The agentic era is getting its own hardware, in three different shapes

Two TPUs for two agentic problems

Google is expanding its eighth-generation TPU portfolio with two new specialized systems, TPU 8tx and TPU 8i Max, aimed at the specific requirements of agentic AI workloads. TPU 8tx is optimized for secure and confidential inference, delivering up to 2 times lower latency compared with TPU 8t, while TPU 8i Max combines two TPU 8i chips on a single package and is designed for massive context windows and high compute density. The split is the story: confidential inference and giant contexts are different engineering problems, and Google is no longer asking one chip to solve both. m:…4752

The confidential half of the story runs through AMD. Fifth-generation AMD EPYC processors power the host infrastructure for Google Cloud TPU 8tx, which AMD says is the first cloud TPU system with hardware-enforced confidential computing. The TPU 8tx system utilizes AMD Secure Encrypted Virtualization to protect data as it moves among the CPU and accelerators, including model weights and customer prompts during execution. TPU 8tx supports enterprise deployment of inference, fine-tuning and reinforcement learning workloads, AMD says, and EPYC processors manage workload orchestration, data preparation and high-throughput communication with the TPU accelerators. m:…8872

Software to make the silicon usable

Google is pairing the hardware with open-source software meant to lower the barrier to using it. The company has open-sourced TorchTPU to allow PyTorch code to run natively on TPUs, and says TPU Sync reduces time to first token by more than 90%, while GKE dynamic slicing can slash job startup times by up to 5 times. The TPU agent suite is currently in preview, Google says, alongside the two new systems built for the agentic era. The software story here is about removing friction: PyTorch code running natively, faster first tokens and faster job starts are the kind of improvements that decide whether a team actually moves a workload onto new silicon. m:…4752

Google says its TPU business has nearly 100 customers, which is a reminder that this is a real product line, not a research demo. Vultr is expanding the same market from the other side: it has launched a 20 MW AMD Instinct GPU supercluster in Chennai and aims to add one gigawatt of computing capacity across India by 2030, with plans to expand in 2027 with AMD Instinct MI455X GPUs and the AMD Helios rackscale solution. Vultr signed an agreement with StartupTN, and eligible startups in Tamil Nadu can receive 20% discounts on Vultr services. The build-out is global, and the customer being courted is increasingly the startup that cannot get capacity elsewhere. m:…4752 m:…0376

Coding moves on-premises

AMD, Dell and Spectro Cloud have launched the AMD Instinct Coder 355, an integrated enterprise platform designed to bring AI-assisted coding on-premises. The platform uses Dell PowerEdge servers, AMD EPYC processors and AMD Instinct MI355X GPUs, and features the GLM 5.3 model for local inference. Spectro Cloud's PaletteAI Inference Launchpad provides intelligent routing for AI requests, defaulting to local models for routine tasks like code generation and documentation while permitting frontier services when justified. The platform supports frontier models including Anthropic Claude, OpenAI GPT and Google Gemini. m:…0951

The pitch is economic and governance-shaped at once. AMD says the platform delivers up to 75% lower total cost of ownership, supports up to 85 developers per node with 50 concurrent users, and achieves up to 98% of frontier model benchmark performance. The company says the approach is intended to help enterprises lower total cost of ownership, improve data governance and gain better control over AI infrastructure spending. The routing logic is the interesting part: most coding work is routine, and routing it locally while reserving frontier models for complex tasks is a cost structure that metered APIs cannot match, if the 98% performance claim holds up on a buyer's own code. m:…0951

Sandboxes are the new unit of compute

E2B's benchmark of NVIDIA Vera hardware shows what agent infrastructure looks like at the sandbox level. E2B provides isolated computers for agents through a simple SDK, and reports that Vera delivered 1.51 times the single-thread compute throughput of its AMD EPYC 9B45 reference, with memory-population throughput of 62.8 GB/s against 17.4 GB/s on the AMD reference. A single dual-socket Vera server sustained between 1,400 and 1,550 concurrent loaded sandboxes within test objectives, E2B says, and completed 22,145 create-and-destroy cycles with zero errors on Vera bare metal. m:…5426

The numbers that matter for agent platforms are the small ones: E2B measured 50 ms warm, sequential sandbox starts on Vera bare metal, and 3.8 ms command round trips at low load. At 1,550 concurrent sandboxes, command round-trip latency was 255 ms at the 95th percentile, and the server sustained 550 concurrent sandboxes when every sandbox ran its work loop continuously. The implication is that the unit of compute for agents is not the GPU or the server, but the isolated sandbox, and the hardware question is how many of them one box can hold without falling over. On that measure, Vera looks like a credible foundation for scaling agent-based cloud environments, E2B says. m:…5426