SoftBank to Invest Up to €75 Billion in France, Targeting 5GW of Data Center Capacity

SoftBank has announced plans to invest up to €75 billion (roughly ¥12 trillion) in data center development in France. The goal is to develop and operate up to 5 gigawatts of additional data center capacity to keep pace with surging AI infrastructure demand. The AI infrastructure race in Europe looks set to heat up even further.

Thaw — Fork a Running LLM Inference Session Like "git branch"

An open-source project called Thaw has been drawing attention on Hacker News. When LLM agents explore multiple branches (RL rollouts, best-of-N, parallel coding, and so on), the same prompt's prefill has traditionally been recomputed for each branch, burning wasted compute.

Thaw snapshots a running inference session (weights, KV cache, scheduler state, and the prefix hash table) so that N child processes can branch from the fork point without any prefill. Think of it as "git branch" for LLMs.

In measured results on an H100 80GB running Llama-3.1-8B, bringing up a warm pool takes 22.3 seconds, after which fork rounds of 4 branches × 64 tokens run with a median of 0.88 seconds — reportedly a roughly 400x speedup compared to a cold boot.

It supports vLLM and SGLang and is released under the Apache-2.0 license.

Is a $6,400 Local LLM Server Actually Worth It Compared to APIs?

A detailed cost analysis of a local LLM server built for around $6,400 (about ¥950,000) has been making the rounds in Reddit's LocalLLaMA community.

Hardware Configuration

× used AMD MI100 32GB: $4,234

ASRock EPYCD8-2T motherboard: $721

W 80+ Platinum PSU: $497

Miscellaneous (CPU, RAM, case, etc.)

The server runs 4 concurrent instances of Qwen3.6 27B.

Daily Throughput

Input tokens: 20.4M/day

Output tokens: 1.32M/day

Cost Comparison (Annual)

| Item | Annual Cost |

|---|---|

| API (OpenRouter equivalent) | $3,701 |

| Local, year 1 (electricity + depreciation) | $2,992 |

| Local, year 2 onward (electricity only) | $770 |

That works out to $708 in savings versus the API in the first year, and more than $2,930 per year from year two onward. That said, the poster notes that the MI100 is a rare GPU with immature optimization, and points out that unified-memory setups like the R9700 or Apple Silicon are easier to work with.

RTX 5060 Ti / 5070 Ti Clearance Prices Surface

Also on Reddit, reports are circulating that Best Buy's in-store-only clearance has the RTX 5060 Ti 16GB available for $300.99 and the RTX 5070 Ti 16GB for $699.99. At steep discounts from regular prices, these are price points worth watching for anyone considering building a local AI setup. Prices are not the same at every store, so you'll need to check your local store for actual pricing.

---

**Sources:**

[SoftBank says it will invest up to €75 billion to build French data centers — TechCrunch](https://techcrunch.com/2026/05/30/softbank-says-it-will-invest-up-to-e75-billion-to-build-french-data-centers/)

[Show HN: Thaw – Git branch for a running LLM — Hacker News](https://github.com/thaw-ai/thaw)

[Cost Analysis of my $6.4k Local LLM Server — Reddit LocalLLaMA](https://www.reddit.com/r/LocalLLaMA/comments/1tsbl9j/cost_analysis_of_my_64k_local_llm_server/)

[PSA: 5060ti 16GB for $300.99 / 5070ti 16GB for $699.99 — Reddit LocalLLaMA](https://www.reddit.com/r/LocalLLaMA/comments/1tse423/psa_5060ti_16gb_for_30099_5070ti_16gb_for_69999/)