Nvidia reported its fourth-quarter results for fiscal 2026: revenue reached $68.13 billion, up 73% year-over-year, with data center revenue at $62.3 billion, up 75% and accounting for over 91% of total revenue; non-GAAP EPS was $1.62. The company guided to approximately $78 billion in revenue for the next quarter (±2%), noting that potential revenue from the China market has been excluded from this outlook. Networking equipment sales surged 263% year-over-year to $10.98 billion, reflecting rising demand for NVLink and Spectrum-X; the Vera Rubin system is scheduled for mass production in the second half of this year, claiming a tenfold performance improvement.

Multiple sources indicate that AMD has entered into a multi-year, multi-generation AI infrastructure agreement with Meta, involving up to 6GW of Instinct GPU deployments, with transaction value reportedly exceeding $100 billion. Terms disclosed include Meta receiving warrants for up to 160 million AMD shares, exercisable based on procurement volume and stock price targets. The first 1GW deployment is expected to begin in the second half of 2026 and may involve co-launching customized MI450 GPUs and EPYC Venice CPUs optimized for Meta workloads. This deal is seen as a key strategic move for AMD to win hyperscale customers and challenge Nvidia's dominance.

Ant Group's inclusionAI team has released and open-sourced two trillion-parameter models from the BaLing series: Ling-2.5-1T (immediate) and Ring-2.5-1T (reasoning). Their core innovation is a '1:7' hybrid linear attention architecture: one layer retains Softmax attention while seven use linear attention within every eight layers, reducing memory access by 90% compared to traditional architectures, increasing generation throughput threefold, and supporting up to 1M tokens context length. The materials claim Ling-2.5-1T leads in long-text understanding and instruction following, while Ring-2.5-1T is optimized for heavy reasoning tasks, demonstrating an engineering path for deploying hybrid linear attention at trillion-scale.

Anthropic announced the acquisition of Vercept to advance Claude's computer use capabilities—enabling perception and interaction within real software environments. Official data shows Claude Sonnet 4.6 achieving 72.5% on the OSWorld benchmark for computer use, a significant increase from under 15% at the end of 2024. Vercept, founded by Kiana Ehsani, Luca Weihs, and Ross Girshick, focuses on stable interface and toolchain interactions across multi-step tasks. Post-acquisition, Vercept will gradually discontinue external product services, with the team joining Anthropic. The integration aims to enhance cross-application workflow execution, codebase operations, and multi-source research synthesis.

Reports suggest the U.S. Department of Defense issued an ultimatum to Anthropic, requiring it to agree to relax usage restrictions on military applications of Claude by a specified deadline—or risk being classified as a supply chain risk and potentially forced to comply under the Defense Production Act. The dispute centers on Anthropic’s longstanding refusal to allow model use in fully autonomous weapons systems or large-scale domestic surveillance. Materials also note the Pentagon previously awarded contracts worth up to $200 million each to several model providers including Anthropic, as defense agencies accelerate the adoption of LLMs into secure networks and operational support workflows.

Inception launched Mercury 2, positioned as a diffusion-based reasoning LLM that replaces autoregressive token-by-token decoding with parallel generation and iterative refinement to reduce end-to-end latency and boost throughput. On NVIDIA Blackwell GPUs, it reportedly achieves output speeds of up to 1009 tokens per second. Pricing is set at $0.25 per million input tokens and $0.75 per million output tokens, supporting 128,000-token context, native tool calling, and schema-compliant JSON output. Mercury 2 is now available for early access with an online chat demo; official benchmarks claim top-tier performance on SciCode, IFBench, and AIME.

Subscription materials indicate OpenAI has added a WebSocket mode to its Responses API, enabling clients to receive model-generated token streams in real time via persistent connections, reducing latency and state management overhead associated with traditional HTTP polling and short-lived connections. This feature targets low-latency use cases such as chat, programming collaboration, and multi-user interaction, allowing smoother streaming experiences where outputs can be generated, displayed, and continued seamlessly. It also simplifies front-end implementation of features like interruption, continuation, and real-time orchestration of tool calls. No pricing changes or regional restrictions were disclosed, but the option is emphasized as a real-time transmission solution for developers aiming to improve interactive responsiveness and engineering integration efficiency.

Multiple reports indicate Anthropic is expanding enterprise capabilities of Claude Cowork by launching plugins and connectors for high-value workflows, along with support for private plugin marketplaces. Connectors integrate common systems including Google Workspace, Slack, DocuSign, and FactSet, with partnerships noted for LSEG and Salesforce. Administrators can centrally manage permissions, monitor usage, and track costs, easing governance challenges across system integrations. The offering is framed as evolving Claude from 'chat' to an orchestratable agent layer, emphasizing Anthropic’s intent to embed within existing enterprise software stacks rather than replace them. Official pricing and availability details were not provided.

Chip startup MatX announced a $500 million Series B round led by Jane Street and Situational Awareness. The company is developing the MatX One processor tailored for large language models, featuring a 'splittable systolic array' designed to adapt to dynamic workloads. Its architecture uses SRAM for storing model weights to minimize latency, while leveraging HBM for KV cache to support long contexts, combined with optimizations such as speculative decoding and block-sparse attention. MatX stated the new funding will be used to complete chip design and fabrication, aiming for tape-out within one year, targeting the increasingly blurred boundary between training and inference compute markets.

New York-based AI accounting startup Basis announced a $100 million Series B round at a $1.15 billion valuation, led by Accel with participation from GV and others. Founded in 2023, the company focuses on using agentic AI to automate client accounting, tax filing, and audit processes. Reports cite its ability to autonomously complete complex tasks such as IRS Form 1065, powered by large models from providers like OpenAI to enable extended autonomous operation. Basis revealed it has established partnerships with about 30% of the 'top 25 accounting firms'. The company plans to use the funds to accelerate platform development and expand its engineering and machine learning teams to meet growing demand driven by chronic talent shortages in the accounting industry.