Startups and Big Tech Challenge Nvidia’s AI Chip Lead

SambaNova this week raised $1 billion at an $11 billion valuation in a financing led by General Atlantic with participation from Seligman Ventures, T. Rowe Price and Capital Group. The round follows an earlier raise that included Intel and a strategic partner, and investors have put billions into inference-focused chip startups in 2026.

Several startups argue that graphics processing units were adapted from gaming and not optimized for inference, the step in which trained AI models respond to user queries. Companies such as Cerebras, Groq and D‑Matrix promote processors and architectures designed specifically to run inference with higher throughput and lower power use. D‑Matrix reports its chips can execute small inference workloads up to 10 times faster while using about five times less energy than standalone Nvidia GPUs. Groq’s architecture drew licensing interest from Nvidia and staff movement between the firms; unconfirmed reports suggested a multibillion-dollar acquisition discussion.

Cloud providers and large AI firms are also building custom chips to reduce reliance on merchant GPUs and to match hardware to software. OpenAI revealed a custom processor developed with Broadcom called Jalapeño. Google is separating training and inference into different chips with TPU8t and TPU8i planned for release later this year. Amazon Web Services has discussed commercializing its Trainium chips after internal demand. Meta has expanded its Meta Training and Inference Accelerator program and deployed the MTIA 300, with additional generations planned through 2027 focused on inference for assistants and recommendations.

Other companies pursue in‑house ASICs or chip partnerships. A Chinese AI developer is building its own processor to lower dependence on foreign GPUs. Anthropic has explored chip collaboration with a major semiconductor firm. Established suppliers are competing as well: AMD has shifted toward data‑center accelerators and AI chips, growing its market value above $840 billion, while Broadcom designs custom processors for large AI customers and signed a multibillion-dollar semiconductor agreement with a major device maker.

Analysts expect Nvidia’s market share to decline gradually once large in‑house ASIC programs scale. KinNgai Chan of Summit Insights Group expects more competition than a year ago and projects that scaled deployments of custom silicon could begin to dent Nvidia’s dominance around 2027. Morningstar analyst Brian Colello projects Nvidia’s share could fall toward the high‑60s by 2030 while overall AI spending rises.

Nvidia remains the market leader and has increased investment in response. The company spent more than $18 billion on research and development in the fiscal year ended January 2026 and is developing new product families, including “Vera” central processors that the company expects will open additional markets and generate roughly $20 billion in revenue by the end of the fiscal year. Nvidia has taken stakes in photonics firms, acquired inference assets, and allowed some rival chips to be integrated alongside its GPUs in server racks.

SambaNova positions its SN40 and SN50 processors to handle the decode portion of inference. According to company chief executive Rodrigo Liang, those chips can perform that part of the workload five to 10 times faster, which frees Nvidia GPUs for training and other tasks. Some service providers are already offering systems that mix rival accelerators with Nvidia hardware to increase throughput and lower operating costs for customers.

Financial results show Nvidia’s data‑center revenue reached a record level, up sharply year over year, while venture funding and in‑house chip programs at hyperscalers continue to expand the range of hardware options available for AI deployment.