NVIDIA Unveils 2.42x Faster Parallel Token Generation; Thinking Machines Hits Hundreds of Millions ARR; Tesla Supercharger Delivers 2 TWh

Key Takeaways

1. NVIDIA's Nemotron-Labs-TwoTower model achieves 2.42x generation speedup with 98.7% quality by parallel token writing.
2. Hugging Face optimizes DeepSeek V4 Pro to reach Sonnet 4.6-level legal benchmark at 1/7 the cost.
3. Thinking Machines' Tinker API reaches hundreds of millions ARR, the highest among neolabs.
4. Tesla Supercharger network delivered 2 TWh in Q2 2026, enabling 60 million charging sessions.
5. Ethan Mollick emphasizes use-case-specific model testing as benchmarks fail to capture real-world differences.
6. Yann LeCun introduces LeVLJEPA, a contrastive-free vision-language pre-training method surpassing CLIP.

1. AI Model Innovations and Performance Gains

1. NVIDIA AI introduced Nemotron-Labs-TwoTower, a diffusion language model that splits a 30B model into two parts to write tokens in parallel, achieving 2.42x generation speedup while maintaining 98.7% quality. — via 1 2
2. Hugging Face shared that DeepSeek V4 Pro was optimized using harness techniques, boosting all-pass rate on the Legal Agent Benchmark from 0% to 5%, reaching Sonnet 4.6 level at 1/7 the cost. — via 1
3. Yann LeCun highlighted LeVLJEPA, the first fully non-contrastive end-to-end vision-language pretraining method, which outperforms CLIP and SigLIP on large datasets without negative samples, momentum, or teacher networks. — via 1
4. NVIDIA's Nemotron 3 Ultra is rapidly growing on Together AI, reaching 35B tokens/day, reflecting community demand for open models. — via 1
5. Hugging Face noted that the White House's Rampart model became the top-ranked token classification model on Hugging Face, signaling a shift toward public organizations owning their own weights. — via 1

2. AI Infrastructure and Commercial Deployment

1. NVIDIA announced that AI is shifting from training to continuous token production, requiring new business models. NVIDIA collaborates with AI clouds to deploy large-scale multi-tenant AI factories through revenue sharing and credit support, providing compute access for startups, model builders, and others. — via 1
2. Elon Musk shared that Tesla's Supercharger network delivered 2 TWh of electricity in Q2 2026, equivalent to ~180,000 US households' annual usage, and enabled 60 million charging sessions. — via 1
3. Deedy reported via Dylan Patel that Thinking Machines' Tinker API (helping post-train large models) has reached hundreds of millions of dollars in ARR, the highest known revenue among ~75 neolabs. The company was valued at $12B and is seeking $50B. — via 1
4. Open AI models (OpenAI GPT OSS and NVIDIA Nemotron) are now available on Amazon Bedrock, with FedRAMP High and DoD IL-4/5 certifications. — via 1
5. Starlink application cases: Japan tests converting fire hydrant signs into disaster communication hubs; Paraguay connects 1,600+ remote schools covering over 50,000 students via Starlink. — via 1 2

3. Agent Systems and Practical Insights

1. Ethan Mollick stressed that users must test models on specific use cases because standard benchmarks fail to capture differences when decision-stacking. For example, Gemini 3.1 Pro lost money running a café while another model succeeded; relying on generic routers or benchmark-based model switching leads to poor outcomes. — via 1 2 3
2. Mollick noted that formalizing agent organization helps understand task delegation between expensive smart agents and cheap weak agents, but there is a severe lack of experience and consensus on optimal workflows for long-running agents. — via 1 2
3. Based on extensive use of Fable, Mollick found that agents may develop internal weird languages and conversation rhythms; the UI is unsuitable for managing autonomous tasks over 5 hours, making real-time observation and intervention difficult. — via 1 2
4. Continuous learning is the biggest barrier to AI adoption: models cannot self-improve and require humans to learn for them, slowing adoption to human processes. — via 1
5. Hugging Face highlighted that agent kernel optimization is the future of on-device inference, demonstrating Fable 5 running Gemma 4 on WebGPU at 255 tok/s, with an open demo. — via 1
6. Hamel Husain shared that DocETL is building an AI-SQL interface suitable for use by Claude Code, Codex, etc., and is starting to work with open-source LLMs. — via 1
7. Hugging Face announced FLARE, a standardized way to report AI flaws, led by multiple universities, emphasizing that open-source AI is more safe. — via 1