Key takeaways

• California Governor Newsom approved several AI bills, including requirements for Foundation Model (FM) providers to disclose training data details by 2026, extending privacy laws to GenAI systems, and mandating AI-generated content disclosures. In addition, the US Department of Commerce published guidance on measuring and managing GenAI risks and collaborating globally on AI standards. This sentiment is echoed in a recent UN report proposing a global AI governance framework.

• The European AI Act became enforceable from August onwards, aiming to govern AI development. Furthermore, the Commission launched a consultation on a Code of Practice for providers of general-purpose AI (GPAI) models to address transparency, copyright-related rules, and risk management. Other AI governance frameworks currently being formed (such as in the UK) will likely be based on the EU AI Act.

• Sixty GenAI ecosystem startups raised USD 5.4 billion, broadly in line with the average quarterly run rate from Q3 2023 to Q1 2024. This was still a drop from the record-breaking Q2 2024 (USD 19.6 billion), inflated by multi-billion-dollar rounds from CoreWeave and xAI. The highlight in Q3 was Illya Sutskever’s startup, Safe Superintelligence, raising USD 1 billion—a rare feat for a company formed four months ago that is yet in the pre-seed stage. 

• Other notable large funding rounds for the quarter included China-based LLM developer Baichuan Intelligence (USD 691 million), chip developer Groq (USD 640 million), and Canadian LLM developer Cohere (USD 500 million).

• This quarter witnessed significant open-source model launches from companies like Mistral AI, Alibaba, and Meta. While their previous focus was on LLMs, several multimodel (Pixtral 12 billion) and fine-tuned models (Codestral Mamba and Mathstral) were introduced by Mistral AI. It also launched an LLM (Mistral Large 2) with 84% pre-trained accuracy. Similarly, Meta unveiled its largest open-source model (Llama 3.1 with 405 billion parameters) to compete with closed models like GPT-4o. This represents a positive shift toward democratizing AI technology, facilitating greater accessibility for developers and businesses. 
• In addition, Q3 also saw several new model launches, including LLM launches such as OpenAI’s o1, Zhipu AI's GLM-4-Flash, AI21's Jamba 1.5, and Fujitsu’s Japanese model, Takane. Fine-tuned model launches were tailored for specific domains, such as NVIDIA’s StormCast and IBM’s Prithvi WxC for weather prediction, Meta’s SAM 2 for object segmentation, and Microsoft’s SpreadsheetLLM for spreadsheet analysis. 

• Meta (VFusion3D) and Stability AI (Stable Fast 3D) released 3D models, while image and 3D tools were prominent within applications. Additionally, Adobe launched GenAI features for content personalization, Midjourney introduced a unified AI image editor, Nextech3D enhanced 3D model visualization capabilities, and Meta launched its text-to-3D generator. This will likely address content creators’ growing interest in experimenting with 3D assets for diverse applications, from video games to product design and visualization.
• GenAI application developers enhanced their products with new underlying models. Video creation tool providers Runway, Luma AI, Hotshot, and Haiper introduced upgraded video capabilities through new models, while Google and xAI enhanced the image-generation abilities of their AI assistants with the launch of Imagen 3 and Grok-2, respectively. Additionally, OpenAI improved its AI assistant's complex problem-solving capabilities with the “o1” series models. These developments underscore the focus of application developers to consistently improve their offerings. 
• During the quarter, cloud solutions and hardware development dominated in infrastructure, with Together AI and introducing private cloud offerings to enhance AI deployment. Zilliz Cloud and Pinecone introduced features for scaling GenAI workloads, while Lambda and focused on model training platforms. Regarding hardware development, IBM’s Spyre AI accelerator chip and ’s SuperCluster for NVIDIA Omniverse could scale AI workloads. Intel’s Xeon 6 CPU with 2x the performance of its predecessor, Gaudi 3 AI GPU designed for large-scale GenAI, and FuriosaAI’s RNGD AI inference chip would improve processing capabilities and deploy complex AI models, leading to faster innovation cycles.

• Other notable updates included Lenovo launching GPU-as-a-service (GPUaaS) to improve AI accessibility, Hugging Face launching inference-as-a-service for AI deployment and expanding its capabilities into LLMOps solutions, Microsoft launching MInference for faster LLM processing, and Amazon SageMaker also adding new capabilities for faster auto-scaling of AI models.

• Notable trends in Q3 included 1) jointly developing new models (Mistral AI and NVIDIA; Cohere and Fujitsu); 2) offering integrated hardware solutions (Oracle and NVIDIA); 3) accessing hardware for model optimization (Sakana AI, Meta, and Google Cloud partnered with NVIDIA); and 4) offering integrated processing and cloud solutions (Oracle and Inspur partnered with NVIDIA; Microsoft and Intel partnered with IBM).
• The trend of integrating GenAI tools into devices continued this quarter, too, as  OpenAI partnered with Wearable Devices to integrate ChatGPT into Apple Watch through the Mudra band and Synchron integrated ChatGPT into its brain-computer interface. 

• GenAI players also reported partnerships with consultancy firms to accelerate adoption. Notable partnerships included Anthropic’s partnership with Deloitte and Accenture’s partnerships with Meta, Amazon, and NVIDIA, which suggest a strategic shift toward developing specialized solutions that address the unique needs of various industries. 

• This quarter saw four M&A deals in the GenAI hardware development space, with AMD reporting two acquisitions: 1) ZT Systems and 2) Silo.AI to improve its infrastructure solutions. NVIDIA also reported two acquisitions: 1) Brev.dev to expand the capabilities of its DGX Cloud service and 2) Octo AI to expand its efforts in machine-learning compilers and cloud infrastructure for AI.

• Emergence of “as-a-service” infrastructure: Developments by Lenovo and Hugging Face and partnerships between major players like Google and NVIDIA indicate a shift toward “as-a-service” infrastructure. Offering scalable, on-demand access to GPU resources and AI capabilities can lower the barriers to entry for businesses and likely enable companies to integrate advanced AI solutions without substantial upfront investments. As a result, we can expect a surge in AI-driven applications across various sectors, from creative industries to enterprise solutions, enhancing productivity and operational efficiency. Moreover, the flexibility of as-a-service offerings allows organizations to adapt quickly to changing market demands and technological advancements. 
• GenAI for web search: The recent integration of GenAI into search engines signals a transformative shift in how users access information online, as seen with initiatives like Microsoft launching AI-powered generative search in Bing and OpenAI’s temporary prototype of AI-powered search engine “SearchGPT” this quarter, following Google’s launch of the "Search Generative Experience" feature last year. By leveraging advanced algorithms, search engines better understand user intent, predict queries, and deliver tailored results that align closely with individual preferences. While the potential for misinformation and hallucinations in the current GenAI systems may pose challenges, fully realized solutions could improve user experiences as the search process becomes less about sifting through links and more about receiving direct, relevant answers. 

• Integrating hardware with cloud solutions: As GenAI models become even larger and more complex, cloud infrastructure plays an increasingly crucial role. In previous quarters, there was a strong focus on developing efficient, low-cost AI hardware solutions. Now, companies are moving toward integrating hardware with cloud solutions, as demonstrated by Google Cloud Run adding NVIDIA GPU support for serverless AI inference and Oracle Cloud Infrastructure integrating AMD Instinct MI300X accelerators with ROCm open software for advanced AI workloads. With this, organizations can expect improved efficiency in processing large datasets, which is crucial for training complex AI models. Innovations in AI-specific cloud services, such as managed AI platforms and AI accelerators, are expected to further streamline AI development, reducing barriers to entry and enabling even broader applications across industries.