Monthly Newsletter #1: Power, Optimize, Connect
Welcome to this month's edition of what I've been thinking about in the technology space. Rather than recap market movements, a few ideas that may seem underweighted in the current conversation.
Energy Islands
Hyperscalers are running into a frustrating reality: they can order the chips and lease the real estate, but they can't get utility connections on any reasonable timeline. Grid infrastructure moves on decade-long timelines while AI capacity demands move on quarterly ones. That mismatch has spawned an entire "bring-your-own-power" ecosystem.
What started as a frantic scramble has become something more permanent. Data centers are shifting from tenants of the grid to their own power utilities. By some estimates, a third of new data center capacity may rely on off-grid power structures by 2030. We're moving toward a world of energy islands, facilities that generate their own power and, in some cases, feed excess capacity back to the grid.
Beyond operators wanting to control their own destiny, the economics are compelling: on-site generation skips the AC-to-DC conversions required by traditional grid delivery, saving 10 to 15% of power. At gigawatt scale, that's not a rounding error.
What's emerging is a broader rotation in how we think about power infrastructure—first in energy sources, then in distribution. This likely isn't the final form. The demand is simply too large for any single solution to satisfy, and constraints keep shifting: today it's grid connection speed, tomorrow it might be emissions regulation or fuel costs. Each new bottleneck opens a door for a different approach.
The Optimization Pivot
The physical constraints outlined above are now a known quantity, which makes the "more is better" narrative feel like an oversimplification. Historically, every era of massive expansion is followed by an intense optimization phase. We saw this in the early days of computing, when developers squeezed entire operating systems into 40MB of memory, a level of rigor that brute-force scaling has temporarily made us forget.
Software has a history of becoming lazy when hardware is cheap. But as we hit the hard limits of energy, and increasingly, capital, that era of bloat may be ending. The shift is already underway: the best AI agents today win not by consuming more compute, but by knowing when not to.
This isn't a call to stop scaling. The expansion continues, and markets will keep rewarding it. But the next wave of winners may be companies that optimize while they scale, not one or the other. Nature offers a useful reference point. The human brain performs the most complex computation in the known universe on about 20 watts. We won't get there overnight, but the direction matters.
The Complexity Paradox
We keep hearing that AI will simplify everything. But inside enterprises, I see the opposite: every new AI tool creates new data flows, every automation creates new edge cases, every team adopting its own solution creates new integration challenges. The bottleneck isn't disappearing; it's shifting from human labor to system coordination.
This is likely why we haven't seen the immediate wave of white-collar displacement some predicted. The hard part isn't getting AI to do a task; it's getting it to function within the messy context of how enterprises actually operate.
For AI to generate value, it needs connectivity across these siloed systems. Context is the moat. The companies that own the connective tissue—integration, orchestration, interoperability, access—will become more valuable. The ones that outlast will be platforms where AI can act on proprietary information that doesn't exist anywhere else. Everything else is just a commodity in different packaging.
As always, I'd welcome your thoughts on any of this.
Until next month,
Jess Xu