Frontier Capability Developments

Google DeepMind announced computer use capability in Gemini 3.5 Flash, enabling the model to interact directly with graphical user interfaces — clicking, typing, and navigating software environments autonomously. This is a meaningful capability expansion because it extends agentic behavior down the cost curve: Flash is Google's speed-optimized, lower-cost model tier, not its frontier capability flagship. The practical implication is that developers building autonomous agents no longer need to pay frontier-model pricing to get GUI interaction. Google DeepMind

This directly targets Anthropic's current competitive advantage. Anthropic has positioned Claude's computer use as a differentiating capability, primarily accessible through its larger, more expensive models. Google deploying equivalent functionality in a flash-tier model compresses the price-performance frontier and forces Anthropic to either accelerate capability diffusion across its own model tiers or defend on quality grounds. The broader pattern here is that agentic capabilities — once considered frontier-only — are now being commoditized into efficient inference models within roughly 12-18 months of first deployment.

OpenAI disclosed Jalapeño, its first custom ASIC for AI inference, designed in partnership with Broadcom. The chip is purpose-built for running large language models at inference time — not training — which is where the overwhelming majority of compute cost occurs at scale. This is a confirmed product announcement, not a roadmap projection, though independent performance benchmarks have not yet been published. The Verge

The strategic logic is straightforward: inference ASICs can deliver significantly better performance-per-watt and cost-per-token than general-purpose GPUs for specific model architectures. OpenAI joins Google (TPUs), Amazon (Trainium/Inferentia), and Meta (MTIA) in building custom silicon to reduce Nvidia dependency and improve unit economics. The Broadcom partnership — rather than a fully in-house design — suggests OpenAI is prioritizing speed to deployment over maximum differentiation. The critical unknown is whether Jalapeño's performance justifies the capital and organizational complexity relative to simply continuing to purchase Nvidia H-series and Blackwell hardware.

Princeton researchers published results showing that reinforcement learning and diffusion models can design radio frequency integrated circuits from scratch, achieving record performance metrics while drastically reducing design time. RFIC design — the engineering of analog circuits for wireless communication — is widely regarded as one of the most difficult specializations in chip engineering, dependent on accumulated expert intuition rather than systematic rules. The research demonstrates that AI can navigate this high-dimensional, physics-constrained design space more effectively than human designers on specific performance dimensions. IEEE Spectrum

The immediate applications span 5G infrastructure, autonomous vehicle radar, and satellite communications — all domains where RFIC performance directly limits system capability. The broader significance is that AI-driven hardware invention is crossing from well-structured digital logic (where AI assistance has been established for several years) into analog and mixed-signal domains previously considered resistant to automation. The researchers note that progress requires large shared chip design datasets and open ecosystems — a bottleneck that industry incumbents have incentive to maintain as a moat.

Qualcomm is acquiring Modular, the startup behind the Mojo programming language and MAX inference engine, for nearly $4 billion. Modular built tooling specifically to make AI inference portable and performant across heterogeneous hardware — Nvidia GPUs, AMD GPUs, Arm CPUs, and custom accelerators. The acquisition price reflects how critical the software layer between AI models and diverse hardware has become as the chip landscape fragments. Wired

For Qualcomm, this is a direct response to Nvidia's CUDA moat. CUDA's developer ecosystem lock-in remains Nvidia's most durable competitive advantage — not the hardware itself. Modular's technology is explicitly designed to route around CUDA dependency. Qualcomm acquiring it positions the company to offer a compelling end-to-end story for on-device and edge AI inference, where Qualcomm's Snapdragon and data center Arm chips compete. The risk is execution: developer ecosystem plays require sustained investment and community trust, and Modular's value is partly its independence.