Arm presented its first complete commercial CPU design at Hot Chips, built on TSMC N3P with dual chiplets (>50B transistors each), up to 136 Neoverse V3 cores at 3.7 GHz, 800+ GB/s memory bandwidth, and 96 PCIe Gen6 lanes at 300W TDP. Demand forecast raised to $2B for 2027-2028.
XPeng Robotics raised over $900M in its first funding round led by IDG Capital, with GaoRong Capital, Tencent, and Alibaba as strategic investors. Funds target mass production of the IRON humanoid robot and global commercial deployment. Record single PE round in China's embodied intelligence sector.
The AI model hub and open-source platform is reportedly in acquisition talks at a $13B+ valuation, having engaged banks to evaluate bids. The potential sale follows surging demand for AI infrastructure; Hugging Face last raised at $4.5B in 2023.
Proposes a stable critic training recipe showing that a well-trained critic can estimate token-level advantages from a single response matching GRPO's multi-sample performance.
Introduces a reward-free continual RL framework for space robots under hardware degradation, removing the need for reward signals during deployment while maintaining online adaptation.
Credal LLMs maintain multiple predictive distributions to separate epistemic ignorance from aleatory ambiguity, improving calibration on ambiguous queries without sacrificing accuracy.
Proposes a low-rank residual compression method reducing communication overhead in split learning while maintaining fine-tuning quality for on-device LLMs.
Derives a physically-motivated optimizer from a response-and-memory model, achieving competitive training loss against Muon and AdamW while providing theoretical grounding for update rules.
Shows that optimizing knowledge-base context artifacts (schema selection, table linking) outperforms model fine-tuning alone for enterprise Text-to-SQL.
A routing system that redirects tail-latency requests to faster surrogate models guided by partial rollout analysis, improving throughput in RL training and evaluation pipelines.
NVIDIA presented its in-house Vera server CPU featuring 88 Olympus-architecture cores, LPDDR5X memory, 96 PCIe Gen6/CXL lanes, and NVLink-C2C interconnects as the CPU backbone of the Vera Rubin AI platform for confidential computing across the full rack.
Xiaomi's in-house Xring O3 flagship SoC on 3nm features a 10-core CPU and 16-core GPU delivering 85% higher graphics performance, 5.22M AnTuTu score, first LPDDR6 memory support at 113.8 GB/s, and 200 TOPS AI tensor performance.
IBM revealed its first dual-ISA processor for IBM Z and LinuxONE at Hot Chips 2026, where each core natively executes both z/Architecture and Arm AArch64 instructions on a 2nm process with 11 Z cores above 5.7 GHz.
Galbot's humanoids demonstrated the world's first live autonomous humanoid tennis match using the LATENT AI framework, tracking balls and executing varied shots without human control over 100 consecutive rallies.
Marvell promotes CXL-based memory expansion using recycled DDR4 modules via Meta's Vistara ASIC, bridging two DDR4 channels for 128GB per chip from recovered server memory as the industry faces a severe DRAM shortage.
An AI coworker platform giving each agent its own browser-based cloud desktop (browser, terminal, files, email, calendar) with persistent memory. Topped Product Hunt on August 24 with 313 upvotes.
Slack shipped a new channel type where teams and AI coding agents (Claude, Devin, GitHub Copilot, Vercel) collaborate in the same shared thread, launching with an Agents and Tools management hub.
Anthropic launched an official learning hub with structured courses and tutorials to bridge AI education and adoption for Claude users.
An autonomous offensive-security agent rewritten in Go, featuring 50+ security skills, a zero-false-positive gate, and automatic CVSS reporting under PolyForm Noncommercial license.
A messaging system providing durable HTTP streams backed by object storage, presented as a lightweight alternative to traditional message queues. Hit 108 points on Hacker News Show HN.
General Intuition, building foundation models for physical AI agents, is in talks at $6B pre-money with Valor Equity Partners, Point72 Ventures, and Seven Seven Six joining existing backers in an oversubscribed round for compute and robotics hiring.
The AI model hub is reportedly in acquisition talks at $13B+, having engaged banks to evaluate bids following surging demand for AI infrastructure. Last raised at $4.5B in 2023.
Smart ring maker Oura plans to raise up to $3B in a US IPO next month at a valuation north of $16B, jumping from $10.9B in its 2025 Series E. Backed by Fidelity, ICONIQ, and Whale Rock.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
