Big Tech's AI Data Center Investment: Debt Swells to $350 Billion

According to the LA Times, Big Tech companies have been piling up massive loans to win the AI data center race, with the total already reaching $350 billion (roughly 52 trillion yen). Microsoft, Google, Meta, and Amazon are all rushing to accumulate debt, and AI infrastructure investment has started to take a toll on their financial health.

Relatedly, Bloomberg has published an in-depth analysis of the "circular deals" underpinning the AI boom. It points out that a structure is spreading in which investments in AI startups flow back to Big Tech in the form of cloud service contracts, increasingly blurring the line between investment and revenue.

Meanwhile, Axios reports that Big Tech's transparency is being called into question as the AI boom expands. In particular, disclosure of data on power consumption and water usage remains insufficient, making it difficult to accurately assess environmental impact.

All of these articles show that sustainability and governance challenges are mounting behind the scenes of the AI race, and pressure from regulators and investors is likely to intensify going forward.

From the Community: Layer Expansion of a Gemma Model Succeeds — "extGemma4"

An interesting project called `extGemma4-40_5B` has been announced in Reddit's LocalLLaMA community. It is an attempt to expand capacity by adding layers to a Gemma 4 31B fine-tuned model; after learning from a previous failed attempt, the retry has succeeded.

The key was how the new layers were initialized. The conventional approach was to insert new layers in a "do-nothing" (pass-through) state and then train them, but this had a problem: gradient signals don't propagate, so training doesn't progress. This time, the new layers were initialized with a blend of the outputs of adjacent layers, so that they perform "some kind of processing" from the start and can receive feedback immediately.

They also noted that the choice of where to insert the layers, and leaving the model's internal "volume" settings (normalization parameters) untouched, were important as well.

In validation, the model was compared against the original across 10 fields including physics, biology, history, and philosophy: it matched the original in 9 fields, and in 1 field (relativistic energy calculations in physics) the expanded model actually outperformed the original. As an experiment demonstrating that a fine-tuned model doesn't collapse even when capacity is surgically added, it is drawing plenty of attention in the community.

NVIDIA Shares a Method to Ease HBM Bottlenecks in JAX-based LLM Training

NVIDIA's developer blog has introduced a host offload technique for easing HBM (high-bandwidth memory) bottlenecks in LLM training with JAX.

In large-scale model training, HBM capacity on the GPU tends to become a constraint, and offloading data to host memory can improve training stability and throughput. As a practical optimization technique, it's a useful reference for teams using JAX.

Running Local LLMs on an Intel Core Ultra 285HX — CPU/iGPU Benchmark Results

Also on Reddit's LocalLLaMA, benchmark results have been shared from running local LLMs on a mini PC setup: the MS-02 (Intel Core Ultra 285HX, 64GB RAM).

Comparing Qwen3-30B, Qwen3.6-35B, and Gemma 4 26B across three backends — Vulkan, SYCL, and CPU-only — surprisingly, CPU-only turned out to be the fastest. For example, Qwen3.6-35B (IQ4_XS) ran at 14 tk/s on CPU-only, 12 tk/s on SYCL, and 0.5 tk/s on Vulkan.

The poster points out that "a lot of the information out there said Vulkan is faster than SYCL, but on newer ARC-based iGPUs the order is reversed," and it seems the community is still exploring the optimal settings for Intel's latest iGPUs. The Qwen models deliver perfectly practical speeds on CPU alone, making them a promising option for GPU-less local LLM environments.

---