GoogleFinlandNuclear energy
Google bought half a nuclear plant’s output because autocomplete got complicated
Google announced Wednesday that it will invest at least €13 billion in Finland over the next two years to expand AI and digital infrastructure—its largest AI-related investment in Europe.
Finland offers abundant low-carbon electricity, a relatively cool climate, strong grid infrastructure and political enthusiasm for large technology investment. Google’s answer to the remaining electricity problem is unusually direct.
Its 22-year power purchase agreement with Fortum covers up to 50% of the Loviisa nuclear power station’s capacity. Fortum says the contract provides economic certainty for extending and upgrading the plant through 2050. The companies also plan to explore additional nuclear and renewable generation.
For decades, computing became more important partly because each calculation became cheaper. AI is now becoming important enough that its aggregate demand can justify extending a nuclear power plant.
The cloud used to abstract away the computer. Now the cloud needs us to remember where electricity comes from.
Somewhere behind Gemini is apparently a reactor.
Source: Reuters ↗
OpenAISamsungSemiconductors
OpenAI is increasingly trying to own the answer to “what does the model run on?”
OpenAI says it has made significant progress working with Samsung Electronics on next-generation chips. Harrison Kim, general manager of OpenAI Korea, described joint chip production and research as one of the areas where cooperation has advanced most strongly, although neither company disclosed detailed specifications Wednesday.
This does not emerge from nowhere. OpenAI announced in June that it had designed its first custom AI processor, Jalapeno, with Broadcom, with TSMC expected to manufacture it. Samsung and SK Hynix previously signed letters of intent to supply advanced memory for Stargate data centers.
Custom silicon has become almost inevitable for frontier AI companies. If you spend tens or hundreds of billions of dollars running a specific class of workload, shaving even modest percentages from power consumption, memory traffic or inference cost can justify extraordinary engineering effort.
Samsung is not merely a supplier. OpenAI says Samsung employees use ChatGPT across research, development, marketing and sales in one of its largest enterprise deployments worldwide.
The AI stack is becoming circular: Samsung helps build the machines that run OpenAI. OpenAI helps Samsung employees design and sell Samsung products.
At some point “software company” stops being a very useful description.
Source: Reuters ↗
United StatesChinaModel distillation
America says China is copying the homework. China says it learned the material.
The U.S. government accused six Chinese AI companies Tuesday of conducting what officials called industrial-scale malicious distillation of American AI models. The companies named include DeepSeek, Moonshot AI and Alibaba. U.S. officials say they used outputs from systems developed by OpenAI, Anthropic, Google and SpaceX to accelerate Chinese model development.
Distillation itself is not exotic. It generally involves training a smaller or cheaper model using outputs generated by a larger system. The dispute concerns scale, authorization and intent.
American officials argue that Chinese companies extracted valuable behavior from expensive U.S.-trained systems rather than reproducing the underlying research and compute investment independently. China rejects the accusations and says its progress reflects domestic scientific and technological development.
The timing is delicate. Presidents Donald Trump and Xi Jinping are expected to meet later this month, while officials are also preparing discussions concerning AI safety. Washington says the alleged copying has potential military and cybersecurity consequences.
Traditional export controls work reasonably well when the valuable object is a machine. You can count lithography systems. You can restrict GPUs. A model exposed through an API can potentially transfer some capabilities simply by answering enough carefully chosen questions.
Countries are trying to protect an industrial advantage whose product can partially describe itself to competitors.
Source: Reuters ↗
