Sep 3, 2026 earnings call — Edge AI platform expansion and new X7 accelerator
Reported fiscal Q2 revenue of $108.1 million and announced X7 AI accelerator plus two partner deals.
Revenue $0.1B (est $0.1B) · EPS $0.18 (est $0.17) · Next-quarter revenue outlook $0.1B
Key quotes
"This SoC is sampling now and expected to land initial design wins in edge infrastructure applications where it can serve as an AI coprocessor for host processors such as ARM or x86."
"Combined, these 2 partnership plan to drive a significant amount of incremental revenue over the next 7 years through customers who have largely been unserved by us so far."
"Ambarella itself is also facing rising supply chain costs, and we plan to pass this cost to our customers to maintain our long-term gross margin target of 59% to 62%."
Analyst Q&A 10
X7 accelerator and robotics SAM
Q (Dylan Ollivier · Susquehanna): Is the new X7 chip bundleable with your existing SoC or does it target a different stack? Also, is the Physical AI / humanoid opportunity behind the higher SAM and are there new design wins?
A: The X7 is an accelerator that can be bundled with any host, including Ambarella chips, and it lets customers upgrade AI performance without redesigning the board. Management said the Physical AI opportunity is a big part of the SAM increase, they added more design wins and raised the revenue target, and Louis noted a quadruped robotic dog using the CV72 this quarter.
"So yes, Chris, for the X7, this is -- chip is an accelerator, which can be bundled with any host, including our own chip." — Fermi Wang (CEO)
Memory shortage and partner revenue timing
Q (Joseph Moore · Morgan Stanley): How is the memory supply issue affecting customer demand and your business? And how quickly will the broader ecosystem partnerships (distribution/resellers) start to drive revenue?
A: They are monitoring memory closely and are comfortable with Q3 guidance but see uncertainty for Q4 as they work to secure enough memory. For ecosystem partners, small design wins can generate revenue next year, but management said meaningful revenues from those partnerships will take 2 to 3 years and they expect about $0.5 billion with each partner.
"Right. So when I say 2 to 3 years, we talk about meaningful revenues." — Fermi Wang (CEO)
CapGemini, Macnica roles and Cooper use
Q (Tore Svanberg · Stifel): What early use cases will CapGemini and Macnica help with and will they work with your Cooper platform or bring their own software? Who will help with edge infrastructure sales?
A: Both partners will use the Cooper software platform; that platform and Ambarella's mature SoC were key reasons they chose to work together. Macnica is already winning design wins in drones, retail channels and manufacturing (small, fragmented starts), while CapGemini targets large enterprise customers; management expects the two partners to be complementary and to help penetrate edge infrastructure markets faster.
"In fact, that's a key driver for them to select to work with us because they see a very mature software platform they can immediately tackle on and start building around it, generating infrastructure for their own product line." — Fermi Wang (CEO)
Gross margin target and $9m charge
Q (Quinn Bolton · Needham & Company): Will the Macnica/CapGemini partnerships change your long-term gross margin target? And was the $9 million charge from a semi-custom cancellation and does it affect semi-custom revenue timing?
A: Management said the long-term gross margin target remains 59% to 62% today and they will inform investors if that changes. The $9 million was not a semi-custom opportunity; it was a development project with an automotive autonomy customer and the termination was finalized in Q2.
"Right. So today, I think our long-term gross margin is still 59% to 62%." — Fermi Wang (CEO)
Memory efficiency and X7 memory size
Q (Kevin Cassidy · Rosenblatt): Are you gaining interest because your products use less DRAM than GPU-based embedded solutions? What is the memory architecture inside X7?
A: Management said memory shortages are hurting everyone, but customers come to Ambarella primarily for power efficiency and lower cost, not just memory savings. For X7, they said the accelerator needs much less memory — for example, it only needs 4 megabytes to run a large language model — and current design wins target a 4 to 5 watt power envelope.
"For example, we only need 4 megabytes memory for the accelerator running large language model." — Fermi Wang (CEO)
X7 competitors and semi-custom model
Q (Sujeeva De Silva · ROTH Capital): Is X7 competing mainly with edge GPUs or with other AI chips? And what is driving demand for semi-custom projects now?
A: They said the market is crowded — NVIDIA and Qualcomm have similar products and there are 'probably 50 start-up companies' doing similar chips — and the key differentiator is power efficiency plus mature hardware and software. On semi-custom, Ambarella builds to customer specs but ensures the spec can be sold to other non-competing customers, and they try to use as much of their own IP blocks as possible in semi-custom chips.
"Right now, well, in addition to NVIDIA and Qualcomm having similar products in this market space, there are probably 50 start-up companies doing similar chips." — Fermi Wang (CEO)
Memory impact on margin and seasonality
Q (Liam Pharr · BofA): How much memory cost inflation are you absorbing and does it hurt gross margin? Also, how much of Q3 guidance is seasonality vs demand and what about Q4 and the full-year guide?
A: Management said memory price does not impact Ambarella's gross margin because they do not buy or resell memory; the risk is that higher memory costs could reduce customer volumes and thus Ambarella's revenue. They said the outlook is uncertain due to memory constraints, they are monitoring customers for Q4, and barring memory impact Q4 should follow regular seasonality.
"Right. So first of all, the memory price doesn't impact our gross margin." — Fermi Wang (CEO)
Robotics traction and training help
Q (Auguste Richard · Northland): Are you seeing traction in domain controllers versus perception modules for robots? And can you help customers with training and 3D mapping for robots/humanoids?
A: They said robot systems today look like early autonomous cars: many customers buy perception modules first and demo first-generation products, while a move to domain controllers will happen over time and Ambarella has solutions for both. For training, Ambarella provides a platform to collect data and services to auto-label data; some customers use their system for 3D mapping and automated labeling.
"I think this is how we are at with the current robots. We see a lot of customers are rushing out their first-generation product by putting individual components together to demo their capabilities." — Fermi Wang (CEO)
Sizing methodology for partners
Q (Martin Yang · Opp): You sized potential revenue from CapGemini and Macnica similarly. Did you use the same method to arrive at those dollar figures or different approaches?
A: Management did not provide a detailed method in that immediate answer and instead invited the team to elaborate. They effectively deferred and asked to proceed with the fuller response.
"Yes. Go ahead." — Fermi Wang (CEO)
How models differ for Macnica vs CapGemini
Q (Muneyb Minhazuddin): How did you model revenue for Macnica and CapGemini and are the approaches different?
A: The company said the models differ: Macnica is a volume play aggregating many small and midsized customers (distribution), while CapGemini is a value play targeting large enterprises with complex solutions deployed at scale. They expect faster timelines on the distribution side and longer timelines for the large, complex enterprise deals, and they modeled these outcomes over a multi-year build-out.