McKinsey touts ROI, OpenAI builds inference monster, data centers drink the ocean

According to McKinsey, enterprise AI is “on the way to return on investment”. The cautious phrasing—neither “has arrived” nor “delivers”, but “on the way”—reveals more than it hides: consulting firms identify pressure from boards to stop burning POC budgets and start showing a profit line, yet the metrics behind the headline remain vague.

Meanwhile, in the silicon layer, OpenAI is preparing its Jalapeño chip: 128 cores in a single package, 1.7 exaFLOPS and 27 terabytes of HBM. On paper this gives Altman an initial advantage over Nvidia’s Blackwell, and perhaps even over the future Rubin. At the same time, early performance tests by Nvidia on Groq’s LPU 3, which ran Gemma 4 31B, paint an optimal scenario for next-gen dataflow accelerators—a $20 billion bet that has not yet been proven under real production loads.

Physical infrastructure does not wait for benchmarks. US data centers have slashed their water footprint over the past decade, and those numbers were collected before the current AI boom, meaning today’s situation is significantly worse. Fragmented reporting between providers, authorities and regulators (silo-ed reporting) makes it hard to get a full picture, and the EPA is considering dropping the public-notice requirement for minor pollution permits, leaving the decision to states and local authorities.

While hardware inflates and infrastructure chokes, someone profits from the mess: “AI slop” is good for businesses if they know what they’re doing. The lack of accountability from those who generate generic content at commercial scale turns into an opportunity for those who can filter, curate or train on the noise—a business model built on chaos rather than its solution.