Ramp has launched Router, a service designed to give developers one endpoint for multiple AI models and tools for managing model cost and performance. Ramp says it spent three years operating and improving the underlying technology on its own production workloads, reducing its AI costs by 30%. Router Strategies lets developers set cost and performance priorities for different request types or use Ramp’s benchmarked default settings. Ramp says the product applies tools it developed to lower its own AI spending to other businesses’ AI workloads. Router is free through 2026; users pay list price for the tokens they consume and receive their first $26 in credits, subject to offer terms.

Modular announced at ModCon 2026 that Mojo 1.0 is fully open source under the Apache 2.0 license, making its compiler and tooling available for developers to extend and port to new platforms. Mojo 1.0 provides a stability guarantee intended to prevent code written against the release from breaking, and native Windows support is in development through a collaboration with Microsoft’s Windows team. Mojo previously supported macOS and Linux, while Windows users relied on WSL. Modular Cloud is now generally available through console.modular.com, offering OpenAI-compatible shared inference endpoints with pay-per-token pricing and dedicated isolated deployments. The company said the service has handled OpenRouter production traffic under the ModelRun name and serves MiniMax’s M3 model on a dedicated deployment at billions of tokens per minute. M3 combines a one-million-token context window, multimodal capabilities, and MiniMax Sparse Attention. The Modular Platform now supports AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra, and Qualcomm Dragonfly accelerators, in addition to CPUs and NVIDIA and AMD GPUs. Modular said the integrations required more than 10 times less engineering effort than traditional hardware enablement. MAX will become source-available through an open alliance program, and its license no longer restricts device usage. Modular said it will continue supporting hardware that competes with Qualcomm’s platforms.

Nvidia no longer supports NVLink on consumer GPUs, with the last supported model being the RTX 3090. NVLink bridges currently range in price from $200 to $400, and tests are being conducted to evaluate the benefits of NVLink for AI workloads using dual RTX 3090s.

NVIDIA Vera CPU Benchmarks: Olympus Cores Delivering The Best Performance Ever Seen On ARM NVIDIA's Vera data center CPU isn't ramping up until later this year but I recently had the opportunity to try out this new ARM-based CPU designed for agentic AI workloads. NVIDIA's Vera CPU with its in-house-designed Olympus CPU cores ends up packing a heavy-hitting punch with competitiveness to Intel/AMD x...