AI Chip Startups: The Companies Challenging Nvidia

Nvidia's graphics processors power most of the world's AI, and the company has become one of the most valuable in history. That success has drawn billions of dollars into startups trying to build faster, cheaper or more efficient alternatives. Some design chips specialised for running AI models; others use light instead of electricity to move data; others bet on open chip designs that companies and countries can license. This guide covers the private AI chip startups in our database, the major players outside it, how their approaches differ and why competing with Nvidia is so hard.

Each company in the table links to a full profile with sourced valuation, funding and investors. Public chipmakers such as Nvidia, AMD, Intel and Broadcom, and startups that have listed, such as Cerebras, are discussed for context.

Private AI chip companies in our database

Why AI chips are a gold rush

Training and running AI models requires enormous numbers of calculations, mostly multiplying large grids of numbers called matrices. GPUs, originally built for video games, turned out to be very good at this, and Nvidia's CUDA software made them easy for researchers to use. As AI spending exploded after 2022, demand for Nvidia's chips far outstripped supply, prices soared and the company's revenue multiplied. AI labs, cloud providers and enterprises have strong incentives to find alternatives: to lower costs, reduce dependence on one supplier, and get chips better suited to specific workloads. That's the opening AI chip startups are pursuing.

The biggest shift is from training to inference. Training a model happens once, on huge clusters; inference, running the model to answer requests, happens billions of times a day and keeps growing as more people use AI and as reasoning models and agents generate more tokens per task. Inference has different requirements, such as low latency, high throughput and low cost per answer, which creates room for specialised chips. Our guide to AI infrastructure companies explains how inference platforms and GPU clouds fit into this market.

The main approaches

Specialised inference chips (ASICs)

Application-specific integrated circuits are designed for one kind of workload. Etched has taken this to the extreme with Sohu, a chip that runs only transformer models, the architecture behind today's large language models. By hard-wiring the transformer into silicon, it aims to deliver far more tokens per second per dollar than GPUs. Its valuation rose from about US$5 billion to US$10.3 billion within months in 2025 and 2026, and it says it has more than US$1 billion in booked orders. The risk is that if AI moves to a different architecture, a transformer-only chip loses its edge.

Groq, another inference chip company, built its Language Processing Unit for very fast responses and offered cloud access to it; in late 2025 it agreed a large licensing deal with Nvidia, under which key staff moved to Nvidia. SambaNova and others also design chips and systems for AI workloads.

Wafer-scale and large-chip designs

Cerebras built a chip the size of an entire silicon wafer, putting enormous amounts of computing and memory on a single piece of silicon to avoid the overhead of connecting many smaller chips. It went public in 2026 in one of the largest technology IPOs of the year. Its success showed that public investors will back AI chip challengers with real revenue, encouraging private competitors.

Photonic interconnect

As models are trained across tens of thousands of chips, moving data between them becomes a bottleneck. Lightmatter uses light rather than electrical signals to connect chips, offering much higher bandwidth with lower energy use. Rather than competing with Nvidia directly, photonics companies aim to be part of next-generation AI systems built by chip designers and cloud providers. Lightmatter was valued at US$4.4 billion in 2024.

Open and licensable designs (RISC-V)

Tenstorrent, founded in Toronto and led by legendary chip architect Jim Keller, combines AI processors with high-performance CPUs based on RISC-V, an open instruction set that anyone can use without royalties. It sells hardware and also licenses its designs to companies and governments that want their own chips, such as automakers and national chip programmes. It raised about US$693 million at a reported valuation of about US$2.6 billion in 2024.

In-house chips from big tech

The largest buyers of AI chips are designing their own: Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia and Meta's MTIA. These custom chips, often designed with partners like Broadcom and Marvell, compete with both Nvidia and startups for the same workloads. Startups must therefore offer something the hyperscalers can't easily build themselves, or sell to the many companies that can't design their own.

Why competing with Nvidia is so hard

• Software: CUDA and Nvidia's libraries have been refined for nearly two decades. Developers and frameworks are optimised for them, so new chips must make porting models easy.
• Pace: Nvidia releases new architectures roughly every year, often with large performance gains, moving the target for challengers.
• Systems, not just chips: Nvidia sells complete systems, including networking, memory and racks, which simplifies buying for data centres.
• Supply chain: access to advanced manufacturing at TSMC and to high-bandwidth memory is limited, and Nvidia is the largest customer.
• Customer risk aversion: data centre buyers are cautious about betting on unproven suppliers for critical infrastructure.

Where startups can win

Despite these barriers, there are openings. Inference workloads are more standardised than training, which makes specialised chips more viable. Many buyers want a second source to negotiate prices and secure supply. Energy limits in data centres reward chips and interconnects that do more per watt. Countries want sovereign chip capabilities and may prefer licensable designs. And as AI usage grows, even a small share of the market can support a large business.

How AI chip startups make money

• Selling hardware: chips, cards, servers or full racks sold to cloud providers, AI labs and enterprises.
• Cloud services: running their own chips in data centres and selling access by the token or hour.
• Licensing: selling chip designs and intellectual property to other companies that manufacture them.
• Engineering services: helping partners design custom chips.

Hardware businesses typically have lower gross margins than software, require large upfront investment in chip design and manufacturing, and face long sales cycles. That's why AI chip startups raise very large rounds before they have significant revenue.

How to evaluate an AI chip company

• Are there independent benchmarks on real models, not just vendor claims?
• How easy is it to run popular models without major engineering work?
• Has the chip reached volume production, and on what manufacturing process?
• Are there named customers with deployed systems, not just letters of intent?
• How does performance per dollar and per watt compare with the latest Nvidia generation?
• What happens if AI model architectures change?

Investors in AI chips

AI chip startups attract a mix of venture investors such as Sequoia Capital and Andreessen Horowitz, growth investors such as T. Rowe Price and Fidelity, strategic investors such as Samsung, LG, Hyundai and AMD, and trading firms such as Jane Street, which use massive computing and have deep technical expertise. Sovereign and state-linked funds also invest as countries pursue chip independence. The diversity of investors reflects how strategic the sector has become.

Geopolitics and export controls

US export controls restrict sales of the most advanced AI chips and chipmaking equipment to China, and China is investing heavily in domestic alternatives from companies such as Huawei and Cambricon. These policies shape where startups can sell, where they can manufacture and who can invest in them. They also encourage other countries, including Japan, South Korea, India and Gulf states, to develop their own chip capabilities, creating demand for licensable designs and partnerships.

What to watch

Watch for independent benchmarks of new inference chips, volume shipments and named customers, Nvidia's response with inference-focused products, more licensing and acquisition deals between startups and incumbents, and further AI chip IPOs following Cerebras. Our guide to the most valuable AI startups covers the model companies that create demand for these chips.

Rebellions and South Korea's AI chip push

Rebellions, South Korea's leading AI chip startup, designs inference chips intended to deliver better performance per watt than general-purpose GPUs. It merged with SK Telecom's Sapeon in 2024 and raised about US$400 million in a pre-IPO round in March 2026 at about US$2.34 billion, led by Mirae Asset and the Korea National Growth Fund. Its rise shows how governments are backing domestic AI chip companies as part of sovereign AI strategies, and it adds an Asian contender to a field long dominated by US firms.