After ‘eye-opening’ early exposure tom the internet, here’s what time in tech has taught Spirent’s Aniket Khosla so far.
What would you describe as your most memorable achievement?
I can’t pinpoint a specific moment, but what stands out to me is the enduring relevance I’ve maintained in an industry known for its rapid pace of change. The fact that I’ve persevered, stayed current, and been actively involved in this dynamic environment is something I find truly remarkable. It’s not a singular event but rather the ongoing journey of adaptation and staying abreast of advancements that I find most memorable.
What first made you think of a career in technology?
From a young age, I’ve always liked to tinker with things. From the time I was in 9th and 10th grade, I used to take things apart and put them back together. I was very fascinated by how things work. Pursuing my engineering degree at a prestigious institute in India further fueled my interest, particularly in computer networking.
Growing up in India, I was exposed to the internet early on, which was eye-opening as I delved into understanding the interconnected systems. Exploring how the internet worked and how devices communicated captivated me. I knew at a very young age that’s where I wanted to focus on for my career. I really wanted to dig in and understand how things worked, and eventually have an influence and shape how they would work in the future.
What style of management philosophy do you employ with your current position?
I have a very open and laid-back management approach. I empower my teams to perform their roles autonomously and I encourage direct communication to foster collaboration and understanding.
I avoid micromanaging and trust my teams to carry out their responsibilities. The way they accomplish tasks and the time it takes to complete them don’t really matter to me. I have an open-door policy and I think interpersonal communication is very important.
Rather than engaging in lengthy email exchanges, I prioritize quick phone conversations. I encourage questions and curiosity, emphasizing that no question is too trivial. I promote a culture of asking and seeking clarity when needed.
What do you think is the current hot technology talking point?
Clearly, the hottest topic out there right now is AI. AI has been around for a while, but with the initial release of ChatGPT, it’s taken the world by storm. In my view, we are just scratching the surface of what AI can achieve. We often draw parallels to the iPhone, which revolutionized the tech landscape, altering how products were created and services delivered. I believe AI is at that early point right now, where it’s just taking off. Everyone is talking about it all the time, and I think we still haven’t realized the potential this has to change the world over the next 5 to 10 years
What are some of the challenges with the AI data center and how do organizations overcome them?
Currently, the landscape of AI data centers differs significantly from the traditional data centers managed by major cloud providers, with most AI data centers being physically or logically distinct entities. AI data centers are expensive to build, with NVIDIA dominating the market with their GPUs, switches, and InfiniBand technology, which comes with a premium price tag.
Hyperscalers are spending billions of dollars building AI data centers which are very complex and will be challenging to maintain and operate. One of the big bottlenecks is the Ethernet fabric that connects these AI data centers. Right now, they’re prone to packet loss and latency, which essentially means that a lot of those expensive GPUs are sitting idle.
We’ve seen some customers try to do early testing on AI data centers, with the Ethernet fabric before they deploy, to make sure they’re getting the most from all of the equipment and GPUs they’ve invested in. It’s challenging to set up the testbed and run it with consistent results to evaluate network performance. A lot of expertise is required to be able to do any kind of fabric testing for these AI data centers.
Throwing more GPUs at the problem is not the solution. The key is optimization (including the network) for maximum efficiency. This requires a continuous, proactive test and verification process to ensure organizations get the most out of the infrastructure as they scale out.
What is Spirent doing along those lines?
Spirent has been a leader in the Ethernet test space for over two decades. We set a challenge to ourselves, to see if we could come up with solutions that could take the complexity out of fabric testing in a pre-production environment. What we really set out to do is give our hardware almost an AI personality, where it would emulate these GPUs and AI workloads on Spirent equipment, without our customers having to go out and buy real GPUs and real servers.
What we’ve done is built these capabilities into our products, where our hardware can now emulate these GPUs in a lab, and our customers can run these tests on the fabric to make sure that there are no performance bottlenecks, which waste GPU usage. We’ve spent the last year developing a solution to emulate and mimic these AI workloads on our equipment in the lab, so that when our customers actually go out to deploy these massive AI clusters, they’re getting the best performance possible out of their AI data center.
What do you see developing in the next 12 months with the AI data center?
NVIDIA has made announcements regarding its upcoming next-generation products, notably the Blackwell line, which promises twice the power and computing capacity of its current H100s while consuming considerably less power. A lot of AI data center environments are power hungry and so it’s a priority for NVIDIA.
At the same time, hyperscalers like Amazon and Meta are venturing into chip development. This shift towards in-house chip production by hyperscalers aims to reduce reliance on single vendors for comprehensive AI infrastructure solutions. Expect a lot of innovation coming out of the hyperscalers, expect significant breakthroughs in energy and power consumption so that these things can continue to scale out the way they’ve been scaling out.
Also, there has to be some disruption in the large language models. AI data centers are running large language models, which is why they need so much capacity. People will start asking, “Do these models need to be so large? Can they be broken down into smaller chunks, with specific areas of expertise?”
Finally, we’re starting to see AI-as-a-Service being offered. Enterprise customers, who don’t necessarily want to build their own AI data centers, will start to rent these services from companies like CoreWeave. Someone else will build a lab and you will just buy the service, like AWS did for the traditional data center. Organizations will offer their AI data centers as a service to enterprise customers so they won’t need to build their own.