April 19, 2026 – AI Daily Recap

Saturday brought another wave of significant moves across the AI landscape. From massive infrastructure deals to new model releases and bold product launches, here’s everything you need to know.

OpenAI Commits $20 Billion to Cerebras in Landmark Chip Deal

OpenAI has doubled down on its relationship with AI chipmaker Cerebras, committing more than $20 billion over three years to purchase servers powered by Cerebras chips. The deal, first reported by The Manila Times, expands an earlier $10 billion agreement signed in January. As part of the arrangement, OpenAI will receive warrants that could give it up to a 10% equity stake in Cerebras, plus contribute $1 billion toward data center construction. The timing is notable: Cerebras has simultaneously filed for a Nasdaq IPO targeting a $35 billion valuation. The deal underscores a broader industry shift as inference — running trained models at scale — is expected to account for two-thirds of all AI compute spending by the end of 2026.

Anthropic Releases Claude Opus 4.7 with Major Engineering Gains

Anthropic has released Claude Opus 4.7, the latest flagship in its Claude model family. The update delivers what Anthropic calls “notable improvement” on advanced software engineering tasks, particularly on the most difficult problems. Vision capabilities have been tripled, now processing images up to 2,576 pixels on the long edge, which opens the door for more precise computer-vision agents and diagram analysis. A new xhigh effort level for extended reasoning gives developers finer control between depth and latency. The model is available across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at existing pricing: $5 per million input tokens and $25 per million output tokens.

NVIDIA Launches Nemotron 3 Super: Open Model, Massive Efficiency

NVIDIA has released Nemotron 3 Super, a 120B-parameter model that activates only 12B parameters at inference time using a hybrid Mamba-Attention Mixture-of-Experts architecture. The result is striking: up to 2.2x higher throughput than competing 120B models, a native 1M-token context window, and 5x the throughput of its predecessor. Pretrained on 25 trillion tokens using NVIDIA’s NVFP4 format optimized for Blackwell GPUs, the model targets complex agentic workflows including multi-step coding, cybersecurity triage, and tool calling. Crucially, Nemotron 3 Super is fully open — weights, datasets, and training recipes are all available for developers to customize and deploy on their own infrastructure.

Perplexity Launches “Personal Computer” Agentic OS for Mac

Perplexity has launched “Personal Computer,” an agentic AI assistant for Mac that can autonomously manage files, navigate apps, research the web, and execute multi-step workflows. Available to Perplexity Max subscribers ($200/month), the feature orchestrates teams of agents across more than 20 frontier models. Perplexity recommends running it on a Mac mini for persistent, always-on operation. All actions are auditable and reversible, with a kill switch for human oversight. The move marks a significant escalation in the race to build truly autonomous desktop agents.

Factory Raises $150M at $1.5B Valuation for Enterprise AI Coding

Factory, the enterprise AI coding startup, has raised $150 million in a round led by Khosla Ventures with participation from Sequoia Capital, Insight Partners, and Blackstone, reaching a $1.5 billion valuation. Founded by Matan Grinberg and Eno Reyes, Factory builds “Droids” — AI agents that handle not just code generation but also testing, review, documentation, and deployment. Hundreds of thousands of developers at organizations including NVIDIA, Adobe, MongoDB, and Zapier already use the platform daily. The raise signals continued investor confidence in enterprise-grade AI coding tools built for compliance, legacy code, and security requirements.

Anthropic Publishes Subliminal Learning Research in Nature

Anthropic has published a paper in Nature examining how large language models can transmit hidden traits — including preferences and subtle biases — through subliminal data signals. The research investigates how LLMs might inadvertently acquire and propagate characteristics that aren’t explicitly present in their training objectives, raising important questions for AI alignment and safety. The findings are particularly relevant as models become more capable and are increasingly used in high-stakes decision-making contexts where subtle behavioral biases could have outsized impact.

The Big Picture

This week’s developments paint a clear picture of an industry entering its infrastructure scaling era. OpenAI’s $20 billion Cerebras commitment and the chipmaker’s simultaneous IPO filing show that the economics of inference are now driving the biggest deals in tech. Meanwhile, the model releases from Anthropic and NVIDIA point in different but complementary directions — Opus 4.7 pushes the capability frontier on complex engineering tasks, while Nemotron 3 Super focuses on making powerful models dramatically more efficient to run. Perplexity’s desktop agent and Factory’s enterprise coding Droids represent the application layer catching up, turning raw model power into autonomous workflows. And Anthropic’s Nature paper on subliminal learning serves as a timely reminder that as these systems grow more capable, understanding their hidden behaviors becomes ever more critical.

Sources

• OpenAI to Spend More Than $20 Billion on Cerebras Chips — The Information
• OpenAI to spend more than $20 billion on Cerebras chips — The Manila Times
• AI chipmaker Cerebras files to go public — CNBC
• Claude Opus 4.7 — Anthropic
• Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI — NVIDIA Blog
• Introducing Nemotron 3 Super — NVIDIA Developer Blog
• Perplexity Launches Personal Computer for Mac — MacRumors
• Perplexity’s Personal Computer AI assistant launches on Mac — 9to5Mac
• Factory hits $1.5B valuation — TechCrunch
• AI coding startup Factory raises $150M — TechFundingNews
• Subliminal Learning in Large Language Models — Nature
• AI News Briefs Bulletin Board for April 2026 — Radical Data Science