Palantir Technologies has announced a new sovereign AI operating system reference architecture developed in partnership with Nvidia. The system is designed to provide organizations with a turnkey AI data center environment that integrates hardware infrastructure, data management, and AI application deployment.

The platform, called the Palantir AI OS Reference Architecture (AIOS-RA), combines Nvidia’s AI hardware and software stack with Palantir’s enterprise data and AI platforms. The companies said the architecture delivers a production-ready system capable of running advanced AI workloads while maintaining strict data sovereignty and operational control.

The reference architecture is built on Nvidia’s Enterprise Reference Architectures and is optimized to run Palantir’s full software ecosystem. This includes the company’s Artificial Intelligence Platform (AIP), Foundry data platform, Apollo deployment system, Rubix security framework, and AIP Hub.

Full-Stack AI Data Center Design

The system integrates Nvidia’s Blackwell Ultra AI infrastructure, including servers equipped with eight Blackwell Ultra GPUs and Spectrum-X Ethernet networking designed for AI training and inference workloads.

Alongside the hardware layer, the architecture includes a hardened Kubernetes-based compute environment for running Foundry services such as data cataloging, pipeline development, and analytics workloads.

Palantir’s Rubix platform provides a zero-trust security model for Kubernetes infrastructure, while the Apollo system automates deployment and lifecycle management across distributed environments.

The architecture also incorporates Nvidia’s AI Enterprise software stack, CUDA-X libraries, Nemotron open AI models, and Magnum IO data acceleration tools to improve training and inference performance.

Supporting Sovereign and Edge AI Deployments

The companies said the architecture is particularly suited for organizations operating in sensitive environments where data control and regulatory compliance are critical. These include government agencies, defense organizations, and industries with strict data residency requirements.

By enabling on-premise and edge deployments, the sovereign AI architecture allows organizations to run AI workloads locally without relying on public cloud infrastructure. This approach can also reduce latency for mission-critical applications and enable AI deployments in geographically distributed environments.

“From our first deployment with the United States government and in every deployment since, our software has had to meet the moment in the most complex and sensitive environments where customers must maintain control,” said Palantir Chief Architect Akshay Krishnaswamy.

Nvidia’s enterprise AI platforms team said the collaboration reflects growing demand for integrated AI infrastructure solutions.

“AI is redefining the infrastructure stack — demanding, latency-sensitive and data-sovereign environments require a full-stack architecture,” said Nvidia Vice President Justin Boitano.

The companies said the joint architecture is designed to help organizations convert large volumes of operational data into actionable intelligence while maintaining control over infrastructure, models, and applications.

Google has released Gemini Embedding 2, its first fully multimodal embedding model built on the Gemini architecture. The model is now available in public preview through the Gemini API and Vertex AI platform.

The new system expands beyond traditional text embeddings by mapping multiple data types into a single vector space. Gemini Embedding 2 can process text, images, video, audio, and documents, enabling applications that require semantic understanding across different media formats.

Embeddings are numerical representations of content that allow AI systems to understand relationships between pieces of information. They are widely used in tasks such as semantic search, recommendation systems, clustering, and retrieval-augmented generation.

By integrating multiple modalities, the model allows developers to simplify pipelines that previously required separate systems for different types of data. According to Google, the unified embedding space can capture relationships between media types, improving the accuracy of multimodal search and analysis.

Multimodal Capabilities and Input Support

Gemini Embedding 2 supports a wide range of inputs across several formats. The model can process up to 8,192 tokens of text, six images per request in PNG or JPEG format, and up to 120 seconds of video in MP4 or MOV format.

The model can also ingest audio directly without requiring transcription and can embed PDF documents of up to six pages. Developers can combine multiple input types within a single request, such as pairing images with text prompts to generate richer semantic representations.

This multimodal capability allows AI systems to analyze complex datasets where information appears in different formats. For example, developers could use a photo or video clip as a search query to retrieve related media assets or documents.

Flexible Embedding Dimensions and Performance Gains

Like earlier Google embedding models, Gemini Embedding 2 uses Matryoshka Representation Learning, a technique that compresses information across nested vector dimensions. This allows developers to adjust embedding sizes depending on performance and storage requirements.

The model’s default vector size is 3,072 dimensions, though developers can scale outputs to smaller dimensions such as 1,536 or 768 for more efficient storage and processing.

Google said the model demonstrates improved performance across text, image, video, and speech tasks compared with previous embedding models. The company noted that the model’s multimodal capabilities could enable new applications in areas such as media search, sentiment analysis, and data clustering.

Early partners are already using the technology to build advanced search tools. Paramount Skydance reported that the model improved its ability to locate video assets based on text queries, allowing the system to identify visual expressions and retrieve related media more accurately.

Gemini Embedding 2 is now accessible through Google’s developer tools and can integrate with AI frameworks including LangChain, LlamaIndex, and vector databases such as Weaviate, Qdrant, and ChromaDB. Google said the model is designed to serve as a foundation for next-generation multimodal AI systems that work with diverse forms of data.

Meta has unveiled four new custom chips designed for artificial intelligence workloads as part of its expanding data center infrastructure. The processors are part of the company’s Meta Training and Inference Accelerator (MTIA) family, a line of chips built to handle specific AI tasks within Meta’s platforms.

The announcement marks the latest step in Meta’s push to reduce reliance on external chip vendors by designing its own silicon optimized for internal workloads. According to Meta Vice President of Engineering Yee Jiun Song, custom chips allow the company to improve price-performance efficiency across its data centers while diversifying its hardware supply chain.

“This also provides us with more diversity in terms of silicon supply and insulates us from price changes to some extent,” Song said.

Meta first revealed the MTIA architecture in 2023 and released a second-generation version in 2024. The new lineup significantly expands the platform as the company scales its AI infrastructure to support recommendation systems and generative AI features across its products.

Four New Chips Target Different AI Workloads

The first of the new chips, MTIA 300, has already been deployed in Meta data centers. It is designed to train smaller AI models used for core platform tasks such as ranking and recommendation algorithms. These systems determine which posts, ads, and videos appear in users’ feeds across services including Facebook and Instagram.

Three additional chips are currently in development. The MTIA 400 is nearing deployment after completing testing, while the MTIA 450 and MTIA 500 are scheduled to become operational by 2027.

Unlike the MTIA 300, the upcoming chips will focus on generative AI inference workloads such as generating images and videos from text prompts. However, Meta said the processors will not be used to train large language models.

Song noted that Meta plans to release new chips roughly every six months as the company rapidly expands computing capacity. Each chip generation is expected to remain in service for more than five years.

Data Center Expansion and AI Infrastructure

The new processors will support Meta’s massive data center expansion. The company is currently building a large facility in Louisiana and additional centers in Ohio and Indiana. Reports also indicate that Meta is exploring leasing space at a major AI data center site in Texas.

The custom chips are manufactured by Taiwan Semiconductor Manufacturing Company, though Meta did not confirm whether production will occur at the company’s new fabrication facilities in Arizona.

Meta’s in-house silicon strategy follows a broader industry trend among major technology companies developing application-specific integrated circuits, or ASICs, for AI workloads. These chips are typically smaller and more energy-efficient than general-purpose GPUs but are optimized for narrower tasks.

Google pioneered this approach with its Tensor Processing Units in 2015, and Amazon followed with custom chips for its cloud services in 2018. Unlike those companies, Meta uses its MTIA chips exclusively for internal operations rather than offering them through a public cloud platform.

Despite building its own silicon, Meta continues to rely heavily on external hardware suppliers. The company recently signed agreements to deploy millions of Nvidia GPUs and up to six gigawatts of AMD GPUs across its data centers.

Song acknowledged that securing high-bandwidth memory remains a potential constraint as AI infrastructure spending increases across the technology industry. However, he said Meta has diversified its supply chain and believes it has secured the resources required for its planned deployments.