SoftBank's €75B French DC bet; GitHub Copilot's token-billing revolt

SoftBank said it will commit up to €75B to French data centers — the largest single European AI-infrastructure pledge yet, and a clear pricing in of the EU's sovereign-compute push. GitHub shifted Copilot to token-based billing and triggered a sharp developer backlash, the first taste of how the same metering shift that's coming to every coding agent will be received. Consumer AI hardware kept multiplying: Meta is reportedly developing an AI pendant, TechCrunch published a positive long-form on Gemini Spark, and the browser wars in 2026 are now openly an AI-agent fight. The research feed is thin — a Saturday signature — but the constrained-decoding unification paper and a sharp probe of whether VLMs actually do spatial reasoning are both worth a look.

7 papers 5 news 3 sources ← Latest

News

5 items

SoftBank bets €75B on French AI infrastructure

SoftBank pledged up to €75B to build French data centers — by a wide margin the largest single European AI-infra commitment. The bet lines up with the EU's sovereign-compute push and reframes Europe as a viable third pole of the global AI build-out, not just a regulator.

News TechCrunch AI

SoftBank says it will invest up to €75 billion to build French data centers

SoftBank pledged up to €75B for French AI data centers — the largest single European AI-infra commitment to date.

commitment up to €75B

Why it matters

Largest European AI-infra commitment from a single backer; reframes Europe as a real third pole of the build-out.
Pairs with the EU's sovereign-compute push and the Pope's encyclical to make Europe the most active non-US AI policy theater this quarter.
Concentrates AI capacity in France specifically — meaningful for data-residency-sensitive SMB buyers.

Source →

Developer-tool billing flips, devs notice

GitHub moved Copilot to token-based billing and the developer reaction was loud. Combined with TechCrunch's roundup of 2026's browser alternatives — now openly an AI-agent fight — the developer-facing AI surface is being repriced and rebundled in the same week.

News TechCrunch AI

'What a joke': GitHub Copilot's new token-based billing spurs consternation among devs

GitHub's switch to token-based Copilot billing landed badly with the developer base — the unit-economics conversation hits a $X/month subscription comfort zone.

Why it matters

First mass-market metered-pricing shock for AI coding tools; sets expectations for the rest of the category.
Practitioners who paid flat now see line items per request — and dislike it visibly.
Anthropic's per-seat Claude Code pricing model gets a comparative advantage at the SMB end.

market products code

Source →

News TechCrunch AI

As the browser wars heat up, here are the hottest alternatives to Chrome and Safari in 2026

TechCrunch surveys the 2026 browser landscape — AI-agent integration is the differentiator now, not just speed or privacy.

products agents

Source →

Consumer AI hardware keeps multiplying

Meta is reportedly working on an AI pendant — another wearable-agent surface. TechCrunch's positive long-form on Gemini Spark gives Google's agent the cleanest review yet. The consumer-AI-hardware category is no longer experimental, even where individual products feel awkward.

News TechCrunch AI

Meta is reportedly developing an AI pendant

Meta is reportedly building an AI pendant — adding another wearable form factor to its Ray-Ban-plus-agents stack.

products agents

Source →

News TechCrunch AI

I put Google's 24/7 AI assistant Gemini Spark to work, and it's actually pretty useful

TechCrunch hands-on with Gemini Spark — a positive review of Google's always-on AI assistant.

products agents

Source →

Papers

3 items

Spatial-reasoning shortcuts and unified decoding

A sharp paper probes whether VLMs actually do 3D spatial reasoning or just exploit statistical shortcuts in natural images; another reconciles free-form and constrained decoding into a single framework that doesn't suppress reasoning early.

Paper Hugging Face

Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

Probes whether VLMs' spatial-reasoning scores reflect real 3D understanding or natural-image statistical shortcuts — using minimal contrastive pairs.

vision-language evaluation interpretability

Source → Arc

Paper Hugging Face

Thinking Before Constraining: A Unified Decoding Framework for Large Language Models

Unifies free-form and constrained decoding so models can reason before format constraints kick in — addresses the structured-output reasoning tax.

Why it matters

Direct fix for the 'JSON mode kills my reasoning' problem in agent toolchains.
Plug-in for any pipeline producing structured outputs from chain-of-thought models.

reasoning tools inference

Source → Arc

Paper Hugging Face

Reflective Prompt Tuning through Language Model Function-Calling

Automates prompt optimization through reflective function-calling — addressing how brittle hand-tuned prompts remain.

fine-tuning tools

Source → Arc

Also today

Paper · Hugging Face Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation — Rethinks how uniform discrete diffusion is parameterized via a leave-one-out denoiser and absorbing-state reformulation.
Paper · Hugging Face Towards Consistent Video Geometry Estimation — Pushes video geometry estimation toward temporally consistent 3D — useful for downstream world-model and robotics work.
Paper · Hugging Face Geometry Matters: 3D Foundation Priors for Learning Semantic Correspondence — Uses 3D foundation priors to learn semantic correspondence — better generalization on unseen object categories.
Paper · Hugging Face Convex Low-resource Accent-Robust Language Detection in Speech Recognition — Accent-robust language detection for ASR pipelines in low-resource settings.