Breaking the GPU stronghold: emerging competition in AI infrastructure (www.kearney.com)

Nvidia’s long-standing monopoly on AI compute is finally being challenged as cost pressures, supply constraints and strategic lock‑in push buyers toward alternatives. High Nvidia gross margins (~75%) and a tightly integrated stack (chips, DGX/HGX systems, CUDA/cuDNN software, and cloud channels like DGX Cloud and CoreWeave) have motivated server OEMs, hyperscalers and startups to build or adopt competing solutions. Analysts project Nvidia’s share could fall from ~90% today to ~70% by 2030, driven by a $400B+ market opportunity and squeezed OEM margins (from ~10–12% to 3–4%). Competition is emerging on three fronts: rival GPUs (AMD’s MI350X/MI450X with HBM3E and FP4/FP8 support plus improving ROCm; Intel’s delayed Gaudi 3), hyperscaler custom silicon (Google TPU v7, AWS Trainium2/Inferentia2—reported up to 80% inference and 50% training cost savings—Microsoft’s Maia, Meta’s MTIA), and architecture-first startups (Groq’s deterministic LPU v2 for sub‑millisecond token latency, Cerebras’ wafer‑scale WSE‑3 for trillion‑parameter training, SambaNova’s SNL40 and SambaFlow for rack‑scale dataflow). These challengers don’t yet match Nvidia’s full‑stack maturity but target specific bottlenecks—latency, memory bandwidth, power and cost—making “good enough” or specialized solutions increasingly viable for many production workloads and enterprise deployments. The result: more choice, lower costs for some use cases, and a gradual unbundling of the GPU‑centric AI infrastructure stack.