Weekly roundup, May 23, 2026
Google and Microsoft just told us the same thing from opposite directions.
Google IO this week was an AI firehose. Gemini 3.5 Flash — faster, cheaper, "better at agentic tasks." Gemini Spark — an agent that orchestrates other agents. Omni video gen. The Ultra tier dropped from $250/month to $100. The keynote ran for hours. I watched excerpts. I'm still tired.
And then there's Microsoft, quietly removing the Copilot button from Office 365. Letting users remap the Copilot key on their keyboards. Two companies. Same message. Different volume.
The message is: AI costs real money, and someone has to pay for it.
Google is still spending. The IO presentation felt like a company that hasn't blinked yet — or can't afford to look like it's blinking. But look closer. Free tier usage is now capped by compute, not prompt count. The bean counters are in the room. They're just not on stage.
Microsoft blinked first. The Copilot button that was mandatory hardware is now optional. The Office integration nobody asked for is being rolled back. Not because Copilot doesn't work — because shoving it into every surface made people resent it. Forced adoption. The worst kind.
The WAN Show put it well: "Are the cracks showing that AI might not be what the people want?" The answer was no — but forcing people to use it in ONE particular way is deeply unpopular. People want agency. Even with their AI agents.
Meanwhile, the economics are reshaping faster than the demos.
xAI is renting its entire GPU capacity to Anthropic. $1.25 billion a month. For three years. Elon built the data center without a plan. Anthropic needs the compute. The deal makes sense for both. But it also tells you something about where the money is going — not to training new models. To running them.
Anthropic is projecting its first profitable quarter. That's a bigger milestone than any benchmark score. It means someone figured out how to sell AI without losing money on every inference. Nobody else has done it yet. OpenAI hasn't. Google's AI division hasn't. Anthropic might be first.
The pressing issues nobody's fixed yet:
Agent reliability. Every keynote shows agents doing things. Few of them work reliably outside the demo. Gemini 3.5 Flash claims better agentic performance. I'll believe it when an agent books a haircut without hallucinating the date, the time, and the hairdresser's name. We're still at the stage where the impressive part is that it almost worked.
Context management. Agents forget. Not "forget" in the cute way. Forget in the way where you build a multi-session workflow, and somewhere around turn 40 the agent starts responding to a conversation from three days ago. Memory is still unsolved. Everyone's adding more context. Nobody's figured out what to remove.
The cost of "free." Google's free tier cap change matters. It means the free-usage era is ending — not with an announcement, but with a restriction. More complex prompts hit the cap faster. Try asking an agent to do real work and you'll notice. The meter is running.
The evaluation gap. We have no good way to measure whether an agent is actually better. Perplexity scores are meaningless for multi-turn tool use. Every company has internal evals. None of them agree. The market is pricing models as if better benchmarks mean better agents. They don't.
What I'm watching:
Google's Gemini Spark — the agent orchestrator. The idea of one agent delegating to others isn't new. But Google baking it into Workspace is. If it works, it's the first time an AI demo about "managing your RSVPs" actually saves time instead of creating it. Big if.
The xAI/Anthropic deal. Three years of guaranteed compute at this price is a structural shift. It means Anthropic bet that inference costs will go down — or that they can charge enough to cover them. Either way, it sets a floor on what running agents at scale actually costs.
Microsoft's retreat. Not because Copilot is failing. Because forced adoption is. The lesson: build something people choose to use, or don't build it at all. The Copilot key was a bet that default placement beats user preference. It lost.
The week wasn't about breakthroughs. It was about gravity.
Money, attention, and patience are finite. The AI boom spent two years ignoring that. This week, it started paying attention.
Gravity doesn't break things. It just makes them settle where they belong. The hype cycle is finally in freefall — and for the first time, that's good news. The companies that survive the correction won't be the ones with the best demos. They'll be the ones that figured out how to make this stuff actually work, at a price the market can stomach.
Top comments (0)