Ambarella (NASDAQ:AMBA) used an appearance at the 28th Annual Needham Growth Conference to outline its transition from a video processing supplier for human viewing to an edge AI semiconductor company focused on machine perception and on-device inference. In a fireside chat moderated by Needham semiconductor analyst Quinn Bolton, CEO Fermi Wang described how the company’s product roadmap has progressed from CNN-based AI accelerators to transformer-capable architectures, and how that shift is influencing opportunities in security, automotive, robotics, and a new edge infrastructure category.
From video processors to edge AI SoCs
Wang said Ambarella was founded in 2004 and spent its first decade building video processors aimed at human viewing, with GoPro as a well-known customer. He said the company’s strategic shift began around 2012, when it redirected investment toward silicon architectures optimized for convolutional neural networks (CNNs), which Ambarella believed would form the foundation of video analytics and AI.
Transformer-based products and the “next wave”
Wang said a subsequent inflection came after the publication of transformer-related research, which Ambarella believed would become foundational for autonomous driving and other applications. He said the company began designing its third-generation architecture to support transformers in addition to CNNs.
Wang said the company is now ramping its third-generation CV7-series silicon, including CV72 and CV75, which entered production with customers last year and began ramping in the third and fourth quarters. He emphasized that CNNs and transformers will “coexist,” rather than transformers fully replacing CNN workloads.
Looking ahead, Wang pointed to the CV3 family for autonomous driving and robotics and the N1 family for edge infrastructure as key platforms for larger language-model use cases, though he said these markets are still forming as customers evaluate how generative AI will affect their businesses. He said Ambarella’s power efficiency for AI inference is a factor in customer interest and expects increased design wins over the next year or two.
Model sizes, power, and edge infrastructure use cases
Wang described performance ranges across Ambarella’s third-generation lineup. He said the low-end CV75 is a two-watt chip that Ambarella has demonstrated running a two-billion-parameter DeepSeek reasoning model in real time, calling it a breakthrough for edge power envelopes. He said CV75-class devices could address models in roughly the 500 million to 2–3 billion parameter range, and he characterized the cost at roughly $20 “plus-minus.”
At the high end, Wang said the N1 family can run 34-billion-parameter models and could scale to 70 billion parameters or higher depending on capacity, adding that the company envisions a roadmap that could ultimately span from hundreds of millions of parameters up to 100 billion.
On techniques such as distillation and mixture-of-experts approaches, Wang said these could make edge applications more powerful by tailoring models for specific verticals. He indicated that Ambarella’s role is enabling customer applications rather than building the applications itself.
For edge infrastructure, Wang said the N1 “AI box” concept centers on aggregating multiple edge endpoints—such as an installed base of cameras—streaming video into an on-premises box to “upgrade” capabilities without replacing cameras. He gave examples such as hotels or retail environments, where video feeds could enable generative AI-based analytics and data collection beyond traditional security use cases. He said the first edge infrastructure design win goes into production in the second quarter of this year and described the N1-665 as having a three-digit average selling price (ASP), the highest ASP in the company’s portfolio, while maintaining similar gross margins to the broader company profile.
Business drivers and market commentary
Wang said fiscal 2026 performance benefited from more than enterprise security. While he described enterprise security as “very healthy,” he highlighted two markets that ramped in ways that surprised the company:
- Telematics/fleet management: Wang cited “Samsara type of a market,” describing deployments of edge AI (including configurations without cameras) to add value in fleet management, which he said boosted automotive-related growth beyond expectations.
- Portable video: Wang said the category has broadened beyond action cameras into multiple camera types, including 360-degree, wearable, and drone cameras, with many ramping last year.
For fiscal 2027, Wang said growth in those markets should continue, driven by both units and ASPs, but he also noted uncertainty in the market and said the company expects to provide official guidance during its year-end announcement in February.
CES highlights: new products, partners, and semi-custom ASICs
Wang outlined several CES takeaways. First, he said Ambarella expanded its portfolio, which he described as a family of 15 AI chips spanning roughly $15 to $400 in ASP. He highlighted a new CV7 chip that he said delivers 2x to 2.5x higher AI performance with lower power consumption, positioning it as a potential upgrade path for existing CV5 customers.
Second, he said Ambarella introduced an additional go-to-market strategy that complements direct sales by engaging partners such as GSIs and ISVs, which he said is important for scaling into fragmented markets like robotics and for developing edge infrastructure applications. Wang said the company is opening up a platform approach so customers and software vendors can evaluate performance and build applications without requiring Ambarella’s direct involvement for every engagement.
Third, Wang said Ambarella is pursuing semi-custom/custom chip opportunities with large customers who want to incorporate their own differentiated IP without building silicon from scratch. He said Ambarella has its first design win for a 2-nanometer chip, with the customer paying NRE and committing certain value. He added that Ambarella will engage in these projects when the customer is a market leader in a large opportunity and when more than 90% of the chip’s IP comes from Ambarella. Wang said the company could likely support about one such ASIC project per year with current R&D resources, and he indicated the company’s long-term corporate gross margin target of 59% to 62% would still apply under the model described.
On automotive, Wang characterized 2025 as a weak year for the sector overall, citing delays among Western OEMs in Level 2 and Level 2+ decisions as they reassess software strategies and competitive pressure from Chinese OEMs. He said Ambarella continues its business development efforts and believes the company’s autonomous-driving investments will also apply to broader robotics markets. He also said Ambarella has developed two large models supporting end-to-end AI capabilities today and is open to licensing software either as a black box or source code, potentially extending beyond automotive into drones and robotics.
Asked about major Level 2+ design wins, Wang referenced a prior near-win with Volkswagen that did not materialize and said there are no guarantees. He said customers have responded positively to Ambarella’s hardware and software, but he cited company scale versus larger competitors as a recurring challenge and said Ambarella is focused on higher-level engagement to avoid being ruled out late in the process.
About Ambarella (NASDAQ:AMBA)
Ambarella, Inc is a global semiconductor company headquartered in Santa Clara, California, specializing in video compression, image processing and computer vision technologies. The company designs low-power, high-definition system-on-chip (SoC) solutions that enable the capture, processing and streaming of video in a variety of embedded applications. Ambarella’s platforms combine advanced video encoding, multi-core central processing units and hardware accelerators to deliver high-resolution imaging with low power consumption.
Ambarella’s product portfolio caters to multiple markets, including security and surveillance, automotive vision, wearable cameras, drones and robotics.
