Hubblo

Life cycle assessments of NVIDIA graphics processors used for artificial intelligence, by Hubblo and TIDE: the carbon footprint of each card or module, with its manufacturing, distribution and end of life stages.

Non-profitSpecialistCopyleft licence

Publisher's page Browse 38 factors

Factors

Values

One per year and method

Countries

2 regions, sub-national included

Edition

2026-03-11

published 11 Mar 2026 · all versions

Last loaded

8 Oct 2026

Report a newer version

Overview

The studies come from the Alt-Impact project (ADEME, CNRS, Inria) and were carried out by TIDE and Hubblo. They cover nine products: eight NVIDIA GPUs, from the GTX 1080 Ti and P100 to the A100, L4 and GH200, plus one GH200 baseboard that the publisher calls a GPU rack.

The results are published in the GPU-Embedded-Environmental-Impacts repository on GitHub, one workbook per product. The method is described in an academic article, and Hubblo runs an online calculator built on the same work. The full report for ADEME is announced but not yet published.

At a glance

Publisher: Hubblo
Kind: Non-profit
Scope: SpecialistOne category holds at least 5% of its factors: Manufactured goods, with 100.0%. A generalist needs 4.
Home region: France
Edition: 2026-03-11
Licence: Copyleft licence (cc-by-sa-4.0)

Open data

How this works

What open-climate.ai standardised

over the publisher's own fileHow we standardise

One category axis

38 of 38 factor values mapped to one category

Learn more 6 checks behind this

Coded regions

2 regions as ISO or registry codes

37 take the publisher's default region, stated rather than assumed

Learn more 3 checks behind this

Canonical units

1 canonical unit in 1 unit families, raw unit kept

Learn more 4 checks behind this

Warming potentials

None is converted: every value keeps the publisher's GWP.

38 factor values left on the publisher's GWP

Learn more 7 checks behind this

8 more things standardised

Declared boundaries

6 boundary codes

10 factor values state in, out or varies for each scope term

Learn more 7 checks behind this

Parts and parents

29 factor values grouped under 9 parents

Use a part on its own, or the parent when you want the whole.

Learn more 3 checks behind this

Summed totals

1 total summed from its printed parts

Learn more 3 checks behind this

Data quality

38 of 38 factor values rated on one quality tier

Learn more 7 checks behind this

EPDs and suppliers

38 factor values naming a supplier

Learn more 1 check behind this

Languages

English: 100.0% of factor values

French: 0.0% of factor values

Learn more 11 checks behind this

Traceable to the source

54 source rows archived

17 rows skipped, each with a reason

Learn more 5 checks behind this · 1 known limitation

Stable addresses

Slugs never break

Learn more

Audit checks

121 passed1 known limitation3 enforced when loadedlast run 8 Oct 2026

Checks we run on every load to keep the data above trustworthy. Counts are for this library; the grey figure is the same check across all libraries. Point at the i to see why a check matters.

What the statuses mean How checks are counted Why we check How checks are grouped

Needs attention

Values traced to one source row agree with each other
Known limitation · 1 in this library · 2 in all libraries
Why it matters Each value traces back to one source row. When values that disagree share a row, it can back only one of them, which weakens that audit trail.

One category axis

Learn more

Every source category maps to our vocabulary
Enforced when loaded
Why it matters A factor with no category drops out of every category filter and comparison. The load stops rather than serve one.
The publisher's classification is kept
Passed
Why it matters You can still find a factor the way the publisher filed it, and check our mapping against theirs.
Every classification level has a valid parent
Passed
Why it matters A broken tree hides factors from browse and puts them in the wrong category totals.
Every product code comes with the codes above it
Passed
Why it matters In product codes like HS or CPC, a search for a chapter should reach every code inside it. A missing level hides the factor from that search.
A publisher's own classification is used only by that publisher
Passed
Why it matters A publisher's own tree only makes sense inside its library. Filed under another publisher's codes, a factor would show up in categories that do not describe it.
A standard classification code has one label in every library
Passed
Why it matters Codes like NACE or NAICS are meant to link libraries. If one code carries two labels, you cannot tell whether both libraries mean the same activity.

Coded regions

Learn more

Every region is a known code
Enforced when loaded
Why it matters Free text like “France continentale” cannot be filtered or joined to your own data. A code can.
No named region is filed under a broader one
Passed
Why it matters A French factor filed as World would be applied to the wrong country, or lose to a global value in a search for France.
Every region sits in the world hierarchy
Passed
Why it matters A search for Europe has to reach France. A region outside the hierarchy is invisible to broader searches.

Canonical units

Learn more

Every unit is a known code
Enforced when loaded
Why it matters Activity data only multiplies correctly when the unit is exact. An unknown unit stops the load.
One source unit maps to one unit
Passed
Why it matters If one of the publisher's units were read as two different units, factors that share it would give inconsistent footprints.
A unit label means the same in every library
Passed
Why it matters Comparing two libraries is only safe when a label like t.km means tonne-kilometre in both.
A unit label that names no basis folds one way
Passed · 1 in all libraries
Why it matters A label like GJ says nothing about the heating value behind it. If the same bare label folds to two different units, two factors that look alike are not comparable.

Declared boundaries

Learn more

Every boundary states what it includes
Passed
Why it matters Cradle-to-gate and full life-cycle values can differ several times over. Without a stated boundary you cannot tell which you are adding up.
Boundary scope terms are known codes
Passed
Why it matters Consistent terms let you keep exactly the life-cycle stages your inventory needs.
No two factors for the same activity partly overlap in scope
Passed
Why it matters Two totals that share some stages but not all cannot safely be added or swapped: adding them counts the shared stages twice.
Separate stages of one activity are linked as parts
Passed
Why it matters Stages served as unrelated factors give no total, so you may use one stage and miss the others.
Every part of a factor states its boundary
Passed
Why it matters A part with an unknown boundary cannot be told apart from the other parts, so you cannot pick out, say, the combustion share of a total.
Every biofuel counted with its biogenic CO2 says how much
Passed
Why it matters Some publishers count the CO2 a burnt biofuel releases and others leave it out. Search tells the counting ones apart by their biogenic CO2 figure, so a factor without it would be offered to someone who asked to leave that CO2 out.
No recycling credit is hidden inside a life-cycle total
Passed
Why it matters A negative total may be a credit for avoided emissions. Filed as a life-cycle burden, it would quietly subtract emissions from your footprint.

Warming potentials

Learn more

Methane and nitrous oxide figures match the stated GWP
Passed
Why it matters If a gas's CO2e is not its mass times the stated GWP, the GWP label is wrong, and converting the value to another GWP version would give a wrong result.
Every convertible value has its AR6 version
Passed
Why it matters Reporting now asks for AR6. Mixing AR5 and AR6 shifts methane-heavy totals by several percent.
CO2e totals leave out what their basis excludes
Passed
Why it matters A value's basis names the carbon pools its total leaves out. When the gas breakdown shows the total counting one of them anyway (biogenic CO2, land-use change, or biogenic methane on a fossil-only value), the label or the total is wrong.
Values that exclude a carbon pool list the gases behind the total
Passed · 50,523 in all libraries
Why it matters With no breakdown by gas, nothing can confirm that a fossil-only or biogenic-CO2-excluded total really leaves those pools out. The value rests on the publisher's word alone.
Standardised figures keep every gas the publisher lists
Passed
Why it matters The comparable figure we serve must carry each gas the publisher reported. A lost gas makes the breakdown disagree with the total and hides that gas from you.
Every value states what its CO2e total includes
Passed
Why it matters For a biofuel, a total with or without biogenic CO2 can differ widely. Without that convention you cannot tell which one you are comparing.
A library with no AR6 version at all has a recorded reason
Passed
Why it matters Values that cannot be converted to AR6 cannot be compared with AR6 figures. That has to be a documented decision, not something missed at load.

Parts and parents

Learn more

Every part links to a valid parent
Passed
Why it matters A part that points nowhere gets summed twice or dropped when you use the parent total.
No factor is an unlinked part of another factor
Passed
Why it matters If one stage of a total is also served on its own, beside that total, adding both counts the same emissions twice.
Every factor points to its top parent
Passed
Why it matters Search groups parts under their top factor. A stale link shows a part as a stand-alone result.

Summed totals

Learn more

Summed totals equal their parts on every gas
Passed
Why it matters Where a total is the sum of its parts, the two must agree on every gas. Otherwise your footprint changes depending on whether you use the total or the parts.
Every summed total has a classification when its parts do
Passed
Why it matters A total with no classification drops out of every category filter, even though its parts are filed there.
Summed totals keep their parts' classification
Passed
Why it matters A total takes a classification only when all its parts share it. Otherwise it lands in a category its parts do not support, or misses one they all share.

Data quality

Learn more

Quality ratings fall inside their scale
Passed
Why it matters An out-of-scale rating cannot be compared, so the value would be ranked arbitrarily.
Every zero value is explained
Passed
Why it matters A zero can be real (contracted renewable power) or a gap in the data. A gap read as real erases emissions from your total.
No default value is withheld for poor quality
Passed
Why it matters A default rated too poor to serve is hidden from search and exports, so that factor is missing for you even when it is the right one.
No implausibly high per-kilogram value is rated reliable
Passed
Why it matters Above 1,000 kgCO2e per kg is often a unit slip, such as grams read as kilograms. Rated reliable, such a value could inflate a footprint a thousandfold.
Every quality level gives a recognised reason
Passed
Why it matters The reason explains why a value got its quality level. An unrecognised one cannot be filtered on or explained to you.
A zero that is only a gap in the source is withheld
Passed
Why it matters A zero with no quality signal, or an implausible regulatory default of zero, is more likely missing data than a real zero. Served, it would erase emissions from your total.
The library is not rated entirely on one fallback quality level
Passed
Why it matters When every value gets the same level because the publisher gives no quality signal, the level tells you nothing and cannot rank one value above another.

EPDs and suppliers

Learn more

A supplier named in the source is linked to the factor
Passed
Why it matters If the supplier is named but not linked, you cannot filter by supplier or tell a company's own figure from a generic one.

Languages

Learn more

Every factor has an English name
Passed
Why it matters Every factor can be searched and cited in one shared language, whatever the publisher's.
Translations match the current source text
Passed · 302 in all libraries
Why it matters When the publisher rewords a label, an old translation can describe a different activity than the one priced.
Translations follow a corrected source text
Passed · 4 in all libraries
Why it matters When we correct an English label, a French or German label translated from the old wording can still describe what the old wording said.
Recorded translations survive a reload
Passed
Why it matters A reload rewrites a library's labels; a translation we recorded and did not put back leaves that factor without a label in that language.
Every served factor has a name
Passed
Why it matters A factor with no name in any language shows as a blank row: you cannot search for it or tell what it measures.
Factors are searched on their English name where one exists
Passed · 22 in all libraries
Why it matters Search matches your query against one name per factor. A name kept in the publisher's language, when an English one exists, is harder to find in English.
Every language with a main name also has a full name
Passed
Why it matters Without the full name, the detail that tells similar factors apart is missing in that language, so two different factors can look identical.
Full names are built from their parts in the same language
Passed
Why it matters A full name must read as the main name followed by its detail. Translated separately, it can drift and describe a slightly different activity.
Every text is labelled with its real language
Passed
Why it matters A French name labelled English would be shown to English readers and matched against English searches.
Every description is available for display and translation
Passed
Why it matters A description missing from the per-language text store is not shown in your language, and never gets translated.
Every usage note is available for display and translation
Passed
Why it matters Names and notes are served from one text store per language. A note missing from that store is not shown, and never gets translated.

Traceable to the source

Learn more

Values traced to one source row agree with each other
Known limitation · 1 in this library · 2 in all libraries
Why it matters Each value traces back to one source row. When values that disagree share a row, it can back only one of them, which weakens that audit trail.
No duplicate factors are served
Passed
Why it matters Identical duplicates split search results and invite counting the same factor twice.
No factor value claims two things that cannot both be true
Passed
Why it matters For example, a purchaser price that leaves out retail or delivery. One of the two claims is wrong, and you cannot tell which.
No energy factor answers a question about packaging or recycling
Passed · 2 in all libraries
Why it matters Packaging and recycling credits belong to goods. On a value per kilowatt-hour they usually mean the boundary was read from the wrong kind of factor. A declared product sold per kilowatt-hour can still carry them, so each case is reviewed.
No factor value the publisher has withdrawn is still served
Passed · 10,480 in all libraries
Why it matters The publisher's end date for the value has passed. The figure may be out of date, so look for a newer one before you use it.

Other checks

One activity keeps one family across its units
Passed · 7 in all libraries
Why it matters When a publisher prints the same activity per kg and per kWh, the two factors belong to one activity family, with the unit as a choice. Filed as separate activities, they look unrelated, and nothing says they measure the same thing.
An activity's grouping key names only the activity
Passed · 6 in all libraries
Why it matters If the key also carried the unit, country or scope, one activity would split into several, and search would not show you its other versions side by side.
A Creative Commons credit names its licence
Passed
Why it matters CC BY asks you to say which licence applies. A credit that leaves it out makes your attribution incomplete when you copy it.
No note of ours appears in the credit or licence notice
Passed
Why it matters Our own notes on how we handle the data would end up in the credit you reproduce, as if the publisher had written them.
The credit line carries no terms of use
Passed
Why it matters You copy the credit line as it stands. Terms of use belong in the licence notice, which you read rather than reproduce.
Every impact assessment method is a recognised code
Passed
Why it matters Free text in place of a code splits one method into several spellings, so filtering by method misses some of its values.
Green-electricity zeros are not rated below market-based peers
Passed
Why it matters A green-contract zero is the contract's own figure. Marked down for being zero, it would lose to less relevant values when you look up contracted supply.
An avoided credit is linked to the burden it belongs to
Passed
Why it matters Served on its own, a credit reads as a factor in its own right. Linked beside its burden, it applies only where it belongs and is never subtracted twice.
Every value is typed as specific, generic or average
Passed
Why it matters The type says how closely a figure fits the product you bought. Any other word would be invisible to a filter on it.
The market-based value is the default
Passed
Why it matters The default then matches the figure most Scope 2 reports lead with, and a green-tariff zero is never the default when another figure exists.
Validity dates are the publisher's own, never guessed
Passed
Why it matters A start date we invented looks just like one the publisher printed, and would hide correct factor values from a search for what applied on a given date.
Every district-heat factor says whether the network's heat loss is included
Passed
Why it matters Heat is lost in the pipes between the plant and the building. Only the publisher can say whether a factor divides by heat produced or heat delivered, so a factor that leaves it unsaid is listed until its library's method is read.
Quality breakdowns use their scheme's axes
Passed
Why it matters A score on an axis the scheme does not define, or off its scale, cannot be read, so the breakdown behind a rating would mislead.
No battery-electric vehicle carries tailpipe emissions
Passed
Why it matters An electric vehicle has no tailpipe. A non-zero combustion figure means its electricity emissions sit in the wrong place, or are counted twice.
The text of each estimated electricity total matches its current figures
Passed
Why it matters The description of an estimated electricity total states its year, mix, quality and the parts it adds up. If those moved, the text would explain a different number than the one served.
Every estimated factor carries its description in each language
Passed
Why it matters An estimated factor is one we computed, not one a publisher printed. Its description says so and names the data behind it. Without it, you could not tell the estimate from a published figure.
The text of each estimated factor matches its current figures
Passed
Why it matters An estimated factor's description states the year, shares and quality behind it. If those moved, the text would explain a different number than the one served.
A total built on an estimated part is marked as estimated
Passed
Why it matters If the total lost the mark, you would read a figure we partly computed as one the publisher printed.
Every estimate is built only from data we may redistribute
Passed
Why it matters An estimate derived from restricted data would pass that data on under our open licence, which its owner never allowed.
No estimated factor is rated above the data it was built from
Passed
Why it matters An estimate cannot be more reliable than its inputs. Rating it higher would lead you to trust it more than a published figure.
No estimate sits beside a total of unclear scope
Passed
Why it matters When the publisher does not say whether its total includes a stage, our estimate of that stage may count it twice. These cases are listed for review.
No estimate repeats a stage the publisher's total already includes
Passed
Why it matters If the publisher's figure already covers, say, grid losses, adding our estimate of those losses would count them twice and overstate your emissions.
Every factor we complete has all of its estimated parts
Passed · 9 in all libraries
Why it matters If one part, such as grid losses, is missing where the others were estimated, a total built from them silently leaves that part out.
No estimate is served where the publisher printed the figure
Passed
Why it matters When the publisher states the number itself, theirs is the one to use. An estimate beside it would give you two answers to the same question.
We only estimate for publishers whose rule allows it
Passed
Why it matters Some publishers print every part themselves, or are deliberately left as they are. An estimate added to them would mix our numbers into theirs, against the rule set for that library.
Every estimated factor still matches the inputs it was computed from
Passed
Why it matters When an input is revised or withdrawn, the estimate built on it is out of date until it is recomputed.
No estimated total competes with a published one
Passed · 24 in all libraries
Why it matters Where the publisher prints a total for the same activity, region and year, an estimated total beside it would give you two answers to the same question.
Every factor with a current value has a default
Passed
Why it matters A factor with values but no default is left out of the served catalogue, so you would never find it in search or exports.
Every factor has a value, in a library that publishes values
Passed
Why it matters A factor listed with no number looks like a match in search but gives you nothing to calculate with.
The text of each fuel supply-chain estimate matches its current figures
Passed
Why it matters Its description states the year, the figures, the ratio and the quality behind the estimate. If those moved, the text would explain a different number than the one served.
No fuel supply-chain estimate is rated above what its method allows
Passed
Why it matters Each way of estimating a fuel's supply chain has a highest rating it can earn, and an estimate never outranks the combustion figure it builds on. A better rating would lead you to trust it more than it deserves.
Every fuel estimate is built only from data we may redistribute
Passed
Why it matters An estimate derived from restricted data would pass that data on under our open licence, which its owner never allowed.
Every fuel eligible for a supply-chain estimate is classified
Passed
Why it matters The supply-chain estimate depends on which fuel is burned. A fuel nobody has classified gets no estimate, so you would miss that part of its emissions without knowing why.
No fuel estimate is served where the publisher printed the supply-chain figure
Passed
Why it matters When the publisher states the figure itself, theirs is the one to use. An estimate beside it would give you two answers to the same question.
No supply-chain estimate is served for a fuel we deliberately left out
Passed
Why it matters Some fuels are left out on purpose, for instance when no published supply-chain figure fits them. An estimate on one of them would be a number we decided not to stand behind.
Every fuel supply-chain estimate still matches the figures it was computed from
Passed
Why it matters The estimate scales the publisher's combustion figure by a supply-chain ratio taken from other published data. When either is revised or withdrawn, the estimate is out of date until it is recomputed.
Every tagged gas has a known GWP
Passed
Why it matters A kilogram of one named gas is restated in AR6 with that gas's own GWP. A tag the GWP table does not carry would leave the value unconverted, or convert it with the wrong number.
Every fixed group of regions lists what it covers
Passed
Why it matters A group like “Europe outside the EU” with no member list cannot be placed, so its factor values never match a search for any country in it.
Where both exist, the residual mix is shown before the green-tariff zero
Passed
Why it matters Green electricity counts as zero only for buyers who hold the matching certificates. Showing zero by default would understate everyone else's emissions.
Every green-electricity zero is a market-based value
Passed
Why it matters A zero from a green contract only holds for the buyer of that contract. Filed as location-based, it would put a zero on the physical grid.
Every electricity factor states its mix
Passed
Why it matters Without its mix you cannot tell a location-based figure from a market-based one, and Scope 2 reporting needs both side by side.
The newest edition of each value is the one marked current
Passed
Why it matters If an older edition stayed marked as current, you would get superseded numbers, or two answers for the same factor.
Every publisher has a rule on whether we estimate its missing parts
Passed
Why it matters For electricity and fuels we can estimate parts a publisher leaves out, such as grid losses or the fuel supply chain. Without a rule, a missing part could be a gap or a deliberate choice, so nothing is estimated until one is set.
Factor descriptions describe the factor, not the whole library
Passed · 1 in all libraries
Why it matters When the same long text is repeated on every factor, it tells you nothing about the one you are looking at and hides what is specific to it.
Every library's licence has its rights recorded
Passed
Why it matters Without them, the library page cannot tell you what you may do with the data: share it, change it or use it commercially.
No fossil electricity mix has a zero fuel supply chain
Passed
Why it matters Electricity with fossil emissions always has a supply chain behind it: extracting and transporting the fuel. A zero there means the chain was priced from the wrong mix. Green power is the exception: its chain covers only biomass, and is a real zero on a grid that burns none.
Every green-electricity zero is exactly zero and says why
Passed
Why it matters A green-tariff factor is zero by rule. Any other number, or a zero with no stated reason, would mean the rule was applied wrongly.
Green-electricity zeros cover power generation only
Passed
Why it matters Certificates cover the power plant's emissions, not grid losses or the upstream fuel chain. A zero on those stages would understate your Scope 2 and Scope 3 figures.
Only a total we computed may mix GWP vintages
Passed
Why it matters A published row states one GWP vintage. A mix of AR5 and AR6 can only come from a total assembled out of several rows, and saying so tells you the number was computed, not published.
Spend factors state their price year and basis
Passed
Why it matters A 2019 euro is not a 2024 euro, and a purchaser price includes margins a basic price does not. Without both, spend footprints drift.
One default value per factor and year
Passed
Why it matters Two defaults for one year give two answers to the same lookup.
Every value has a number, a GWP version and a reference year
Passed
Why it matters Without the number there is nothing to multiply. Without the GWP version or the year, you cannot tell whether two values are comparable.
No factor description is too long to display in full
Passed · 28 in all libraries
Why it matters Text beyond the limit is cut short on the factor page, so part of what the publisher wrote shows only in the download.
The library-wide answers on what a factor covers are up to date
Passed
Why it matters Some questions, such as whether a figure includes transport, are answered once for a whole library. If the stored answers fall behind, a factor's details can tell you something we no longer state.
Every library-wide answer applies to at least one factor
Passed · 3 in all libraries
Why it matters An answer about a category or stage the library no longer covers changes nothing you see. It is listed so it can be removed.
Every excused arithmetic disagreement is stated
Passed
Why it matters Where a publisher's own total does not equal its parts and we serve it anyway, the exception is listed on every audit run rather than left in code where nobody sees it.
Every reference year falls between 1960 and next year
Passed
Why it matters A year outside that range points to a typo or a misread column, and would put the value in the wrong period when you filter by year.
The parts of a total are never listed twice under one label
Passed
Why it matters A group that repeats a part label is a list of terms, not a split of the total, so adding its rows up would count some emissions twice.
Every residual mix has its green-tariff counterpart
Passed
Why it matters Market-based Scope 2 reporting needs both: the residual mix for ordinary contracts and a zero for certified green power. With one missing, you have to find it elsewhere.
Every publisher's quality rating system is read
Passed
Why it matters If we do not read a publisher's rating system, its values are treated as unrated and marked down, although the publisher rated them.
Every factor is filed under the most specific category its source allows
Passed · 1,852 in all libraries
Why it matters Some factors sit in a broad category because their source covers several narrower ones. They are listed so a finer mapping can place them more precisely where the source allows it.
Every factor has a category
Passed
Why it matters A factor with no category drops out of every category filter and comparison. Without this check, a library with no category mapping would load unlabelled, and nothing would say so.
Every categorised factor has a GHG Protocol scope
Passed · 30 in all libraries
Why it matters Without a scope, you cannot tell whether the factor belongs in Scope 1, 2 or 3 of your inventory, and you have to place it yourself.
Single-gas values have their AR6 version
Passed · 6 in all libraries
Why it matters A kilogram of one refrigerant is worth its GWP, so it can be restated in AR6 exactly. Blends and untagged gases stay on their published vintage, which can move a leak's footprint by several percent or more.
No factor answers a question that does not arise for it
Passed · 6 in all libraries
Why it matters Some questions only make sense for some kinds of factor. An answer where the question should not arise is either misplaced, or a sign the kind is too coarse for that factor, so each case is reviewed.
A part covering the same scope as its total is linked to it
Passed
Why it matters When a part and the total beside it cover exactly the same scope, nothing in the numbers tells them apart. Linking them stops you adding a part to the total that already contains it.
Every device total that includes its use also serves the embodied part
Passed · 112 in all libraries
Why it matters A device footprint that includes years of electricity use is not the manufacturing number you need for a purchase, and if you also count the device's electricity you count its use twice. Devices whose publisher prints no use share are listed as known gaps.
A year in a factor's name matches its reference year
Passed
Why it matters A name like “Grid 2019” on a 2023 value usually means a yearly series was flattened, so you could apply one year's figure to another.

How the audit works

What can I do with this data?

Copyleft licence · credit the publisher

Your rights

Copyleft licence · cc-by-sa-4.0

Use it to calculate my own footprint: Yes
Publish my footprint results: Yes, if
you credit the publisher
you say what you changed
you share the result under the same licence
Share the file with my customers: Yes, if
you credit the publisher
you say what you changed
you share the result under the same licence
Put the factors in a product I sell: Yes, if
you credit the publisher
you say what you changed
you share the result under the same licence
Modify it or combine it with my own data: Yes, if
you credit the publisher
you say what you changed
you share the result under the same licence

Named licence, checked on 17 Jul 2026

CC BY-SA 4.0 grants CC BY's rights; §3(b) adds that Adapted Material You Share carries a BY-SA licence (same elements, this version or later), so the obligation travels with anything blended from it.

Credit line

Lees-Perasso E., Ekchajzer D., Roussilhe G., De Latour T. (2026), Analyses de cycle de vie de GPU pour l'intelligence artificielle, ADEME. Licensed under CC BY-SA 4.0. Normalised by open-climate.ai, https://open-climate.ai/factors/libraries/hubblo/ (CC BY-SA 4.0)

Not legal advice: the publisher's terms prevail.

Full terms Where this answer comes from

How to read these answers

Coverage

What the library covers, by category and by boundary. Click a slice to drill in, then a sub-category or a boundary to filter the factors below.

By category

9factors

Manufactured goods9100.0%

Click a slice to see what is under it. A name opens its category page; its swatch hides it from the ring.

By boundary

9factors

Click a slice to see what is under it. Click a name to hide it from the ring.

Browse factors

top 0 matches

Saved	Unfold	Factor	Region	Boundary	Unit

Versions

How this works

Version	Published	Loaded	Status	Source
2026-03-11	11 Mar 2026	8 Oct 2026	Served	Source

Methodology

How this works

Each value is a study's climate change result in kg CO2e per product: one card, one module or one baseboard, as printed.
The studies are bottom-up life cycle assessments, built component by component in the EIME software (version 6.3) with the EF 3.0 indicator set of the Product Environmental Footprint method. The L4 study's documentation declares conformity with ISO 14040 and 14044, the ILCD entry-level requirements and the PEF method.
Climate change is GWP100 from IPCC 2013 (AR5), as printed with every result. Values stay on AR5: each result is a CO2e total, not a split by gas, so it cannot be restated on AR6.
The total is net: fossil, biogenic, and land use and land use change. The studies print these three parts and we serve them with the total.
The studies print five stages: manufacturing (cradle to gate), distribution (transport to the customer), installation, use and end of life. Installation is zero for all nine products and use is zero for eight, so for those eight the total is the embodied footprint: manufacturing, distribution and end of life. Each total states that transport to the customer is included.
The A100 SXM4 40GB is the only study with a use phase: 400 W for three years, 10,512 kWh, on the French low-voltage electricity mix of 2018. Use is 702.35 of its 862.78 kg, and its total states that the use phase is included.
The reference year is the data year printed for the L4 (2020). The other studies print none, so the year the study was exported from EIME is used: 2025 for six cards, 2026 for the GH200 and its baseboard.
Every value is specific: the study of one NVIDIA product by the authors. The results are not environmental product declarations.

What we import

The nine result workbooks of the repository as of 2026-03-11: 37 printed factors, one value each. Nine are product totals and 28 are life stage parts, each part linked to its total. Every printed total equals the sum of its stages.
One more factor is ours: the A100 SXM4's embodied subtotal, 160.43 kg, the sum of its printed manufacturing, distribution and end of life parts. It sits between those three parts and the printed total, so 38 factors in all.
The 17 stages printed as zero (installation for all nine products, use for eight) are skipped.
NVIDIA is shown as the supplier of every factor.
Production spans several countries, so every factor is placed in the world as a whole. The one exception is the A100 SXM4 use stage, placed in France, where its electricity is modelled.
No classification code is published. Every factor carries the open-climate.ai category Electronics and ICT equipment, under manufactured goods.
The studies give no description per factor. Each stage part is labelled with the name of its stage.
The 18 other PEF indicators (water use, resource use, toxicity and so on) are kept with each value in our archive, not served.
Not imported: the results per component, the bills of materials and inventories, the elementary analysis of three cards, and the calculator's parametric model with its factors per die area, per circuit board area and per kg of casing or heatsink.
Names are in English, as published.

Licence

How this works

Copyleft. The repository is published under CC BY-SA 4.0, which allows commercial use and redistribution with attribution. The L4 study's documentation repeats it: free of charge for all users and uses, under CC BY-SA 4.0. Attribution: "Lees-Perasso E., Ekchajzer D., Roussilhe G., De Latour T. (2026), Analyses de cycle de vie de GPU pour l'intelligence artificielle, ADEME. Licensed under CC BY-SA 4.0." Last verified 2026-10-07.

ShareAlike applies: values derived from or blended with these studies must be shared under the same licence. Do not mix them into a result you publish under a more permissive licence.

What can I do with this data?

Credit line: Lees-Perasso E., Ekchajzer D., Roussilhe G., De Latour T. (2026), Analyses de cycle de vie de GPU pour l'intelligence artificielle, ADEME. Licensed under CC BY-SA 4.0.

cc-by-sa-4.0

Limitations

All 38 values are rated medium quality. Eight of the nine studies publish no quality score; the L4 study's PEF data quality rating ("Excellent quality") is kept in our archive and not used for the grade. See how data quality is graded.
The A100 SXM4 total includes three years of use on French electricity and is not comparable with the other eight totals. Its embodied footprint is the sum of its manufacturing, distribution and end of life parts, 160.43 kg, served as a subtotal. If you report the card's electricity separately, use the subtotal, not the total.
Nine products only, all NVIDIA. The H100 is covered only as part of the GH200 and, as a standalone PCIe card, by an estimate in the publisher's calculator, which is not imported.
For most products the reference year is the year of the study, not a data year.
The GH200 and its baseboard print a weight of 1 kg, a placeholder. Do not use it as the product's mass.
The L4 documentation notes that relevant flows are missing from the land use part of climate change.
Until the ADEME report is published, the method is documented only by the workbooks and the academic article.

Changelog

2026-10-07: edition of 2026-03-11 loaded.
2026-10-08: the A100 SXM4's embodied subtotal served beside its total.

Highlighted factor pages

A selection of Hubblo factors, with each value, boundary, source and derivation.

Manufactured goods Electronics and ICT equipment

Baseboard GH200 GLOBAL · unit · Cradle to grave, embodied
A100 SXM4 40GB GLOBAL · unit · Cradle to grave
P100 GLOBAL · unit · Cradle to grave, embodied
GTX 1080 Ti GLOBAL · unit · Cradle to grave, embodied
Titan RTX GLOBAL · unit · Cradle to grave, embodied

Know a newer edition?

If Hubblo has published an edition we do not serve yet, tell us where to find it.