[AINews] New AI Infra unicorns: Exa, Modal, TurboPuffer

2026-05-22 · Source: Latent.Space - Www.latent.space · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

This AI News recap for May 20-21, 2026, details significant advancements across AI infrastructure, models, agents, and specialized applications. Three AI infrastructure companies achieved major milestones: Turbopuffer reached \$100M ARR and profitability with less than \$1M raised, Exa secured a \$250M Series C at a \$2.2B valuation, and Modal raised \$355M in Series C funding at a \$4.7B valuation. Research updates include RAEv2 showing >10x faster convergence for unified vision, NVIDIA's Gated DeltaNet-2 outperforming other linear attention models at 1.3B parameters, and insights into data filtering. Agent development saw Codex expanding with Appshots and remote Mac control, while Gemini 3.5 Flash topped APEX-Agents-AA benchmarks and broadened its tool integration. Developer infrastructure is converging on retrieval, streaming, and sandboxes. Compute remains a strategic bottleneck, with HBM growing to 63% of AI chip component spending by Q4 2025. Multimodal, biology, and robotics also saw progress, including Runway's Edit Studio and Hugging Face's open-source LeRobot Humanoid.

Key takeaway

For AI Scientists and Machine Learning Engineers evaluating new model architectures or agent systems, you should prioritize exploring RAEv2 for vision tasks and NVIDIA's Gated DeltaNet-2 for language modeling, given their reported performance gains. Consider integrating agent harnesses like physics-intern to boost existing model capabilities. Furthermore, if you are building agent infrastructure, focus on robust retrieval mechanisms, secure sandboxes, and efficient streaming protocols, as these are becoming critical for scalable and reliable deployments.

Key insights

The AI ecosystem is rapidly maturing, marked by significant infrastructure funding, model advancements, and agent-centric product development.

Principles

• "Boring" AI infrastructure drives significant wealth creation.
• Harnesses provide substantial model capability gains.
• Agent design should prioritize single-agent systems initially.

Method

RAEv2 improves by summing K encoder layers and reformulating REPA as internal self-guidance. Gated DeltaNet-2 decouples erase/write in linear attention using channel-wise gates. Mechanistic interpretability suggests clustering SAE features by joint firing patterns.

In practice

• Integrate hybrid BM25 + vector retrieval in agent infra.
• Deploy sandbox Auth Proxies for agent security boundaries.
• Explore 3D-printed open-source humanoids for robotics.

Topics

• AI Infrastructure
• AI Agents
• Model Architectures
• Compute Economics
• Multimodal AI
• Open Robotics

Best for: AI Engineer, Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• [AINews] New AI Infra decacorns: Fireworks, Baseten (with OpenRouter on the way) — The AI landscape is rapidly maturing, shifting focus from raw model power to robust inference infrastructure and sophisticated agentic harnesses.
• AI Infrastructure War: Meta’s Agent Network, NVIDIA’s Gigawatt Deal & Oracle’s AI Surge — The AI race is shifting focus from models to infrastructure, platforms, and novel data-efficient methods.
• not much happened today — AI performance increasingly depends on context pipelines and open harnesses, not just model weights.
• [AINews] not much happened today — AI progress spans non-invasive BCIs, efficient inference, agent orchestration, and open-weight model commercialization, driven by infrastructure and evaluation.
• not much happened today — Rapid advancements in non-invasive brain-to-text, efficient AI inference, and sophisticated agent orchestration are reshaping AI development.
• The TechBeat: We brought Hermes Agent to iMessage, even on Linux and Windows (6/25/2026) — The tech landscape is rapidly evolving with AI-driven innovations and infrastructure challenges across diverse domains.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.