Nvidia is facing mounting competition from both hyperscale cloud providers and traditional chip rivals as the AI computing market shifts from training to inference workloads. At its GTC 2026 conference in San Jose, the company unveiled new hardware aimed at maintaining its lead, even as industry dynamics evolve rapidly.
The strongest challenge comes from cloud giants accelerating their in-house silicon programs. Google, Amazon, Microsoft, and Meta are all developing custom chips designed to reduce reliance on Nvidia’s GPUs. Google’s latest TPU, Ironwood, delivers 4.6 petaflops of FP8 compute per chip and scales to thousands of units. Amazon recently began production of its Trainium 3 on TSMC’s 3nm process, touting a 50% cost reduction compared with equivalent Nvidia-based instances. Meta plans four new chip generations, while Microsoft introduced its Maia 200 inference processor.
Much of this custom silicon is being built in partnership with Broadcom, which now dominates the AI ASIC design market. Counterpoint Research projects Broadcom will hold about 60% of that market by 2027, with AI server ASIC shipments among hyperscalers expected to triple over the same period. The company’s AI revenue reached $8.4 billion last quarter, doubling year over year.
Nvidia’s response centers on its newly announced Vera Rubin platform and Groq 3 inference processor, developed through a $20 billion technology deal with startup Groq. The company claims these chips deliver up to 10 times higher inference throughput per watt than existing systems. CEO Jensen Huang projected cumulative orders could reach $1 trillion between 2025 and 2027, bolstering the company’s long-term outlook.
However, inference economics are reshaping the market as specialized hardware gains traction. Nvidia’s software ecosystem remains a major advantage, but AMD and major cloud providers are investing heavily in alternatives, signaling a more competitive era for the AI chip leader.
Read Article: OpenAI Plans Desktop ‘Superapp’ Combining ChatGPT, Codex, and Browser
