Artificial Intelligence Insights

Research — April 17, 2025 Generative AI Digest: NVIDIA GTC 2025, IBM acquisitions 451 Research Recent weeks have been busy for GenAI, with an array of announcements, including a plethora surrounding NVIDIA Corp. GTC 2025. Developments in the space include multiple acquisitions, new model launches and a range of global regulatory updates. This report covers significant new product releases, funding and M&A activities, and regulatory changes from the past few weeks. Although Anthropic PBC's Claude 3.7 and OpenAI's GPT 4.5 exhibit advancements over previous versions, the naming conventions imply that, despite significant investment in model training, these efforts have not yet produced models that signify a new era of capabilities — at least not any ready for public release. The increasing resources these foundation model providers have reportedly invested in training, coupled with the relatively modest gains against benchmarks, may raise concerns about the return on AI infrastructure investment. Additionally, GPT-4.5 is significantly more expensive to operate than its 4o predecessor. This concern should be contextualized with two important points. First, the scaling paradigm extends beyond training. These newly released models might offer a stronger foundation for building reasoning models, which use more resources during inference, but will potentially enhance performance against the most challenging tasks. Second, both OpenAI LLC and Anthropic emphasize a greater focus on real-world applications rather than benchmark performance. While it might be tempting to view the de-emphasis on benchmarks as an excuse, there is concern that chasing benchmark scores can lead to poorly balanced models, and that many of these benchmarks may not accurately reflect performance for most enterprise tasks. Product releases and updates OpenAI announced GPT 4.5, its latest frontier model. Rather than emphasizing benchmarks, the release focused on the model's personality, better writing style and addressing practical challenges. While not a "reasoning" model (in the manner of the o1 or o3 model families), which means it does not perform as well for some science, coding and mathematics tasks, GPT 4.5 may form the foundation for an enhanced set of reasoning models. The model was presented as "hallucinating less" and having deeper world knowledge. Amazon.com Inc. released Alexa+ in February, an upgrade to the company's assistant that introduces agentic capabilities, including the ability to "memorize" information and take advantage of connected devices. All Amazon Prime members will get Alexa+ for free, but it can also be purchased with a separate monthly subscription. Alexa+ is described as "model agnostic" — leveraging Amazon's own models as well as Anthropic and other third-party models. International Business Machines Corp. announced expansions to its Granite AI model family, including a new vision language model, a more efficient Guardian safety model and chain-of-thought capabilities for its updated 2B and 8B language models. Tencent Holdings Ltd. released a new AI model, Hunyuan Turbo S. The Chinese technology giant suggests its model offers faster query responses than competing models. The announcement explicitly drew comparisons to Hangzhou DeepSeek Artificial Intelligence Co. Ltd.'s V3 model — claiming comparable performance for knowledge, math and reasoning tasks. This was one of many AI releases from Chinese technology companies this month, including Alibaba Group Holding Ltd. releasing R1-Omni, which is reportedly able to identify emotions from video footage, among other visual analysis tasks. Cohere Inc. announced Command A, a foundation model designed for low latency that can be deployed on just two GPUs. The 111 billion-parameter model was presented as being effective for AI agents and calling external tools. Google LLC released Gemma 3 (the latest version of its open-source model family), Gemini Robotics and Chirp 3. Gemma 3 performs well in user preference rankings, and even its largest variant (27 billion parameters) reportedly requires only a single GPU to run. Gemini Robotics is the company's new vision-language-action model, with Google DeepMind showcasing robotic arms moving fruit into containers and playing tic-tac-toe. The company also announced that synthetic voice generation product Chirp 3 would be added to Vertex AI platform. Lenovo Group Ltd. announced the ThinkEdge SE100 at MWC25, an entry-level AI inferencing server. It is intended for SMBs needing more accessible and affordable AI. It is an edge AI offering, with the server 85% smaller than a standard 1U server. Industries Lenovo cited case studies for include manufacturing, healthcare and retailers. Salesforce Inc. announced an update with Agentforce 2dx, adding an array of new orchestration capabilities. A major theme with the release was connectivity, including the ability to integrate agents into processes — exposing the Agentforce API and introducing MuleSoft Agentforce connector. At NVIDIA GTC 2025, significant attention was paid to new GPU Blackwell Ultra, which NVIDIA positioned as speeding up inference and reasoning tasks, and the design appears to have significantly more HBM3e memory than its Blackwell predecessor. The road map also included Vera Rubin chips, which will be released in the second half of 2026, alongside a claim that this will be followed by the launch of Vera Rubin Ultra in 2027. DGX AI computers, robotics foundation model GR00T N1 and some quantum announcements at the inaugural Quantum Day for the conference, also generated headlines. Further announcements around GTC included Equinix Inc. announcing that it will be the first to offer NVIDIA Instant AI Factory, a managed service featuring the new NVIDIA DGX GB300 and DGX B300 systems. There were also announcements from storage companies integrating NVIDIA AI Data Platform into their data storage technologies, including IBM, Pure Storage Inc. and Hitachi Vantara Corp., among others. In addition, DataStax Inc. introduced Astra DB Hybrid Search, leveraging NVIDIA NeMo Retriever Text Reranking to improve search relevance. Hewlett Packard Enterprise Co. Private Cloud AI saw the addition of new blueprints to speed up AI project development, and Dell Technologies Inc. made a spread of announcements around Dell AI Factory with NVIDIA. Verizon Business announced a generative AI assistant offering to help small businesses manage customer interactions. The assistant is a text-based capability that answers customer questions from a knowledge base and moves requests it can't currently answer to an employee. This month also featured several 3D generation announcements. Stability AI Ltd. introduced the Stable Virtual Camera, a new model that enables movement through a 3D scene using just an image prompt. Additionally, Roblox Corp. launched its Cube model, which allows text prompts to generate 3D meshes. By making this tool open source, Roblox has made it accessible for creators to use outside of its ecosystem. Funding and M&A IBM announced multiple acquisitions over the past few weeks. One was the pickup of DataStax, a NoSQL database vendor known for providing commercial support atop Apache Cassandra. DataStax had expanded into generative AI with a separate vector-supported database (AstraDB), and had acquired Langflow, which had been a key enabler for the company's GenAI stack strategy called RAGStack. IBM also completed its acquisition of HashiCorp Inc., an infrastructure management and security vendor. It explicitly called out the ability for HashiCorp's products to "automate and secure the infrastructure that underpins hybrid cloud applications and generative AI" in its announcement. Synthetic data generator Gretel Labs Inc. has reportedly been acquired by NVIDIA, at a nine-figure sum that exceeds Gretel's most recent valuation of $320 million. Video generator Natural Synthetics Inc., which did business as HotShot, has been acquired by X.AI Corp. The company provided foundation models, supporting text-to-video and text-to-GIF models. HotShot was a very small team, numbering just four developers when its flagship HotShot model entered preview in 2024. Process automation vendor UiPath Inc. acquired British startup Peak AI Ltd. on March 12. The stated objective of UiPath was to enhance its "vertical AI solutions strategy." Peak AI focuses on AI for product inventory and pricing decisions. Peak AI had raised $120 million since its 2014 incorporation. Anthropic has secured $3.5 billion in funding, led by Lightspeed Ventures with participation from several new and returning investors, at a post-money valuation of $61.5 billion. The funds will be used to advance the development of next-generation AI systems and expand compute capacity. Additionally, the foundation model provider plans to accelerate its international expansion, particularly into Asia and Europe. Image generator Bria Artificial Intelligence Ltd. declared a $40 million series B round. The startup, which focuses on commercial safety with a proprietary attribution engine, saw investment from Red Dot Capital, Maor Investment, Entrée Capital, GFT Ventures, Intel Capital and In-Venture. Reflection AI Inc., which suggests it is building a "superintelligent autonomous coding system," announced a $105 million series A. The company, which appears to have not yet released a product, says the round was co-led by returning investor Charles River Ventures and new investor Lightspeed Ventures. BlackRock announced that x.AI and NVIDIA are joining the AI Infrastructure Partnership fund, and that GE Vernova Inc. and NextEra Energy Inc. will be collaborating with the partnership to provide energy for the projects. Ilya Sutskever's startup Safe Superintelligence Inc. has reportedly raised a further $2 billion, led by Greenoaks Capital Partners. The startup does not yet have a product and is focused on the long-term goal of "superintelligence." Politics and regulations US President Donald Trump took aim at the $52.7 billion CHIPS Act, a law designed to build the domestic semiconductor industry. In a speech to the US Congress, he referred to the act as "horrible, horrible" as part of a broadside on the bipartisan bill, with some semiconductor companies likely concerned that it may lead to walking back on disbursements, and a clear transition from carrot to stick. Reports indicate that the UK government is delaying its AI regulation plans. This delay is seen by many as an effort to better align with the US administration's stance on the issue. However, it also likely reflects the intention of UK Prime Minister Keir Starmer's administration to strengthen the domestic AI industry. One of the key mandates in the now-delayed bill was the requirement for large AI models to be assessed by the UK's AI Security Institute. The Japanese government approved an AI misuse bill, requiring companies to cooperate with a range of AI measures. A noteworthy component of the draft law is that if individual rights are violated, the government will launch an investigation and instruct the firm to take corrective measures; if the act is seen as malicious, the government will disclose the name of the business to the public. The draft law did not explicitly cite penalties for bad practice. A task force led by Prime Minister Shigeru Ishiba will oversee AI policy and investment plans. The EU AI Act's implementation is ongoing, with Article 5, which addresses prohibited AI practices, now in effect. This article bans AI applications designed to significantly manipulate behavior to cause harm and prohibits the expansion of facial recognition databases through indiscriminate scraping of facial images from the internet or CCTV footage. Meta Platforms Inc. faces ongoing copyright disputes, including new legal action from three publishing trade groups in Paris for economic "parasitism." The company is also embroiled in several AI copyright lawsuits in the US. Notably, a federal judge recently allowed the prominent Kadrey v. Meta Platforms Inc. case to proceed — a case brought by prominent US authors — although some claims were dismissed. The Cyberspace Administration of China has announced new labeling requirements. Starting in September 2024, all content generated by AI will have to be explicitly labeled as such in China. Gain access to our full news & research coverage and the industry-specific data that informs our insights This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global. 451 Research is a technology research group within S&P Global Market Intelligence. For more about the group, please refer to the 451 Research overview and contact page.