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

• Nvidia is making AI development more accessible with RTX AI PCs taking center stage at GTC 2025. The company is introducing on-device model training, reducing reliance on the cloud, while launching prepackaged AI models like Llama 3.1 and YOLOX for seamless PC-based AI integration.

• Anthropic is closing in on a massive $3.5 billion funding round at a $61.5 billion valuation. This latest raise nearly doubles the company’s worth as it unveils Claude 3.7 Sonnet, designed to improve AI reasoning and speed, further intensifying competition in the LLM space.

• Microsoft is warning that new U.S. AI chip export restrictions could backfire, pushing allied nations toward Chinese alternatives. With China actively courting restricted markets, Microsoft is urging policymakers to revise the rules to maintain U.S. leadership in AI.

• OpenAI has launched GPT-4.5, a more efficient and refined version of GPT-4, though it is not classified as a frontier model. It enhances writing, pattern recognition, and emotional nuance while reducing hallucinations, and will roll out to ChatGPT Pro users before expanding to other tiers and Microsoft’s Azure AI Foundry.

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 + Software

RTX AI PCs take center stage at NVIDIA GTC 2025. Link.

• NVIDIA’s GTC 2025 will highlight how RTX AI PCs and workstations optimize AI workloads, from digital humans to small language models.

• Sessions will cover AI development tools, hardware-aware optimizations, and on-device model training for efficiency and control.

• NVIDIA RTX GPUs power local AI development, reducing reliance on cloud services while maintaining performance and security.

• The launch of NVIDIA NIM microservices introduces prepackaged AI models, including Llama 3.1 and YOLOX, for seamless PC-based AI integration.

Tech giants race to develop custom AI chips, but TSMC remains indispensable. Link.

• Companies like Google, Microsoft, and OpenAI are designing custom AI chips to reduce reliance on NVIDIA and cut costs.

• Despite the shift, Taiwan’s TSMC remains the dominant manufacturer, fabricating the majority of AI chips worldwide.

• ASML’s extreme ultraviolet lithography plays a critical role in producing advanced semiconductors, with IBM and Meta exploring AI-driven materials innovation.

• TSMC’s $65 billion investment in U.S. fabs signals an effort to localize production while maintaining its industry leadership.

AI-powered tool automates cell analysis in microscopy. Link.

• Researchers adapted Meta’s Segment Anything model for microscopy, retraining it on 17,000 images with 2 million manually annotated structures.

• The resulting tool, μSAM, can segment tissues, cells, and organelles across diverse imaging conditions without manual annotation.

• Already in use for projects like hearing restoration and cancer research, μSAM dramatically reduces analysis time from weeks to hours.

• This breakthrough enhances both biological research and medical diagnostics by making complex image segmentation widely accessible.

Models

Tencent launches Hunyuan Turbo S to challenge DeepSeek’s AI dominance. Link.

• Tencent's new AI model, Hunyuan Turbo S, claims faster response times than DeepSeek-R1, answering queries in under a second.

• The model’s reasoning and math skills reportedly match DeepSeek-V3, the chatbot that has outpaced ChatGPT in global downloads.

• DeepSeek’s rapid success has intensified competition among Chinese tech giants, prompting Tencent and Alibaba to accelerate AI development.

• Tencent highlights Turbo S’ significantly lower usage costs, reflecting pressure from DeepSeek’s open-source and low-cost strategy.

Consumer AI Applications

Amazon unveils Alexa+, a generative AI-powered assistant free with Prime. Link.

• Alexa+ enhances conversational AI, enabling more natural interactions and deeper personalization based on user preferences and history.

• The assistant introduces 'agentic capabilities,' allowing it to autonomously navigate the web and complete complex tasks like booking repairs.

• Alexa+ integrates with thousands of services, from smart home devices to shopping and entertainment platforms.

• Priced at $19.99/month, it is free for Prime members and will roll out in waves starting with select Echo devices.

Startup Funding & Valuations.

AI startups outpace SaaS growth, Stripe argues ‘LLM wrapper’ criticism is misguided. Link.

• Stripe reports AI startups are scaling faster than SaaS, with the top 100 AI companies reaching $5M ARR in 24 months, vs. 37 months for SaaS in 2018.

• Examples like Cursor ($100M revenue), Lovable ($17M ARR in 3 months), and Bolt ($20M ARR in 2 months) highlight AI’s rapid monetization.

• Stripe refutes the ‘LLM wrapper’ label, emphasizing industry-specific AI tools' long-term economic impact.

• Stripe’s 2024 payment volume hit $1.4T (+38%), with a $91.5B valuation confirmed via a tender offer.

Anthropic nears $3.5B fundraise at $61.5B valuation. Link.

• AI startup Anthropic is reportedly raising $3.5 billion, nearly doubling its valuation to $61.5 billion, with backing from Lightspeed, General Catalyst, and Bessemer.

• The round surpasses the $2 billion target reported last month, highlighting continued investor appetite for AI startups.

• Anthropic's funding push comes as OpenAI seeks a valuation of up to $300 billion in its own raise.

• The company also unveiled Claude 3.7 Sonnet, an AI model aimed at improving speed and reasoning capabilities.

Perplexity AI reportedly launching $50M fund for early-stage AI startups. Link.

• Perplexity AI is creating a $50 million venture fund targeting U.S.-based pre-seed and seed AI startups, with outside LPs providing most of the capital.

• The company recently closed a $500 million round, tripling its valuation to $9 billion, backed by SoftBank, Nvidia, and Jeff Bezos.

• Perplexity is expanding its product lineup, announcing a new web browser, Comet, alongside its AI-powered shopping assistant.

• AI investments dominated 42% of U.S. venture capital in 2024, with industry leaders calling it the start of the 'Agentic Age.'

Together AI raises $305M to expand Nvidia-powered AI cloud. Link.

• Together AI secured $305M in Series B funding, led by General Catalyst, valuing the company at $3.3B.

• The company is scaling AI infrastructure with 200MW power capacity and deploying Nvidia Blackwell GPUs across North America.

• A partnership with Hypertec includes a 36,000-GPU cluster, with plans to scale beyond 100,000 GPUs.

• Founded in 2022, Together AI has rapidly grown, raising $100M in 2024 before this latest funding round.

Regulation and Legal

Microsoft warns US AI chip export rule could boost China’s AI dominance. Link.

• Microsoft argues Biden’s AI Diffusion rule limits chip access for 150 ‘middle-tier’ nations, pushing allies toward non-US supply chains.

• The Brookings Institute and CSIS warn the policy could backfire, fostering resentment and accelerating open-source AI development favoring China.

• Microsoft urges Trump to simplify restrictions while maintaining security controls to prevent AI chip diversion.

• With China courting restricted nations, Microsoft sees an urgent need to revise the rule to preserve US AI leadership.

Safety and Ethics

Hacker exploits AI software to ruin man's life and leak Disney data. Link.

• Disney employee Matthew Van Andel unknowingly downloaded malware hidden in an AI image generator, leading to a massive data breach.

• A hacker used the malware to access Van Andel’s password manager, leaking Disney's internal data and exposing his personal information.

• Disney fired Van Andel after the breach, citing alleged policy violations he denies, costing him $200,000 in bonuses and healthcare.

• The case highlights the risks of AI-related malware and the dangers of weak cybersecurity practices, including failing to enable two-factor authentication.

OpenAI

OpenAI launches GPT-4.5 but clarifies it’s not a frontier model. Link.

• GPT-4.5, OpenAI’s largest LLM, improves computational efficiency by 10x over GPT-4 but lacks ‘frontier’ advancements seen in o1 and o3-mini.

• The model enhances writing, pattern recognition, and emotional nuance while reducing hallucinations compared to GPT-4o.

• OpenAI trained GPT-4.5 using a mix of traditional and new supervision techniques, including synthetic data from its o1 reasoning model.

• Initially available to ChatGPT Pro users, it will expand to Plus, Team, Enterprise, and Edu users, as well as Microsoft’s Azure AI Foundry.

• OpenAI has launched GPT-4.5, a more efficient and refined version of GPT-4, though it is not classified as a frontier model. It enhances writing, pattern recognition, and emotional nuance while reducing hallucinations, and will roll out to ChatGPT Pro users before expanding to other tiers and Microsoft’s Azure AI Foundry.

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