# No Hype: Phonon, 75 MB of local speech that wins in Spanish

> Phonon brings local speech with a commercial licence, NVIDIA admits memory per GPU barely grows, and Europe protests against data centres.

- Published: 2026-11-06
- Author: Jacobo González Jaspe
- Language: en
- Tags: news, opinion, no-hype, ai
- Canonical URL: https://vorluxai.com/en/blog/sin-humo-2026-11-06-en/
- Source: VORLUX AI (J4SGON S.L.), https://vorluxai.com/

---

Hugging Face opens Phonon, a synthetic voice that fits in 75 MB and runs on your own machine. NVIDIA explains why memory rules, Meta fears the ballot box and Europe takes to the streets over data centres.

## Phonon: local speech that wins in Spanish

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-06-1.webp)
*Illustration generated with AI on our own machine.*


**What happened.** Two weeks ago Hugging Face opened Phonon, its small text-to-speech model that runs on your own device ([Hugging Face](https://huggingface.co/blog/phonon-tts)). It comes in two sizes, about 65 MB and 75 MB. Phonon 31m is Apache 2.0, fine for commercial use; the other models are CC BY-NC. It speaks English, French, German, Spanish and Portuguese, with more than 100 built-in voices. It was downloaded more than 200,000 times in its first two weeks.

**What we think.** The news for us is the licence, not the size: only the 31m goes into a product you charge for, and the 40m stays non-commercial ([Hugging Face](https://huggingface.co/blog/phonon-tts)). On intelligibility, Phonon ranks first in German and Spanish on the Open TTS Leaderboard. That measures whether it is understood, not whether it sounds natural; listen before you put it in front of a client.

**The good bit.** Its inference framework, ptts, supports more than 10 backends across CPU, GPU and NPU ([Hugging Face](https://huggingface.co/blog/phonon-tts)).

**What you can do.** Run Phonon 31m on your machine with your own business text. If you build product, start with [local AI for developers](https://vorluxai.com/en/developers/).

## NVIDIA: memory per GPU barely grows

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-06-2.webp)
*Illustration generated with AI on our own machine.*


**What happened.** From Blackwell Ultra to Vera Rubin, NVIDIA's HBM capacity per GPU rose roughly 3% ([NVIDIA](https://blogs.nvidia.com/blog/memory-full-stack-codesign/)). On the AgentX DeepSeek V4 Pro workload, GB300 NVL72 delivers up to 10x higher throughput per dollar than the H200 platform at 100 tokens per second per user, with HBM per GPU roughly doubling. Agentic trajectories can process up to 3 million tokens in some cases.

**What we think.** The vendor itself says memory per GPU barely moves between generations; the gain comes from designing hardware and software together ([NVIDIA](https://blogs.nvidia.com/blog/memory-full-stack-codesign/)). Your box has the same problem at a smaller scale: long context eats memory before the model does. We do the sum before we buy.

**The good bit.** Long-context agents now have a reference figure: up to 3 million tokens per trajectory ([NVIDIA](https://blogs.nvidia.com/blog/memory-full-stack-codesign/)).

**What you can do.** Measure what your model needs at your real context length in the [VRAM calculator](https://vorluxai.com/en/tools/vram-explorer/).

## Meta fears the midterms will slow its data centres

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-06-3.webp)
*Illustration generated with AI on our own machine.*


**What happened.** Meta president Dina Powell McCormick said the Democratic wins in the midterm elections will make it a lot harder to build data centres ([TechCrunch](https://techcrunch.com/2026/11/05/metas-warns-that-election-backlash-against-data-centers-could-make-ai-expansion-harder/)). Meta has data centres in 23 states, and five of them are planned as projects of 1 gigawatt or more. She said the company has been building them for 20 years. Its consumer AI agent Muse has 10 million users today.

**What we think.** When a company with 20 years of building data centres says the next one gets harder, local opposition is already in the budget ([TechCrunch](https://techcrunch.com/2026/11/05/metas-warns-that-election-backlash-against-data-centers-could-make-ai-expansion-harder/)). Your sum is smaller: a server in your office needs nobody's planning permission.

**The good bit.** Meta says it out loud instead of hiding it ([TechCrunch](https://techcrunch.com/2026/11/05/metas-warns-that-election-backlash-against-data-centers-could-make-ai-expansion-harder/)).

**What you can do.** If your plan leans on someone else's capacity, read what the [EU AI Act guide](https://vorluxai.com/en/eu-ai-act/) asks of you and what you could bring in-house.

## Europe protests against data centres

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-06-4.webp)
*Illustration generated with AI on our own machine.*


**What happened.** Protesters in 15 European countries are planning two days of action against data centres this week ([The Guardian](https://www.theguardian.com/world/2026/nov/05/activists-across-europe-protests-ai-against-datacentres)). More than 100 groups are joining the "Stop Dirty Data Centres" events on Friday and Saturday. The Öko-Institut of Germany says AI data centre CO2 emissions will be six times the 2023 level by 2030. In the UK, 315 data centres are queueing for a grid connection: 73 gigawatts of demand against a national peak of 45GW.

**What we think.** The UK queue asks for more power than the whole country's peak; that figure carries more weight than any banner ([The Guardian](https://www.theguardian.com/world/2026/nov/05/activists-across-europe-protests-ai-against-datacentres)). A small model on a machine you already own never joins that queue.

**The good bit.** The grid-queue figure is public and easy to compare ([The Guardian](https://www.theguardian.com/world/2026/nov/05/activists-across-europe-protests-ai-against-datacentres)).

**What you can do.** See where Spain's market stands in [Spain's AI market in 2026: the real figures and where to start](https://vorluxai.com/en/blog/spain-ai-market-opportunity-2026-en/).

## Meta pauses its new AI terms

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-06-5.webp)
*Illustration generated with AI on our own machine.*


**What happened.** Meta announced its updated terms are paused and will not take effect on November 20, 2026 ([404 Media](https://www.404media.co/meta-generative-ai-commercial-terms-on-pause/)). They were first set for October 30. A Norwegian court granted a temporary injunction after a case backed by the Norwegian Media Businesses' Association and 180 Norwegian media companies. Nieman Lab reported that Meta removed the word "train" from a list of allowed use cases.

**What we think.** Deleting "train" from a list says nothing about what happens to the data; the pause is real, and a court won it ([404 Media](https://www.404media.co/meta-generative-ai-commercial-terms-on-pause/)). If your text matters, keep it where you decide.

**The good bit.** A court and 180 media companies moved the date ([404 Media](https://www.404media.co/meta-generative-ai-commercial-terms-on-pause/)).

**What you can do.** List which of your data ends up in other people's services, and check how [Ollama vs vLLM for 1 to 10 users](https://vorluxai.com/en/blog/ollama-vs-vllm-small-office-measured-en/) runs the same work in-house.

## In one line
Phonon 31m is the local voice you can ship in a product; the rest of the family is non-commercial only.

## Work with us

If something in this column belongs in your business, on models that run on your own hardware rather than someone else's cloud, [tell us what you need](https://vorluxai.com/en/contact/) or see how we work in [consulting](https://vorluxai.com/en/consulting/).
