Nvidia's Vera CPU heads to orbit, OpenAI unveils Jalapeño chip, and Perplexity goes local
Nvidia partners with SpaceX for orbital data centers using the new Vera CPU. OpenAI reveals its custom Jalapeño chip, while Perplexity launches a local-first agent platform. Thomson Reuters builds a proprietary legal model, and Meta prepares its 'Hatch' agent platform.
Podcast В· 2 min
Nvidia's Vera CPU and SpaceX's orbital data centers
Nvidia has officially launched its new Vera CPU, a chip specifically designed to handle the 'busywork' of agentic AI—coordinating tools, running code, and managing databases. SpaceXAI has signed on as the first high-profile customer, planning to deploy the Vera architecture to power its next generation of AI agents. Elon Musk revealed that SpaceX intends to take this hardware into orbit, with plans to establish 'Starmind' data centers by late 2027. The Vera platform, which features 88 'Olympus' cores and up to 1.2 terabytes per second of memory bandwidth, is built to address the bottleneck where expensive GPUs sit idle waiting for CPU instructions. By optimizing for agentic task completion, Nvidia claims the chip is up to 1.8x faster than traditional x86 alternatives. This move signals Nvidia's aggressive expansion into the standalone CPU market, targeting a $200 billion opportunity. For SpaceX, the orbital data center project represents a significant bet on the future of infrastructure. While critics have previously dismissed space-based computing as impractical, the partnership with Nvidia provides the necessary hardware density and efficiency to make the concept viable. The space-hardened racks are designed to be simpler, lighter, and more cost-effective than terrestrial equivalents, potentially flipping the economics of orbital compute in the coming years.
OpenAI unveils custom 'Jalapeño' AI chip
OpenAI has published the first benchmark results for 'Jalapeño,' a custom AI chip developed in partnership with Broadcom. Designed specifically for running AI models rather than training them, the 700-watt chip reportedly outperforms Nvidia's flagship 1,200-watt systems in both speed and power efficiency. OpenAI credits its Astra model and Codex for assisting in the chip's design, which took nine months from concept to manufacturing-ready status. Despite the strong performance, OpenAI does not plan to sell the chip commercially. Hardware VP Richard Ho stated that the company has sufficient internal demand to utilize the entire supply. The chip is expected to reach OpenAI's data centers later this year, with production ramping up through 2027. Two additional generations of the chip are already in development. This development mirrors the strategy of other major labs like Google, which has long benefited from its in-house TPUs. By owning its silicon, OpenAI aims to reduce the cost of token generation and create a tighter, more personalized design loop for its future models, reducing its reliance on third-party hardware for inference.
Perplexity and Nvidia launch 'Portable Computer'
Perplexity and Nvidia have introduced 'Portable Computer,' a new on-device agent platform that allows users to run AI locally on Nvidia's DGX Spark consumer hardware. The system offers a privacy-focused alternative to cloud-based agents, allowing users to choose between Qwen 3.8 27B or Perplexity’s own PPLX 27B models for local tasks, with the ability to call upon cloud models when necessary. This release is particularly significant for its cost and privacy implications. By moving the AI workload from the cloud to the device, users can perform tasks without consuming credits, provided the work stays local. The software installs in a single click on the DGX Spark desktop supercomputer, with support for other PCs expected soon. The launch follows reports that Nvidia is in talks to invest billions in Perplexity at a valuation exceeding $30 billion. This move positions Perplexity to capture the growing demand for local-first AI, turning frontier models into an occasional backup rather than the default, and leveraging Nvidia's hardware dominance to secure its market position.
Meta reportedly preparing 'Hatch' agent platform
Meta is reportedly set to launch 'Hatch,' a new consumer-facing AI agent platform, within the coming weeks. The platform is designed to complete tasks on behalf of users and will feature a premium tier priced competitively with top-tier plans from OpenAI and Anthropic. Alongside the platform, Meta is expected to release its next flagship model, codenamed 'Watermelon,' in October. This move marks a strategic shift for Meta, which has historically relied on free, ad-supported services. By introducing a paid agent platform, the company is attempting to monetize its AI capabilities directly through consumer subscriptions. The success of this initiative will depend on the agent's ability to handle complex, multi-step tasks effectively. Industry analysts are watching closely to see if Meta can successfully transition its massive user base to a paid AI model. The platform's ability to integrate with existing social ecosystems while providing genuine utility will be the key differentiator in an increasingly crowded market for personal AI agents.
Thomson Reuters builds proprietary legal AI model
Thomson Reuters has introduced its first homegrown AI model, developed by retooling Alibaba's open-source Qwen and training it on decades of the firm's legal content. The two-year project cost $40 million in computing power and staff, with the most recent training run costing $450,000. The model has been tested internally and reportedly benchmarks ahead of Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5 in specific legal tasks. CTO Joel Hron emphasized the strategic importance of this move, comparing renting AI models to renting a house, where the user builds no long-term equity. By building its own model, Thomson Reuters aims to retain value and control over its specialized knowledge base. An open-weights version of the model will be made available to researchers. This development highlights a growing trend among large enterprises with deep, proprietary data sets. As training costs decrease and open-source base models improve, companies are increasingly choosing to build custom solutions rather than relying solely on third-party APIs, aiming for better performance and long-term cost efficiency.
Accelerated Understanding launches physics-based AI
Caltech professor Anima Anandkumar and engineer Benedikt Jenik have launched 'Accelerated Understanding,' a startup focused on training AI to forecast physical world evolution rather than predicting text. The company's models utilize 'neural operators'—a physics-based design that tracks events through 3D space and time, departing from the traditional transformer architecture. In initial tests, the model processed 5 trillion data points in a single run, which the founders claim is 5 million times the capacity of top models from Google and Anthropic. The company is targeting enterprise applications, including chip material development, extreme-weather prediction, and robotics. The founders notably turned down an offer of top positions and a 35% stake in Jeff Bezos’ 'Prometheus' project to pursue this venture. This decision underscores their conviction that physics-based modeling is the superior path for the next generation of AI, positioning the startup as a direct competitor to the massive, text-heavy models currently dominating the industry.
Chinese hackers increase attacks using DeepSeek
Research from TeamT5 indicates that Chinese state-backed hacking groups have more than doubled their attack volume after integrating open-source AI models, specifically DeepSeek, into their operations. The researchers report that these groups favor DeepSeek due to its combination of high performance and relatively loose cyber guardrails compared to other frontier models. The hackers are reportedly using the AI to assist in writing break-in code, raiding corporate email systems, and mapping targets. While security concerns have largely focused on the potential misuse of top-tier models, this report highlights the significant risk posed by cheaper, more accessible models that lack robust safety restrictions. This trend aligns with warnings from the AI Security Institute, which noted in May that autonomous AI cyber capabilities are advancing rapidly. The ability of malicious actors to leverage open-source AI to scale their operations presents a complex challenge for cybersecurity, suggesting that the focus may need to shift toward securing the models themselves rather than just the infrastructure they target.