The Agentic Post
Breaking
Black October 2026: Every Major AI Security Incident, Explained  ·  ARTEX: The Open-Source AI Hacking Tool Behind October’s Biggest Attacks  ·  Chinese Hackers Used AI Agents to Hit 2,250 Microsoft 365 Accounts  ·  PHP2Shell Is Now Being Exploited, and Attacks Began Before the Patch  ·  OpenAI-Backed Thrive Buys ADT for $15.3 Billion to Rebuild It With AI  ·  The AI Data Center Backlash Hits Minnesota and Australia in One Week  ·  
Home/Hardware & Robotics/Data Centers
Liquid Cooling in AI Data Centers

Liquid Cooling in AI Data Centers

Data Centers

An explainer on why liquid cooling has become a near-standard requirement in AI data centers, and how it connects to the broader power constraint on AI capacity.

AP
Staff
September 25, 2026 · 1 min read

Liquid cooling has gone from a specialty option to a near-standard requirement in AI data centers, driven by a simple physical problem: modern chips generate more heat per square foot than air alone can remove fast enough.

Traditional data centers used air cooling because server power density was manageable with fans and airflow alone. High-density AI accelerator racks draw far more power in the same footprint, generating heat air cooling genuinely can’t remove fast enough to keep chips at safe temperatures.

Rather than blowing air across components, liquid cooling circulates coolant directly to or near the chip, since liquid conducts heat away far more efficiently than air per unit of volume. This lets racks pack in far more compute per square foot than air cooling would allow safely.

The newest AI accelerator platforms increasingly ship with liquid cooling built in as a core design requirement, not an optional add-on, since current chip power density essentially requires it. A real shift from just a few years ago, when it was reserved for a small subset of the highest-density deployments.

This connects to the same underlying constraint covered in our explainer on data center energy use: power availability, not chip supply, is the real bottleneck, and cooling is a meaningful part of a facility’s total power draw. See the International Energy Agency’s own data on demand.

Related

Modular Data Centers, Explained
Data Centers

Modular Data Centers, Explained
AI Data Centers and Water Usage
Data Centers

AI Data Centers and Water Usage
Where the Biggest AI Data Centers Are
Data Centers

Where the Biggest AI Data Centers Are
Up Next
Gemini 4 Is in Post-Training, and Google Wants It Out Early

Gemini 4 Is in Post-Training, and Google Wants It Out Early

Big Tech

Google DeepMind chief Koray Kavukcuoglu said Gemini 4 has entered early post-training and that he hopes to ship it much earlier than the end of 2026, after a September in which OpenAI and Anthropic both shipped flagship models.

AP
Staff
September 25, 2026 · 2 min read

Google DeepMind is targeting a Gemini 4 launch before the end of 2026, and possibly well before. Speaking at The Information’s AI Agenda Live Summit, DeepMind chief Koray Kavukcuoglu said the flagship model has entered the early phase of post-training and that he hopes to ship it "much earlier" than year end. No date, pricing, or capability claims have been published.

What does early post-training actually mean?

In plain terms, the expensive part is done. Pre-training is where a model learns from the bulk of its data, and it is the longest and most compute-heavy stage. Post-training is the shaping phase: fine-tuning, reinforcement learning, safety work, and tuning how the model behaves in conversation and tool use.

Saying a model is in early post-training is a real signal that it exists and works, not that it is weeks away. Post-training can run for months, and it is also where most of the safety evaluation happens. The honest read is that Gemini 4 is real and progressing, and that "much earlier than year end" is an aspiration from an executive, not a committed ship date.

Why the timing pressure is obvious

September has been relentless. OpenAI shipped GPT-6 Astra on September 3, then GPT-6 Sol and Luna on September 22 at half the previous price. Anthropic shipped Claude Opus 5.5 the same day, 90 minutes earlier. DeepSeek shipped V4.1 Flash on September 10 with open MIT-licensed weights and cached input at three tenths of a cent per million tokens.

Google’s most recent releases have been in the Gemini 3.x line, including 3.8 Flash and the restricted-access 3.8 Flash Cyber covered in our report on the Fairwind Program. Competitive against mid-tier models, but not a new flagship generation while two rivals shipped one in the same month.

Who is Kavukcuoglu, and why does his word carry weight?

He has run DeepMind’s daily operations since the August 2026 reorganisation in which Demis Hassabis stepped back to become chairman and Alphabet chief scientist, covered in our report on the leadership reshuffle. Kavukcuoglu was DeepMind’s CTO and Google’s chief AI architect for 13 years before that, and reporting at the time indicated he had already been shaping Gemini’s direction for roughly a year. Gemini 4 is the first flagship generation fully under his operational control, which makes its reception a referendum on that change.

What to watch for

Whether Google leads with capability or price. Both of its main rivals just competed on cost per task rather than raw benchmark scores, and Anthropic paired a stronger model with a 40 percent cost reduction. A Gemini 4 that only wins on benchmarks, without a matching efficiency story, would be answering last year’s question. Also worth watching: whether it ships with the kind of restricted-access safeguards Google applied to Flash Cyber, given how the September governance debate has developed.

Reported by The Information from its AI Agenda Live Summit.

Powered by CookieAdmin