Japan Firm Eyes Production of Plastic Fiber for AI Centers

A Japanese company is preparing to bring high-speed plastic optical fiber (POF) cables to mass production, targeting the explosive demand for data center interconnects driven by artificial intelligence workloads. The move signals a potential disruption in the $15 billion data center cabling market, where traditional glass fiber has dominated for decades.

Unlike conventional glass optical fiber, plastic fiber cables promise easier installation, greater flexibility, and significantly lower costs — advantages that could reshape how AI infrastructure is built and scaled across the globe.

Key Takeaways

• A Japanese manufacturer is moving high-speed plastic optical fiber cables from R&D to production stage
• Plastic fiber cables are lighter, more flexible, and cheaper to manufacture than traditional glass fiber
• The technology targets short-range, high-bandwidth connections inside AI data centers
• AI training clusters require massive internal bandwidth, creating a bottleneck that plastic fiber could help solve
• Production timelines suggest commercial availability could begin within the next 12 to 18 months
• The global data center interconnect market is projected to exceed $20 billion by 2028

Why Plastic Fiber Could Transform AI Data Centers

Modern AI data centers are not like traditional cloud facilities. Training large language models such as GPT-4, Gemini, or Claude requires thousands of GPUs working in parallel, exchanging enormous volumes of data at speeds measured in terabits per second. The internal cabling that connects these GPU clusters — often called intra-data center interconnects — has become one of the most critical bottlenecks in AI infrastructure.

Traditional glass fiber excels at long-distance, high-speed data transmission. However, within the confined spaces of a data center, glass fiber presents challenges: it is rigid, fragile, requires skilled technicians for installation, and demands expensive connectors and splicing equipment.

Plastic optical fiber addresses many of these pain points. Made from polymers such as polymethyl methacrylate (PMMA) or perfluorinated materials, POF cables can bend at tighter angles without signal loss, weigh substantially less, and can be terminated with simpler, lower-cost connectors. For AI facilities that may contain hundreds of thousands of individual cable runs across GPU racks, these advantages translate into meaningful savings in both time and money.

Japanese Innovation Targets a Growing Market Gap

Japan has long been a leader in advanced materials science, and its expertise in polymer chemistry positions its companies well to capitalize on this opportunity. Companies like Asahi Kasei, Toray Industries, and several specialized manufacturers have invested heavily in next-generation plastic fiber research over the past decade.

The specific company moving toward production is reportedly leveraging a perfluorinated graded-index polymer fiber design. This approach allows plastic fiber to achieve data rates previously thought impossible for non-glass materials — potentially reaching 100 Gbps per channel or higher over short distances of up to 100 meters. That range is ideal for intra-rack and inter-rack connections inside AI training clusters.

Compared to standard multimode glass fiber, which typically operates at similar speeds but at higher material and installation costs, perfluorinated POF could offer a 30% to 50% reduction in total cabling expenses. For hyperscale operators spending hundreds of millions of dollars on a single AI campus, those savings add up quickly.

The AI Infrastructure Bottleneck Explained

To understand why this matters, consider the scale of modern AI training infrastructure. A single NVIDIA DGX SuperPOD — a popular AI training platform — connects hundreds of H100 or B200 GPUs through a dense fabric of high-speed links. Each GPU communicates with its neighbors at rates of 400 Gbps to 900 Gbps, and the total aggregate bandwidth within a cluster can exceed petabits per second.

The physical cabling required to support these connections is staggering:

• A mid-size AI training cluster may require 50,000 to 100,000 individual cable runs
• Each cable must support at least 400 Gbps with minimal latency
• Cable management in dense GPU racks creates airflow and cooling challenges
• Installation and maintenance of glass fiber in these environments is labor-intensive
• Damaged glass fiber cables require specialized repair, increasing downtime risk

Plastic fiber's inherent flexibility and durability could alleviate several of these issues. Its lighter weight improves airflow management in densely packed racks, while its resistance to breakage reduces maintenance overhead.

How This Compares to Existing Alternatives

Plastic optical fiber is not the only technology vying to solve the AI interconnect challenge. Several competing approaches are under active development:

• Active optical cables (AOCs): Currently the dominant solution, combining glass fiber with integrated transceivers, but expensive at scale
• Copper direct-attach cables (DACs): Low cost but limited to very short distances (under 5 meters) and high power consumption
• Co-packaged optics (CPO): An emerging approach that integrates optical components directly into chip packages, but still years from widespread deployment
• Silicon photonics: Promising technology backed by Intel and others, but primarily focused on transceiver-level innovation rather than cabling
• Hollow-core fiber: Ultra-low latency glass fiber variant under development by companies like Lumenisity (acquired by Microsoft), but extremely expensive

Plastic fiber occupies a unique middle ground in this landscape. It does not require the expensive transceivers of AOCs, offers greater reach than copper DACs, and is far closer to commercial readiness than co-packaged optics. For AI data center operators looking for a solution they can deploy in 2025 or 2026, POF may represent the most practical upgrade path.

Industry Giants Are Watching Closely

The major hyperscale operators — Microsoft, Google, Amazon Web Services, and Meta — are collectively spending over $200 billion on AI infrastructure in 2025 alone. Any technology that reduces the cost or complexity of that buildout attracts intense interest.

Microsoft's acquisition of Lumenisity in 2022 demonstrated that major cloud providers are willing to invest directly in novel fiber technologies. Google has similarly invested in custom networking hardware for its TPU-based AI clusters. If plastic fiber can prove its performance claims at production scale, partnerships or supply agreements with these hyperscalers could follow rapidly.

Japanese manufacturers also have an advantage in the Asian data center market, where companies like SoftBank, NTT, and Sakura Internet are building new AI-focused facilities. Japan's government has committed over $10 billion to domestic AI infrastructure development, creating a ready market for locally produced components.

Technical Challenges Remain Before Mass Adoption

Despite its promise, plastic optical fiber still faces hurdles on the path to widespread adoption. The most significant is signal attenuation — plastic fiber absorbs more light than glass, limiting its effective range. While perfluorinated polymers dramatically reduce this absorption compared to standard PMMA, glass fiber still holds a clear advantage for distances beyond 100 meters.

Other challenges include:

• Temperature sensitivity: Plastic fiber can deform at high temperatures common in dense GPU environments
• Standardization: Industry standards bodies like the IEEE and TIA have not yet fully codified specifications for high-speed POF in data center applications
• Ecosystem maturity: Transceivers, connectors, and test equipment optimized for POF are less widely available than glass fiber equivalents
• Perception: Data center engineers have decades of experience with glass fiber and may resist switching without extensive validation

The Japanese manufacturer reportedly addresses the temperature issue through advanced polymer formulations that maintain stability up to 85°C — sufficient for most data center environments with adequate cooling. Standardization efforts are also underway, with several industry consortia evaluating POF specifications for AI workloads.

What This Means for the AI Supply Chain

If high-speed plastic fiber reaches mass production successfully, the implications extend beyond simple cost savings. The technology could democratize AI infrastructure by making high-performance networking more accessible to smaller operators, enterprise AI labs, and emerging market data centers that lack the capital for premium glass fiber deployments.

For AI developers and businesses, this translates to potentially lower cloud computing costs as hyperscalers pass infrastructure savings downstream. For the broader supply chain, it opens new manufacturing opportunities and reduces dependence on the concentrated glass fiber supply chain, which is dominated by a handful of producers including Corning, Prysmian, and Furukawa Electric.

Looking Ahead: A Production Timeline Takes Shape

The Japanese company reportedly plans to begin pilot production in late 2025, with full-scale manufacturing capacity coming online by mid-2026. Initial products will likely target 400 Gbps and 800 Gbps applications over distances up to 50 meters — the sweet spot for GPU-to-switch connections in AI training clusters.

If successful, this could mark the beginning of a broader shift in how AI data centers are built. The combination of lower costs, easier installation, and sufficient performance for short-range AI interconnects makes plastic fiber a compelling proposition in a market desperate for scalable solutions.

The race to build AI infrastructure is intensifying, and the cables that connect it all together may prove just as important as the chips they serve. Japan's bet on plastic fiber could be the quiet revolution that helps the world keep pace with AI's insatiable appetite for bandwidth.