March 2026 Summaries

NVIDIA's GTC 2026 conference in San Jose highlighted significant advancements in AI and GPU computing, drawing attention to pivotal developments such as the introduction of DLSS 5 and the shift from generative to agentic AI systems. The conference emphasized the evolving focus on AI inference and the transformation of AI datacenters into "AI Factories," which integrate training and inference for enhanced cost efficiency and reliability. With the launch of NemoClaw, NVIDIA aims to enhance trust in agentic AI frameworks, facilitating their adoption at scale by addressing security concerns. The event also underscored NVIDIA's renewed presence in the Chinese market with the resumption of H200 chip manufacturing, following U.S. government licensing. These developments, alongside innovations in AI-rendering technologies for gaming, signal a broader impact on the global AI market and the growing demand for GPU power in creative and industrial applications.

Andrej Karpathy has introduced autoresearch, a framework that enables an AI agent to autonomously conduct machine learning research experiments, optimizing model architecture and hyperparameters without human intervention. By deploying a single H100 GPU overnight, the framework can execute around 100 experiments, each lasting five minutes, allowing the agent to modify, train, evaluate, and decide on the retention of improvements. This innovation addresses the bottleneck in ML research by automating the repetitive experiment loop, thus freeing researchers to concentrate on more complex tasks requiring human intuition. The framework features a simplified implementation of nanochat, a small GPT model trained on FineWeb-Edu, which the AI agent adapts within its constraints. A guide for running autoresearch on Vast.ai includes instructions for setting up the necessary environment and monitoring the experiments, requiring only a short setup time and accounts on Vast.ai and Claude Code.

Google Colaboratory, commonly known as Google Colab, is a browser-based platform that allows users to write and execute code in a Jupyter notebook environment without local setup, making it popular for Python-based machine learning and data analysis projects. While it offers benefits like accessibility, Google Drive integration, and collaboration, users face limitations such as dynamic usage caps, idle timeouts, and restricted GPU access, even in paid plans. In contrast, Vast.ai provides a more flexible and powerful cloud GPU platform with customizable hardware options, no time restrictions, and reliable performance, making it an attractive alternative for more demanding workflows. Vast allows projects to run uninterrupted with scalable GPU configurations and simplifies the process of running Jupyter notebooks in the cloud, offering a stable platform for real-time collaboration without the constraints found in Google Colab.

NVIDIA's Nemotron-Cascade-2-30B-A3B is a cutting-edge reasoning model designed for high performance in math and programming competitions, achieving gold-medal-level scores with just 30 billion total parameters and 3 billion active per token, markedly fewer than competing models. This model, which fits on a single 80 GB GPU, integrates built-in chain-of-thought reasoning and tool-based capabilities for math and code, and is compatible with OpenAI's API. It can be efficiently deployed on Vast.ai using vLLM, offering significant cost savings over traditional cloud providers by utilizing A100 or H100 GPUs. The deployment process involves setting up a Vast.ai account, securing an API key, and configuring the environment for the model to serve as an OpenAI-compatible API server, which can be accessed to test its reasoning capabilities. Nemotron-Cascade-2's efficiency and performance make it an attractive solution for demanding reasoning tasks, offering a practical blend of capability and cost-efficiency.

Vast.ai simplifies the deployment of AI projects by offering a diverse Model Library filled with prebuilt templates that cater to various modalities and use cases, including image, text, audio-video, and multimodal AI workflows. These templates provide ready-to-run environments for popular AI tools and allow users to maintain full control over runtime configurations, ensuring efficient and customizable setups. Highlighted models include FLUX.2-Dev for advanced text-to-image generation, LTX-2 for synchronized audio-video production, WAN 2.2 for cinematic video creation, ACE-Step V1 for fast music generation, Dia 1.6B for realistic dialogue synthesis, DeepSeek V3.2 Exp for long-context reasoning, and Qwen3.5 397B A17B for complex multimodal tasks. The platform's offerings allow developers to quickly launch and iterate on AI projects, supported by templates that enhance memory usage, processing efficiency, and output quality across various domains.

SkyPilot, an open-source framework developed by UC Berkeley, simplifies running AI workloads across multiple cloud providers by abstracting cloud infrastructure management. It integrates with over 25 cloud services, including Vast.ai, known for offering cost-effective on-demand GPU compute options. Through a unified interface, SkyPilot automates tasks such as resource discovery, provisioning, cost optimization, and failover, ensuring efficient allocation of resources. The framework supports advanced Vast-specific configurations, allowing users to leverage features like instance selection and spot pricing for cost savings. The guide details how to set up and deploy AI models, specifically using the example of deploying the DeepSeek R1 model on a Vast.ai H100 GPU, while emphasizing the importance of proper configuration and management of credentials, instances, and deployments. Additionally, it provides troubleshooting tips and outlines limitations, such as the lack of support for multi-node clusters and object store mounting, while encouraging users to explore further resources and documentation for maximizing the utility of SkyPilot.

Alibaba's Qwen 3.5 model family is an advanced hybrid architecture that combines Gated DeltaNet and standard attention to enable significantly faster inference speeds, achieving up to 8.6x faster performance at a 32K context and 19x at 256K while maintaining strong reasoning capabilities. The medium models released in February 2026 include the Qwen3.5-122B-A10B, Qwen3.5-35B-A3B, and Qwen3.5-27B, each with different configurations of parameters and VRAM requirements, with the MoE models activating only a fraction of their total parameters per token to achieve large-model quality at a reduced inference cost. These models are Apache 2.0 licensed, allowing for easy deployment without needing HuggingFace tokens. The guide provides detailed instructions for deploying the Qwen3.5-35B-A3B model on Vast.ai using SGLang, emphasizing the efficient use of resources such as an 80 GB GPU and specifying settings like memory allocation and context length to optimize performance. With the reasoning-parser qwen3, the model can separate its reasoning process from the final answer, making it suitable for complex tasks, and various quantization options are available for different hardware configurations, including consumer GPUs like the RTX 4090.

Liquid AI has introduced LFM2-24B-A2B, a groundbreaking 24 billion parameter foundation model that activates only 2.3 billion parameters per token, making it one of the most aggressively sparse mixture-of-experts models available. This model diverges from traditional transformer-only models by employing a hybrid architecture that incorporates gated short convolution blocks instead of predominantly using attention layers, with only 10 out of 40 layers utilizing grouped query attention while the rest are convolutional. This innovative design allows the model to maintain competitive performance across various benchmarks such as MMLU-Pro and GPQA Diamond, despite its low active parameter count. Liquid AI, stemming from MIT, has scaled this architecture from 350M to 24B parameters, following log-linear scaling laws, and the LFM2-24B-A2B stands as the first in its family to be broadly applicable. The model's deployment is facilitated by day-one support for vLLM, SGLang, and llama.cpp, and it can be efficiently run on platforms like Vast.ai due to its compact weight of approximately 48 GB at BF16, negating the need for quantization or multi-GPU setups.

Vast.ai has emerged as a leading vendor in the AI compute infrastructure sector, recognized by corporate spend platforms Ramp and Brex for its significant growth based on actual corporate expenditures rather than fundraising or media coverage. Ramp's March 2026 report identified Vast.ai as a breakout growth vendor, particularly in the "Agent hosting" category, highlighting a shift in AI workloads from prototypes to production systems. Meanwhile, Brex ranked Vast.ai #11 among the fastest-growing software vendors of 2025. Over the past year, Vast.ai has seen a dramatic increase in new signups and active paying customers, alongside substantial revenue growth, driven by its marketplace offering faster, cheaper, and flexible GPU compute solutions. The company attributes its success to its model-agnostic, cloud-agnostic, and contract-free platform that allows teams to deploy quickly and scale efficiently. As the AI infrastructure landscape evolves, Vast.ai is focusing on enhancing reliability, autoscaling, and integrations for agent workloads to maintain its growth trajectory and meet the needs of its expanding user base.

Claude Code is an advanced AI tool for software development but is limited by usage restrictions and costs. An alternative is running open-source coding agents like Ralph on Vast.ai, costing around $1.50 per hour, allowing continuous operation without API limitations. Ralph autonomously progresses through software projects by selecting user stories, writing and testing code, and addressing failures until completion. Using the Qwen3-Coder-Next model on Vast.ai provides cost-effective, scalable GPU resources and greater control over the development environment, supporting security and compliance needs. Ralph is well-suited for overnight workflows and can handle tasks like building CLI applications, implementing REST APIs, generating test suites, and creating web scrapers. The Overnight Ralph Loop guide details the setup process for running Ralph on Vast.ai, offering an affordable solution for sustained software development.