The AI Rundown by Lightscape Partners - 1/30/25

Good morning and welcome back to another edition of The AI Rundown by Lightscape Partners.

• DeepSeek is making waves in AI hardware and efficiency, challenging long-held assumptions about GPU optimization. By bypassing Nvidia’s CUDA in favor of low-level PTX programming, DeepSeek achieved a 10x efficiency gain over Meta on similar hardware, raising concerns among investors about future AI chip demand. Meanwhile, their cost-efficient AI model has triggered a broader tech sell-off, with enterprise software firms like SAP eyeing potential benefits from cheaper AI infrastructure.

• ElevenLabs has secured a massive $180 million funding round at a $3.3 billion valuation, signaling continued confidence in AI-driven audio. The company is expanding beyond voice synthesis into conversational AI, audiobook publishing, and multimodal tools. As AI-generated media continues to evolve, ElevenLabs is also ramping up safety measures with digital signatures and stricter content moderation.

• In the public sector, OpenAI has launched ChatGPT Gov, a secure AI platform tailored for U.S. government agencies. With over 90,000 government users already leveraging AI for translation, research, and administrative tasks, OpenAI is positioning itself as a critical player in government tech adoption. The company also faces an escalating dispute with DeepSeek, suspecting that the Chinese AI startup improperly used OpenAI-generated data to train its models—adding another layer to the ongoing battle over AI training data rights.

Stay tuned as we explore these stories and their implications for the future of AI, technology, and innovation.

If you haven’t yet, please support the newsletter by subscribing!

Hardware

DeepSeek optimizes Nvidia GPUs beyond CUDA, disrupting AI hardware assumptions. Link.

• DeepSeek trained a 671B-parameter MoE model using Nvidia H800 GPUs with 10X efficiency gains over Meta by leveraging low-level PTX programming instead of CUDA for key functions.

• The team reconfigured GPU resources, allocating 20 streaming multiprocessors for server communication and implementing fine-grained thread optimizations.

• These modifications pushed Nvidia’s hardware to its limits but required extreme engineering expertise and are difficult to maintain.

• Investors reacted sharply, fearing reduced demand for high-performance AI hardware, though industry leaders argue AI will always consume as much compute as available.

Models

DeepSeek R1 launches on Azure AI Foundry and GitHub. Link.

• DeepSeek R1 is now available on Azure AI Foundry, joining a catalog of over 1,800 models for enterprise AI integration.

• The model enables cost-efficient AI reasoning, allowing businesses to scale applications with minimal infrastructure investment.

• Microsoft has implemented rigorous safety evaluations, including red teaming and content filtering, to ensure responsible AI use.

• Developers can quickly deploy DeepSeek R1 via Azure’s model catalog, with serverless endpoints and GitHub resources for seamless integration.

Enterprise Applications

DeepSeek's AI shakes up enterprise software and hardware costs. Link.

• DeepSeek's AI assistant, built without costly Nvidia chips, triggered a tech sell-off but could benefit enterprise software firms like SAP.

• SAP CEO Christian Klein sees DeepSeek as part of a shift where AI hardware becomes commoditized, reducing costs for software-driven AI applications.

• Analysts debate its impact on enterprise tech, with concerns for Oracle's Project Stargate and Adobe's GenAI monetization.

• Uncertainty remains over DeepSeek’s true capabilities and whether its cost-efficient model can sustain AI infrastructure demands.

Funding & Valuation

ElevenLabs raises $180M at $3.3B valuation to expand AI audio capabilities. Link.

• The AI audio startup secured funding from a16z, ICONIQ, and others to enhance expressive voice models and multimodal AI.

• ElevenLabs plans to expand its conversational AI tools, integrating speech-to-text and telephony support.

• The company is moving into consumer-facing products, testing audiobook publishing and AI-generated podcasts.

• Despite past misuse concerns, ElevenLabs is prioritizing safety with digital signatures and content moderation efforts.

Rad AI raises $60M Series C to expand generative AI in healthcare. Link.

• The $60M round, led by Transformation Capital, values Rad AI at $525M, just seven months after a $50M Series B.

• Rad AI’s generative AI solutions, used by nearly 50% of US radiology providers, aim to reduce radiologist burnout and improve efficiency.

• Its AI-driven workflow tools cut dictated words by 90% and improve follow-up rates for critical findings like potential cancers.

• With over $140M raised to date, Rad AI plans to accelerate hiring and expand its impact on healthcare AI adoption.

SafelyYou raises $43M to expand AI-driven predictive care in senior living. Link.

• The Series C round, led by Touring Capital, brings SafelyYou’s total funding to over $100M, fueling expansion of its AI-powered senior care solutions.

• SafelyYou Respond™, which reduces falls by 40% and ER visits by 80%, is now trusted by 50% of the largest senior care providers.

• New offerings like SafelyYou Clarity™ and SafelyYou Aware™ enhance care measurement and predictive analytics, improving efficiency and outcomes.

• A three-year study found the AI-driven system extends resident stays by over four months, reducing turnover and boosting occupancy rates.

Safety + Ethics

DeepSeek AI left sensitive database exposed, raising security concerns. Link.

• Chinese AI startup DeepSeek exposed over 1 million log lines, including secret keys and chat history, due to an unprotected ClickHouse database.

• The vulnerability allowed full database control without authentication, raising concerns over AI security practices.

• DeepSeek has since closed the exposure but faces scrutiny over its data handling and ties to China, with regulatory inquiries in Italy and Ireland.

• OpenAI and Microsoft are investigating whether DeepSeek illegally used OpenAI’s API to train its models via distillation.

Data Centers + Energy

DeepSeek's AI efficiency shakes up data center power forecasts. Link.

• Chinese AI startup DeepSeek claims its models use far less electricity, challenging projections of surging U.S. data center power demand.

• Analysts warn this could disrupt energy sector investments, with AI-exposed energy stocks taking a hit.

• While DeepSeek lowers training costs, increased AI use could still drive overall electricity consumption higher.

• The uncertainty highlights the need to reassess long-term energy infrastructure planning before committing to major new power projects.

OpenAI

OpenAI launches ChatGPT Gov for U.S. government agencies. Link.

• ChatGPT Gov offers federal, state, and local agencies a secure, self-hosted AI solution through Microsoft Azure’s commercial and government clouds.

• The platform includes GPT-4o, administrative controls, and custom AI tools while meeting strict compliance requirements like FedRAMP High.

• Over 90,000 government users have already leveraged ChatGPT for translation, research, and administrative efficiency.

• OpenAI aims to expand ChatGPT Gov to classified environments, reinforcing AI’s role in national security and public sector innovation.

OpenAI claims DeepSeek may have improperly used its model data. Link.

• OpenAI suspects Chinese rival DeepSeek used a tactic called distillation to train its AI model on OpenAI-generated data.

• DeepSeek’s unexpectedly strong and low-cost model disrupted U.S. markets, contributing to a drop in Nvidia’s stock.

• OpenAI is working with the U.S. government to protect its technology, with White House officials monitoring the issue.

• The dispute highlights the AI industry’s ongoing battles over data usage, as OpenAI itself faces lawsuits over its own training practices.

Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to [email protected].