On September 3, 2026, Broadcom updated VMware Cloud Foundation to 9.1.1 with multi-tenant model sharing in general availability, a gateway to more than 150 models, and a 2.6x jump in Kubernetes scale. For inference on regulated data, compare the total cost of a consolidated on-premises GPU estate against a hyperscaler’s per-token price before you commit.
Stable kernels 7.2.5 and 6.18.51, released September 11, 2026, fix video output for Blackwell GPUs in nouveau and close a memory leak and a use-after-free in ksmbd. Update first if you expose SMB shares or run a Blackwell card.
On 26 August 2026, AWS and NVIDIA announced two million additional GPUs to be deployed by 2028, NVIDIA Vera CPUs, NVHBM memory, and 100,000-GPU federal AI factories. This is no longer a silicon supply deal: AWS is now co-engineering interconnect and memory with NVIDIA. Here is what it changes for a cloud engineer.
University of Toronto researchers published GPUThor, a Rowhammer attack that bypasses the ECC on NVIDIA Ampere GPUs to trigger denial of service or a root-level privilege escalation. Anyone sharing GPUs or running untrusted CUDA has to redraw their trust boundary.
On August 5, 2026, OpenCost teamed up with llm-d to ship the first per-token inference cost tracking for Kubernetes. Splitting allocation cost from usage cost ends the flawed math that was wrongly justifying self-hosted LLMs.
The three dominant GPU terminals of 2026 all offer buttery-smooth rendering and near-instant keyboard response. Raw speed is no longer the differentiator — what matters is whether you want ruthless minimalism, an integrated power-user platform, or a programmable terminal that speaks Lua.
In April 2026, Vultr announced that its Nvidia GPU infrastructure costs 50 to 90% less than equivalent offerings from AWS, Google Cloud, and Azure. Startups and SMBs priced out of hyperscaler margins now have a credible alternative — built around AI agents and transparent per-GPU pricing.
Released on 22 April 2026, Kubernetes 1.36 graduates Dynamic Resource Allocation to general availability. GPUs are no longer an opaque integer count — they become attribute-aware, partitionable resources the scheduler can reason about natively.
