Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Ren� Jonker, Chief Product Officer, Soitec
AI is scaling faster than the physical infrastructure built to support it. Across the industry, our data centers are seeing unprecedented power consumption, an issue that only evolution in semiconductor materials will solve.
In 2026, as data centers continue to face constraints with bandwidth, power, heat, and instant connectivity, the next wave of innovation won’t be models or GPUs. It will be the infrastructure they run on – with silicon photonics at its core.
Prediction #1: AI Scale Makes Data Movement the New Bottleneck
AI models are now being trained across thousands of computers at once. According to Andreessen Horowitz, frontier model training requires hundreds of thousands of GPUs – meaning datacenters must not only supply enough power for these workloads but also ensure the underlying infrastructure can support the speed required for these GPUs to communicate in real time.
However, the speed of AI training doesn’t depend solely on the semiconductor chips. As models become more mature, larger, and more distributed, the real spotlight shifts to the interconnect – the bandwidth and latency of the network fabric that links all these GPUs together. This is the system responsible for moving data between nodes, synchronizing models, and ensuring GPUs aren’t left waiting on missing information.
Even as GPUs get faster, the overall output can’t improve if information can’t move seamlessly and quickly between them. In 2026, the true differentiator for AI at scale won’t be raw compute, but advanced interconnect solutions that can move massive datasets at extremely high speeds with minimal delay.
Prediction #2: Expanding Model Sizes Will Continue to Compress Technology Refresh Cycles
GPUs aren’t the only part of AI contributing to infrastructure strain. Across the industry, AI model sizes are growing at an exponential pace, and that acceleration shows no signs of slowing.
In January of this year, Chinese AI startup DeepSeek released DeepSeek-R1, a model with nearly 671 billion parameters. Just months later, the company followed with DeepSeek-V4 MoE, a 1-trillion-parameter model, the largest open-access MoE model to date. That kind of scale jump within a single year underscores how quickly the industry’s definition of “large” is shifting.
This unprecedented growth is why, in 2026, expanding model sizes will continue to compress technology refresh cycles. Datacenters will be forced to double available bandwidth roughly every two years, accelerating the industry-wide transition toward advanced optical interconnects capable of supporting these workloads.
Prediction #3: Density and Power Limitations of Copper Shifting Connectivity Architecture in AI Clusters Towards Silicon Photonics
The strain on our data centers worldwide is undoubtedly intensifying. A recent report from the IEA underscores this, estimating that global electricity consumption from data centers is expected to nearly double over the next five years, rising from 415 TWh in 2024 to 945 TWh by 2030.
This is especially true for data centers with scalable architectures. The density and power limitations of copper will, in tandem, limit bandwidth, thermal performance, and system density. This is where silicon photonics (SiPho) comes in. This is already a key material to enable high-speed, energy-efficient optical connectivity; and in 2026, we’ll see more mainstream adoption of SiPho extending from pluggable modules to Co-Packaged Optics (CPO), where the optical engine is integrated directly with the XPU to minimize signal loss and latency.
Prediction #4: Silicon Photonics and CPO Will Solve the Data Center Energy Problem
As AI clusters grow larger, a significant portion of cluster energy is not solely consumed by the GPUs themselves but the networking layer connecting these accelerators. It’s a constant wave of connection, data transfer between thousands of accelerators and this is why optical networking is becoming so critical.
In 2026, SiPho and CPO will be the technologies that drive down power consumption across these networks. SiPho is so critical because it reduces heat and power loss by moving data as light instead of electricity. CPO integrates optical connectivity directly into the processor housings to increase both bandwidth and energy efficiency. To support this shift, SOI (silicon-on-insulator) will increasingly become the baseline material platform, offering lower energy loss, higher bandwidth, and better thermal performance.
Next year, efficiency will matter as much as raw speed. AI networking will be judged on how much performance it can deliver per watt, and SiPho, CPO, and SOI will be the foundational technologies enabling that shift.
Prediction #5: The Overlooked Materials Shaping the Future of AI
AI accelerators generate significant heat and require higher electricity consumption to run efficiently. This puts a strain on traditional silicon-based power electronics and we’re seeing industry shift toward advanced wide-bandgap materials like Silicon Carbide (SiC) and GaN platforms, which are foundational technologies for the semiconductor and energy markets.
Next year, these materials will be the foundation for scaling the next era of AI, enabling data centers to operate at extremely high temperatures without sacrificing performance or increasing energy grid strain.
As we look to 2026, we’ll see a heightened spotlight on the semiconductor materials powering the next wave of AI. The industry breakthroughs won’t be AI models, GPUs, or more data center sites, but instead the infrastructure that makes all of this possible. The companies that invest in this will be the ones that scale AI efficiently, and at the pace this technology now demands.
##
ABOUT THE AUTHOR
Ren� Jonker, Executive Vice-President, Chief Product Officer, he has over 20 years of semiconductor industry experience.
After beginning his career at Philips Semiconductors, he transitioned to NXP, leading an RF IC development site in the US, and continued to work on FE technology from Singapore. Ren� later joined Nexperia in Hong Kong as head of Operations and subsequently became Operations Manager for Automotive at ams Osram in Singapore.
His international experience spans Europe, the United States, Hong Kong and Singapore, where he is based, showcasing his proficiency in managing multi-site and multicultural teams.
