Deep|GOOG’s Optical Edge: OCS, Gemini 3, Model Progress, and the Coming Storage Crunch
Why We Believe Google's TPU Will Be a Major Winner in the AI Infrastructure Race
The competition in large-scale AI models is consolidating into a three-way race among Google, OpenAI, and Anthropic. While OpenAI and Anthropic excel in consumer adoption and coding/agent workflows, Google is leveraging its unique TPU + Optical Circuit Switching (OCS) architecture to achieve decisive advantages in scalability, cost, and performance. The upcoming release of Gemini 3 — expected to outperform rivals across multimodality, cost efficiency, and ultra-long-context reasoning — positions Google to potentially claim leadership in overall model capabilities by year-end.
At the same time, a shift from pre-training to mid-training and the explosive rise of multimodal models are driving exponential data generation, creating unprecedented demand for enterprise-grade SSD infrastructure. These dynamics suggest a new cycle of storage shortages, capacity expansion, and supply chain reconfiguration — with NAND flash and advanced interconnect technologies as key beneficiaries.
Google’s Advantages in Model Technology and Hardware Architecture
In the coming month, the most anticipated model is Gemini 3, which we expect to be the standout release of the year. It is poised to deliver across speed, cost, and quality. Over the past few weeks, the Nano Banana model has already drawn widespread attention, signaling a significant leap forward in Google’s overall model capabilities. Over the last two years, Google has gone from being seen as an “AI laggard” to steadily catching up with OpenAI and Anthropic — and by year-end, it could well emerge as the leader in overall model performance.
Industry chatter — including from Semianalysis and others — suggests Gemini 3 is receiving strong early reviews. Its multimodal abilities, cost efficiency, ultra-long-context reasoning, and advancements in coding and agentic tasks are expected to see major improvements. The most critical factor is its exceptionally low cost. Compared with rival models such as GPT-5 and Claude 4.1, Gemini 3 promises superior performance at a lower price point.
The breakthrough is powered by Google’s TPU + OCS architecture. With low-latency, high-bandwidth interconnects, lower costs, and support for massive context windows, this stack may prove to be Google’s decisive edge in surpassing Nvidia and OpenAI.
While Google’s TPU architecture (e.g., v4) has been around since 2021–2022, its potential is only now being realized. In the interim, Google faced setbacks in algorithms and data strategy — losing key talent and struggling with compliance hurdles that limited use of synthetic and agent-generated data.
Now, with talent flowing back into Google and industry know-how converging across algorithmic approaches, the playing field on algorithms and data has leveled. This makes compute capacity the decisive differentiator — and Google’s infrastructure advantage could be the factor that sets it apart.
At present, Google’s TPU architecture appears to be temporarily ahead of Nvidia’s. Specifically, Google’s Ironwood generation of TPUs (also referred to as V6P or V7) can interconnect 9,216 TPU chips via OCS bandwidth to function as a single system, creating a massive 1.77 PB directly addressable HBM memory pool. By contrast, Nvidia’s GB300 series offers only 20.7 TB of scalability per rack — nearly two orders of magnitude smaller. Even Rubin CPX, which will not reach volume production until the end of next year, scales only to 100 TB, still leaving Google with a 17x advantage.
This 3D Torus topology combined with OCS is the architectural foundation that enables Google to maintain a hardware lead and drive step-function improvements in model capabilities.
Future Development Pathways Based on Optical Scale-Up Technology
Scaling up through optical interconnects is widely recognized as the definitive future path, aimed at overcoming the physical limits of copper wiring to unlock higher bandwidth. At present, four main technology approaches are emerging:
Nvidia’s CPO + OIO Approach
Nvidia is driving co-packaged optics (CPO) combined with optical I/O (OIO). While CPO offers power savings, it faces hurdles such as low yields, a closed ecosystem, and customer concerns over lock-in. Large-scale adoption remains difficult. Moreover, the power savings are negligible relative to high-power GPUs, and Nvidia’s high margins dilute its cost advantage. OIO-based scale-up is expected to arrive between 2027–2029. The approach involves embedding OIO within racks, co-packaging GPUs and optics. However, combining expensive GPUs with fragile optical components results in high costs and reliability challenges.
Huawei Cloud’s Pluggable Optics Solution
Historically, pluggable optical modules were considered unsuitable for scale-up. Huawei challenged this view with its LPOD technology and 6,812 pluggable modules, proving the feasibility of the design. While it comes with some cost and power trade-offs, Huawei has demonstrated stable performance that even outpaces Nvidia in certain respects. Meta has also been exploring pluggable optics for scale-up.
Google’s 3D Torus + OCS Architecture
Google has adopted a 3D Torus topology combined with OCS (optical circuit switching) interconnects. This design offers exceptional scalability and lower costs by eliminating the need for electrical switches. The 3D Torus connects nodes in three-dimensional rings, delivering lower latency and stronger network performance. Its limitation is the lack of full non-blocking, all-to-all communication, as messages must traverse multiple hops. However, with OCS integration, Google can achieve dynamic reconfiguration and rapid deployment, dramatically improving scalability. For ultra-large model training (e.g., Gemini), Google has used ring all-reduce to replace all-to-all communication, validating the approach. Ironwood has already been deployed at scale, putting Google temporarily ahead of Nvidia’s GB300. With Google’s TPU v8 and Nvidia’s Rubin both slated for next year, Google may maintain its hardware lead. As algorithms and data converge across players, the competition is shifting back to compute and interconnect architectures — with Meta, Microsoft, and others now also exploring OCS. Long-term, optical and electrical interconnect technologies are likely to co-develop in parallel.
Regarding 3D Torus topology, we first covered this in detail in our January report on CLS and CRDO. Click the article below to learn more.
Broadcom’s Open Strategy
Broadcom is pursuing a more open path, favoring Ethernet for scale-up and scaling back its commitment to CPO. Instead, it is collaborating with players like Meta to advance pluggable optics solutions. Emerging technologies such as LPO and NPO could also shape the next stage of development.
Optical Circuit Switching (OCS) Interconnect Architecture
OCS (Optical Circuit Switch) is a passive device — essentially an optical switch — that directly connects optical modules to TPUs, eliminating the need for traditional electrical switches. The key advantage of this architecture is that it decouples complex signal and data processing, offloading part of that functionality to the TPU and optical module. As a result, optical modules gain in technical sophistication, value contribution, and margins, while partially substituting for the role of electrical switches.
Advantages of OCS
