The AI boom is entering a new phase, and the shift is happening faster than many expected. The race is no longer just about training bigger models. It’s about running them efficiently at scale. That shift is pulling billions of dollars into a new class of AI chip startups aiming straight at Nvidia’s strongest position.
For years, Nvidia’s GPUs have powered the rise of modern AI. Chips originally built for gaming became the backbone of model training and deployment. That advantage turned Nvidia into the most valuable company in the world and gave it a grip on the infrastructure layer of AI that few could challenge.
Now that the grip is being tested.
Nvidia AI chip rivals attract record funding as competition heats up
In 2026, AI chip startups have raised $8.3 billion globally, CNBC reported, citing data from Dealroom. The number is expected to climb further if funding momentum holds. Investors are no longer treating AI chips as a niche category. They’re backing them as a core part of the next wave of computing.
Credit: CNBC
The shift comes down to one word: inference.
Training models was the first phase of the AI boom. Running those models in real-world applications is the next. Every chatbot response, recommendation, or AI-generated output depends on inference. That’s where costs stack up and performance starts to matter more than raw scale.
Startups see an opening.
The argument is simple. GPUs were not built for inference workloads at scale. They work, but not efficiently enough when millions of requests hit production systems. New architectures promise faster responses, lower energy use, and significantly lower costs.
“Inference is dominant now, and the existing GPU architecture wasn’t built for it in ways that matter most at scale,” Patrick Schneider-Sikorsky, director at the Nato Innovation Fund (NIF), which has invested in U.K. AI chip startup Fractile, told CNBC.
That thinking is driving a wave of experimentation across the chip stack. Founders are rethinking everything from how data moves through a processor to how memory is accessed, and even how light can replace electricity in moving information.
NVIDIA is not standing still. The company is still moving aggressively to defend its position. In December, it acquired assets from inference startup Groq in a $20 billion deal. Months later, it committed $4 billion to companies working on photonics. Over its last financial year ending January 2026, Nvidia spent more than $18 billion on research and development.
That level of investment makes it clear: Nvidia sees the same shift everyone else does.
At the same time, investors are willing to fund challengers at scale. In the U.S., Cerebras Systems raised $1 billion earlier this year. MatX, Ayar Labs, and Etched have each secured $500 million rounds. In Europe, Axelera and Olix have each raised more than $200 million, with several others preparing nine-figure rounds.
“It’s no longer a niche bet,” said Carlos Espinal, managing partner at Seedcamp, which backed chip startup Vaire Computing. “It’s becoming a core part of how people think about AI infrastructure.”
The funding surge isn’t happening in isolation. A new group of AI chip startups is emerging, each targeting a specific weakness in Nvidia’s GPU model, from latency and energy consumption to the cost of scaling inference.
Top AI Chip Startups Challenging Nvidia in 2026
The companies drawing investor attention are not all chasing the same goal. Each is attacking a different bottleneck in how AI systems run at scale.
Groq is building Language Processing Units designed for fast, predictable inference, delivering low-latency responses for large language models where speed matters most.
Cerebras Systems has taken a different path with wafer-scale chips that pack massive compute onto a single piece of silicon, reducing the need for complex distributed systems.
Lightmatter is betting on photonic computing, using light instead of electricity to move data, cutting energy use and improving speed.
Tenstorrent, led by Jim Keller, is building AI processors around RISC-V, offering a more open and flexible approach across workloads.
SambaNova Systems is focused on reconfigurable architectures that adapt across training and inference, giving enterprises more control over deployment.
Untether AI and d-Matrix are tackling one of AI’s biggest bottlenecks by placing compute closer to memory, reducing data movement, and improving efficiency.
Hailo is targeting edge environments, enabling AI models to run on devices with limited power and space.
Celestial AI is developing optical interconnects that link compute and memory more efficiently, aiming to push past current performance limits.
Etched and Taalas are taking a more specialized approach, building chips designed for specific AI models to maximize performance and reduce cost.
Together, these companies reflect a broader shift in how AI infrastructure is being built. The focus is moving away from general-purpose compute toward designs that are tightly aligned with how AI is actually used in production.
What ties all of these efforts together is a shared bet: that the future of AI won’t be decided by who can train the biggest model, but by who can run it most efficiently at scale.
That’s the pressure building around Nvidia.
The company still holds a commanding lead, with a deep software ecosystem, strong developer loyalty, and the financial muscle to invest across the stack. No startup has come close to matching that combination.
But the market is changing. As demand for inference grows, the cost of running AI becomes a central concern for every company deploying it. That opens the door for alternatives that can deliver better performance per dollar.
The result is a new phase in the AI chip race. Less about raw power. More about efficiency, cost, and control.
And for the first time in years, Nvidia is facing a wave of well-funded rivals built for exactly that moment.
