Meta increased capital expenditures for 2026 to $60-65B, allocating most spending to AI data center infrastructure including compute, networking, and storage systems. The company spent $48B in 2025.
AMD released ROCm 6.3 open software platform with updated libraries for PyTorch, TensorFlow, and JAX frameworks. The release includes MIGraphX 2.11 inference engine and enhanced memory optimization for MI300 accelerators.
Cisco launched Nexus 9000 switches with 51.2 Tbps throughput capacity, targeting distributed training workloads across GPU clusters. The networking gear includes remote direct memory access capabilities to reduce training bottlenecks.
Stanford researchers demonstrated video discriminator models trained on human task footage achieve 20%+ higher success rates on unseen robotic tasks compared to robot-only training data. The approach uses Something-Something dataset clips combined with robot interaction episodes.
Scientists evaluated Kolmogorov-Arnold Networks against standard architectures, finding KANs require fewer parameters for symbolic formula tasks but show mixed results on image classification. The architecture uses learnable activation functions on network edges rather than fixed neurons.
Researchers proposed TAPINN neural networks with time-adaptive pattern inference, showing improved performance on temporal prediction tasks. The architecture adjusts inference patterns based on input sequence characteristics.
Autonomous vehicle teams are implementing explainable AI systems that provide passengers with decision rationale through audio, visualization, or haptic feedback. The approach aims to increase rider trust by surfacing factors like detected obstacles or route selection logic.
Medical imaging applications deployed deep learning models for analyzing diagnostic scans, while trading firms adopted vision systems for processing market data visualizations. Enterprises are moving foundation models from research to production environments.
The infrastructure buildout reflects growing compute requirements as organizations scale from prototype to deployment. GPU memory capacity, interconnect bandwidth, and cooling systems are constraining factors for training runs exceeding 10,000 accelerators.
Hardware vendors are releasing annual product cycles aligned with hyperscaler purchasing timelines, competing on performance-per-watt metrics as power costs become a larger portion of total cost of ownership.