Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Steve Woo, distinguished inventor and fellow at Rambus
Packaging and interconnects in the spotlight
As compute gets faster each generation, the industry is fighting an uphill battle as package sizes grow and greater numbers of chiplets and HBM stacks and other memories are integrated into larger super chips. Packaging and IOs will further entrench themselves as important enabling technologies and key citizens in chip and system architecture designs. Performance, power-efficiency, and reliability for I/O designs and pin-field engineering will grow in importance as data movement requirements dictate bandwidth needs, which in turn impact physical design constraints, power delivery, and thermal management.
Breakthrough year in robotics: Humanoid & mobile manipulation robots move from demos to real deployments
Expect a noticeable uptick in bipedal/mobility robots doing structured, repetitive tasks in logistics and automotive assembly lines-complementing the million+ mobile robots already in warehouses. Figure AI reported months-long shifts on a BMW body line; Agility Robotics’ Digit is expanding pilots beyond Amazon to GXO and LatAm players; 1X’s Neo is being positioned for industrial use. With scale will come a new focus on safety, provenance, and cybersecurity-researchers are already flagging risks like “botnets on legs.”
Memory supply bottlenecks inside and outside the data center
Data center investments continue to rise with AI, and in particular agentic AI. Greater demand will put pressure on supply chains as data centers consume more memory. Non-data center markets will face headwinds in memory supply as a result.
Memory will be in focus as applications consume higher capacities and bandwidths, driving demand for newer DRAM technologies like HBM4 and GDDR7 as they ramp. Higher capacity and bandwidth memory technologies like stacking, multi-PAM signaling, and CXL-attached pools will see growing interest and adoption to keep accelerators and other computing infrastructure fully fed with data.
Inference to get dramatically cheaper
Inference gets dramatically cheaper and more capable, enabling agentic AI to move into edge systems where designers must balance performance with robust security.
- Advances in Specialized Silicon – Purpose-built accelerators and optimized architectures for AI inference are reducing cost per operation, making high-performance inference increasingly viable outside of hyperscale data centers.
- Energy Efficiency Gains – Improvements in power consumption per inference task means edge devices can run more sophisticated models without prohibitive thermal or battery constraints.
- Economies of Scale and Process Shrinks – Mature 3nm and emerging 2nm nodes, combined with chiplet-based designs, will drive down costs while maintaining application performance targets.
On-device AI becomes a default design choice (“mixed compute” is the new cloud)
By 2026, most premium PCs and phones will ship with NPUs at approx. 40-45 TOPS and OS features that include local inference for speed, privacy, and cost control-while bursting to the cloud for heavier jobs. Microsoft’s Copilot+ PC program (Windows 11, NPU-first features) and Qualcomm’s Snapdragon X platforms demonstrate the hardware baseline; Apple’s Apple Intelligence roadmap expands on-device LLMs, Live Translation, visual intelligence and developer access to the local foundation model.
Expect privacy-centric, battery-efficient experiences (search, transcription, image editing, offline copilots) to become table stakes, with enterprise policy deciding what runs on device vs. in cloud.
Post-quantum cryptography (PQC) will appear in products and networks
With NIST’s PQC standards finalized (FIPS 203/204/205) and NSA’s CNSA 2.0 timelines pressing national systems to migrate, vendors began enabling hybrid Kyber/ML-KEM in TLS across browsers and CDNs. By 2026, PQC-ready libraries, device stacks and policies are expected to be standard-and RFPs will start requiring crypto agility and PQC roadmaps. Short-term friction (larger TLS ClientHello, middlebox breakage) is being resolved; Cloudflare, Google and others have switched to ML-KEM + X25519 in production.
##
ABOUT THE AUTHOR
Steve Woo, distinguished inventor and fellow at Rambus
I was drawn to Rambus to focus on cutting-edge computing technologies. Throughout my 15+ year career, I’ve helped invent, create, and develop means of driving and extending performance in both hardware and software solutions. At Rambus, we are solving challenges that are completely new to the industry and occur as a response to highly sophisticated and advanced deployments. I’m excited about the new opportunities as computing is becoming more and more pervasive in our everyday lives. With a world full of data, my job and my fellow inventors’ job will be to stay curious, maintain an inquisitive approach, and create technologically superior solutions that seamlessly intertwine with our daily lives. After an inspiring workday at Rambus, I enjoy spending time with my family, being outdoors, swimming, and reading.
