TL;DR
Big AI buyers are starting to build their own chips just as governments and communities push the real energy and political costs of data centers back onto operators.
Public markets are finally questioning AI cash burn, so the whole space is trading less like magic software and more like a leveraged infrastructure bet where compute, power, and geopolitics decide who survives the next drawdown.
Key Events
Report
The AI trade is starting to look like late-stage dot-com: custom chips everywhere, infra hitting hard limits, and equity markets finally asking about cash flow.
Underneath the noise, three levers matter for your P&L this quarter: who controls compute, what AI infra really costs, and how fast the bubble air leaks out.
OpenAI launched its first custom LLM chip Jalapeño with Broadcom and TSMC, touting top-tier performance-per-watt for ChatGPT-style inference and parity with Nvidia Blackwell and Google TPUs.
The chip went from concept to working silicon in about nine months using OpenAI’s own models for design, though some observers argue it largely repackages Broadcom ASIC IP.
Qualcomm is buying Modular for $4B, adding the Mojo language and MAX inference stack to earlier bets on Alphawave and Ventana and talks with Tenstorrent to own edge AI silicon+software.
Nvidia is countering with near-zero-water liquid cooling and domain-specific agents like BioNeMo and Metropolis, even as customers report H100/B200 reliability problems and sharp AI-chip price spikes in China.
AI-exposed stocks have sold off from Wall Street to Asia, with names like Micron whipsawing as investors reassess how much of the AI boom is already priced in.
Analysts increasingly compare the AI run-up to the dot-com bubble and warn a correction could echo past crises, while SoftBank’s Masayoshi Son dismisses bubble talk as “blasphemy.” OpenAI has reportedly missed revenue forecasts, faces electricity bills in the millions, and is said to require “hundreds of billions” to stay viable at current compute intensity.
Anthropic’s CEO warns a $1T compute era could push AI firms toward bankruptcy by 2027 just as enterprises pull back, stressing the need to turn “claimed revenue” into durable contracts.
In Washington, a proposed bill would make tech firms, not local ratepayers, directly cover the energy costs of their AI data centers.
On the ground, AI data-center build-out is running into a “human bottleneck” of scarce skilled labor plus mounting community resistance over noise, water use, and thin local economic benefits.
Communities and regulators are demanding granular transparency on energy and data practices, highlighted by Meta pausing an employee-tracking program and related AI-training project after an internal leak exposed how staff data were used.
Meanwhile AI-driven power demand is already turning electric utilities into growth stocks and is expected to raise customer costs as grids are upgraded to feed data-center loads.
What This Means
Capital, compute, and politics are finally colliding: markets are starting to price AI like a leveraged infra trade, while states and communities renegotiate who controls—and who pays for—the chips, power, and models underneath. The live decision is how much of your exposure sits in high-burn frontier labs versus the infra, energy, and open ecosystems that are likely to survive a hard reset.
On Watch
Interesting
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
Sources
Key Events
On Watch
Interesting