billion in committed capital, the data centers being poured right now in Virginia and Texas and Singapore — is predicated on current chip architecture and its associated infrastructure remaining the dominant paradigm long enough to generate returns on investments that will not fully depreciate for a decade or more. The assumption embedded in that capital allocation is continuity. The photonics transition challenges that assumption directly.

Deloitte's 2026 semiconductor industry outlook states the risk plainly: every generation of chips becomes substantially more efficient, "likely making the incumbent installed base more of a liability than an asset." This is not a hypothetical observation about the distant future. It is a description of a process already underway, in a market where the efficiency gains from the photonic transition are measured not in marginal percentages but in factors. CPO doesn't improve interconnect energy efficiency by ten percent. It improves it by fifty percent. At that magnitude, the total cost of ownership calculus for existing infrastructure becomes unfavorable not gradually but suddenly, as soon as the alternative is available at scale.

The companies most exposed to this dynamic are not primarily the chip designers — NVIDIA has clearly seen this coming and is positioning itself to lead the photonic transition rather than resist it. The exposure is concentrated in the infrastructure layer: the companies that manufacture and operate the copper-based switching and networking equipment that fills current data centers, the companies that financed their buildouts against expected useful lives that may prove optimistic, and the companies whose manufacturing processes are tuned for the current architecture and face expensive retooling to participate in the next one.

The strategic response of the intelligent incumbents — NVIDIA, TSMC, GlobalFoundries, Broadcom — has been acquisition and vertical integration at a pace that indicates genuine urgency. The companies that are not acquiring and not integrating are the ones worth watching with concern.

The Longer Horizon: When Light Does the Computing

The discussion above covers the interconnect layer of the photonic thesis — using light to move data between electronic chips that still do their computation with electrons. This is the near-to-medium term story, the one with product deployments in 2026 and 2027 and clear market size projections through 2031.

The longer and more speculative horizon is photonic computing itself: using light not just to move data but to perform calculations. The physics case for this is compelling. Light-based computation can perform certain classes of mathematical operations — specifically the matrix multiplications that are the core operation of neural network inference — at speeds and energy efficiencies that electronic computing cannot approach. Nosova has described this as the intersection she finds most exciting: not just photonics for interconnect but photonics for computation, doing calculations with light and achieving gains in speed and energy efficiency that represent a qualitative rather than quantitative change in what AI hardware can do.

The honest assessment of timeline, offered by investors closest to this work, is that optical computing power is not a 2026 story and is not clearly a 2028 story either. Lightmatter, the photonic interconnect and computing company backed by serious venture capital and technical credibility, focuses its near-term commercial roadmap on interconnects while treating full photonic compute as the longer horizon. ORCA Computing's PT-3 commercial system, targeting launch in 2026, claims computational output equivalent to approximately 180 GPUs for certain specific tasks — a proof of concept, not a market replacement, but a real demonstration that the direction of travel is not merely theoretical.

The silicon photonics market is projected to grow from $2.3 billion in 2026 to $7 billion in 2031 and $17.8 billion by 2035, at a compound annual growth rate of 25.3 percent. The CPO market specifically is projected at roughly 30 percent CAGR through 2034. These are the numbers for the interconnect layer alone, before photonic computing adds its own layer on top.

The Risks: What the Bullish Framing Leaves Out

The supply chain constraint on indium phosphide is one risk. The others deserve equal honesty.

Manufacturing photonic chips is substantially harder than manufacturing conventional semiconductor devices. The fabrication process requires controlling light wavelengths at nanometer scales in manufacturing environments of extraordinary precision. Yield optimization — the process of reducing the percentage of chips that fail quality testing — is at an earlier stage than it is for mature electronic semiconductor processes. Integration of optical and electronic components in the same package requires solving thermal management, signal integrity, and physical alignment challenges simultaneously, and solving them in volume production conditions, not laboratory conditions. The photonics industry carries significant talent gaps relative to traditional semiconductor manufacturing, and the specialized knowledge base required — spanning optics, electronics, materials science, and advanced packaging — does not exist in depth.

Standardization is a further complication. The optical networking industry is navigating competing form factors and interface standards simultaneously — pluggable optics, onboard optics, co-packaged optics — and the absence of dominant standards slows enterprise adoption, increases integration complexity for system builders, and creates winner-take-most dynamics in sub-segments where the winner is not yet clear.

On the investor side, the most visible entry points in the space — Lumentum (LITE) and Coherent (COHR), the two companies NVIDIA just committed $4 billion to — were already trading at 202 times and 251 times earnings respectively before the NVIDIA announcement. The photonics theme is not undiscovered. The question for public market investors is not whether the trend is real but whether the valuation already reflects it, and whether the timeline uncertainty means that being early is functionally the same as being wrong for a period that tests conviction.

Industry insiders within the photonics supply chain also warn, against the bullish consensus, that reliability, packaging yields, manufacturing capacity, and data routing architectures may each become the next hard constraint on AI scaling in sequence, just as copper, power, DRAM, and NAND each became the bottleneck in their turn. The photonic transition resolves the copper bottleneck. It does not resolve all future bottlenecks simultaneously.

Where the Investable Opportunity Actually Sits

The layered structure of this thesis matters for investment positioning. The analysts closest to this space, including the BankChampaign analysis of the photonics theme from March 2026, have articulated the layers clearly: the computing power layer of AI infrastructure has already been priced into markets — that's the NVIDIA trade of 2023 and 2024. The photonics interconnect layer is being priced now — that's the Lumentum/Coherent trade, the GlobalFoundries trade, the supply chain trade. The optical computing layer is still early, still speculative, and represents the longest-duration bet.

For the investor with pre-IPO access — the "favored investor category" — the most interesting terrain is exactly where Nosova has focused her own entrepreneurial attention: the intersection of electronics and photonics at the component and system level, companies building the enabling technology rather than the finished product. The substrate manufacturers controlling the InP supply. The epitaxial wafer growers. The co-packaged optics module makers at the early commercialization stage. The photonic integrated circuit designers. The companies solving the packaging problem that everyone agrees is the immediate technical gating factor for mass CPO deployment.

The large established players — TSMC, GlobalFoundries, NVIDIA, Broadcom — are the quality-and-safety play, already large, already valued for growth, likely to benefit but with returns that reflect the market's existing awareness of the thesis. The asymmetric opportunity, the kind that generates the 10x rather than the 2x, sits in the companies solving specific technical problems in the supply chain that have not yet been priced by the public market.

The Structural Conclusion

The energy argument is ultimately the one that closes this debate, because it is not a financial argument — it is a physics argument. Global electricity consumption by high-performance computing is expected to rise to between 620 and 1,050 terawatt-hours by the end of 2026, with the higher estimate representing roughly Germany's annual electricity demand. The Department of Energy projects data center power demand to double or triple within four years. Governments in the United States, Europe, and Asia are beginning to treat data center energy consumption as a national infrastructure crisis, not a corporate operating expense problem.

An architecture that cuts interconnect energy consumption by fifty percent is not a nice-to-have in that environment. It is a requirement that the physics of the situation will eventually impose regardless of the capital that has already been committed to the alternatives. The copper wall is real. Light is faster, cooler, and the supply of photons is not constrained by a mining operation in a geopolitically complicated country.

The transition will not happen at the pace the most optimistic projections suggest. It never does. The manufacturing challenges are real, the standardization debates are genuine, and the InP supply constraint is a near-term ceiling on how fast the ecosystem can scale. But the direction is not in question, the capital committed by the companies with the best visibility into their own roadmaps confirms it, and the investment window for the supply chain layer is open right now rather than at the point where every analyst on television has already told everyone else to buy.

The AI infrastructure we are building today is real and necessary and will generate returns. It is also the last generation of this architecture. The next one moves at the speed of light, and the companies positioning for it are worth understanding in detail.

Jonathan Brown (A.A.Sc., B.Sc) writes about cybersecurity infrastructure, privacy systems, the politics of AI development and many other topics at bordercybergroup.com and aetheriumarcana.org. Border Cyber Group maintains a cybersecurity resource portal at borderelliptic.com . He works from a custom-built Linux platform (SableLinux™) which is currently under development and fully documented at https://github.com/black-vajra/sablelinux.

If you would like to support our work, providing useful, well researched and detailed evaluations of current cybersecurity topics at no cost, feel free to buy us a coffee! https://bordercybergroup.com/#/portal/support