The News
At re:Invent 2025, AWS CEO Matt Garman delivered a sweeping keynote centered on agentic AI, custom model training through Nova Forge, major GPU and Trainium hardware expansions, and a deeper framing of AWS as the infrastructure layer powering billions of future agents. The keynote emphasized customer outcomes, full-stack optimization, and a shift from AI assistants to production-grade autonomous systems.
Analysis
AWS Repositions the Cloud Around Agentic AI at Scale
AWS is no longer talking about generative AI in abstract terms; it is anchoring the future of cloud computing on agentic systems. Garman repeatedly framed agents as the next platform shift, comparable in impact to the internet and early cloud adoption. This message aligns with ECI and theCUBE Research findings that organizations expect AI systems to evolve from task automation into continuous reasoning engines that operate across application boundaries.
AWS sees this shift as inseparable from infrastructure. The keynote tied agentic adoption to global-scale compute, predictable inference economics, and strong governance. This is a combination enterprises repeatedly cite as the top blockers to operationalizing AI beyond pilots.
Hardware Becomes a Strategic Differentiator Again
The hardware portion of the keynote underscored a thesis AWS has been reinforcing for years, noting that AI performance, cost, and reliability depend on vertical integration. Garman highlighted deep co-engineering with NVIDIA, operational rigor on GPU clusters, and growing demand for ultra-scale compute for frontier models.
Several notable moments signaled AWS’s long-term posture:
- Expansion of UltraServer architectures using NVIDIA GB300 systems
- Continued success of Trainium 2 (already a multibillion-dollar business)
- General availability of Trainium 3 Ultra Servers (first 3 nm AI chip in AWS Cloud)
- A public preview of Trainium 4 with major jumps in compute and bandwidth
Developers increasingly face pressures around inference latency, workload portability, and training speed. AWS’s hardware narrative is aimed squarely at those pain points as organizations shift from experimentation toward always-on, multi-model infrastructures.
AI Factories and Sovereign Deployment Models
One of the most substantial announcements was AWS AI Factories, a dedicated AI-region-like capability customers can deploy in their own data centers. This hybrid model acknowledges two accelerating realities:
- Highly sensitive workloads (including biotech, defense, and regulated industries) require sovereignty that public regions alone cannot satisfy.
- Organizations frequently need cloud-native tooling and automation without relocating data, power, or compute to shared infrastructure.
This aligns with ECI research showing that sovereignty and locality are among the top constraints for enterprise AI, particularly as agent-based workflows begin touching operational, financial, and mission-critical systems.
Nova 2 and Nova Forge Shift the Model Landscape
If hardware is the foundation, Nova 2 and Nova Forge are the centerpiece of AWS’s model strategy. Nova 2 extends AWS’s control over price-performance, offering reasoning-oriented models optimized for production, not experimentation.
Nova Forge represents a deeper shift where AWS is enabling customers to inject proprietary datasets directly into pretraining checkpoints mid-stream, not post-hoc, as with traditional fine-tuning. This “open training model” capability aims to target a real developer challenge of getting models to understand internal processes, historical knowledge, edge cases, and domain-specific semantics without catastrophic forgetting.
AWS positioned Nova Forge as a way to produce proprietary “Novellas,” custom models built on Nova scaffolding but tailored to a company’s private intelligence. For teams who struggle with the limits of RAG-only approaches, this may prove significant.
AgentCore Matures With Policy and Evaluations
AgentCore emerged as the backbone for AWS’s agentic ecosystem. Garman acknowledged what developers already know, that building reliable agents is hard because nondeterministic systems require new guardrails, new governance, and new testing paradigms.
AWS announced two key features:
- Policy: Real-time deterministic controls at the gateway level, allowing teams to specify which actions agents may or may not take.
- Evaluations: Continuous quality scoring and drift detection across correctness, helpfulness, safety, and appropriateness.
AI systems that behave well at low scale often behave unpredictably under load, with poor traceability. Guardrails must shift closer to infrastructure and away from prompts or application code.
Enterprise Voices Reinforce the Pattern
Large enterprise guests Sony and Adobe echoed the same story that success depends on controlling data, productionizing models, and integrating agents into existing operational flows without disrupting reliability or compliance. Their experiences underscore AWS’s message that agents will require:
- Reliable access to structured and unstructured data
- Consistent governance against runaway autonomy
- Multi-model flexibility
- Foundations that scale from experimentation to mission workloads
Across industries, teams are transitioning from “AI features” to AI-as-a-core-architecture, which aligns strongly with re:Invent’s agent-first framing.
Looking Ahead
AWS is steering the ecosystem toward a cloud model where agents are the dominant application type, and infrastructure, models, and governance mechanisms converge to support them at planetary scale. If developers adopt AgentCore, Nova Forge, and Trainium-class hardware as intended, AWS could become the default substrate for autonomous, reasoning-driven applications.
But the long-term impact depends on how easily these components weave into existing DevOps pipelines, security frameworks, and application architectures. The industry appears ready for this pivot, as indicated by rising investment in AI automation, multi-model strategies, and cloud-native governance, but developer experience and operational simplicity will determine adoption velocity.
Over the next year, watch for how fast teams begin shipping real agentic workloads, how Nova Forge shapes proprietary model strategies, and how multi-cloud and sovereign requirements influence the competitive landscape across hyperscalers.
Agentic AI Becomes a Full Stack, Not a Feature
Akuity Expands GitOps Delivery to Multi-Cloud & Multi-Environment Targets
LocalStack Hits Record Adoption as Local-First Cloud Development Accelerates
Inside Amazon’s New Framework for Measuring Developer Productivity
Agentic AI, Nova Forge & Cloud-Scale AI
