The end of cheap AI, Google's messy I/O, and the shift from doing work to managing it

The era of heavily subsidized AI usage is ending, and prices are about to get real. Google showed off a sprawl of new products without much clarity. And the way knowledge work actually gets done is quietly changing underneath all of it.

• AI vendors start charging real money for heavy usage
  My take: For about six months, power users have been getting 10x to 20x more value in tokens than their subscriptions cost. That subsidy is now ending. Anthropic, GitHub, and others are moving heavy and automated usage to pay-per-use pricing, and some teams are seeing estimated bills jump from a few hundred dollars to five figures. This is not a temporary pricing experiment. It is the predictable result of demand for AI outstripping the supply of compute. If you have built workflows or budgeted on the assumption that AI usage is nearly free, revisit those numbers now. Treat AI spend like any other input cost with an ROI attached, not a flat subscription.
• Google's I/O event delivered a confusing pile of products
  My take: Google announced a lot: a new video model called Omni, a personal agent called Spark, an updated coding tool, and a new Gemini model. The problem was that almost nobody could explain how it all fits together, and several products are not even available yet. Google still has real advantages, mainly 900 million Gemini users and deep integration into tools you already use. But if you are choosing AI tools for your business, Google is not the safe bet for clarity right now. Pick the vendor whose product roadmap you can actually understand and plan around.
• Companies are rewarding employees for using more AI
  My take: Firms like Amazon, Meta, and Disney have built leaderboards and dashboards tracking AI usage, and yes, some employees game them. The cynical take is that this proves AI is hype. The better read is that nobody yet knows the best ways to use AI agents, and the only way to find out is to experiment. Most of that experimentation will not show up as direct revenue this quarter, and that is fine. It is research and development at the individual level. Encourage your team to experiment and burn some tokens on dead ends, but ask people to show what they built and learned, not just how much they spent.
• AI work is moving from your laptop to your phone
  My take: OpenAI put its coding agent inside the mobile app, so you can start, review, and steer tasks from your phone while the work runs on a machine back at the office. The real shift here is not the device. It is that more knowledge work is becoming about managing agents that do the work rather than doing it yourself. Your job increasingly becomes setting up the conditions, then reviewing and approving. If you manage a team, start thinking about where the new bottleneck is: it is no longer how fast work gets produced, it is how fast a person can review and approve it.
• Access to the best AI models is getting more restricted
  My take: Anthropic's powerful security model was released to only a short list of approved companies, and OpenAI did the same with its equivalent. For security and compute reasons, the assumption that everyone gets equal access to the best AI is breaking down. That matters strategically: if your competitive position depends on having the same AI capability as everyone else, that is no longer guaranteed. Pay attention to which tier of access you actually have, and do not assume the frontier model you can use today will be the one you can use next year.
• The 'AI will destroy all the jobs' narrative is being challenged
  My take: A lot of the loudest job-loss predictions come from the people building and selling AI, which is unusual and worth noting. The more grounded view is that AI is genuinely disruptive but on a longer and messier timeline, partly because it turns out to be expensive and partly because real businesses move slowly. AI also creates demand: cheaper services reach new customers, and entirely new service models become possible. For planning purposes, do not freeze hiring because 'AI will replace them anyway.' Plan for disruption that is real but slower and more specific than the headlines suggest.