#AI Startup

Anthropic is set to continue supplying its AI model Claude to the NSA, despite being flagged as a supply chain risk by the Pentagon. This decision is influenced by the NSA's lack of access to Nvidia's latest Grace Blackwell chips, while Anthropic's 'Mythos' model operates on older hardware. Notably, the contentious 'any lawful use' clause has been excluded from the current agreement.

Anthropic's decision to continue supplying Claude to the NSA, despite being flagged as a supply chain risk, highlights the ongoing demand for AI solutions in national security. This could signal opportunities for builders and PMs to create AI models that can operate on older hardware, while investors should note the implications of government contracts on AI companies' growth trajectories.

Deepseek has made its 75% discount on the V4-Pro model permanent, pricing input tokens at $0.435 per million, making it 11.5 times cheaper than GPT-5.5 and over 34 times cheaper for output tokens. This aggressive pricing strategy could significantly impact Western AI providers reliant on token consumption.

Deepseek's decision to make its 75% discount on the V4-Pro model permanent, pricing output tokens over 34 times cheaper than GPT-5.5, signals a potential shift in the competitive landscape. Builders and PMs should consider how this cost advantage could enable new applications and business models, while investors may need to reassess the valuation of traditional AI providers facing increased pricing pressure.

Alibaba's Qwen team has launched Qwen3.7-Max, an AI model that autonomously optimized code for its custom chip over 35 hours, outperforming competitors like Claude Opus 4.6 and DeepSeek V4 Pro. The model also demonstrated its capabilities by controlling a four-legged robot.

Alibaba's Qwen3.7-Max autonomously optimized code for its custom chip over 35 hours, outperforming competitors. This development signals a significant advancement in AI's capability to enhance hardware efficiency, which could lead to more powerful and cost-effective solutions for builders and investors in tech infrastructure.

OpenAI's shift towards agent-based products marks a significant change in AI development, with DeepSeek's pricing strategy making their V4 Pro model 75% cheaper, outperforming competitors like GPT-5.5 by 19x. Meanwhile, Gemini 3.5 Flash shows mixed feedback despite improvements, and new protocols like MCP simplify operational complexities for AI infrastructure.

OpenAI's transition from Model Labs to Agent Labs signifies a strategic pivot towards more interactive AI systems, which could reshape product development priorities for builders and PMs. Additionally, DeepSeek's V4 Pro model being 75% cheaper while outperforming competitors highlights a growing price-performance advantage in the AI market, attracting investor interest in cost-effective AI solutions.

NVIDIA's MAISI framework synthesizes high-resolution 3D medical images, enabling scalable data generation for AI models. The NV-Generate-MR-Brain model utilizes the MR-RATE dataset, the largest open-source multimodal MRI dataset, to enhance training and generalization in medical imaging AI.

NVIDIA's MAISI framework for synthesizing high-resolution 3D medical images allows builders and PMs to generate vast datasets efficiently, improving AI model training and generalization in medical imaging. For investors, this development signals a significant advancement in healthcare AI, potentially leading to more robust applications and faster market adoption.

A specialized 3-billion-parameter model outperformed all tested commercial APIs, including Claude Opus 4.6, by scoring 0.911 on a Brazilian Portuguese OCR benchmark while costing approximately fifty-two times less per million pages. This challenges the prevailing assumption that larger models are always superior, emphasizing the importance of task-specific training.

The development of a specialized 3-billion-parameter model that outperformed larger commercial APIs on a Brazilian Portuguese OCR benchmark highlights the potential cost-effectiveness and efficiency of task-specific training. Builders and PMs should consider optimizing models for specific tasks rather than defaulting to larger models, while investors may find opportunities in specialized AI solutions that deliver better performance at lower costs.

Huxe, an audio generation app by ex-NotebookLM developers, is shutting down just after Spotify launched a competing feature. The app will be removed from stores, and user data will be deleted after seven days, highlighting the competitive landscape in consumer AI.

The shutdown of Huxe, an audio generation app, following Spotify's launch of a competing feature signals the intense competition in the consumer AI space. Builders and PMs should note the rapid market shifts and the importance of differentiation, while investors may need to reassess the viability of startups in crowded niches.

At Google I/O, CEO Demis Hassabis emphasized AI's potential in science, showcasing WeatherNext's life-saving capabilities while hinting at a shift towards agentic AI systems like Gemini for Science, which may redefine scientific research. Despite ongoing development of specialized tools like AlphaFold, Google is increasingly prioritizing general-purpose AI that can autonomously contribute to scientific advancements.

Google's shift towards general-purpose AI systems like Gemini for scientific research signifies a potential transformation in how scientific advancements are made, moving from specialized tools to more autonomous solutions. Builders and PMs should consider the implications for product development and integration, while investors may want to explore opportunities in AI platforms that facilitate this new approach to scientific inquiry.

HMD launched the Vibe 2 5G smartphone in India, preloaded with Sarvam's Indus AI chatbot, which supports 22 Indic languages. The device, priced at ₹10,999 ($114), aims to tap into the Indian market where HMD's smartphone share is currently negligible, while the Indus app has seen only 293,000 downloads since its launch.

HMD's launch of the Vibe 2 5G smartphone in India, preloaded with the Indus AI chatbot, signals a strategic move to penetrate a competitive market by addressing local language needs. For builders and PMs, this highlights the importance of integrating AI solutions that cater to diverse user demographics, while investors should note the potential for growth in underserved markets.

New AI unicorns Exa, Modal, and TurboPuffer have achieved significant funding milestones, with Exa raising $250M at a $2.2B valuation and TurboPuffer reaching $100M ARR. Additionally, advancements in AI models like RAEv2 and Gated DeltaNet-2 are pushing the boundaries of performance in language modeling and vision tasks.

The emergence of new AI unicorns like Exa and TurboPuffer, alongside advancements in models such as RAEv2, signals a growing investment landscape and innovation in AI infrastructure. Builders and PMs should consider leveraging these cutting-edge technologies to enhance their products, while investors may find new opportunities in these rapidly scaling companies.

Daytona's CEO Ivan Burazin discusses the company's remarkable 74% month-over-month growth, achieving 850,000 daily runs through innovative Bare Metal Sandboxes and Reinforcement Learning evaluations. The introduction of their new Agent Cloud is set to further enhance performance and scalability for AI agents.

Daytona's introduction of the Agent Cloud and its 74% month-over-month growth signal a significant advancement in AI scalability and performance. Builders and PMs should consider how these innovations can enhance their own AI projects, while investors may see this as a strong indicator of market demand and potential returns in AI infrastructure.

Google DeepMind is launching the Accelerator program in the Asia Pacific to address environmental risks, leveraging AI technologies to enhance sustainability efforts. This initiative aims to support startups and organizations focused on innovative solutions for climate challenges, fostering collaboration and knowledge sharing in the region.

The launch of the Google DeepMind Accelerator program in Asia Pacific specifically targets environmental risks, providing a platform for startups to innovate in sustainability. Builders and PMs can leverage this initiative to access resources and expertise, while investors may find new opportunities in climate-focused technologies that align with global sustainability goals.