Super Micro Computer reported results for the first quarter of fiscal 2027, with revenue of $19.9 billion, above market expectations of $16.4 billion. Adjusted earnings per share were $1.81, and net income was $1.594 billion. The company raised its second-quarter revenue guidance to $19.5 billion to $21.5 billion and its full-year guidance to $72 billion to $82 billion. It said it is expanding capacity to meet data center demand from cloud service providers and technology companies.

In its updated IPO filing, Anthropic disclosed revenue of $33.24 billion for the first nine months of 2026, up 1,200% year over year, a net loss of $104 billion, and compute and infrastructure commitments of about $895.4 billion. The company also said it may be approaching its definition of recursive self-improvement, in which AI participates in developing more capable models.

Meta raised $15 billion through a bond offering, attracting orders of about $58 billion, or roughly 3.9 times the offering size. The company expects AI spending to exceed $740 billion over the next three years. Since 2025, its bond financing has totaled about $70 billion, second only to Alphabet and Amazon among major tech companies.

Anthropic has open-sourced FlashProve, a cryptographic verification system that lets users verify whether a response came from a precommitted model without disclosing the model weights. In tests with Llama 4 Maverick, a single NVIDIA H100 could generate proofs at a rate of about 110 tokens per second. A proof for a 1,024-token request was about 19 MB and could be verified on a standard CPU in about 17 seconds. The system is currently a research prototype; generating proofs consumes 70 to 95 times as much GPU time as inference itself.

SenseTime officially released SenseNova 6.8 Flash, a model that supports cross-session long-term memory and proactive interaction. It can link new information with users' previously recorded needs for reminders and task collaboration. In design scenarios, the model can generate dynamic presentations and 3D scenes, connect to tools such as Figma and Blender, and output design files that users can continue editing.

Tencent has open-sourced the WeVisOne-VL vision model, which has 4 billion parameters and supports text recognition, object detection, counting, and pixel-level segmentation. A lightweight decoder with about 9.6 million parameters handles fine edges and alpha channels, eliminating the need for additional segmentation models such as SAM. On some official benchmarks, it outperforms 8B and 9B models.

Irregular Security CEO Dan Lahav said cyberattacks targeting Anthropic, OpenAI, Meta, and Google stemmed from an internal testing configuration error: security isolation measures had not been set up correctly, allowing AI models to leave the sandbox and access external systems. He attributed the incident to process-management failures, not to the models actively bypassing safety mechanisms.

South Korea announced its 2027 spending plan, totaling 820.9 trillion won, up 12.8% year over year. Of that, 21.3 trillion won is earmarked for AI and semiconductors, and 4.7 trillion won for developing sovereign AI models. The government also plans to train 2,000 specialists through courses offered in partnership with companies such as Samsung and SK hynix, while providing employment opportunities.

OpenAI developed Pando, a Git backend, to handle the increase in code repository pushes and pulls driven by AI agents while preserving existing development workflows. The system supports parallel updates across branches and reuse of packfiles across clients. In production, it reduced median push latency by about 80% and transfers nearly 1 PB of Git data per day. Built in Rust, the project went from concept to production in about 10 weeks.

Google has introduced LiteRT Tensor API, a lightweight C++ interface that lets developers build computation graphs one operation at a time and control model architecture, optimization, and accelerator execution. The API supports directly loading checkpoints and Hugging Face safetensors weights, as well as zero-copy buffer sharing across CPU, GPU, and NPU. In a SAM 2 multi-object segmentation example, it ran 1.5 times faster than Apple MLX on macOS Metal.