Bad Harmonics on the US Grid

This Bloomberg article looks at the problems that the increased demand for electricity by data centers, primarily due to AI, is having on the electrical grid in the US. The demand for electricity will increase by 16% in the US in the next 5 years, which is three times the estimated increase made one year ago. This rise is unprecedented, and was not even seen in times of population boom. A data center is roughly equivalent to 10’000 homes, and the rise is coming at a time that the grid infrastructure is aging and in need of repair, extreme weather situations are more common, and people are using more electrical equipment (notably cars). There are fears of power outages and increased costs for consumers. One problem is that AI’s energy consumption is more like a sawtooth graph than a smooth line. Current grids are not designed to handle such load fluctuations. In Northern Virginia, electrical suppliers foresee separating the infrastructure that supplies data centers from equipment sending power to homes.

Another key problem mentioned is “bad harmonics”. These are distortions to the electrical energy (sine-ware) signals sent across high-voltage lines from power stations. Sustained distortions that exceed 8% can damage electrical devices in the home (refrigerators, TVs, etc.), and the article estimates that this can lead to billions of dollars in damage. Sudden surges can even provoke fires. The study shows that areas close to data centers are particularly affected, notably in Northern Virginia whose data center capacity is twice that of Beijing. The study was made from readings by 770’000 home sensors across the US between February and October 2024.

The Google CEO, Sundar Pichai, has told employees “to be relentlessly focused on unlocking the benefits of [AI] technology and solve real user problems”. The focus in the coming year will be on Gemini AI and its derivatives like the App on consumer devices as well as Project Astra – Google’s AI agent project. Google is facing fierce competition from OpenAI and Perplexity, but it believes that there is still time and development to be done to come out ahead. The company recently launched Jules, the AI-based coding assistant, and NotebookLM and Project Mariner (a Google Chrome extension) were launched earlier this year. The Gemini App allows to gain access to several tools including Google’s chatbot. The company believes that the App could become the 16th Google application to obtain half a billion users. Outside of technology issues, the company is under pressure from the US Department of Justice which is accusing Google of enforcing a monopoly on on-line advertisements, and the company may be forced to split away the Chrome browser into a separate company.

The U.S.-China Economic and Security Review Commission was given the mandate to “monitor, investigate, and report to Congress on the national security implications of the bilateral trade and economic relationship” between the US and China. It published its findings in November which notably calls on the US Congress to fund a “Manhattan Project-like program dedicated to racing to and acquiring an Artificial General Intelligence (AGI) capability” because “China’s rapid technological progress threatens U.S. economic and military leadership and may erode deterrence and stability in the Pacific”. China is currently facing an economic downturn, notably due to the collapse of the property market, and the commission believes that this has been a pretext for General Secretary Xi Jinping to establish increased authoritarianism, and that “it has become unlikely that anyone could dissuade Xi should he decide to take actions that risk igniting a catastrophic conflict”.

China is investing heavily on technological research and is competing in several areas with the US. The report admits that China has taken the lead in battery energy storage systems, including the lifecycle of access to minerals for fabrication to the production of batteries needed for electrical vehicles. The commission is recommending an import ban on energy products that require remote servicing or monitoring. In the area of quantum, China is seen as leading in the field of quantum communications while the US maintains its lead in quantum computing and quantum sensing. The Commission is urging Congress to maintain and enforce export restrictions.

This article cites a report from the Sophos cybersecurity company which finds that 67% of healthcare organizations were the target of ransomware attacks in 2024, compared to 34% in 2021. Encryption is not a panacea against these attacks as 53% of healthcare organizations that were attacked, and which had encrypted their data, still paid the ransom. Recovery times are also increasing, with only 22% of attacked organizations fully recovering from an attack within one week (compared to 54% in 2022). The medium ransom payment is valued at 1.5 million USD.

In this context, the US Department of Health and Human Services is proposing a modification to the Health Insurance Portability and Accountability Act (HIPAA) of 1996. Healthcare organizations will now be obliged to compile a technology asset and network map inventory, carry out a compliance audit each year, ensure all patient records are encrypted, use multi-factor authentication, deploy malware on all machines and remove software that has no relevant purpose, implement network segregation, as well as implement and test controls for backup and restoration. Organizations will also be obliged to define a procedure for the restoration of data within 72 hours.

This report relates the results of a pulse survey of 53 Chief Information Security Officers (CISOs) on their perception of the use of AI. The CISOs came from a group of CISOs that operates as an angel investor syndicate and advises on the use of innovative cybersecurity technologies. There is a wide acceptance of the need for new tools because of the increasing pressure on CISOs. The report cites a previous report at the organization which found that 63% of SOC analysts are suffering from burnout. Among the challenges to adopting AI, 66% of CISOs consider risks to data privacy as the main challenge; 60% cite the shortage of cybersecurity talent and 51% cite misaligned priorities from management with a poor understanding of AI. Another risk for 49% of CISOs is the presence of inflexible technologies such as legacy systems in the infrastructure which link poorly to AI systems. Overall, 94% are concerned about the increased pressure from management to use AI. Nonetheless, 74% believe that faster decision-making will introduce more benefits for SOC management than risks.

This VentureBeat article reviews some of the key features of OpenAI’s o3 model, announced last week and expected to be available at the end of January. The model has been in the news since it scored 75.7% on the ARC benchmark under standard computing conditions – and 87.5% in a high computing setting. The previous high-score was 53% obtained by Claude 3.5. One aspect of the model is program synthesis which is the ability of the model to combine patterns, algorithms, or methods learned during pre-training. This allows the model to solve coding and logic puzzle challenges that are more advanced than those observed during training. Another feature is its use of chain-of-thought (CoT) reasoning at inference time – the idea is that the model proposes several possible answers, and CoT is used to choose the best of these. This approach is seen as analogous to how humans find solutions to problems. The evaluator model is trained on expert-labeled data. This is also seen as a potential weakness of the approach because the multiple intermediate solutions are judged on internal metrics, rather than on real-world equivalent problems, and a computation requires over one million tokens. Critics do not see the approach as being able to scale, with Google DeepMind’s Denny Zhou saying that reliance on reinforcement learning (RL) scaling and search (of options) was a potential “dead end”.

This TechCrunch article reports on a discussion journalists had with 20 venture capitalists about their predictions for 2025. The article notes that their predictions for 2024 had been somewhat over-optimistic for AI take-up in enterprises as corporate budgets remained tight, and enterprises still feared risks related to model hallucinations. For 2025, the VCs expect companies to focus on better data management for their AI projects, and to increase adoption of AI coding agents, notably for tasks like re-platforming existing mainframe applications. The VCs expect more investment by companies next year as management puts pressure on departments to get leverage from AI. They also foresee more AI focus on behind-the-scenes work like accounting and revenue cycle management. VCs also hope that 2025 brings a better understanding of what pricing models are the most useful in the AI domain (consumption-based, outcome-based, …). Among the areas where VCs expect to invest in 2025 are enterprise resilience (following the impact of the Crowdstrike attack), data sovereignty (driven by regulatory requirements and geopolitical concerns), agent frameworks, and AI technology that reduces business friction and drives enterprise-value (as opposed to driving revenue and reducing costs). Apart from AI, the VCs cite quantum, gaming, energy (including nuclear fission and fusion), and multi-cloud deployment as domains for investment next year. The AI companies expected to thrive in 2025 are those working on solutions to reduce sales and procurement cycles, as well as those linked to the defense industry.