No visible vendor switches
There were no public competitive‑switch signals in the last 48 hours — no mentions of migrations, cancellations, or evaluations away from NVIDIA, AMD, Intel, Trainium/Inferentia, TPU, or Cerebras. Publicly tracked churn and switch risk remain low for now. (x.com)
Baseten’s engineering blog shows active integration with NVIDIA’s Dynamo inference stack and a March 16, 2026 post describing Dynamo-driven KV cache routing and performance gains on Baseten’s platform. (baseten.co) Baseten’s broader financing activity included a $300M round reported in January 2026 that industry coverage says featured a $150M strategic investment from NVIDIA, reinforcing a formal commercial link rather than a move away from NVIDIA hardware. (marketscreener.com) Modal’s documentation and product posts continue to present serverless GPU runtimes (examples call out A100/H100 usage) and product guidance for attaching GPUs via its Python SDK as of late 2025–2026. (modal.com) NVIDIA’s Rubin platform entered production rollouts in early 2026 and the company names CoreWeave among initial Rubin providers, signaling new supplier capacity for customers that prefer NVIDIA’s next‑gen stack. (nvidianews.nvidia.com) Hyperscaler-level shifts remain the bigger multi‑vendor story: Anthropic formally announced a plan on Oct. 23, 2025 to expand use of Google TPUs (up to one million TPUs), and recent reporting documents large OpenAI–AWS Trainium commitments tied to multi‑gigawatt capacity. (anthropic.com) Some startups show explicit alternative alignments — Lamini has been publicly tied to AMD’s Instinct program and the Lamini team was hired into AMD in mid‑2025, while Cerebras announced an AWS + Cerebras disaggregated inference solution on March 13, 2026 — facts that explain why switches, when they happen, are often announced as strategic vendor partnerships rather than short notice churn. (crn.com)
