Evanesca vs Groq: raw speed, or models run in Europe

At a glance

Each Groq cell links to the Groq page it comes from, read on 6 October 2026; each Evanesca cell links to the page that describes it.

Speed and price

Groq designs its own accelerators, the LPU, and runs models on racks of 256 of them source. Its model list gives a speed for each model: about 500 tokens per second for GPT-OSS 120B and 1,000 for GPT-OSS 20B source. Evanesca makes no speed or latency promise; if the speed of each answer is what you buy, Groq is built for it.

Both serve OpenAI's open-weight GPT-OSS 120B and Qwen3.8 27B, which Groq offers as a preview “for evaluation purposes only”; on Evanesca, Scaleway runs both in Paris, and OVHcloud takes over in Gravelines if Scaleway fails. GPT-OSS 20B, which Groq also lists, is paused on Evanesca for now, so it is left out. Groq lists Llama 3.3 70B too, but for “Contact Sales”, with no public price, so it is left out here source.

Groq publishes its prices in US dollars, before tax; Evanesca's are in euros, margin and 20% VAT included. So the verdict compares both before tax: Evanesca's price divided by 1.2, and the dollar converted at the ECB reference rate of 6 October 2026, 1 EUR = 1.1269 USD (ECB). A gap under 2% names no winner.

Groq is cheaper on input and on output for every model in the table: on GPT-OSS 120B, Groq costs 37% less on both input and output; on Qwen3.8 27B, Groq costs 15% less on input and 23% less on output. Price is one difference among others: speed, data and law separate the two just as much.

What happens to your prompts

« By default, Groq does not retain customer data for inference requests. »

Groq makes two exceptions: features that need storage to work, such as batch jobs and fine-tuning, and logs kept “to troubleshoot system failures or investigate abuse”, for up to 30 days. Any customer can switch those logs off with zero data retention. Usage metadata is always kept, without inputs or outputs source. Its services agreement adds that Groq may not use inputs or outputs for training unless you grant permission source.

Evanesca stores no prompt and no answer: they are held in memory while the answer is written, your history stays in your browser, and only billing data — tokens and cost — is kept. The operator side is stated, not switchable: OVHcloud, which runs both GPT-OSS models, keeps “only the data required for billing purposes” source; Scaleway, which runs other models on Evanesca, may keep and read requests that cause errors or look abusive, for up to 14 days source; Mistral AI keeps its models' inputs and outputs 30 days to monitor abuse source. Each answer's passport names its operator.

Contract, data location and the CLOUD Act

A customer in the EEA or Switzerland signs with Groq UK Limited in London, under English law and the courts of London; everywhere else, the contracting party is Groq LLC, in Mountain View, California source. Groq lists data centres on four continents, including Vantaa in Finland and London source.

The contract's law and the data's location are separate questions. Groq's own page puts retained customer data — the 30-day logs, batch files, fine-tuning sets — in Google Cloud buckets in the United States, with standard contractual clauses for transfers source. In the US, the CLOUD Act lets authorities require a provider under US jurisdiction to disclose data in its “possession, custody, or control” source.

On Evanesca, every model is run by a European company: OVHcloud, Scaleway, Mistral AI or Infomaniak, with no US parent, so a US order cannot reach those requests directly — or AKI.IO, a German company whose shareholders are not published. Evanesca keeps none of their content to hand over. The sovereignty page sets out who controls each operator.

Where Groq is the better choice

• Speed. Groq publishes the throughput of each model and builds its own hardware for it source.
• A free plan to try models before paying, within rate limits source.
• Zero data retention open to every customer, which also switches off the 30-day abuse logs source.
• More than chat: speech-to-text and text-to-speech models (models), batch and flex processing on paid plans (plans).
• It runs today, from data centres on four continents; Evanesca has not opened yet.

For real-time voice agents, interactive tools and anything where seconds matter, Groq's published figures are an argument Evanesca cannot match.

Where Evanesca fits better

• Models run in Europe, by European operators with no US parent, by default and on every request sovereignty.
• No prompt stored by Evanesca: of your messages, only billing data — tokens and cost — is kept privacy.
• 17 models, all run in Europe, all with public prices, from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google models.
• Prices in euros with VAT included, no subscription, and credit that never expires pricing.
• A chat with a passport under every answer, and a signed Art. 28 DPA for Business workspaces business.
• The same models in an Agent mode, for colleagues who don't write code: it searches the web, writes documents and creates images Agent mode.

The API is OpenAI-compatible on both sides; Evanesca's covers chat completions, the model list and embeddings docs.

How to choose

• Latency first — voice, autocomplete, live agents: Groq.
• Personal, client or regulated data that you want processed in Europe, with no copy kept by the provider: the models Evanesca runs in Europe.
• A prototype on a budget: Groq's free plan, or €2 of credit on Evanesca, whichever models you need.

Because both use the OpenAI request format, you can route by data type in your own code: sensitive traffic to Evanesca, speed-critical traffic to Groq.

How this page was checked

Groq's pages were read on 6 October 2026. Evanesca's prices come from its catalogue and every verdict above is computed from them, with the dollar at the ECB reference rate of that day. If something has changed, tell us.

• Groq — supported models, prices and speeds
• Groq — rate limits, free plan
• Groq — your data
• Groq — services agreement, last modified 22 June 2026
• Groq — platform and data centres

Frequently asked questions