AWS re:Invent wrapped up last week, and as usual, the announcement volume was overwhelming. Hundreds of new features, services, and updates – most of which won’t matter for most enterprises.

I’ve spent the week going through the announcements. Here’s what actually deserves your attention if you’re making enterprise AI decisions.

The Big Picture: Amazon Plays Catch-Up

Let’s be direct: Amazon has been behind Microsoft (OpenAI partnership) and Google (Gemini) in the generative AI race. This re:Invent was about closing that gap.

The strategy is clear: if you can’t beat OpenAI at models, beat them at infrastructure. AWS is positioning Bedrock as the enterprise-grade foundation for AI deployment, regardless of which models you choose.

This is a reasonable strategy. Enterprises care about security, compliance, integration, and total cost of ownership – areas where AWS has deep expertise.

Amazon Bedrock Updates

Bedrock is AWS’s managed service for accessing foundation models. The key updates:

Model expansion: Anthropic’s Claude 3, Meta’s Llama 2, and others are now available alongside Amazon’s own Titan models. This model optionality is genuinely valuable – you’re not locked into one provider.

Agents for Bedrock: The ability to build AI agents that can take actions – query databases, call APIs, execute multi-step workflows. This competes directly with OpenAI’s Assistants API and Microsoft’s Copilot Studio.

Knowledge Bases for Bedrock: A managed RAG (retrieval-augmented generation) service. Upload your documents, point Bedrock at them, and it handles the embedding, indexing, and retrieval.

Guardrails for Bedrock: Automated content filtering and safety measures. Given growing concern about AI outputs in enterprise settings, this is timely.

My take: These updates make Bedrock a credible enterprise AI platform. If you’re already on AWS and haven’t evaluated Bedrock recently, it’s worth another look. The gap with Azure OpenAI has narrowed significantly.

Amazon Q: AWS’s Enterprise Assistant

Amazon Q is AWS’s answer to Microsoft Copilot. It’s an AI assistant that can:

• Answer questions about AWS services and documentation
• Help developers write and debug code
• Connect to enterprise data sources
• Generate reports and analysis

The interesting move: Q connects to business applications like Salesforce, Jira, and ServiceNow. This positions it as more than an AWS admin tool – it’s an enterprise knowledge assistant.

The catch: Q is new and the third-party integrations are limited compared to Microsoft’s ecosystem. If you’re heavily invested in Microsoft 365, Copilot still has a significant integration advantage.

My take: Worth piloting if you’re AWS-native. Not compelling enough to switch from Microsoft if you’re already in that ecosystem.

Amazon SageMaker Updates

For teams doing custom ML development, SageMaker got significant updates:

SageMaker HyperPod: Distributed training infrastructure that can span thousands of GPUs with automatic failure recovery. Essential for training large models.

SageMaker Canvas updates: Low-code ML model building, now with generative AI capabilities. The target is business users who want to build models without coding.

Model governance features: Tracking lineage, managing model versions, documenting decision-making processes. Important for regulated industries.

My take: SageMaker remains the most comprehensive ML platform for custom development. The HyperPod capabilities are particularly impressive if you’re training large models. For most enterprises using off-the-shelf models via Bedrock, these SageMaker updates matter less.

Trainium2 and Inferentia2

AWS announced their next-generation custom AI chips:

Trainium2: Training-focused chips offering 4x performance over first-gen at similar cost. AWS claims 50% better price-performance than comparable Nvidia hardware.

Inferentia2: Inference-focused chips for serving models in production.

Why this matters: GPU availability has been a genuine constraint for AI projects. AWS building custom chips means more reliable capacity and potentially better economics.

The caveat: Custom chips require software adaptation. Models optimised for Nvidia GPUs may need work to run efficiently on Trainium/Inferentia.

My take: Interesting for large-scale deployments where you control the full stack. For most enterprises, the model availability in Bedrock matters more than the underlying chips.

Security and Governance

Some less flashy but important announcements:

Private model fine-tuning: Train custom versions of foundation models without data leaving your VPC. Critical for regulated industries.

Expanded audit logging: Track every AI interaction for compliance purposes.

Data residency controls: Explicit controls over where data is processed and stored.

These aren’t exciting features, but they’re the table stakes for enterprise adoption. AWS is clearly listening to enterprise procurement concerns.

What This Means for Enterprise Planning

A few implications:

Multi-cloud AI is viable: With both AWS Bedrock and Azure OpenAI offering comprehensive enterprise AI platforms, organisations can make choices based on existing cloud relationships rather than AI capabilities alone.

Custom AI chips reduce dependency on Nvidia: Long-term, this should improve availability and pricing. Short-term, it’s one more factor in cloud AI economics.

The “build vs. buy” line is shifting: Managed RAG, guardrails, and agent frameworks mean less custom development is needed. Platform capabilities are catching up to what previously required significant engineering.

Don’t ignore the smaller announcements: Security, governance, and integration features often matter more than flashy AI capabilities for enterprise adoption.

My Recommendations

Based on re:Invent:

1. If you’re AWS-native: Evaluate Bedrock seriously if you haven’t recently. The platform has matured significantly.

2. If you’re evaluating clouds for AI: Both AWS and Azure are now credible. Other factors (existing investment, skills, compliance) should drive the decision.

3. If you’re building AI agents: Look at Agents for Bedrock alongside Microsoft’s offerings. The capability gap has closed.

4. If you’re in regulated industries: The security and governance features announced are worth reviewing with your compliance team.

Final Thought

re:Invent 2024 wasn’t revolutionary. It was about AWS catching up and competing credibly in enterprise AI. For customers, that’s good news – competition improves offerings and reduces lock-in risk.

The AI platform wars are far from over, but the options available to enterprises are better than they’ve ever been.