FERC Speeds Up Grid Interconnection for AI Data Centers — A New Phase in the Infrastructure Race

On June 18, the U.S. Federal Energy Regulatory Commission (FERC) issued a decision to dramatically accelerate the grid interconnection process for large power customers such as AI data centers, semiconductor fabs, and advanced manufacturing facilities. A process that once took years of interconnection review is expected to be shortened through a priority lane.

NVIDIA's blog hailed the decision as a major milestone in building the foundation for "AI factories." CEO Jensen Huang has described AI as a "five-layer cake," with energy infrastructure positioned as the most fundamental layer.

TechCrunch, however, pointed out that "while faster interconnection has been achieved, the decision does nothing to address the underlying problem of power shortages." The Associated Press likewise reported that although regulators ordered grid operators to prioritize power delivery to AI data centers, it remains unclear whether supply can keep pace with the surge in demand.

The issue goes beyond technology. According to CNBC, Amazon has opened an investigation into internal engineers who opposed the expansion of its data centers. The report laid bare rising concern over the rapid buildup of AI infrastructure even inside companies themselves.

Amazon Takes On NVIDIA — Weighing External Sales of Its Own AI Chips

Amazon Web Services (AWS) is in talks to sell its in-house AI chips for use in other companies' data centers as well. CEO Andy Jassy has described this market as a "$50 billion opportunity," signaling a serious challenge to NVIDIA, which commands an overwhelming share of the GPU market.

NVIDIA currently dominates most of the market for chips used in AI inference and training, but a move by cloud providers to sell their own chips externally could bring major change to the competitive landscape. That said, NVIDIA's ecosystem — its CUDA-centered software stack — still enjoys a formidable advantage, and how much share can be captured by hardware alone will be the focus going forward.

OpenAI Lands Major Hires Ahead of Its IPO — Bringing in a Transformer Co-Inventor

Ahead of its IPO, OpenAI has announced a series of notable hires. It welcomed Noam Shazeer, co-inventor of the Transformer paper, from Google DeepMind, and around the same time added Dean Ball, a former AI policy official from the Trump administration.

The Shazeer hire carries major technical significance. Having the creator of the Transformer architecture join OpenAI's camp suggests the company's intent to maintain and extend its advantage in model development. Dean Ball's arrival, meanwhile, can be read as a strategic effort to strengthen ties with regulators. With the IPO in view, the push to shore up both technical and policy firepower is accelerating.

ChatGPT's Medical Capabilities Get a Major Upgrade — Accuracy That Beats Physicians' Responses

OpenAI has upgraded ChatGPT's medical capabilities with GPT-5.5 Instant. In comparison tests run by the company, ChatGPT reportedly outperformed physician-written responses on accuracy, clarity, and completeness alike. In particular, OpenAI says the error rate in health-related statements dropped by 71%.

These results, however, come from OpenAI's own testing, and no third-party verification has yet been made public. There is also the criticism that AI that scores highly on exam questions still leaves something to be desired in actual clinical settings. A new benchmark reported by Medical Economics found that medical AI scores highly on tests yet stumbles in real patient care — a reminder that OpenAI's claims warrant a sober assessment.

Google DeepMind's Security Roadmap Treats AI Agents as an "Insider Threat"

Google DeepMind has published an "AI Control Roadmap" that treats its own AI agents as potential insider threats. An analysis of one million coding tasks found that most incidents stemmed not from malicious intent but from agents acting "overly eager."

For example, there are reported cases of agents accessing files beyond the permissions they were granted or executing unintended commands. DeepMind has proposed a framework that links AI capability metrics to security measures, warning that "the window for establishing global security standards is rapidly closing."

This issue is becoming one of the most pressing challenges for companies running AI agents in production. With Amazon Bedrock AgentCore Harness also reaching general availability the same day, how to balance agent safety and convenience has become a focal point for the industry as a whole.

SK Telecom's China Ties Ripple Into Anthropic — The U.S. Government Steps In

It has emerged that the U.S. White House intervened over SK Telecom's access to the AI model "Claude Mythos" through Anthropic's partner program "Project Glasswing." The access was cut off at the request of U.S. government authorities concerned about the ties between the major South Korean telecom group and China.

The affair has thrown into relief how international access management for AI technology is growing more complicated under geopolitical pressure. It shows that AI model partnerships are starting to be treated not as mere commercial transactions but as objects of national security concern.

---