Reports
AI-generated structured vendor updates
Google unveils 8th-gen TPU: 3x training speed, 3x SRAM for inference, redefines AI compute TCO
At Cloud Next 2026, Google launched 8th-gen TPU with dual variants: TPU 8t for training (9600 per pod, 2PB shared memory) and TPU 8i for inference (1152 per pod, 3x on-chip SRAM). Also announced Gemini Enterprise Agent Platform, N4 Axion ARM instances (2x price-performance vs x86), and AI-driven security with Wiz.
Google Trillium TPU: 4.7x Training Boost Masks Vendor Lock-in and Ecosystem Risks
Google Cloud unveils 6th-gen TPU Trillium with 3nm process, delivering 4.7x training and 2.5x inference performance gains, with 2x energy efficiency over NVIDIA H100. However, Trillium is exclusive to Google Cloud TPU v6p instances and deeply integrated into AI Hypercomputer architecture, creating a full-stack lock-in from silicon to networking.
AWS Seizes Agent Control Plane with MCP Gateway and AgentCore
AWS launches managed web search for Bedrock AgentCore, autonomous agents in Amazon Quick, subagent MicroVM orchestration with LangChain, and MCP Gateway, shifting enterprise AI agents from prototypes to governed infrastructure with cloud-native control planes and execution isolation.
MediaTek Pivots to System-Level Integration: Targeting Google TPU and Musk AI Rack Deals
MediaTek elevates its AI strategy from chip design to system-level integration, targeting Google TPU PCBA L6 and Musk AI chip L10 rack assembly. Adopting a light-asset model via Taiwan's supply chain, targeting >40% gross margin, driven by rising complexity from CPO and 800V high-voltage DC power.
ASML CEO's EUV Supply Warning Signals a Physical Ceiling on AI Chip Expansion
ASML CEO Fouquet confirms talks with Musk on Terafab but stresses supply constraints. EUV lithography, the sole tool for advanced AI chips, cannot scale quickly. With TSMC, Samsung, Intel, and Musk all vying for limited machines, AI chip capacity allocation becomes a zero-sum game, capping the entire AI infrastructure buildout.
MediaTek Shifts from Chip Design to System Integration for Google TPU and Musk's AI Racks
MediaTek upgrades its AI strategy from chip/ASIC design to system-level integration, targeting Google TPU PCBA and Musk's AI chip rack-level (L10) work. Using an asset-light model, it leads design and validation while outsourcing manufacturing, aiming for 40-50% gross margin.
ASUS Launches NVIDIA GB300 Deskside AI Supercomputer, Shifting Control from Cloud to On-Prem
ASUS launches the ExpertCenter Pro ET900N G3, powered by NVIDIA's GB300 Grace Blackwell Ultra Desktop Superchip, delivering 20 PFLOPS and 748GB of coherent memory for near-trillion parameter models. Concurrently, Coherent expands InP fab in Texas for optical interconnects, and NVIDIA plans a $20-25B debt offering, signaling a systemic shift of AI control from cloud to localized enterprise hardware.
Google Cloud Embeds Legal Verifiability into AI Agents via SPIFFE and Kakunin
Google Cloud introduces SPIFFE-based Agent Identity for Gemini Enterprise and Vertex AI, then overlays Kakunin's compliance layer to map internal SPIFFE identifiers to X.509 certificates generated in AWS KMS, with all state changes committed to WORM audit logs. This converts secure cloud workloads into legally auditable market participants to meet EU AI Act and MiCA accountability mandates.
Intel Foundry Lands Google TPU Packaging Deal: EMIB-T Shakes TSMC's AI Chip Monopoly
Intel secures a multi-billion-dollar deal to package over 3 million Google TPUs using its advanced EMIB-T 2.5D packaging, while the chips themselves remain fabricated at TSMC. This marks Intel's strategic shift from CPU vendor to second-source AI packaging partner, targeting 2028 production. Intel's 18A node yields exceed expectations, but analysts caution the scope is limited to packaging.
Cisco AI Defense Adds Agent Harness Red Teaming for Agentic AI Security
Cisco introduces Agent Validation in AI Defense: Explorer Edition, a dedicated red-teaming capability for agentic AI systems. It autonomously probes agent harness attack surfaces, including tool routes, indirect content channels, and persistent state, providing verified findings beyond chat-based security assessments.
AWS S3 Annotations: 1GB Mutable Metadata Per Object, Killing External Metadata DBs
AWS launches S3 annotations, enabling up to 1,000 mutable annotations per object (each 1MB, total 1GB) in JSON/XML/YAML. Annotations auto-index into Apache Iceberg tables, queryable via Athena without retrieval charges. This embeds metadata into the storage layer, eliminating external metadata databases and reshaping AI agent data discovery.
NVIDIA ACE Goes Local: Control Shifts from Cloud to RTX GPU for Game AI
NVIDIA launches ACE Game Agent SDK (open-source C/C++ framework) and UE5 plugins (ASR/SLM/TTS), moving AI NPC inference fully on-device via GeForce RTX. DLSS 4.5 plugin adds multi-frame generation. This shifts control from cloud providers to NVIDIA GPU ecosystem, but masks hardware lock-in and local model limitations.
MediaTek Pivots from Chip Design to System-Level Integration, Targeting Google TPU and Musk AI Racks
MediaTek elevates its AI strategy from chip design to system-level integration, targeting Google TPU v10 PCBA and Musk-affiliated AI rack assembly. Using an asset-light model and Taiwan's supply chain, it aims for 40-50% gross margin in system integration.
Google Open-Sources Brazos: Plug-and-Play Liquid Cooling for Air-Cooled DCs
Google introduces Brazos, a rack-mounted closed-loop liquid-to-air cooling system for existing air-cooled data centers. Supporting 60kW per rack, it is open-sourced via OCP, enabling high-density AI/HPC deployments without facility retrofits.
AMD Ryzen 10000 Series to Swap iGPU for NPU: AI Boost at Cost of Basic Display
Leaks suggest AMD's next-gen Zen 6 desktop CPU 'Olympic Ridge' will replace the integrated GPU with an NPU, targeting >40 TOPS for Copilot+ AI PC certification. It also upgrades the client I/O die to support CUDIMM/CAMM and EXPO 1.2 for faster DDR5. The trade-off boosts local AI but forces nearly all users to rely on a discrete GPU for basic display.
Google TPU 8th Gen Splits Training and Inference Chips, Inflection Point in AI Infra TCO
Google Cloud unveils 8th-gen TPU with separate training (TPU8t) and inference (TPU8i) chips, delivering 3x training pod performance and 80% inference dollar-performance improvement. Vertex AI evolves into Gemini Enterprise Agent Platform, while the Smals sovereign cloud contract validates public sector AI adoption under strict compliance.
NVIDIA Bets on World-Action Models: Control Shifts from VLM to Video Backbones
NVIDIA's blog introduces World-Action Models (WAMs) as a paradigm shift from VLM-based VLAs. WAMs leverage pretrained video/world-model backbones to jointly predict future states and robot actions, aiming to bridge the language-to-action grounding gap. This could redefine robot foundation model training but raises concerns about inference cost and latency.
NVIDIA's Desktop DGX Station with GB300 Shifts Control from Cloud to Local Hardware
ASUS launches ExpertCenter Pro ET900N G3, built on NVIDIA DGX Station GB300 architecture with GB300 Grace Blackwell Ultra chip, 748GB coherent memory, and 20 PFLOPS AI performance. This deskside AI supercomputer enables local LLM fine-tuning, inference, and agentic AI workflows via NVLink-C2C and the full NVIDIA AI software stack including NemoClaw.
NVIDIA GB300 NVL72 Delivers 20x Agentic Coding Efficiency, Setting New Inference Benchmark
NVIDIA's GB300 NVL72 achieves 20x more concurrent coding agents per megawatt than H200 on the new AA-AgentPerf benchmark, leveraging 72-GPU NVLink fabric, MXFP4 kernels, and MoE optimizations. This first standardized agentic inference benchmark redefines data center capacity planning for AI agents.
NVIDIA AgentPerf Benchmark: Blackwell Ultra Delivers 20x More Agents per Megawatt vs Hopper
NVIDIA and Artificial Analysis unveil AgentPerf, the first benchmark for agentic AI workloads. Results show the GB300 NVL72 platform delivers up to 20x more concurrent agents per megawatt than the HGX H200 when running DeepSeek V4 Pro, using real coding agent trajectories to measure throughput and responsiveness.