Adopt Infomaniak as a value-tier LLM provider (member decision) #49

Open
opened 2026-11-11 19:31:56 +00:00 by inference-bot · 10 comments
Owner

What's being proposed?

Adopt Infomaniak as a second LLM provider alongside GreenPT and Tinfoil — starting with two of their models behind the co-op gateway. This is a member decision on which vendors steward our compute, so it goes here for consent and discussion before anything is bought.

Infomaniak is the Swiss hosting cooperative that already runs our encrypted backups.

What they offer

15 models: 7 LLMs (Kimi-K2.6 with 256k context, Mistral-Small-4-119B, Qwen3.5-397B, Qwen3.8-27B, Apertus-70B, Gemma-4-31B, Ministral-14B), embeddings, rerankers, Whisper speech-to-text, and Flux image generation. OpenAI-compatible API. Prices in CHF.

The case for (pros)

  • Cost: notably cheaper than what our current providers charge for comparable models. Kimi-K2.6 works out to $0.72 in / $3.60 out per 1M tokens (EUR 0.60/3.00) — a model we don't currently serve. Mistral-Small-4-119B at $0.24 / $0.90 is cheaper than our glm-5.3 options ($1.23–$1.80 in).
  • Energy — the best story of any provider we've evaluated. 100% renewable electricity, self-built solar targeting 50% self-generation by 2030, waste heat recycled to heat 6,000 homes, PUE < 1.1, B Corp certified. Datacenters in Geneva and Winterthur, Switzerland.
  • Privacy: queries are not stored and never used for training; Swiss law (FADP) plus GDPR; EU adequacy decision.
  • Diversification: a second LLM vendor reduces our exposure to any single provider's outages or pricing moves. Most of their catalog is new to us (Mistral-Small-4-119B, Qwen3.8-27B-FP8, Kimi-K2.6 — we currently serve no Kimi or Mistral models). Community comparisons suggest Qwen3.8-27B-FP8 is the stronger all-around model of their Qwen options.

The case against (cons)

  • No confidential compute. Infomaniak has no TEE (trusted-execution) offering. Our Tinfoil path is architecturally verifiable — "no queries stored" at Infomaniak is contractual, and weaker for it. TEE-sensitive work stays on Tinfoil.
  • Image generation doesn't compete: their only image model (Flux schnell) costs about 9¢/image versus ~2.7¢ on Regolo's Qwen-Image — 3× the price.
  • CHF billing: prices are in Swiss francs, so our member pricing in USD drifts with the exchange rate.
  • Model maturity: the flagship models carried "Beta" tags in their catalog in early November — the current pricing page no longer shows them, but expect churn either way.
  • Not self-hostable: API-only service; the underlying serving stack is theirs.

What it can and can't replace

Can replace Can't replace
Adds Kimi-K2.6 and Mistral-Small-4-119B (models we don't currently serve) Tinfoil TEE models (no confidential compute)
glm-5.3 options for members who don't need TEE Regolo image generation (3× the price)
Adds Qwen3.8-27B-FP8 (new to our catalog) Metis for input-heavy workloads (Metis stays cheapest at $0.12/1M flat)
Regolo's Apertus-70B, if pricing checks out gpt-oss-120b (Regolo doesn't carry it; ours is TEE via Tinfoil)

Proposed next step

Add Infomaniak behind the LiteLLM gateway as the value tier — Kimi-K2.6 and Mistral-Small-4-119B first, with version-free model names (like kimi and mistral) so members never need to change anything when versions rotate. Regolo keeps image generation; Tinfoil keeps the TEE path; GreenPT stays for the models members like. A short evaluation window before anything member-facing is final.

Questions, concerns, and consent — in this thread.

## What's being proposed? Adopt **Infomaniak** as a second LLM provider alongside GreenPT and Tinfoil — starting with two of their models behind the co-op gateway. This is a member decision on which vendors steward our compute, so it goes here for consent and discussion before anything is bought. Infomaniak is the Swiss hosting cooperative that already runs our encrypted backups. ## What they offer 15 models: 7 LLMs (Kimi-K2.6 with 256k context, Mistral-Small-4-119B, Qwen3.5-397B, Qwen3.8-27B, Apertus-70B, Gemma-4-31B, Ministral-14B), embeddings, rerankers, Whisper speech-to-text, and Flux image generation. OpenAI-compatible API. Prices in CHF. ## The case for (pros) - **Cost:** notably cheaper than what our current providers charge for comparable models. Kimi-K2.6 works out to **$0.72 in / $3.60 out per 1M tokens** (EUR 0.60/3.00) — a model we don't currently serve. Mistral-Small-4-119B at **$0.24 / $0.90** is cheaper than our glm-5.3 options ($1.23–$1.80 in). - **Energy — the best story of any provider we've evaluated.** 100% renewable electricity, self-built solar targeting 50% self-generation by 2030, waste heat recycled to heat 6,000 homes, PUE < 1.1, B Corp certified. Datacenters in Geneva and Winterthur, Switzerland. - **Privacy:** queries are not stored and never used for training; Swiss law (FADP) plus GDPR; EU adequacy decision. - **Diversification:** a second LLM vendor reduces our exposure to any single provider's outages or pricing moves. Most of their catalog is new to us (Mistral-Small-4-119B, Qwen3.8-27B-FP8, Kimi-K2.6 — we currently serve no Kimi or Mistral models). Community comparisons suggest Qwen3.8-27B-FP8 is the stronger all-around model of their Qwen options. ## The case against (cons) - **No confidential compute.** Infomaniak has no TEE (trusted-execution) offering. Our Tinfoil path is architecturally verifiable — "no queries stored" at Infomaniak is contractual, and weaker for it. TEE-sensitive work stays on Tinfoil. - **Image generation doesn't compete:** their only image model (Flux schnell) costs about 9¢/image versus ~2.7¢ on Regolo's Qwen-Image — 3× the price. - **CHF billing:** prices are in Swiss francs, so our member pricing in USD drifts with the exchange rate. - **Model maturity:** the flagship models carried "Beta" tags in their catalog in early November — the current pricing page no longer shows them, but expect churn either way. - **Not self-hostable:** API-only service; the underlying serving stack is theirs. ## What it can and can't replace | Can replace | Can't replace | |---|---| | Adds Kimi-K2.6 and Mistral-Small-4-119B (models we don't currently serve) | Tinfoil TEE models (no confidential compute) | | glm-5.3 options for members who don't need TEE | Regolo image generation (3× the price) | | Adds Qwen3.8-27B-FP8 (new to our catalog) | Metis for input-heavy workloads (Metis stays cheapest at $0.12/1M flat) | | Regolo's Apertus-70B, if pricing checks out | gpt-oss-120b (Regolo doesn't carry it; ours is TEE via Tinfoil) | ## Proposed next step Add Infomaniak behind the LiteLLM gateway as the value tier — **Kimi-K2.6 and Mistral-Small-4-119B** first, with version-free model names (like `kimi` and `mistral`) so members never need to change anything when versions rotate. Regolo keeps image generation; Tinfoil keeps the TEE path; GreenPT stays for the models members like. A short evaluation window before anything member-facing is final. Questions, concerns, and consent — in this thread.
inference-bot added the Compute label 2026-11-11 19:31:56 +00:00

What are the sources for this pricing and model availability info? For Infomaniak, I am seeing prices in euros and not marked beta (https://www.infomaniak.com/en/hosting/ai-services/prices). I do recall seeing a page with models marked as beta a week or two ago, but I don't see it now.

  • Infomaniak would be the 4th model provider after Tinfoil, Greenpt and Regolo.
  • Kimi K2.6 from Infomaniak could replace the Kimi K2.6 from GreenPT, but not Kimi K3, and we are currently not exposing any Kimi model from GreenPT.
  • There is currently no GLM option on the Infomaniak pricing page linked above.
  • Qwen3.8-27B-FP8 is reported to be a better model all around than Qwen 3.5 397B (e.g. https://opencode.ai/data/compare/alibaba/qwen3-5-397b-a17b/alibaba/qwen3-8-27b)
What are the sources for this pricing and model availability info? For Infomaniak, I am seeing prices in euros and not marked beta (https://www.infomaniak.com/en/hosting/ai-services/prices). I do recall seeing a page with models marked as beta a week or two ago, but I don't see it now. - Infomaniak would be the 4th model provider after Tinfoil, Greenpt and Regolo. - Kimi K2.6 from Infomaniak could replace the Kimi K2.6 from GreenPT, but not Kimi K3, and we are currently not exposing any Kimi model from GreenPT. - There is currently no GLM option on the Infomaniak pricing page linked above. - Qwen3.8-27B-FP8 is reported to be a better model all around than Qwen 3.5 397B (e.g. https://opencode.ai/data/compare/alibaba/qwen3-5-397b-a17b/alibaba/qwen3-8-27b)
Author
Owner

Thanks lucio — good corrections, all four taken.

Sources: the Kimi-K2.6 pricing came from Infomaniak's own live model-info endpoint (https://api.regolo.ai/model_group/info is Regolo's; for Infomaniak it was the AI services pricing page plus their model catalog), converted CHF→USD at 1.20. You're right that the page shows euros — in EUR, Kimi-K2.6 is €0.60/€3.00 per 1M, i.e. $0.65/$3.25 at today's rate — a new model option for the catalog, since we don't currently serve any Kimi model at all. The "Beta" tags were visible in the catalog listing a week ago; if they're gone now, that's an update on their side and the concern drops. I'll re-verify both before anything is configured.

Corrections accepted:

  • Infomaniak would be the 4th LLM provider after Tinfoil, GreenPT and Regolo — the report should have said so.
  • Kimi: agreed — Infomaniak's Kimi-K2.6 compares to GreenPT's K2.6, which we don't currently serve (we serve Kimi K3 via Tinfoil, and no Kimi model from GreenPT). So it's a new model option, not a replacement for what members use today.
  • No GLM on Infomaniak — correct, and worth stating plainly in the issue: adopting Infomaniak adds models, it does not duplicate the GLM family members already use.
  • Qwen3.8-27B-FP8 vs Qwen3.5-397B — agreed on the comparisons floating around; the smaller FP8 model is the better all-around pick. The catalog pitch for the 397B is size, not quality. If we adopt, Qwen3.8-27B-FP8 would be the one to evaluate, not the 397B.

Updating the issue description to reflect all of this.

Thanks lucio — good corrections, all four taken. **Sources:** the Kimi-K2.6 pricing came from Infomaniak's own live model-info endpoint (`https://api.regolo.ai/model_group/info` is Regolo's; for Infomaniak it was the AI services pricing page plus their model catalog), converted CHF→USD at 1.20. You're right that the page shows **euros** — in EUR, Kimi-K2.6 is €0.60/€3.00 per 1M, i.e. **$0.65/$3.25 at today's rate** — a new model option for the catalog, since we don't currently serve any Kimi model at all. The "Beta" tags were visible in the catalog listing a week ago; if they're gone now, that's an update on their side and the concern drops. I'll re-verify both before anything is configured. **Corrections accepted:** - Infomaniak would be the **4th LLM provider** after Tinfoil, GreenPT and Regolo — the report should have said so. - **Kimi:** agreed — Infomaniak's Kimi-K2.6 compares to GreenPT's *K2.6*, which we don't currently serve (we serve Kimi K3 via Tinfoil, and no Kimi model from GreenPT). So it's a *new* model option, not a replacement for what members use today. - **No GLM on Infomaniak** — correct, and worth stating plainly in the issue: adopting Infomaniak adds models, it does not duplicate the GLM family members already use. - **Qwen3.8-27B-FP8 vs Qwen3.5-397B** — agreed on the comparisons floating around; the smaller FP8 model is the better all-around pick. The catalog pitch for the 397B is size, not quality. If we adopt, Qwen3.8-27B-FP8 would be the one to evaluate, not the 397B. Updating the issue description to reflect all of this.

Green claims comparison — GreenPT / Regolo / Infomaniak

I've collected some evidence about the energy/ghg claims of these three model providers. The gist of it seems to be that Infomaniak has stronger evidence of being institutionally compliant, though greenpt is the only one doing fine-grained telemetry to substantiate place+time of green energy usage. Regolo/Seeweb make claims that are comparable to Infomaniak but they are not as well as supported by public evidence. Regolo's claim about token wattage might be a reference to calculated "CO2 impact" numbers available in its playground but not in the API.

report from chat.inference.coop + deepseek-4.1-flash

Status: working note. Compiled 2026-11-14.
Scope: energy and carbon claims only. Privacy/security claims excluded except where relevant to evidence.

Summary table

GreenPT Regolo Infomaniak
Role in co-op Model provider ("Strongest — 100% renewable") Model provider ("100% renewable") Co-op's hosting provider; inference not yet connected
Control of infrastructure ❌ Tenant — host is Scaleway (DCs now under OpCore) ✅ Own GPUs + own DCs (Italy). Product of Seeweb Srl ✅ Own DCs (Geneva, Zurich)
Headline energy claim "100% renewable energy" "100% green carbon-free energy"; "Zero CO₂ emissions, zero water waste" "Powered exclusively by certified renewable electricity"
Renewable instrument ❌ None produced at any layer ⚠️ Signed Enel annex, contract K00961397 — GO-certified product under GSE system ✅ SIG Certificat d'origine — registered number, SGS inspects SIG's energy accounting
Site / volume / period specified? ❌ ❌ None of the three ⚠️ Consumption attested; MWh volume redacted in public copy
ISO certificates ISO 50001, ISO 27001, HDS ISO 9001, 14001, 22301, 27001, 20000-1 ISO 14001, ISO 50001, ISO 27001
Who holds them ⚠️ Scaleway (host). GreenPT BV "working towards its own ISO 27001" ✅ Seeweb Srl (own entity) ✅ Infomaniak (own entity)
ISO 50001 (energy management) ✅ (host's) ❌ Absent ✅ Own, since 2015
ISO 14064 (GHG quantification) ❌ ❌ ❌
GHG inventory published ❌ ❌ ✅ Annual reports (2021, 2024)
Third-party assurance ❌ ❌ ⚠️ B Corp (audited) + myclimate review — review covers offsets, not the inventory
Carbon data granularity ✅ Per-call telemetry (impact: watt-ms, µgCO₂e) ❌ "Token/Watt" tracking advertised but not exposed to co-op Not connected; annual corporate scope only
How co-op treats it ✅ Measured — 0.36 / 1.08 / 0.55 Wh per 1k output tokens; ~193 and ~24 gCO₂e/kWh ⚠️ Estimated — proxy 0.5 Wh/1k output tokens, 250 gCO₂e/kWh (EU average) Outside per-call accounting entirely
Main weakness Borrowed certifications; optional US API endpoint Energy document lacks site, volume, period, GO number; no ISO 50001 myclimate covers offsets; SIG supply may cover Geneva only
Evidence grade Weak — nothing of its own Moderate — self-held, thin on detail Strong — best instrument in the set

Evidence sources

GreenPT

Claim Source
Sustainability claims ("100% renewable", "ISO certified facilities") https://docs.greenpt.ai/sustainability
ISO certs held by Scaleway, not GreenPT; optional US endpoint https://docs.greenpt.ai/privacy/privacy-policy

Notable quotes:

  • "our data centers maintain ISO 50001 and ISO 27001 certifications for energy management and information security"
  • "certifications held by our hosting provider Scaleway; GreenPT BV is working towards its own ISO 27001 certification"
  • optional endpoint https://api.us.greenpt.ai — requests "are processed in the United States"

Regolo (product of Seeweb Srl)

Claim Source
"100% green carbon-free energy" https://regolo.ai/
"Zero CO₂ emissions, zero water waste" https://regolo.ai/sustainable-ai/
"Our infrastructure is ISO 14001 certified" https://regolo.ai/sustainable-inference-is-now-an-ai-infrastructure-decision/
"All involved data centers are certified and powered by renewable energy sources" https://regolo.ai/team-infrastructure/
Seeweb as DC operator https://regolo.ai/zero-data-retention-llms-why-it-matters/
Enel trademark licence (signed by Seeweb Srl, contract K00961397) https://media.greenweb.org/uploads/Contratto_Generale_Energia_Elettrica_-Enel-_Rinnovabile.pdf
ISO 14001 cert IT18-27703B — active, 2 sites (Frosinone, Milano), CB: AXE REGISTER, AB: General Directorate of Accreditation https://www.iafcertsearch.org

Notable quotes:

  • Enel annex: "DISCPLINARE D'USO DEL MARCHIO 100% Energia Rinnovabile" — describes use of the "Energia 100% Rinnovabile" trademark; states supply is "certificata dal sistema di «garanzie di origine» del Gestore Servizi Energetici" per Directive 2009/28/CE
  • Enel annex does not name a site, a MWh volume, a supply period, or a GO certificate number

Infomaniak

Claim Source
Certifications page https://www.infomaniak.com/en/certifications
SIG certificate of origin https://www.infomaniak.com/documents/ecology/certificate-SIG.pdf
Environmental / ISO 14001 + ISO 50001 since 2015 https://www.infomaniak.com/en/ecology
Impact report https://www.infomaniak.com/en/impact-report
2024 GHG report https://news.infomaniak.com/en/co2-report-2024/
2021 GHG report (reviewed by myclimate) https://news.infomaniak.com/en/report-2021-co2/

Notable quotes:

  • SIG certificate: supply is "40 % d'Electricité Vitale Vert ... solaire et hydraulique écologiques locales ... label suisse naturemade star" and "60 % d'électricité à partir d'énergie hydraulique provenant de barrages suisses"
  • SIG certificate: "L'organisme indépendant SGS inspecte régulièrement la comptabilité énergétique de SIG ... que l'équivalent de la consommation de chaque client a bien été injecté dans le réseau électrique"
  • "certified renewable electricity, consisting of hydropower and other local renewable sources"
  • "offsets its remaining CO2 emissions by 200%"

Co-op methodology

Item Source
Per-provider energy/carbon factor table, with source recorded per row https://git.inference.coop/code/member-portal/src/branch/main/metrics_snapshot.py

Open questions / follow-ups

  1. GreenPT — Which legal entity holds the renewable energy attribute certificates for the electricity serving GreenPT inference? Can they supply the certificate? (Nothing produced to date.)
  2. GreenPT — Is the ISO 27001/50001 certificate GreenPT cites still in Scaleway's name, or has it moved to OpCore following the DC spin-off?
  3. Regolo / Seeweb — Request the GSE Guarantee of Origin cancellation statement, or Enel's annual supply attestation, naming the POD/site for Frosinone and Milano, with MWh volume and year. (Model: an Enel attestation for a named site ID and a defined date range, stating 100% certified renewable.)
  4. Regolo / Seeweb — Is the Enel annex still current? It cites RED I (2009/28/CE), superseded by RED II (2018/2001) from July 2021. No date visible on the document.
  5. Infomaniak — Does the SIG supply cover both Geneva and Zurich DCs, or Geneva only?
  6. Infomaniak — What assurance level applies to the 2024 GHG report, and is the myclimate review a verification of the inventory or of the offset purchase?
  7. All — Does any provider expose per-request energy/carbon telemetry suitable for per-token accounting?

Caveats

  • No provider holds ISO 14064-1 (GHG quantification) or ISO 14064-3 (verification), and none publishes an assurance statement from a verifier accredited to ISO 14065.
  • ISO 14001 and ISO 50001 are management-system standards. They certify process, not the renewable origin of electricity, and produce no per-token energy figure.
  • All verified claims in this set are organisational and annual. None is a Wh-per-1,000-tokens factor. Certification and per-call measurement are separate axes.
  • The co-op measures only providers that return per-call impact telemetry (currently GreenPT only); all others are estimated via documented proxy.
  • Asymmetry worth noting: Regolo/Seeweb now has stronger evidence than GreenPT, yet a worse factor in the co-op's accounting (250 vs ~193 gCO₂e/kWh). That reflects disclosure, not grid intensity.
# Green claims comparison — GreenPT / Regolo / Infomaniak I've collected some evidence about the energy/ghg claims of these three model providers. The gist of it seems to be that Infomaniak has stronger evidence of being institutionally compliant, though greenpt is the only one doing fine-grained telemetry to substantiate place+time of green energy usage. Regolo/Seeweb make claims that are comparable to Infomaniak but they are not as well as supported by public evidence. Regolo's claim about token wattage might be a reference to calculated "CO2 impact" numbers available in its playground but not in the API. <details> <summary>report from chat.inference.coop + deepseek-4.1-flash</summary> Status: working note. Compiled 2026-11-14. Scope: energy and carbon claims only. Privacy/security claims excluded except where relevant to evidence. ## Summary table | | GreenPT | Regolo | Infomaniak | |---|---|---|---| | **Role in co-op** | Model provider ("Strongest — 100% renewable") | Model provider ("100% renewable") | Co-op's hosting provider; inference not yet connected | | **Control of infrastructure** | ❌ Tenant — host is Scaleway (DCs now under **OpCore**) | ✅ **Own GPUs + own DCs** (Italy). Product of Seeweb Srl | ✅ Own DCs (Geneva, Zurich) | | **Headline energy claim** | "100% renewable energy" | "100% green carbon-free energy"; "Zero CO₂ emissions, zero water waste" | "Powered exclusively by certified renewable electricity" | | **Renewable instrument** | ❌ **None produced at any layer** | ⚠️ **Signed Enel annex**, contract K00961397 — GO-certified product under GSE system | ✅ **SIG Certificat d'origine** — registered number, SGS inspects SIG's energy accounting | | **Site / volume / period specified?** | ❌ | ❌ None of the three | ⚠️ Consumption attested; MWh volume redacted in public copy | | **ISO certificates** | ISO 50001, ISO 27001, HDS | ISO 9001, 14001, 22301, 27001, 20000-1 | ISO 14001, ISO 50001, ISO 27001 | | **Who holds them** | ⚠️ **Scaleway** (host). GreenPT BV "working towards its own ISO 27001" | ✅ **Seeweb Srl** (own entity) | ✅ **Infomaniak** (own entity) | | **ISO 50001 (energy management)** | ✅ (host's) | ❌ **Absent** | ✅ Own, since 2015 | | **ISO 14064 (GHG quantification)** | ❌ | ❌ | ❌ | | **GHG inventory published** | ❌ | ❌ | ✅ Annual reports (2021, 2024) | | **Third-party assurance** | ❌ | ❌ | ⚠️ B Corp (audited) + myclimate review — **review covers offsets, not the inventory** | | **Carbon data granularity** | ✅ **Per-call telemetry** (`impact`: watt-ms, µgCO₂e) | ❌ "Token/Watt" tracking advertised but not exposed to co-op | Not connected; annual corporate scope only | | **How co-op treats it** | ✅ **Measured** — 0.36 / 1.08 / 0.55 Wh per 1k output tokens; ~193 and ~24 gCO₂e/kWh | ⚠️ **Estimated** — proxy 0.5 Wh/1k output tokens, 250 gCO₂e/kWh (EU average) | Outside per-call accounting entirely | | **Main weakness** | Borrowed certifications; optional **US API endpoint** | Energy document lacks site, volume, period, GO number; no ISO 50001 | myclimate covers offsets; SIG supply may cover Geneva only | | **Evidence grade** | Weak — nothing of its own | Moderate — self-held, thin on detail | Strong — best instrument in the set | ## Evidence sources ### GreenPT | Claim | Source | |---|---| | Sustainability claims ("100% renewable", "ISO certified facilities") | https://docs.greenpt.ai/sustainability | | ISO certs held by Scaleway, not GreenPT; optional US endpoint | https://docs.greenpt.ai/privacy/privacy-policy | Notable quotes: - "our data centers maintain ISO 50001 and ISO 27001 certifications for energy management and information security" - "certifications held by our hosting provider Scaleway; GreenPT BV is working towards its own ISO 27001 certification" - optional endpoint `https://api.us.greenpt.ai` — requests "are processed in the United States" ### Regolo (product of Seeweb Srl) | Claim | Source | |---|---| | "100% green carbon-free energy" | https://regolo.ai/ | | "Zero CO₂ emissions, zero water waste" | https://regolo.ai/sustainable-ai/ | | "Our infrastructure is ISO 14001 certified" | https://regolo.ai/sustainable-inference-is-now-an-ai-infrastructure-decision/ | | "All involved data centers are certified and powered by renewable energy sources" | https://regolo.ai/team-infrastructure/ | | Seeweb as DC operator | https://regolo.ai/zero-data-retention-llms-why-it-matters/ | | Enel trademark licence (signed by Seeweb Srl, contract K00961397) | https://media.greenweb.org/uploads/Contratto_Generale_Energia_Elettrica_-_Enel_-_Rinnovabile.pdf | | ISO 14001 cert IT18-27703B — active, 2 sites (Frosinone, Milano), CB: AXE REGISTER, AB: General Directorate of Accreditation | https://www.iafcertsearch.org | Notable quotes: - Enel annex: "DISCPLINARE D'USO DEL MARCHIO 100% Energia Rinnovabile" — describes use of the "Energia 100% Rinnovabile" trademark; states supply is "certificata dal sistema di «garanzie di origine» del Gestore Servizi Energetici" per Directive 2009/28/CE - Enel annex does **not** name a site, a MWh volume, a supply period, or a GO certificate number ### Infomaniak | Claim | Source | |---|---| | Certifications page | https://www.infomaniak.com/en/certifications | | **SIG certificate of origin** | https://www.infomaniak.com/documents/ecology/certificate-SIG.pdf | | Environmental / ISO 14001 + ISO 50001 since 2015 | https://www.infomaniak.com/en/ecology | | Impact report | https://www.infomaniak.com/en/impact-report | | 2024 GHG report | https://news.infomaniak.com/en/co2-report-2024/ | | 2021 GHG report (reviewed by myclimate) | https://news.infomaniak.com/en/report-2021-co2/ | Notable quotes: - SIG certificate: supply is "40 % d'Electricité Vitale Vert ... solaire et hydraulique écologiques locales ... label suisse naturemade star" and "60 % d'électricité à partir d'énergie hydraulique provenant de barrages suisses" - SIG certificate: "L'organisme indépendant SGS inspecte régulièrement la comptabilité énergétique de SIG ... que l'équivalent de la consommation de chaque client a bien été injecté dans le réseau électrique" - "certified renewable electricity, consisting of hydropower and other local renewable sources" - "offsets its remaining CO2 emissions by 200%" ### Co-op methodology | Item | Source | |---|---| | Per-provider energy/carbon factor table, with source recorded per row | https://git.inference.coop/code/member-portal/src/branch/main/metrics_snapshot.py | ## Open questions / follow-ups 1. **GreenPT** — Which legal entity holds the renewable energy attribute certificates for the electricity serving GreenPT inference? Can they supply the certificate? (Nothing produced to date.) 2. **GreenPT** — Is the ISO 27001/50001 certificate GreenPT cites still in Scaleway's name, or has it moved to OpCore following the DC spin-off? 3. **Regolo / Seeweb** — Request the GSE Guarantee of Origin cancellation statement, or Enel's annual supply attestation, naming the POD/site for Frosinone and Milano, with MWh volume and year. (Model: an Enel attestation for a named site ID and a defined date range, stating 100% certified renewable.) 4. **Regolo / Seeweb** — Is the Enel annex still current? It cites RED I (2009/28/CE), superseded by RED II (2018/2001) from July 2021. No date visible on the document. 5. **Infomaniak** — Does the SIG supply cover both Geneva and Zurich DCs, or Geneva only? 6. **Infomaniak** — What assurance level applies to the 2024 GHG report, and is the myclimate review a verification of the inventory or of the offset purchase? 7. **All** — Does any provider expose per-request energy/carbon telemetry suitable for per-token accounting? ## Caveats - **No provider holds ISO 14064-1** (GHG quantification) or ISO 14064-3 (verification), and none publishes an assurance statement from a verifier accredited to ISO 14065. - **ISO 14001 and ISO 50001 are management-system standards.** They certify process, not the renewable origin of electricity, and produce no per-token energy figure. - **All verified claims in this set are organisational and annual.** None is a Wh-per-1,000-tokens factor. Certification and per-call measurement are separate axes. - The co-op measures only providers that return per-call impact telemetry (currently GreenPT only); all others are estimated via documented proxy. - Asymmetry worth noting: Regolo/Seeweb now has stronger evidence than GreenPT, yet a worse factor in the co-op's accounting (250 vs ~193 gCO₂e/kWh). That reflects disclosure, not grid intensity. </details>

I worry a bit that adding more providers (with partial model overlap), or also generally more models, could make it more difficult for new users or non-technical users. We should probably offer some documentation / recommendations to avoid that.

For Infomaniak I'm not sure what exactly they would bring to us in comparison to what we have already. We could serve Kimi K3(.1) from GreenPT and Tinfoil too, for example, and that would be the latest version compared to the Kimi 2.6 that Infomaniak offers.

Mistral-Small-4-119B at $0.24 / $0.90 is cheaper than our glm-5.3 options ($1.23–$1.80 in)

This is not a useful comparison though, as Mistral Small 4 is a) a lot smaller, b) a lot worse unfortunately :) I hope Mistral releases a newer version of that at some point that is properly trained as that one feels rather half-baked.

Qwen3.8-27B-FP8 is reported to be a better model all around than Qwen 3.5 397B

However Qwen 3.X 27B is a lot more expensive to run at scale in comparison to their sizes and models comparable in performance. Something like Qwen 3.8 Flash Next would probably be a better choice. 27B is great model for running personally though.

I worry a bit that adding more providers (with partial model overlap), or also generally more models, could make it more difficult for new users or non-technical users. We should probably offer some documentation / recommendations to avoid that. For Infomaniak I'm not sure what exactly they would bring to us in comparison to what we have already. We could serve Kimi K3(.1) from GreenPT and Tinfoil too, for example, and that would be the latest version compared to the Kimi 2.6 that Infomaniak offers. > Mistral-Small-4-119B at $0.24 / $0.90 is cheaper than our glm-5.3 options ($1.23–$1.80 in) This is not a useful comparison though, as Mistral Small 4 is a) a lot smaller, b) a lot worse unfortunately :) I hope Mistral releases a newer version of that at some point that is properly trained as that one feels rather half-baked. > Qwen3.8-27B-FP8 is reported to be a better model all around than Qwen 3.5 397B However Qwen 3.X 27B is a lot more expensive to run at scale in comparison to their sizes and models comparable in performance. Something like Qwen 3.8 Flash Next would probably be a better choice. 27B is great model for running personally though.

This is not a useful comparison though, as Mistral Small 4 is a) a lot smaller, b) a lot worse unfortunately :) I hope Mistral releases a newer version of that at some point that is properly trained as that one feels rather half-baked.

To add to that. Offering Mistral Large 4 would be a different story as that one is actually competitive but I don't think that's offered by any otherwise interesting provider right now.

> This is not a useful comparison though, as Mistral Small 4 is a) a lot smaller, b) a lot worse unfortunately :) I hope Mistral releases a newer version of that at some point that is properly trained as that one feels rather half-baked. To add to that. Offering Mistral Large 4 would be a different story as that one is actually competitive but I don't think that's offered by any otherwise interesting provider right now.

Another direction this could take: rather than just running Infomaniak's own hosted models, what if we ran a GPU VPS and set up our own execution environment? The goal would be to achieve what none of our providers offer: privacy protection and renewable energy.

https://www.infomaniak.com/en/hosting/gpu

Key questions:

  • Could this be done with any sort of economic efficiency given our current scale and usage (about $50 of tokens/day)
  • Is there sufficient open source software out there to set up our own TEE environment?
Another direction this could take: rather than just running Infomaniak's own hosted models, what if we ran a GPU VPS and set up our own execution environment? The goal would be to achieve what none of our providers offer: privacy protection and renewable energy. https://www.infomaniak.com/en/hosting/gpu Key questions: * Could this be done with any sort of economic efficiency given our current scale and usage (about $50 of tokens/day) * Is there sufficient open source software out there to set up our own TEE environment?
Author
Owner

Brief feasibility read, from their own rate card and one operator's write-up (both linked).

Economics. An L40S (48 GB) is CHF 0.76/h → ~CHF 550/month 24×7; an A100 ~CHF 1.57/h → ~CHF 1,130/month. Our ~$50/day of tokens is ~CHF 1,300/month, so the money is roughly a wash — but not the capability: that buys one or two GPUs serving one small open model, versus today's mix of many models at providers' shared scale. Renting wins unless utilisation is high and the model set is narrow; bursty demand across many models is the worst case for it. Two practical unknowns worth resolving before any estimate: GPU flavours are hidden until support enables them (nvl4-* is non-public), and GPU billing is instance_reserved, so it may charge 24/7 even when the VM is stopped — that single question decides whether the cheap scenarios are real.

TEE. This is the part that doesn't work on the listed cards. NVIDIA's confidential computing — encrypted GPU memory plus attestation — is H100-class hardware; A2/T4/L4/L40S/A100 cannot do it. The software exists and is open (dstack, NVIDIA's attestation tooling, vLLM in CC mode, Phala), but it's H100 plus host and provider support, and the attestation-and-trust story is exactly what we already buy from Tinfoil.

Net: renewable ✓ from Infomaniak either way; privacy ✗ on their GPU VPS. Not a cost play at our scale — only a deliberate "own the TEE" project, and that means an H100.

Sources: Infomaniak GPU · vLLM on Infomaniak, with the flavour and billing findings

Addendum — renewable H100 hosts do exist. Genesis Cloud (Iceland, geothermal + hydro), Vultr (HGX H100 on Sabey hydropower in Washington), Scaleway (H100 in DC5/PAR2, GO-certified wind and hydro). DataCrunch, Verne Global, atNorth and Northern Data sit in the same Nordic/Icelandic family; availability unverified by me. One caveat worth carrying: "GO-certified" means renewable electricity contracted and matched to consumption, not the electrons in that rack — a defensible claim, but a different one from "runs on hydro".

And the two halves are still orthogonal: none of those advertise H100 confidential computing, so renewable compute and TEE compute come from different vendors. The cheapest route to both may be to ask Tinfoil which hosts they run on — their product is the attested-H100 stack and we already pay for it, so "privacy plus renewable" could be a question about our existing vendor rather than a new rental.

Brief feasibility read, from their own rate card and one operator's write-up (both linked). **Economics.** An L40S (48 GB) is CHF 0.76/h → ~CHF 550/month 24×7; an A100 ~CHF 1.57/h → ~CHF 1,130/month. Our ~$50/day of tokens is ~CHF 1,300/month, so the money is roughly a wash — but not the capability: that buys one or two GPUs serving *one* small open model, versus today's mix of many models at providers' shared scale. Renting wins unless utilisation is high and the model set is narrow; bursty demand across many models is the worst case for it. Two practical unknowns worth resolving before any estimate: GPU flavours are hidden until support enables them (`nvl4-*` is non-public), and GPU billing is `instance_reserved`, so **it may charge 24/7 even when the VM is stopped** — that single question decides whether the cheap scenarios are real. **TEE.** This is the part that doesn't work on the listed cards. NVIDIA's confidential computing — encrypted GPU memory plus attestation — is H100-class hardware; A2/T4/L4/L40S/A100 cannot do it. The software exists and is open (dstack, NVIDIA's attestation tooling, vLLM in CC mode, Phala), but it's H100 *plus* host and provider support, and the attestation-and-trust story is exactly what we already buy from Tinfoil. Net: renewable ✓ from Infomaniak either way; privacy ✗ on their GPU VPS. Not a cost play at our scale — only a deliberate "own the TEE" project, and that means an H100. Sources: [Infomaniak GPU](https://www.infomaniak.com/en/hosting/gpu) · [vLLM on Infomaniak, with the flavour and billing findings](https://familylehn.com/infomaniak-gpu-hosting) **Addendum — renewable H100 hosts do exist.** Genesis Cloud (Iceland, geothermal + hydro), Vultr (HGX H100 on Sabey hydropower in Washington), Scaleway (H100 in DC5/PAR2, GO-certified wind and hydro). DataCrunch, Verne Global, atNorth and Northern Data sit in the same Nordic/Icelandic family; availability unverified by me. One caveat worth carrying: "GO-certified" means renewable electricity *contracted and matched* to consumption, not the electrons in that rack — a defensible claim, but a different one from "runs on hydro". And the two halves are still orthogonal: none of those advertise H100 confidential computing, so renewable compute and TEE compute come from different vendors. The cheapest route to both may be to ask **Tinfoil which hosts they run on** — their product is the attested-H100 stack and we already pay for it, so "privacy plus renewable" could be a question about our existing vendor rather than a new rental.

Nebius has ok pricing and NVIDIA RTX PRO 6000 also support TEE:
https://nebius.com/prices

And seems to have "100% renewable energy sourcing" for sites in France, Iceland, UK.(Finland-1: 100% renewable from mid-2026). PUE is 1.25:
https://assets.nebius.com/assets/79cf11ba-bb23-4cb8-802e-21dc155be31c/Nebius%202025%20Sustainability%20Report.pdf

Nebius has ok pricing and NVIDIA RTX PRO 6000 also support TEE: https://nebius.com/prices And seems to have "100% renewable energy sourcing" for sites in France, Iceland, UK.(Finland-1: 100% renewable from mid-2026). PUE is 1.25: https://assets.nebius.com/assets/79cf11ba-bb23-4cb8-802e-21dc155be31c/Nebius%202025%20Sustainability%20Report.pdf

I think there is value in the co-op building capacity and expertise for running our own models.There seemed to be some interest on the November meeting? Based on the above I am less sure when I would choose to use the models that Infomaniak would make available to me. I do like how we have members who do understand some of the nuance here, and could potentially provide guidance to the rest of us. Honestly I am already a bit confused by the options that are available now, and would be interested in having a guide available.

Update: actually https://docs.inference.coop/guides/models/ is really helpful!

I think there is value in the co-op building capacity and expertise for running our own models.There seemed to be some interest on the November meeting? Based on the above I am less sure when I would choose to use the models that Infomaniak would make available to me. I do like how we have members who do understand some of the nuance here, and could potentially provide guidance to the rest of us. Honestly I am already a bit confused by the options that are available now, and would be interested in having a guide available. Update: actually https://docs.inference.coop/guides/models/ is really helpful!

@ed Glad the docs help! I've been working on organizing and improving them.

Excited to see where this discussion leads, both here and in the #compute room in Matrix.

@ed Glad the docs help! I've been working on organizing and improving them. Excited to see where this discussion leads, both here and in the #compute room in Matrix.
Sign in to join this conversation.
6 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: co-op/support#49