OpenAI announced the completion of an $110 billion fundraising round, with a pre-money valuation of approximately $730 billion (some reports indicate a post-money valuation of around $840 billion). Amazon will contribute up to $50 billion ($15 billion upfront, with the remainder tied to milestones), while Nvidia and SoftBank will each invest $30 billion. The two companies also signed a multi-year partnership: AWS will become the sole third-party cloud distributor for Frontier models and jointly develop a 'Stateful Runtime' on Bedrock; OpenAI commits to using approximately 2GW of Trainium compute and plans to adopt Trainium3/4 in the future.

The Trump administration has directed federal agencies and military contractors to cease collaboration with Anthropic and designated it a 'supply chain risk,' allowing a maximum six-month transition period. The conflict arose when the Pentagon demanded unrestricted access to Claude for 'all lawful purposes,' but Anthropic refused, citing prohibitions against use in mass surveillance of U.S. citizens and fully autonomous weapons, stating that current safety boundaries are insufficient to permit such access. The incident may impact its government contracts and corporate client evaluations, setting a new precedent for negotiations between AI vendors and governments over usage red lines.

Google Cloud is offering a preview of the image generation and editing model Nano Banana 2 via Vertex AI and Gemini API, targeting enterprise creative and content production. Officially described as enhancing speed under the Flash lineage, the model improves instruction following, visual consistency, and multilingual text rendering, supporting 2K/4K super-resolution and multiple aspect ratios. It includes built-in SynthID watermarking and supports C2PA content credentials to enhance traceability and compliance disclosure of generated content; integration with various Google product lines will roll out gradually.

Perplexity has launched an autonomous task agent called 'Computer,' priced at $200 per month for Max subscribers. The product decomposes a single objective into multi-step workflows, routing tasks across specialized models executed within isolated cloud sandboxes equipped with controlled browsers, file systems, and validation tools—shifting focus from 'answering questions' to 'executing tasks.' Multiple sources report it can orchestrate around 19 models to handle complex combinations of research, content generation, multimedia creation, and script processing, emphasizing secure execution in isolated environments.

Truffle Security disclosed that Google API Keys intended for frontend use (e.g., Maps keys) could be 'silently upgraded' to high-privilege credentials capable of accessing sensitive Gemini endpoints when generative language APIs are enabled in a project. Many legacy keys lacking source or scope restrictions remain exposed in public code, enabling attackers to read data or trigger high-cost API billing. Google has tightened default restrictions on new keys and revoked known leaked ones, but urges enterprises to immediately scan, rotate, and refine key permissions to prevent abuse of legacy projects.

AI hardware design platform Flux announced $37 million in new funding: $27 million Series B led by 8VC and a previously undisclosed $10 million Series A (co-led by Outsiders Fund and Bain Capital Ventures). Flux assists with PCB layout, component selection, and testing via natural language prompts, generating full manufacturing-ready outputs entirely in-browser. The company claims over 1 million users have designed nearly 6.5 million devices across wearables, robotics, smart home, and drone applications. Funds will support expansion into more complex electronic applications.

A team including DeepSeek proposed the DualPath inference framework to address the KV-Cache transfer bottleneck in long-context agent scenarios. It introduces a second 'Storage→Decode' loading path: first loading cache into the decode engine, then forwarding it via RDMA to the prefill engine, leveraging idle network interface card (NIC) bandwidth on the decode side to offload I/O pressure. The framework also enables global bandwidth pooling and dynamic load balancing through centralized scheduling, with traffic isolation and priority management to reduce interference from cache migration on computation and communication—pointing toward I/O restructuring in inference systems.

He Kaiming's team introduced GeoPT, a pretraining paradigm using 'Synthetic Dynamics' to transform unlabeled 3D geometry data into learnable physical evolution trajectories: random velocity fields generate synthetic motion, enabling models to learn spatial constraints and coupling dynamics without relying on costly physics labels. The paper claims 20%–60% reduction in simulation data required for equivalent accuracy; optimized cross-algorithm pretraining data generation runs about 10^7 times faster than industrial-grade CFD. Experiments also show consistent downstream performance gains scaling from 3M to 15M parameters and larger datasets.

The open-source agent framework openclaw released version 2026.2.26, focusing on enhanced key and runtime governance: it adds the openclaw secrets external key management workflow, supporting auditing, configuration, application, and hot reloading. The ACP/thread-bound agent is now changed to a single-threaded runtime for improved lifecycle control, with CLI support for account-level routing management. The release prioritizes WebSocket transport by default for Codex and strengthens security measures across platforms—including execution approval, plugin channel HTTP authentication, gateway pairing, and SSRF protection—while fixing stability issues in messaging and file uploads across multiple chat and collaboration platforms.