Microsoft Build consumed most of this week’s cloud news with a coherent agentic push. Two AWS items cut through independently: a redesigned Bedrock console and a Cognito feature that removes a long-standing single-region constraint.
Cobalt 200 Brings Custom Arm Silicon to Azure in Early Access
Microsoft has opened early access to the Azure Cobalt 200 Arm-based VM series, a custom chip on TSMC’s 3nm process claiming 50% generational performance gains over Cobalt 100, optimised for scale-out, Linux-based agentic workloads including inferencing, data pipelines, and API tiers. Cobalt 100 already runs Teams and Defender internally and is deployed across 32 Azure regions, giving Cobalt 200 a clear internal validation path before broader customer availability.
Why it matters: The 50% claim is vendor-measured against Microsoft’s own workload mix. Independent benchmarks against your actual inference and API tier patterns are worth running before this informs any architecture decision.
HorizonDB Enters Public Preview as a Postgres Rebuilt for Agentic Apps
Azure HorizonDB is now in public preview: a fully managed, PostgreSQL-compatible database with a Rust-based storage engine and DiskANN vector indexing, claiming up to three times faster transactions and search than self-managed PostgreSQL. Transactional data, vector indexing, semantic search, and in-database model access are bundled in a single service, with native integration into Microsoft Fabric and Foundry.
Why it matters: If the claims hold, HorizonDB reduces the operational overhead of running separate data stores for transactional, vector, and semantic search workloads. In public preview with an unproven storage engine; evaluate before committing.
Microsoft Foundry Formalises a Production Model Operations Lifecycle
Microsoft used Build to publish a model operations lifecycle for Foundry, spanning selection, evaluation, optimisation, and continuous improvement. The concrete release alongside it is Fireworks AI on Foundry reaching general availability: open model inference through a single Azure endpoint with enterprise SLAs, PTU Data Zone support, and SOC2 readiness, after processing 176 billion tokens across 17 S&P 500 enterprises in preview.
Why it matters: Fireworks AI GA is immediately actionable: high-performance open model inference in the Foundry control plane, enterprise SLAs, no separate vendor contract.
Amazon Bedrock Gets a Redesigned Console Around the Mantle Endpoint
AWS has shipped a new console experience for Bedrock centred on the bedrock-mantle endpoint, which surfaces GPT, Claude, and open-weight models via the OpenAI Responses and Chat Completions APIs alongside the Anthropic Messages API. The redesign adds project-based organisation, side-by-side comparison of up to three models, and live documentation that auto-prefills code snippets with your project’s model ID, endpoint URL, region, and API key.
Why it matters: The bedrock-mantle endpoint is the substantive change: a single AWS surface, IAM boundary, and billing account for GPT, Claude, and open models reduces multi-provider governance complexity considerably.
Amazon Cognito Adds Native Multi-Region Replication
AWS has launched multi-Region replication for Cognito, automatically synchronising user profiles, credentials, and pool configurations to a secondary region. Both regions accept tokens from either endpoint so existing sessions survive failover without forced password resets. Activation requires a multi-region customer managed KMS key and region-agnostic OIDC endpoints; Lambda-based custom auth flows, WAF rules, and log streaming need manual redeployment in the target region. Pricing is $0.0045 per MAU per replica region on Essentials tier, with a 30% M2M token surcharge.
Why it matters: Cognito’s single-region design has been a genuine constraint for resilient architectures in regulated industries. No forced password resets during failover addresses the main user experience liability, but Lambda, WAF, and log streaming each need manual preparation in the target region before this goes live.
Most of what Microsoft shipped at Build is in early access or preview, making this a week for evaluation planning rather than architecture commits. The interesting test is whether Cobalt 200, HorizonDB, and Foundry cohere as a system in practice, not just as a Build keynote narrative. Cognito multi-region is the clearest near-term action item: available today, no preview caveats, addressing a documented gap. For teams managing multiple AI model providers, the Bedrock mantle endpoint is worth evaluating before provider sprawl compounds.
