DeepSeek r2 Leaks!

An internal paper that has been circulating in specialist forums since the weekend suggests that DeepSeek could radically reduce the costs of AI services with its upcoming R2 model.

R2 combines a so-called “mixture-of-experts” architecture with 1.2 trillion parameters, of which only 78 billion are actively calculated. According to the leak, this technology reduces the price per processed token by 97 percent compared to GPT-4 Turbo. The basis is a 5.2-petabyte dataset from legal, financial and patent data; the training runs on Huawei's Ascend-910B chips with 82 percent utilization and achieves around 512 PetaFLOPS - almost half an ExaFLOP.

This opens up two opportunities for the AI community: firstly, the running costs for research, start-ups and open source projects are drastically reduced. Secondly, R2 shows that competitive models are also possible without NVIDIA GPUs - a step towards greater technological independence.

If DeepSeek confirms the figures, developers could soon be using powerful models locally or in inexpensive clouds. Which application would you be the first to try out?

Why it matters: R2 promises affordable high-performance AI and thus paves the way for broader participation in modern AI technology. At the same time, it shifts the balance of power in the global AI hardware race.

Most slide decks look the same—static, boring, and forgettable.

Leave them behind with Prezi AI, the AI presentation software that builds unique, dynamic presentations just from your prompt. It’s fast, intuitive, and designed to fit your ideas perfectly.

Stand out with Prezi’s unique format that uses zooming, cinematic movement to wow your audience.

Try for free today

Morgan Stanley describes concerns about a decline in AI investment as unfounded. Analyst Joseph Moore highlights the continued strong demand for GPUs, particularly for inference tasks, despite recent market weakness in AI stocks. The introduction of more efficient language models such as DeepSeek and new US tariffs have put short-term pressure on companies such as Nvidia. Nevertheless, Moore predicts strong revenue growth for Nvidia once supply bottlenecks, particularly for H20 chips, are resolved. This underlines the structural importance of AI infrastructure for future economic growth.

The Bank of England identifies potential systemic risks from the increased use of AI in banks and insurance companies. While AI can increase efficiency, common weaknesses in widely used models could lead to misjudgements of risk and credit misallocations. There is also a risk that AI-driven trading strategies could lead to synchronized market movements during periods of stress, jeopardizing stability. Dependence on a small number of AI service providers also increases operational risk. The Bank emphasizes the need for flexible and forward-looking monitoring in order to meet these challenges.

Bank of America plans to invest $4 billion in AI and new technologies in 2025, representing almost one-third of its total technology budget. Over 90% of employees are already using the internal AI assistant, leading to a reduction of IT support requests by more than half. AI tools are boosting efficiency in areas such as development, training, customer service, and customer engagement. These investments highlight the strategic importance of AI for future competitiveness and efficiency improvements in the banking sector.

Newer AI models like OpenAI's o3 and o4-mini aren't just improving on benchmarks — they're fundamentally changing what tasks can be fully delegated to them. Unlike earlier versions like GPT-3.5 and GPT-4, these models use integrated tools to independently manage complex workflows. With minimal human input, o3 can now execute detailed tasks from general instructions. This marks a major shift toward AI acting as autonomous collaborators rather than simple assistants.