As agentic workloads multiply, some companies see cloud token bills eat through budgets faster than planned. That pressure is bringing on-premises infrastructure and open-weight models back into enterprise AI planning, Dell and Intel say.
At the 2026 Open Compute Project summit, Meta presented new open networking hardware for AI, including its first liquid-cooled network rack, IcePack, and a 102.4T switch platform, IceCube. It also detailed Flattened Planar Fabric, a new way of connecting GPUs to the network, and MetaRoCE, an open-source AI transport protocol.
Deepgram, which expanded its Flux voice model to 10 languages this year, is pushing voice AI onto air-gapped laptops. As voice becomes a main way of using devices, the company needs compute, storage and partners to support both listening and speaking models.
As companies move AI projects from pilots into production, Dell is changing how it sells to West Coast customers who build much of the AI stack themselves. Infrastructure vendors are finding that these builders call for a different sales approach.
Meta explains how it uses idle inference capacity across its fleet to run extra ads recommendation work. Two elasticity mechanisms let the workload expand onto spare machines without hurting baseline performance.
At the Open Compute Project Global Summit in San Jose, NVIDIA announced new partnerships and contributions, including opening its Vera Rubin architecture. The company argues open standards help build AI infrastructure faster and more reliably as facilities grow from megawatts to gigawatts.