Virtusa CEO Admits Losing Money on AI, But is Betting the House on Chips

While much of the IT services industry is busy figuring out how AI will change software development, Virtusa CEO Nitesh Banga has another interesting industry on his mind: semiconductors.

The company expanded into the semiconductor space last year through the SmartSoC acquisition, a move that Banga believes could position Virtusa for opportunities beyond traditional software services. His ambition is to build capabilities spanning the entire technology stack, from chip design and networks to cloud infrastructure and applications.

But the timing is interesting. Virtusa is making this bet while navigating another major technological shift that is already changing the economics of its existing business.

Banga openly admits some of the company’s AI-led projects are losing money, though the economics are a bit more nuanced.

“We are also losing money,” he tells AIM, explaining how the company is still learning to estimate the cost and complexity of delivering projects using AI. Expensive frontier models, token consumption, and the difficulty of modernising decades-old enterprise systems are making the transition more complicated than simply replacing engineers with agents.

For Virtusa, the two bets are connected. As AI moves from software into physical devices, specialised chips, data centres, and connected systems, Banga believes the opportunity for technology services companies will expand well beyond writing code.

From Chips to Cloud to Applications

Virtusa’s semiconductor ambitions are filling a gap in the company’s existing capabilities. “To me, you got to create a service offering all the way from chip to network to cloud to app,” he says.

The company already has a strong cloud-to-application business, but Banga believes it needs to move closer to the hardware layer as AI applications increasingly extend into the physical world.

He points to medical devices such as defibrillators and sleep apnea machines. Building these products requires expertise across chip design, embedded software, cloud infrastructure, and applications. 

This is where Virtusa sees an opportunity.

SmartSoC brings capabilities in pre-silicon design and post-silicon verification—areas that Banga believes are seeing growing demand from established semiconductor companies as well as startups.

The latter are particularly interesting. As chip development becomes increasingly specialised, smaller companies are building processors for specific workloads, including inference and automotive applications.

Banga believes this creates opportunities for companies that can provide engineering expertise across the semiconductor development cycle. He also sees a larger opportunity emerging around the semiconductor-to-data-centre-to-energy ecosystem.

The Changing Economics of Software

The semiconductor bet comes at a time when Virtusa is also rethinking how it builds and delivers software. The company is moving from the traditional software development lifecycle (SDLC) towards what he calls an agentic development lifecycle (ADLC).

The shift goes beyond deploying coding assistants. Virtusa has been building its own engineering harness to coordinate agents across this lifecycle. But the company has learned that simply asking agents to write code faster can create more problems.

The company initially experimented with vibe coding, only to realise that agents struggled to retain context when developers tried to build additional features on top of previously generated code.

There was also the problem of unnecessarily large codebases and technical debt.

The company has since moved towards specification-based and scaffold-based engineering, with tighter controls over how agents generate and maintain code.

The potential productivity gains are significant. Banga says a traditional Scrum team of 11 or 12 people could potentially be reconfigured into four human orchestrators working alongside seven or eight agents.

But fewer engineers do not necessarily mean cheaper projects. “How do you configure and harness in such a way that you’re not hitting the LLM all the time?” Banga ponders.

Token economics has become an important part of the equation. Virtusa is increasingly deciding which tasks need frontier models and which can be handled by smaller open-source models. In some cases, even deprecated models could be sufficient.

The objective is to avoid using an 80-billion-parameter model when an 8-billion-parameter model can do the job.

The Uncomfortable Economics of AI

This transition is also changing Virtusa’s commercial model.

Banga says around 21% of the company’s revenue comes from AI-driven activities. Virtusa separates this into AI-driven, AI-influenced, and AI-assisted revenue.

The company is also moving away from traditional time-and-material contracts towards outcome-based and fixed-price engagements.

That sounds attractive when AI promises to deliver work faster. But fixed-price contracts also mean services companies bear more risk when projects take longer or cost more than expected.

Banga shares the example of a legacy modernisation project involving a 50-year-old platform built across multiple technologies, including AS400 and PL SQL.

While Virtusa achieved the expected outcomes in two technology areas, it struggled with AS400. Some models couldn't fully address the complexity, forcing the company to invest more and partner with specialised startups.

Why AI Companies Still Need IT Services

With Anthropic, OpenAI, and other foundational model developers increasingly moving towards enterprise deployment, there is a growing concern about whether traditional IT services companies will eventually become redundant.

Banga does not think so.

He draws a parallel with the cloud revolution, when platforms such as Salesforce, Workday, and Oracle were expected to reduce the need for technology services. Instead, services businesses found new opportunities in implementation, integration, and enterprise transformation.

He argues that building a capable AI model is different from making it work inside a large enterprise.

Take KYC onboarding. A retail bank and a custodial bank have very different requirements. While retail KYC might be completed within a day, onboarding for a custodial bank can take 43 days. Bringing that down to 15 days requires an understanding of the underlying business processes, data, and regulatory requirements.

Banga believes model makers will eventually need services companies to scale such deployments. He also believes mid-sized services companies could have an advantage over larger competitors such as TCS, Infosys, and Wipro. Larger companies have more legacy business to protect, creating a greater risk of cannibalisation as they move towards autonomous delivery models.

“The bigger the company is, the more difficult it is to kind of turn the wheel,” he quips.

The Changing Workforce

For all the opportunities across AI and semiconductors, Banga says his biggest concern is whether Virtusa can transform its workforce quickly enough.

The company has around 30,000 employees, and traditional technology roles are already changing. Full-stack Java engineers may need to become agentic engineers, database programmers may need to become AI-assisted engineers, and business analysts may have to evolve into what Banga calls process “reimagineers”.

The company has spent the past 12 months working on this workforce transformation.

Banga says even client conversations have changed. 

"What are you doing in AI? How can you help me?" is how he describes the new opening question. "Are we transforming fast enough to be relevant?"

Virtusa is pursuing a $5 billion ambition by 2030-31, with AI, enterprise transformation, and semiconductors forming important parts of its growth strategy.

Banga believes the services industry will remain relevant, just as it did during the internet and cloud revolutions. For now, Virtusa is trying to figure out how to make money from AI while positioning itself for a world where the biggest opportunities may not stop at software.