OpenAI’s $500B Ohio Bet, LeCun’s JEPA, Nemotron 3, Murati’s Startup, Claude Fable 5

OpenAI Negotiates 10-Gigawatt Ohio Data Center with Nvidia as Guarantor

OpenAI is in talks to lease a planned 10-gigawatt data center in Ohio, a project developed by SB Energy on federal land in Pike County. The site, which previously housed a uranium enrichment facility, could cost at least $500 billion at full buildout. OpenAI would sign a 20-year lease, its largest infrastructure commitment to date, and Nvidia would serve as financial guarantor for both the lease and project financing, a novel role for the chipmaker at this scale. The first phase, expected by 2028, will deliver 800 megawatts. The deal echoes the earlier Stargate initiative with Oracle and SoftBank, which made little progress. Negotiations remain ongoing, and plans could still change. OpenAI also confidentially filed for an IPO this week.

OpenAI wants its biggest data center yet, and Nvidia would back the bill →

Yann LeCun Champions $1 Billion JEPA Initiative to Replace LLMs

Turing Award winner Yann LeCun is leading a $1 billion effort to develop Joint Embedding Predictive Architecture (JEPA), a paradigm meant to overcome the fundamental limitations of large language models. LeCun argues that LLMs rely on statistical language patterns and fail to capture real-world understanding, whereas JEPA learns by predicting abstract representations from raw video data, inspired by how infants interact with their environment. The architecture consists of six modular components, including a Perception Module and World Model, designed to simulate, predict, and act within dynamic environments. LeCun believes that grounded representations are essential for true intelligence, and JEPA’s approach could reshape AI research priorities by prioritizing causal understanding and generalization over scaling language-based systems.

Why Yann LeCun is Spending $1 Billion to Replace LLMs with JEPA →

NVIDIA Nemotron 3 Ultra: 550B Parameters, Million-Token Context, Mixture-of-Experts

NVIDIA has released Nemotron 3 Ultra, a 550-billion-parameter language model with a mixture-of-experts architecture that activates only 55 billion parameters per task, dramatically reducing computational demands. The model features a million-token context window, enabling it to handle complex multi-step workflows in reasoning, coding, and long-term decision-making. It outperforms larger models like GPT-4 and Anthropic Opus on agent-specific benchmarks, including faster token generation and superior results on Pinchbench. Training strategies such as multi-tier policy distillation and fine-tuning with agent-specific datasets enhance adaptability. The model is open-weight, allowing organizations to customize it for automation, research, and customer service applications.

Why NVIDIA’s Nemotron 3 Ultra Outperforms Trillion-Parameter AI Models →

Mira Murati Unveils Thinking Machines Lab’s Real-Time Multimodal Interaction Model

Mira Murati, former OpenAI CTO who left in late 2024, revealed her startup Thinking Machines Lab in her first media interview since founding it. She raised $2 billion in under a year and is building what she calls interaction models: multimodal AI systems that process audio, text, and video simultaneously and collaborate with humans in near real time, without requiring prompts. The first model, named TML-Interaction-Small, will be released publicly later this year. Murati described the approach as a tandem bike rather than an autonomous system advancing on its own. Her background includes mechanical engineering at Dartmouth, product management at Tesla and Leap Motion, and six years at OpenAI where she oversaw ChatGPT, DALL-E, and GPT-4. She also served as interim CEO during Sam Altman’s temporary ousting and later testified in Elon Musk’s lawsuit against OpenAI, criticizing the lack of checks and balances in the company’s governance.

Mira Murati Unveils Her Startup’s A.I. Model in First Interview Since OpenAI →

Anthropic’s Claude Fable 5 Sets New Benchmarks in Coding and Science

Anthropic has released two new models in the fifth Claude generation: Claude Fable 5, a general-purpose model with conservative safety guardrails, and Claude Mythos 5, which drops those restrictions and is available only to select partners. Both share the same base model. On SWE-Bench Pro, which tests real software engineering tasks from public GitHub repos, Fable 5 scored 80.3 percent, far ahead of Claude Opus 4.8 at 69.2 percent and GPT 5.5 at 58.6 percent. On Cognition’s FrontierCode benchmark for demanding coding tasks, Fable 5 achieved 29.3 percent versus 13.4 percent for Opus 4.8. Payment processor Stripe reported that Fable 5 compressed five months of engineering work into days, migrating a 50-million-line Ruby codebase in one day. The model also topped Hebbia’s Finance Benchmark for analytical tasks and set new state-of-the-art results on vision and long-term memory tasks.

Anthropic releases Claude Fable 5 and Mythos 5 with major gains in coding and science →