Anthropic has priced its IPO at $83 a share, offering 980 million shares, with trading expected to begin on Nasdaq on November 19, 2026, under the ticker ANT. OpenAI says its models helped AMD optimize GPU circuits, improving power, performance and area metrics by up to 14% over highly optimized baselines, with some changes already included in designs sent for manufacturing. According to The Wall Street Journal, Palo Alto Networks has agreed to acquire Applied Compute for $4.5 billion in an all-stock deal, not a cash acquisition.
Both use equity, but going public to raise capital and acquiring capabilities send funds in different directions.
Anthropic has priced its initial public offering at $83 a share, offering 980 million shares. Of these, approximately 956 million are being sold by the company and about 23.65 million by existing shareholders. Trading is expected to begin on Nasdaq on November 19, 2026, under the ticker ANT. The company will not receive any proceeds from existing shareholders' sales. Underwriters also have a 30-day option to purchase up to 147 million additional shares.
Read original →According to The Wall Street Journal, Palo Alto Networks has agreed to acquire Applied Compute for $4.5 billion in an all-stock deal. Applied Compute helps enterprises customize open-weight models—AI models whose weights are available to access. The acquisition price exceeds the company's $3.25 billion valuation earlier this month.
Read original →🔍 Analysis: Anthropic's offering combines fundraising for the company with share sales by existing shareholders, while Palo Alto Networks plans to exchange stock for Applied Compute's enterprise model-customization capabilities. In the former, proceeds going to the company must be distinguished from those going to selling shareholders; in the latter, the $4.5 billion deal consideration must be distinguished from cash spending. Assessing capital investment in AI requires more than comparing deal sizes.
Chip design, model training and always-on local tasks place different practical demands on hardware.
OpenAI says its models rewrote circuit-description code to improve power, performance and area metrics for modules in AMD's flagship GPUs by up to 14% over highly optimized baselines. The team used Synopsys chip-design tools to verify that circuit behavior remained unchanged after the rewrites and to evaluate the improvements. Some changes have already been included in AMD designs sent for manufacturing, while others have been added to production design files for use in subsequent versions.
Read original →According to the South China Morning Post, Alibaba CEO Eddie Wu said on Thursday's earnings call that the company is using in-house chips to train its next-generation AI models. Alibaba is accelerating its shift to domestically produced chips. Wu said infrastructure powered by its in-house, domestically produced chips is becoming the foundation of Alibaba's full-stack AI capabilities. The company also reported revenue and profit growth that both exceeded expectations.
Read original →RTX Spark compact desktops from ASUS, Dell and MSI are now on sale, running Windows and targeting local AI workflows. NVIDIA says the devices offer up to 128GB of unified memory and can run 120-billion-parameter models locally, while also supporting image and video generation and model fine-tuning. They are designed for independent, always-on operation: agents and scheduled tasks can keep running even when a laptop is taken elsewhere or used for other work.
Read original →Beyond context-window size, per-task costs, cache memory usage and attention computation determine the cost of working with long texts.
On November 18, Google released Gemini 4 Neon, which supports a one-million-token context window and can process documents, charts, audio and video together. Google says its per-task cost at the default reasoning setting is 32% lower than Carbon's. It offers four reasoning levels so teams can adjust capability and spending to suit each task. Pricing is $3 per million input tokens and $15 per million output tokens, and access is now rolling out to API users.
Read original →Tencent has released the weights for the instruction-tuned and base versions of Hy4 Flash, supporting text and image inputs and a one-million-token context window. Tencent says its inference cache memory usage is one-sixth that of Hy4, while its average per-task cost in the internal WorkIndex evaluation is one-twentieth of Hy4's. The base version is a pretrained model, not an instruction-tuned chat model.
Read original →Zyphra announced that it is applying its ZHA attention method to open-source models including Qwen and Nemotron. The method's computational complexity grows more slowly than the square of context length, and Zyphra says it delivers a 9–60× speedup at 10 million tokens. The company also says the converted models retain their performance on reasoning and agent tasks.
Read original →🔍 Analysis: Gemini 4 Neon and Hy4 Flash both support a one-million-token context window, but the former provides API pricing and task costs relative to Carbon, while the latter offers open weights and emphasizes cache usage and costs in internal evaluations. The same context-window length does not imply the same deployment model. ZHA targets attention computation at even longer context lengths; its reported 9–60× speedup at 10 million tokens cannot be directly equated with end-to-end task cost savings. For developers, comparing options requires separating API fees, local memory requirements and long-sequence computational efficiency, rather than putting 32%, one-twentieth and 9–60× on the same performance leaderboard.
On November 18, Google launched Genie 4 in Project Genie, allowing users to change the weather and add characters through text while exploring, with generated audio that adapts to the scene. A new agent can build worlds through conversation, and a community gallery supports sharing and remixing creations. The new experience is rolling out to eligible Google AI Ultra subscribers, while Pro users can join a waitlist.
Read original →