Sandra Rivera, a Silicon Valley veteran who is the former CEO of Altera, an Intel FPGA spinout, and long-time Intel executive, recently became Chair of the Board of Directors of Paris-based VSORA. VSORA, a technology leader redefining AI inference for next-generation data centers, cloud infrastructure and edge, is focused on addressing high-performance, low-latency inference use cases, a market that is expected to grow to $250B by 2029.
I recently had a chance to talk with Rivera to find out more about what she sees as VSORA’s strengths and opportunities.
Rizzatti: What specific technical bottlenecks in the current AI hardware landscape convinced you that VSORA’s unique architecture provides a solution?
Rivera: One of the things we know now, in terms of the AI industry and the juncture that we’re at, is that the bottleneck for more scalable deployments is not in raw compute—it is absolutely about data movement. And if you look at the current architectures being deployed, they’ve been focused more on heavy training workloads, which GPUs are more suited.
If you look at the problem we’re trying to solve around AI inference—particularly for large models and real-time workloads—we know that performance is not constrained by compute, but by memory bandwidth, latency, and determinism,.
When we look at the problems customers are expressing to us, they want a solution that addresses high throughput, low latency, and a more deterministic set of requirements that they need to deploy in their environments. And VSORA’s architecture directly attacks this problem in terms of how data flows through the system. It minimizes off-chip memory access, maximizes effective bandwidth, and delivers much more predictable latency.
What we’re hearing from customers is that current solutions are increasingly underutilized for inference—perhaps even over-engineered for inference. They’re expensive to scale and power-hungry, whereas the VSORA architecture and product are compelling because they are built specifically to address this bottleneck: high-bandwidth, low-latency, scalable inference.
And these solutions are also quite cost-effective compared to the power-hungry GPUs that were designed to address a different part of the workflow, mainly training.
Rizzatti: Nvidia acquired Groq. Its architecture differs from that of VSORA, but at a 30,000-foot view, the problem it is addressing is similar. Do you view this acquisition as an implicit acknowledgement by Nvidia that the GPGPU is not doing the proper job for LLM processing?
Rivera: The investment by Nvidia in Groq certainly reinforces the position we’ve taken regarding the problem we’re focused on solving. It validates the thesis for the VSORA business, which is that customers are increasingly looking for solutions to address the AI inference problem.
This is about power efficiency, cost per token, and certainly much lower latency and more determinism than you’re able to get from more general purpose computing architectures.
So yes, it’s quite helpful, because the market leader is acknowledging that it is not a one-size-fits-all architecture for addressing all the different elements of an overall AI workflow. Indeed, you need heterogeneous architectures that address different areas of that AI continuum.
For us, it’s a strong validation that the thesis we had—focusing on what is probably the biggest pain point and the largest market opportunity in the coming years—is correct: low-latency, high-performance inference that is cost-effective and power-efficient. And with our innovative architecture, we address those problems head-on.
Rizzatti: In a previous chapter, the VSORA team founded, built, and successfully exited a company called DibCom, an innovator in radio decoding that developed an advanced digital signal processor. In many ways, it laid the groundwork for VSORA. Are you confident this same team can achieve a similar level of success in AI.
Rivera: One of the biggest appeals for me in joining the team—and for prospective customers considering VSORA solutions—is the fact that this is a team that has been working together and delivering products for many years.
They have had 14 successful tape-outs in their history. They have now taped out a very complex logic device on leading process technology, including advanced packaging technology and high-bandwidth memory embedded into the overall platform.
It is not an easy thing for organizations, and certainly for silicon development teams, to come together and deliver complex products. The fact that this team, with this particular leader—the CEO—has done that 14 times before, and now has done it once again in an fast-moving field like AI, really demonstrates the cohesiveness of the team and the deep experience and expertise they have in developing complex silicon products.
I think it is one of the biggest differentiators of VSORA compared with many of the new startups that don’t necessarily have a history of working together or a demonstrated track record of success.
I consider this one of the major positives that prospective customers can feel confident about when choosing VSORA solutions for their AI infrastructure.
Rizzatti: Do you agree or disagree with the “one size doesn’t fit all” narrative that a few major players in AI processing promote? How does VSORA fit into this narrative?
Rivera: As I said earlier, the industry is very much moving toward heterogeneous architectures that address the entire AI workflow.
Even if you look at the role of the CPU as the head node and orchestrator, it handles preprocessing and data cleaning before training. GPUs and highly parallel architectures used effectively for training-heavy workloads with massive frontier models.
Then you move into architectures designed for specific parts of the workflow and different applications. Some focus on media processing, some on networking throughput and others on storage acceleration.
In our case, we are focused on efficient data movement between the processing engine and external memory. We have a unique architecture that addresses the memory wall problem where compute units stall waiting for data, resulting in wasted performance and excessive power draw. Our award-winning, patented architecture enables a very power-efficient, low-latency, and low-cost solution for the AI inference problem.
The need for heterogeneous architectures to address the varying AI workload requirements is not just my belief – you’re seeing a number of strategic partnerships, collaborations, and acquisitions in the industry to support this approach..
Lauro: 2025 was defined by the race to scale. As we look forward to 2026 and beyond, what would you predict to be the main trends?
Rivera: I think the next phase will be defined much more around efficiency, specialization, and economics—going back to having the right architecture for the right part of the workload.
All the data and analyst research show that inference is going to dominate in terms of what enterprises are looking to address for large-scale deployments.
Problems around latency, power efficiency, and deployment costs will matter much more than headline peak benchmark numbers, which is how much of the industry has evaluated solutions to date.
I think we will also see tighter coupling between software and hardware, and architectures specifically designed for inference characteristics will shine compared to solutions repurposed from training accelerators.
In that market landscape, and in meeting customer requirements, VSORA is very well positioned—not because we seek to be different, but because we are aligned with customer pain points and where the market is heading.
We believe we will be a major player in enabling AI to scale sustainably, not just because of the technical solution, but also because of the commercial attractiveness of the offer.
Also Read:
CEO Interview with Naama BAK of Understand Tech
CEO Interview with Dr. Heinz Kaiser of Schott
CEO Interview with Moshe Tanach of NeuReality
Share this post via:
The Risk of Not Optimizing Clock Power