Co-Packaged Optics: A Decade Away From Reality?

CPO Promises Big but Faces a Long Road to Mainstream Adoption

Co-Packaged Optics (CPO) — the technology that integrates photonics directly alongside switching silicon — is generating enormous buzz across the AI infrastructure world. Yet despite dazzling demonstrations at major conferences like OFC, ECOC, and NVIDIA's GTC, conversations with data center managers reveal a sobering reality: full-scale CPO deployment could still be a decade away.

The gap between lab-ready demos and production-grade deployment is not just a matter of engineering. It reflects deep concerns about operational risk, supply chain maturity, and the cultural inertia that governs how billion-dollar data centers adopt new technology.

Key Takeaways

• CPO integrates optical components directly into switch packages, dramatically reducing power consumption and latency compared to traditional pluggable transceivers.
• AI workloads have reignited interest in CPO, pushing bandwidth demands beyond what current architectures can sustain.
• Most data centers — outside a handful of hyperscalers — have not deployed CPO in any meaningful capacity.
• Industry insiders estimate 7-10 years before CPO achieves broad commercial adoption.
• Operational concerns around serviceability, thermal management, and vendor lock-in remain unresolved.
• Standards bodies and suppliers are bullish, but end users remain cautious and skeptical.

Why CPO Is Back in the Spotlight

CPO is not a new idea. The concept of tightly packaging photonics and electronics together traces back years — to IBM's early research on supercomputer interconnects and so-called 'flyover' interconnect architectures. For a long time, CPO was positioned as a broadly applicable innovation for next-generation networking.

What changed everything is the rise of artificial intelligence. AI has become the dominant workload in modern data centers, and its appetite for bandwidth, low latency, and energy efficiency is insatiable. Training a single large language model like GPT-4 or Llama 3 requires thousands of GPUs communicating simultaneously across high-speed fabrics. The electrical interconnects and pluggable optics that serve today's data centers are approaching fundamental physical limits.

Traditional pluggable transceivers — such as QSFP-DD and OSFP modules — consume significant power just moving data from the switch ASIC to the front panel. CPO eliminates much of this electrical path by placing the optical engine directly on or adjacent to the switch die. This can reduce power consumption by 30-50% per port and cut latency significantly, according to estimates from companies like Broadcom and Cisco.

The Technical Promise Is Real — But So Are the Challenges

From a pure performance standpoint, CPO's advantages are compelling. At OFC 2024, multiple vendors demonstrated 3.2 Tbps CPO-based switch prototypes. Broadcom showcased its Bailly CPO platform, while Intel presented silicon photonics solutions targeting 1.6T and 3.2T speeds. Startups like Ayar Labs and Celestial AI are pushing the boundaries of optical I/O with chiplet-based architectures that could complement or even leapfrog traditional CPO approaches.

However, the challenges are equally formidable:

• Serviceability: Traditional pluggable optics can be hot-swapped in seconds. If a CPO optical engine fails, the entire switch module may need replacement — a scenario that terrifies network operations teams.
• Thermal management: Integrating high-power optical components next to hot switch ASICs creates complex thermal challenges that remain unsolved at scale.
• Testing and qualification: CPO modules cannot be tested independently before integration, making quality assurance far more difficult.
• Supply chain fragmentation: The CPO ecosystem requires tight collaboration between foundries, packaging houses, and optical component makers — a level of vertical integration the industry has not yet achieved.
• Standardization gaps: While groups like the Co-Packaged Optics Collaboration (CPOC) and OIF are working on interoperability standards, the ecosystem is still fragmented.

These are not minor engineering hurdles. They represent structural barriers that affect procurement decisions, operational workflows, and the total cost of ownership for data center operators.

The Psychology of Data Center Adoption

Perhaps the most underappreciated dimension of the CPO debate is human psychology. Data center infrastructure buyers are inherently conservative. They operate facilities where a single hour of downtime can cost millions of dollars. Their procurement cycles are measured in years, not quarters.

When you talk to data center managers about CPO, the conversation quickly shifts from technical specifications to risk management. 'What happens when something breaks at 3 AM?' is a far more pressing question than 'How much bandwidth can it deliver?'

This cultural dynamic creates a classic chicken-and-egg problem. Vendors need large-scale deployments to prove reliability and drive down costs. But operators won't deploy at scale until reliability is proven and costs come down. Breaking this cycle requires either a bold first mover — likely a hyperscaler like Google, Microsoft, or Meta — or a forcing function so severe that the risks of not adopting CPO outweigh the risks of adopting it.

Notably, the relationship between infrastructure buyers and their suppliers is also evolving. Hyperscalers increasingly design their own custom ASICs and networking hardware. Companies like Amazon (with its Trainium chips) and Google (with TPUs) have the engineering depth to experiment with CPO in controlled environments. But enterprises and colocation providers — who represent the vast majority of the market — lack this capability.

Where Hyperscalers Lead, Others Follow — Eventually

The most likely adoption path for CPO follows a familiar pattern in data center technology. Hyperscalers adopt first, absorb the early risks, and help mature the ecosystem. Then, 5-7 years later, the technology trickles down to enterprise and mid-market data centers.

We saw this pattern with 400G Ethernet, disaggregated networking, and liquid cooling — all technologies that hyperscalers pioneered years before they became mainstream. CPO appears to be on a similar trajectory.

Currently, the CPO market is estimated at less than $500 million annually. But analysts at LightCounting and Dell'Oro Group project the market could exceed $5 billion by 2030-2032, driven primarily by AI cluster deployments. That 10x growth projection is significant — but it also underscores just how early the market is today.

Meanwhile, pluggable optics are not standing still. The 800G pluggable ecosystem is maturing rapidly, and 1.6T pluggable modules are expected by 2025-2026. These advances reduce the urgency for CPO adoption by extending the life of existing architectures. Every generation of pluggable optics that successfully scales buys the industry another 2-3 years before CPO becomes truly necessary.

The Competitive Landscape Is Heating Up

Despite the long timeline, investment in CPO and related technologies is accelerating. The competitive landscape includes both established players and well-funded startups:

• Broadcom — Leading with its Bailly CPO platform and deep relationships with hyperscalers.
• Intel — Leveraging its silicon photonics expertise from years of R&D investment.
• Ayar Labs — Raised over $250 million to develop optical I/O chiplets that could serve as a bridge to full CPO.
• Celestial AI — Pursuing a 'photonic fabric' approach that uses light for both chip-to-chip and rack-to-rack communication.
• Ranovus — Focused on single-mode CPO solutions targeting AI/ML workloads.
• TSMC and advanced packaging houses — Developing the manufacturing processes needed to integrate photonics into advanced chip packages.

The involvement of TSMC is particularly significant. As the world's largest contract chipmaker, TSMC's investment in photonics integration signals that the semiconductor industry views CPO as a serious long-term bet — not just a niche research project.

What This Means for the AI Industry

For companies building or operating AI infrastructure today, CPO represents both an opportunity and a planning challenge. The technology is not ready for mainstream deployment, but ignoring it entirely would be a strategic mistake.

Practical implications include:

• Hyperscalers should continue pilot programs and work closely with CPO vendors to shape standards and identify failure modes.
• Enterprise data centers should monitor CPO developments but plan their next 2-3 refresh cycles around advanced pluggable optics (800G and 1.6T).
• Investors should view CPO as a long-duration bet — the market will grow substantially, but returns may take 5-7 years to materialize.
• Network engineers should start building familiarity with CPO architectures now, even if deployment is years away, as the operational model will differ fundamentally from pluggable optics.

The AI bandwidth crisis is real. NVIDIA's GB200 NVL72 racks demand 3.6 Tbps of bisection bandwidth per GPU tray. As models scale to trillions of parameters and multi-modal architectures become standard, the pressure on data center interconnects will only intensify.

Looking Ahead: A Decade of Transformation

The consensus emerging from recent industry research is that CPO will not follow a sudden, disruptive adoption curve. Instead, it will unfold gradually over 7-10 years, with distinct phases.

In the near term (2025-2027), expect continued R&D, pilot deployments at 2-3 major hyperscalers, and ongoing standardization work. In the mid term (2028-2030), early commercial products will reach the market, likely targeting AI training clusters first. In the long term (2031-2035), CPO could become the default interconnect technology for high-performance data centers, finally displacing pluggable optics in the most demanding environments.

The journey to CPO is as much about organizational readiness as it is about technology readiness. Data centers must evolve their operational practices, supply chains, and vendor relationships alongside the hardware itself. Those who start preparing now — even if deployment is years away — will be best positioned when the technology finally matures.

CPO is coming. But the data center industry moves at the speed of trust, not the speed of light.