NVIDIA Vera Rubin Delivers 10x Performance Leap; Elon Musk Unveils Grok 4.5 Dominance and Robotaxi Expansion; Major Financial Risks Exposed

Key Takeaways

1. NVIDIA Vera Rubin NVL72 achieves 10x Tokens per MW over Blackwell, with deployments by CoreWeave, Google Cloud, Microsoft, and Oracle; Spectrum-6 102.4Tbps Ethernet switch also ships.
2. Elon Musk reports Grok 4.5 leads benchmarks over GPT 5.5 and Claude Opus 4.8, launches Grok for Excel, and expands Tesla Robotaxi to 7 regions.
3. AI models achieve perfect scores on IMO 2026 (Claude Fable 5, GPT 5.6 Sol, Kimi K3, Axiom Math), signaling frontier models surpass human-level math.
4. Yann LeCun reveals $1.65 trillion in off-balance-sheet debts (GPU contracts, data centers) across Big Tech, with Meta hiding $420B—investors see less than half.
5. Hugging Face releases open-source Cosmos 3 Edge world model, NVIDIA Nemotron audio models, and a Korean MoE model competitive with Minimax M3 and DeepSeek v4 Pro.
6. Mistral AI expands global partnership with Microsoft, enabling controlled AI deployment for regulated industries.

1. Infrastructure and Platform Leaps

1. NVIDIA Vera Rubin NVL72 achieved a 10x Tokens per Megawatt improvement over Blackwell, with real-world performance data released. CoreWeave, Google Cloud, Microsoft, and Oracle Cloud have already deployed the platform. — via 1 2
2. NVIDIA Spectrum-6 102.4Tbps Ethernet switch is now shipping to AI factories, with CoreWeave, Microsoft, Nebius, SpaceX AI, and Tesla as early adopters. — via 1
3. NVIDIA Blackwell Ultra set a world record in pre-training DeepSeek-V3 671B, reaching 1,648 TFLOPs per GPU, approximately 3x performance over the previous generation. — via 1
4. Mistral AI expanded its global strategic partnership with Microsoft, offering controlled frontier AI for enterprises and regulated industries, extending EU AI compute capacity, and providing deployment from cloud to fully disconnected environments. — via 1

2. Model Benchmark Breakthroughs and Open-Source Releases

1. Elon Musk stated Grok 4.5 leads GPT 5.5 and Claude Opus 4.8 on Snorkel AI's professional work benchmarks, excelling in education, law, and healthcare with the lowest error rate, and ranks #1 on Long-Horizon Terminal-Bench for complex agentic tasks. SpaceX engineering data will further train Grok. — via 1 2 3
2. Four AI models (Claude Fable 5, GPT 5.6 Sol, Kimi K3, Axiom Math) scored a perfect 42/42 on the 2026 IMO exam. Fable was fastest with one pass, Sol cheapest, K3 used more attempts, and Axiom proved all solutions in Lean. The result suggests AI far exceeds IMO-level mathematics. — via 1
3. Hugging Face released Cosmos 3 Edge, an open frontier world model (4B parameters with 2B Nemotron reasoner) for on-device robotics, autonomous driving, and vision AI. NVIDIA also launched it. — via 1 2
4. NVIDIA Nemotron open-weights audio-native models (2B and 30B) handle transcription, translation, sound recognition, audio Q&A, TTS, and full speech-to-speech. — via 1
5. Korean company motif released a 13B active / 314B total MoE model (open weights) that performs on par with Minimax M3 and DeepSeek v4 Pro, using per-expert polynomial activation and GDLA differential attention. — via 1
6. Xiaomi-Robotics-1, a robot foundation model trained on 100,000 hours of real-world manipulation, autonomously performs household tasks like folding laundry and washing dishes. — via 1

3. Applications and Real-World Deployments

1. Tesla's Robotaxi service expanded to 7 regions, including newly opened Orlando, Miami, and Tampa, with various driving modes (unsupervised or safety-supervised). The Cybercab will integrate Starlink V5, no steering wheel, large screen for work or 4K streaming. — via 1 2
2. Grok for Excel has launched, supporting financial modeling, market analysis, and chart generation. Users report 99% reliance on Grok 4.5 due to speed and cost efficiency. — via 1 2
3. Sam Altman noted ChatGPT memory after a major June update now works well, with many users feeling the improvement. — via 1
4. Perplexity Computer improved quality: answers citing financial data now show company logos with full traceability to SEC filings and earnings calls; derived data displays complete formulas and calculation chains. — via 1

4. Financial Risks and Policy Insights

1. Yann LeCun highlighted a Nikkei survey finding $1.65 trillion in off-balance-sheet debts (GPU contracts, data center leases, JVs) among Alphabet, Microsoft, Amazon, Meta, and Oracle. Meta's hidden debt is $420B, three times its reported debt; Oracle's grew 30x in four years. Investors see less than half the truth, and if AI demand disappoints, losses would fall on private credit and insurance. — via 1
2. Ethan Mollick argued that no frontier open-weight model exists outside China; the US has no incentives and the EU no willingness, leaving only mid-tier models. He proposes US-China collaboration on common testing standards for security certification. — via 1 2
3. Hugging Face asserted banning open-source AI would hurt defenders 10x more than attackers, making the world more dangerous. Open weights and open models are inherently more secure. — via 1 2
4. Anthropic is offering up to $50,000 in Claude usage credits to accelerate rare disease treatment research, the first targeted call under its AI for Science program. — via 1