AWS re:Invent 2025 made one thing unmistakably clear—the companies that will succeed with agentic AI over the next 24 months aren’t the ones chasing the most sophisticated models. They’re the ones that can actually deploy agents into production against clean, modern infrastructure. For most enterprises, that’s the hard part.
The challenge is structural, not technical. Organisations report spending roughly 30% of engineering capacity on maintaining legacy systems rather than advancing new capabilities. For them, launching an agentic AI initiative looks less like opportunity and more like a budget collision: modernisation consumes resources, AI consumes resources, and neither gets the investment it needs.
AWS’s re:Invent messaging, across infrastructure announcements and AI tooling, suggests a different path forward. Two pieces work together: a vertically integrated infrastructure stack designed to run agentic AI cost-effectively, and AI-powered tooling that actually demolishes legacy code at scale. Neither solves the problem alone. Together, they attempt to answer the question that’s actually blocking enterprise AI: how do you modernise fast enough to feed agents clean data and APIs without consuming your entire technology budget?
Read the complete AWS re:Invent 2025: Key Announcements Roundup
The Infrastructure Thesis: Vertical Integration for AI Economics
AWS introduced three pieces of silicon and infrastructure designed to make agentic AI workloads cheaper to run: Trainium3 for training, Graviton5 for general compute, and AI Factories for on-premises deployments.
Trainium3, built on a 3-nanometre process, claims up to 4x faster training than its predecessor and up to 50% lower training and operating costs compared with equivalent GPU solutions for certain workloads. Trainium2 has already become a multi-billion-dollar business, with CEO Andy Jassy describing it as approaching $2 billion in annual revenue. Trainium3 expands this addressable market by moving beyond the “very large customers” who could afford previous-generation custom chips toward mid-market enterprises seeking cost efficiency.
Graviton5, meanwhile, offers up to 25% better compute performance than prior generations, optimised for workloads that don’t require GPUs. Paired together, these processors form a complete CPU and accelerator stack for enterprises wanting to run AI workloads—both training and inference—without depending on a single hardware vendor’s ecosystem.
ALSO READ: AWS Launches Graviton5, Its Most Powerful Custom CPU for EC2
AI Factories represent the strategic inflection point. These are dedicated AWS infrastructure deployments installed inside a customer’s own data centre, operated by AWS but run exclusively for that customer. They combine Trainium accelerators, NVIDIA GPUs, high-speed networking, and AWS services such as Amazon Bedrock and SageMaker AI in a private environment that behaves like a dedicated AWS Region. The explicit target: regulated industries and governments with strict data residency requirements who cannot move workloads fully into public cloud but still need frontier-scale AI infrastructure.
The economic message is consistent across all three pieces: AWS is betting that enterprises will choose AI platforms based on unit economics and total cost of ownership, not just model sophistication. By integrating silicon, software, and managed services, AWS believes it can offer price-performance advantages that make switching prohibitively expensive once workloads are deployed.
Microsoft offers custom chips (Maia, though less widely deployed than Trainium) and Copilot Studio. Google provides TPUs and AI services. OpenAI has frontier models but no infrastructure or hardware of its own. Each vendor is making different strategic choices about what to build and what to partner on. AWS’s choice is vertical integration, arguing that complete control over the stack is how you compete on cost at scale.
The Modernisation Prerequisite: Clearing the Path
But cheap infrastructure doesn’t solve the problem if enterprises can’t actually feed it clean data and well-designed APIs. That’s where the second piece—modernisation tooling—enters.
AWS Transform, announced at re:Invent as “Transform custom,” is fundamentally an agentic solution to the technical debt problem. The service uses AI agents to automate large-scale code refactoring across thousands of repositories, learning an organisation’s specific transformation patterns and applying them consistently.
The numbers are concrete. AWS claims Transform has already processed 1.1 billion lines of code and saved over 810,000 hours of manual effort across customer deployments. Customers report up to 80% reduction in execution time for modernisation projects. Air Canada modernised thousands of Lambda functions in days, achieving what would have taken weeks or months of manual work. QAD reported transformations that previously required two weeks now completing in three days.
Beyond pre-built runtime upgrades (Java, Node.js, Python, AWS SDKs), Transform custom lets organisations define their own transformation rules using natural language, documentation, and code samples. The agent then learns these patterns and executes them autonomously across an entire codebase—in parallel. For enterprises with 25 years of accumulated institutional patterns, utility libraries, and architectural conventions, this means transforming scattered, manual modernisation into a scalable, repeatable organisational capability.
The scope of what Transform covers has expanded. It now handles full-stack Windows modernisation (SQL Server to Aurora PostgreSQL, .NET Framework to cross-platform .NET, deployment to containers), mainframe “reimagine” scenarios (AI-driven architectural redesign based on business logic extraction), and database migrations where application code, ORM changes, and schema conversion happen as a coordinated workflow.
The workflow remains human-supervised. Organisations review transformation plans, approve steps, and validate results. But the grunt work—the thousands of small changes that make large-scale migrations soul-crushing—is what the agents handle.
The Convergence: Why This Matters
These aren’t separate initiatives. They’re designed to work together.
An enterprise’s actual path forward looks something like this: Use Transform to modernise a critical legacy system (core transaction system, data warehouse, monolithic application). Cost is reduced by 80% versus traditional approaches. Outcome: modern APIs, clean data pipelines, documented patterns. Then deploy agents on Nova models, running on Trainium for inference, powered by Bedrock. Because infrastructure is modernised, agents can operate safely. Because agents are cost-effective, ROI becomes visible quickly.
The compound advantage builds over 18–24 months. Enterprises that modernise early and deploy agents early will have cost, speed, and security advantages that compound. They’ll move faster than late-movers still trapped on legacy plumbing.
This is why AWS framed re:Invent 2025 not as “look at our new model” or “look at our new chip” but as a coherent narrative about how frontier agents get to production at enterprise scale. Models matter. Hardware matters. But the real bottleneck for most organisations isn’t choosing between Claude or GPT-5; it’s deciding whether they’re modernised enough to deploy agents without creating security and operational nightmares.
ALSO READ: Big Tech’s Enterprise AI Initiatives in 2025: A Guide by Business Need
The Execution Question
AWS has the pieces. Whether enterprises actually adopt the full picture—modernising with Transform and deploying on Nova + Trainium infrastructure—remains the open question.
Developers often prefer NVIDIA’s software ecosystem, which has a 10-year head start on community and tooling. Mid-market enterprises might modernise with Transform but then run agents on existing cloud vendor relationships or open-weight models on their preferred platform. Some organisations might embrace Trainium but not modernise.
Similarly, Transform custom is new and ambitious. Early results are promising, but large-scale refactoring of mission-critical systems has always carried risk. The real test is whether a mid-market financial institution or healthcare system will trust an AI agent to refactor millions of lines of code without exhaustive human review at every step.
The implicit stakes of re:Invent 2025, however, are becoming clearer. The enterprises that win on agentic AI won’t win because they picked the best model. They’ll win because they solved the modernisation problem first, positioned themselves to run agents cost-effectively, and iterated on production deployments while competitors were still planning their migration strategies.
For technology leaders, the question isn’t which cloud or which model to choose. It’s whether your organisation is ready to make modernisation a first-class problem and treat agentic AI deployment as a compounding economic advantage rather than an isolated capability. AWS is betting it can make both easier. Over the next two years, that thesis will be tested in the market.
ALSO READ: 6 Revolutionary AI Coding Models Transforming Developer Workflows in 2025
