580 billion USD Spent on Data Centers in 2025

This TechCrunch article presents the San Francisco-based company, Scribe, which has an AI-based platform that analyzes company workflows in order to identify where AI can yield process improvements.

• The founders argue that companies approach AI in an incorrect manner. Companies tend to brainstorm about where AI automation can help, rather than taking a bottom-up approach of analyzing existing workflows: “Without really knowing how work is done, it is really hard to know where to improve it, where to automate it, where agents can help.”.
• The company has documented more than 10 million workflows across 40’000 software applications, which covers 5 million users and 94% of Fortune 500 companies.
• The company claims that its application leads to efficiency savings of 35 to 42 hours per person per month.
• The company just raised 75 million USD in a funding round, yielding a post-money valuation of 1.3 billion USD.

The Chinese AI company DeepSeek has released a model that significantly improves the ability of an AI model to remember information when processing requests.

• AI models currently model conversations with users as text that the model breaks down into tokens. However, this process uses a lot of compute and memory space, which leads to the AI model having to forget earlier elements of the conversation to optimize storage. This phenomenon is known as context rot.
• The DeepSeek model stores history as an image – as if a digital photo of the conversation were taken. Optical character recognition (OCR) is used to extract text from the pictures. The approach allows the model to retain information at a fraction of the computing and memory costs compared to using text.
• Scientists say that approach can become closer to how humans retain memories. Past memories may be blurred but return to focus when needed. Current AI models only remember recent, but not the important details.
• The approach can also help during training. DeepSeek says that their OCR system can generate over 200’000 pages of training data a day on a single GPU.

This VentureBeat article from CapeStart looks at six lessons learned by observing AI projects that have succeeded and failed over the past few years. The main culprits for projects that do not make it into production are poor planning and imprecise goals, rather than AI technologies.

• As in all projects, SMART criteria (Specific, Measurable, Achievable, Relevant, Time-bound) should be used to determine goals. Further, there should be stakeholder adherence from the start. In one example cited, the users did not trust the model developed.
• Another reason for failure is poor quality data. The author cites Python tools like Pandas, Great Expectations and Seaborn for data cleaning and exploratory visual analysis.
• Two other reasons cited are ignoring the IT realities of production environments (e.g., servers unable to adapt to the compute requirements of AI) and failure to maintain the model (detect and counter the effect of data drift – where a model’s performance degrades over time because the data processed deviates statistically from the training data).

This Guardian article looks at some of the investor concerns over OpenAI, and its possibility to seriously generate revenue.

• OpenAI is expected to spend 1.4 trillion USD on computing infrastructure in the next 8 years. Its annual revenue is currently 13 billion USD.
• The company claims that the 1.4 trillion USD can be paid off by future demands for its products. ChatGPT subscriptions currently account for 75% of its income. The service has 800 million weekly users and 1 million business customers.
• OpenAI is also hoping that AI development and chip costs will fall, and that the partnership with Apple on a chip design will help to create revenue.
• The OpenAI chief financial officer recently suggested that the US government might underwrite the company’s chip spending. CEO Sam Altman has denied claims that he will ask for a bailout.
• Meta, Google and Microsoft have similar concerns about their AI spending, but these companies already have existing products generating large revenues.

Tesla CEO Elon Musk launched rumors of a partnership with Intel to produce its 5th generation AI chip AI5.

• Tesla is currently suffering from supply chain bottlenecks. Its current suppliers are Samsung in South Korea and TSMC in Taiwan.
• Musk is considering creating a huge chip-making factory with Intel in the US. The first AI5 chips would be produced in 2026, with full production capacity in 2027.
• Musk claims that the AI5 chip will consume just one-third of the power consumed by Nvidia’s Blackwell chip, and cost only 10% of Blackwell’s manufacturing cost.
• The deal would be good news for Intel, which is currently 10% owned by the US government. The deal also aligns with current US policy on technological and supply chain independence.

OpenAI has released an experimental large language model where it is possible to identify the neurons activated during reasoning.

• The domain that seeks to understand how models reason is known as mechanistic interpretability. It is important so that hallucination correction, transparency, and explainable AI can be achieved.
• The neural networks in today’s large models are dense, meaning that the neurons in one layer are connected to a large number of neurons in the next layer. This means that neurons can be involved in representing a number of concepts (a phenomenon called superposition). It is therefore difficult to attribute neurons to a specific concept.
• OpenAI’s experimental model is sparse, with neurons being connected to relatively few neurons in the next layer. This makes it much easier to attribute the activation of different neurons to specific concepts.
• On the other hand, the model is less powerful than contemporary large models. It is roughly equivalent to GPT-1, released in 2018. This model is far less powerful than the landmark GPT-3 model that launched the popularity of ChatGPT.

This article from TechCrunch summarizes a podcast discussion by journalists on generative AI and climate change.

• The International Energy Agency estimates that the world will spend 580 billion USD on data centers in 2025. This is 40 billion USD more than the amount spent on finding new oil supplies. The data centers are also transforming the landscape in many areas.
• Half of the electricity demand will come from the US, the remainder mainly from Europe and China. Most data centers will be built near large cities, which will place a challenge on the existing, and aging, grid networks. Places like Texas are expected to have rolling blackouts during periods of high energy consumption, like in Summer.
• OpenAI has announced that it will spend 1.4 trillion USD on data centers in the coming years, Meta will spend 600 billion USD and Anthropic 40 billion USD.
• On the other hand, some investors doubt the ability of these companies to pay such large investments. Further, experts suggest that the demand could lead to renewed investments in renewable energies – given the increasing environmental regulation.

This InfoWorld article argues that companies should focus more on model inference performance, rather than on model training performance.

• The International Data Corporation (IDC) estimates that spending on AI infrastructure for inference will surpass training spending in 2025. Models are trained once, but queried many times. Models trained on low-quality data are useless to organizations, whereas an AI inference infrastructure that provides up-to-date quality corporate information is valuable.
• The IDC also estimated that 65% of organizations are trying to run over 50 generative AI use cases in production in 2025. Each use case represents potentially millions of inference calls.
• The article argues that developers need to master retrieval-augmented generation (RAG) pipelines, vector databases, secure APIs for model querying, and cost management for API calls.
• The latest Nvidia GPUs are optimized also for low-cost inference, as are many chips being produced by start-ups.

Google and UCLA research has led to new technique, called supervised reinforcement learning, that improves the performance of reasoning models across mathematical reasoning and agentic software engineering benchmarks.

• A recent contribution to reasoning models has been to include reinforcement learning. Here, a model gets rewarded when it comes up with a correct and verifiable result. However, this approach does not scale to complex problems.
• Another approach is imitation learning, implemented using Supervised Fine-Tuning, where the model learns by observing experts solve problems. However, these models tend to overfit the training data, and so adapt poorly to new problems.
• This paper presents an approach called Supervised Reinforcement Learning. The idea is to break a problem solving challenge into a series of actions, each of which can be rewarded using reinforcement learning. In the training phase, the model articulates its internal reasoning monologue. Each action is rewarded when it corresponds to the action suggested by an expert who solved the problem.

In an interview, Dario Amodei, CEO of Anthropic warned other AI companies of the dangers of a lack of transparency regarding AI.

• He compared the current situation to the tobacco industry which “knew there were dangers, and they didn’t talk about them, and certainly did not prevent them”.
• Amodei warned that AI could lead to the elimination of 50% of white-collar entry-level jobs in the next 5 years.
• He spoke of a compressed 21st century, where “we get 10 times the rate of progress and therefore compress all the medical progress that was going to happen throughout the entire 21st century into five or 10 years”.
• He mentioned that the flip-side of achieving progress on vaccines and other medical breakthroughs is the increased facility by bad actors to create biological weapons.
• Amodei also spoke of risks, such as models being aware when they are being evaluated, attempted blackmail of human overseers, and use of AI to launch cyberattacks. An example is a Chinese state sponsored hacker group which used Claude Code to create cyberattacks on 30 entities worldwide, with several successful intrusions. Only 80% to 90% of the attacks had humans in the loop.