Why The AI Race Is Really An Infrastructure Race

The most dangerous mistake in the AI race is assuming that the smartest model will win it.

For the past few years, much of the conversation has focused on model performance, benchmark scores, investment rounds and eye-catching demonstrations. Yet inside large organizations, the biggest obstacle is often far less glamorous. It is the aging software, fragmented data, weak testing, slow release processes and technical debt sitting underneath the AI strategy.

In a recent conversation with Matthias Patzak, AWS executive in residence and former CTO, and Joel Hron, CTO of Thomson Reuters, one idea came through clearly: the AI race is becoming an infrastructure race.

Companies may have access to similar foundation models, but they do not have equal ability to connect those models to trusted data, monitor agent behavior, control costs, secure workflows and release improvements at speed. That capability is becoming a source of competitive advantage.

The Model Is Becoming The Least Interesting Part

Patzak put it bluntly: “There is no best model right now because every use case, and even different stages in a use case, require different models.”

That challenges the obsession with choosing a single winning model. Enterprises will use a mix of models selected for accuracy, speed, price, privacy or specialist reasoning. More advanced organizations may route requests dynamically at runtime.

The harder question is what the model knows about your business. Does it understand your customers, products, processes, policies and risks? Can it access the right context at the right moment? Can the organization see what an agent is doing and how much the interaction cost?

As Patzak explained, “Infrastructure is really becoming the differentiator and it’s no longer the model.”

I agree. Access to high-performing models will spread. The lasting advantage will come from proprietary data, enterprise context, workflow integration, security, observability and execution discipline.

AI Exposes Every Weakness You Already Have

Patzak described AI as “an amplifier,” which is one of the most useful ways to think about enterprise adoption.

A company with clean data, strong engineering practices and fast decision-making can use AI to accelerate progress. A company with brittle systems, unclear ownership and weak controls may simply produce mistakes faster.

This is especially visible in software development. AI coding tools can generate more code, but increased output puts pressure on testing, security reviews, deployment pipelines and operational monitoring. If those systems are weak, AI can overwhelm them.

Hron made the same point from the Thomson Reuters perspective. The company has existed for more than 150 years and sells more than 100 products across legal, tax, audit, compliance, risk and news. Its technology estate has evolved across many generations.

Thomson Reuters began its modernization journey with AWS close to a decade ago, initially driven by cost efficiency, developer productivity and scalability. Those earlier investments later became an important foundation for AI.

What Thomson Reuters Learned From Modernizing Early

The emergence of agentic AI has raised the stakes again.

Traditional software is usually designed around human behavior. A lawyer might search for a case, read it, consider the findings and then perform another search. An AI agent can perform comparable actions at far greater speed and volume.

Hron explained that an agent using a platform such as Westlaw can create “orders of magnitude more” activity than a human user. Systems designed for hundreds of human interactions may struggle when agents generate hundreds of thousands of searches.

This shift forces companies to redesign applications for agent-native use cases. Token usage, data consumption, compute demand and transaction volumes can rise dramatically when agents operate across enterprise systems.

Hron said Thomson Reuters discovered benefits that were difficult to predict at the start: “As we modernized, we were able to capture upside and opportunities that we did not forecast we would be able to capture when we started.”

Modernization programs are often judged through cost savings alone. AI changes the calculation. The upside may include faster product development, stronger security, new revenue streams, improved scalability and services that were previously impossible.

From Big Bang Projects To Continuous Modernization

Modernization once meant a large, expensive, multi-year transformation program. These projects carried significant risk, uncertain outcomes and a painful gap between investment and reward.

AI is beginning to change the process itself.

Patzak described a shift from modernization as a one-off project to modernization as a continuous capability. AI tools can scan repositories, identify outdated dependencies, flag policy violations, recommend fixes and generate pull requests as part of the development process.

He said AWS customers have seen modernization velocity increase by four times, with cost reductions of 40 to 60 percent in some cases. Those figures will vary by organization and workload, but the direction is clear. Automation can turn migration tasks that once took years into work measured in months or weeks.

Hron offered the example of moving a legacy .NET application to .NET Core. Work that might once have required many months can now happen far faster, provided the surrounding engineering environment is strong.

That condition is crucial. AI needs reliable tests, usable development environments, clear documentation and disciplined release processes. Without them, automated modernization can create a larger cleanup exercise.

The best results combine AI-generated code with unit testing, regression testing, sandboxing, CI/CD, security controls and real-time observability. Hron also described security, infrastructure, accessibility and compliance teams building their own agents so policies can be applied earlier in the development flow.

The Business Question Comes First

A modern infrastructure stack will not rescue a weak strategy.

Many organizations begin with, “Does AI work?” They then move toward small efficiency use cases such as summarizing emails or generating presentations. These experiments can be useful, but they rarely transform a company.

The more valuable questions focus on growth. How could AI change the business model? Which products could become intelligent services? Which customer problems can now be solved in a new way? Which processes should be redesigned from the ground up?

Patzak praised Thomson Reuters for approaching AI with a growth mindset and a willingness to make ambitious bets. Hron extended that idea to individuals, especially software engineers whose value may increasingly come from problem-solving, architecture, judgment and product thinking rather than the volume of code they personally type.

Experimentation also requires tolerance for failure. Patzak’s view was refreshingly direct: “If you’re not failing, you’re not really experimenting.”

This means creating controlled experiments, learning quickly and scaling the ideas that demonstrate real value.

The Cost Of Waiting Is Getting Higher

Leaders often delay modernization because it appears too expensive, too disruptive or too risky. The greater danger may now come from standing still.

Hron said, “The risk of doing nothing is the biggest risk, quite honestly, because I think the cost of waiting in many cases is quite existential.”

AI is lowering the barriers to product development. Smaller companies can build sophisticated services with fewer people, less capital and greater speed. Established enterprises still have powerful advantages, including trusted brands, deep expertise, customer relationships and proprietary data. Those strengths lose value when they remain locked inside systems that cannot move at the pace of the market.

The answer is a continuous modernization discipline linked directly to business priorities.

Start with the workflows carrying the greatest customer value or operational risk. Break large systems into smaller components. Improve testing and observability. Build shared AI services so every project does not begin from scratch. Give teams permission to experiment, while maintaining clear security, governance and accountability.

The winners in the AI race will not necessarily own the biggest model. The leaders in the AI race will be those able to adapt their infrastructure, data, operating model and culture as the technology changes.

And it will change again, sooner than most legacy systems are prepared for.

See how Thomson Reuters modernized on AWS with agentic AI