AI billions shift from training to deployment

Microsoft’s new Frontier unit gets $2.5 billion to embed engineers inside enterprises, while AWS and Anthropic each pledged a similar billion‑dollar forward‑deployment effort. The shift signals the AI race is moving from training ever‑larger models to building custom silicon, dedicated cloud regions, and inference‑optimized stacks that actually deliver value.

Claude models are now generally available in Microsoft Foundry with Azure-native billing, identity, and governance. Customers can only select a US data zone, and Anthropic’s documentation confirms no European‑hosted option exists, meaning EU enterprises cannot meet data‑residency rules. The gap forces European teams to seek alternative LLM solutions.

Cycle unveiled a dedicated EU control plane that stores platform management data and telemetry exclusively in Europe. The offering lets European customers satisfy data‑sovereignty regulations while keeping operations isolated from other regions. It marks a shift toward regional control layers as cloud compliance pressure mounts.

Even as per-token prices drop, multi-step agent workflows can burn hundreds of thousands of tokens per job, inflating spend faster than model discounts. Engineers must redesign pipelines, compressing context, routing simple tasks to cheaper models, and pruning handoffs, to curb token tax and keep AI budgets in check.