Hot Deal Talk

  Wall Street often talks about AI as if it were a software revolution. New models. New algorithms. New applications. The story is usually framed around code. But outside the technology industry, the same phenomenon looks very different. Utilities do not see AI as software. They see it as load . Every new model requires training. Training requires compute. Compute requires servers. And servers require electricity. A lot of electricity. Large AI clusters can consume tens of megawatts of power. The largest data centers now rival the electricity demand of small cities. From a utility’s perspective, AI is not a new type of software product. It is a rapidly growing industrial power demand. That difference in perspective matters. Technology companies think about scaling in terms of code. Utilities think about scaling in terms of infrastructure. Servers require cooling. Cooling requires water or specialized systems. Water access requires permits. Permits require time. Software can scale...

  Chips started the race. Power will decide it. For the past two years, the AI boom has been defined by one word: chips . The first phase of the AI expansion was driven almost entirely by semiconductors. Graphics processing units, especially those produced by companies like NVIDIA, became the core engines of the new AI economy. Every major technology company rushed to secure GPUs. Cloud providers expanded their compute clusters. Startups raised billions to train larger models. In this first phase, the narrative was simple. Better chips meant better AI. And companies that designed or supplied those chips captured enormous investor attention. But silicon was only the beginning. As AI models grow larger and computing clusters expand, a different constraint is becoming visible. The problem is no longer just computational power. It is electrical power . Running large- scale AI systems requires massive amounts of electricity. Modern AI training clusters can consume tens of megawatts of p...

  AI chips can scale quickly. Power infrastructure cannot. The AI boom is often described as a race for better chips. Faster GPUs. Larger models. More powerful computing clusters. But as the industry expands, another constraint is becoming visible. It is not silicon. It is infrastructure. Running large AI systems requires enormous amounts of electricity. Modern AI data centers can consume tens of megawatts of power, and the largest facilities are approaching the energy demand of small cities. That power has to come from somewhere. Before a new data center can operate, it must connect to the electrical grid. And that process is often much slower than building the facility itself. Utilities must evaluate grid capacity. Engineers must study the impact on transmission systems. Substations may need upgrades. New equipment may need to be installed. Then there are permits. Large infrastructure projects require regulatory approvals, environmental reviews, and local plan...

  For many years the technology sector moved in relatively short cycles. A new device launches. Demand spikes. Production ramps. Then the market cools. The AI boom is beginning to look very different. This time the cycle is not limited to software or consumer electronics. It is expanding into infrastructure. And infrastructure cycles tend to last much longer. Training large AI models requires enormous computing clusters. Those clusters require specialized chips. But chips alone are not enough. They must be installed in data centers. And data centers require land, power connections, cooling systems, and network capacity. Each layer takes time to build. Permits must be approved. Grid connections must be secured. Electrical equipment must be manufactured. Facilities must be constructed and tested. These steps slow the cycle down. But they also make the cycle longer. A software boom can expand quickly and disappear quickly. Infrastructure buildouts rarely move tha...