A new report from venture firm a16z highlights a shifting race in generative artificial intelligence (AI). Google’s Gemini, China’s DeepSeek, and even Grok, backed by Elon Musk, are gaining ground on OpenAI’s ChatGPT.
But as these AI rivals advance, there’s an urgent question: how green are their growing footprints? Let’s take a closer look at each of the top three AI’s environmental footprints below.
Competitors Rise: How Google and Grok Are Gaining Ground on ChatGPT
The a16z report maps the top 100 generative AI apps, showing that ChatGPT has strong competition emerging. Google’s Gemini is expanding quickly, and Grok—new but promising—is stepping onto the field, too.

Gemini’s strength comes from Google’s massive infrastructure. Its backing allows faster improvements and better integration across services like search, Gmail, and cloud tools. Gemini’s smooth response and deep context give it a competitive edge.
Meanwhile, DeepSeek earns the third spot because it strikes a middle ground between efficiency and emissions. Much of its footprint comes from running on China’s coal-heavy power grid, which raises its carbon intensity compared to peers with greater access to renewable energy.
Meanwhile, ChatGPT stays strong thanks to its large user base and bold partnerships. OpenAI’s alignment with Microsoft means tight integration in Office, Azure, and more. ChatGPT also supports fine-tuning and plugins, making it more flexible for businesses and developers.

Despite their differences, the report shows all three top models are advancing quickly in user experience, expanding features, and market presence. It marks a growing field, not one dominated by ChatGPT alone anymore.
Watt for Watt: Who’s the Greenest Chatbot? Comparing AI Footprints
As AI usage grows, its environmental impact becomes critical. Let’s compare how these three models fare in energy use and emissions.
OpenAI ChatGPT
ChatGPT sits in the middle of the spectrum. Its exact footprint varies depending on which study you use, but most analyses suggest it consumes more energy and emits more carbon per query than Gemini.
Part of this comes from heavier model sizes and widespread usage. Improvements in hardware efficiency and energy sourcing are bringing numbers down, but its typical footprint is still higher than Google’s.
OpenAI’s Sam Altman claims a ChatGPT query uses as much power as running an oven for about one second. Independent estimates align with this level.
Although a single query uses moderate energy, the rapid growth in usage means overall consumption is significant. U.S. data centers—many of which power AI—could account for up to 8% of U.S. electricity use by 2030.
Greenly, a carbon accounting firm, estimates that using ChatGPT-4 to respond to one million emails monthly could generate 7,138 tonnes of CO₂, equating to about 4,300 round-trip flights Paris–New York per year.

- Energy use per prompt: ~3 Wh (can be lower in some estimates, ~0.3 Wh)
- CO₂ emissions per prompt: ~2–3 g (includes amortized training emissions)
SEE MORE: ChatGPT Hits 700M Weekly Users, But at What Environmental Cost?
Google Gemini
Google has been working to make its AI models more efficient, and Gemini reflects this push. According to Google’s own reporting, text-based queries in Gemini consume very little energy compared to earlier AI systems.
The company highlights dramatic efficiency gains in both energy use and carbon intensity, making Gemini one of the leaner large models when handling short, text-only prompts.
- According to Google, a median Gemini AI text prompt uses just 0.24 watt-hours, emits 0.03 grams of CO₂, and consumes 0.26 milliliters of water—about five drops.
Over the past year, Google claims a 33× reduction in energy use per prompt and a 44× reduction in carbon footprint while improving quality.

Experts warn Google’s method may understate environmental cost by excluding indirect water usage (e.g., power plant cooling) and relying on market-based carbon accounting.
- Energy use per prompt: ~0.24 Wh
- CO₂ emissions per prompt: ~0.03 g
- Water use per prompt: ~0.26 mL
READ MORE: Google Reveals the Environmental Cost of Gemini AI Query
DeepSeek R1
DeepSeek’s reasoning models work well with long, complex prompts. This makes them more energy-intensive than regular chat models.
DeepSeek hasn’t shared its exact CO₂ figures. However, benchmarking shows that its energy use per query is much higher than competitors. This is especially true for tasks that require multi-step reasoning or coding. This places DeepSeek at the high end of per-query emissions.
A recent academic study found that models like DeepSeek-R1 use more than 33 Wh per long prompt—over 70× the energy of smaller models like GPT-4.1 Nano. Large-scale inference, with 700 million queries daily, could use as much electricity as 35,000 U.S. homes. It would also need a forest the size of Chicago to offset its carbon emissions.
- Energy use per long reasoning prompt: >33 Wh
- CO₂ emissions per prompt: Likely an order of magnitude higher than ChatGPT (depends on grid mix): ~2–4 g
At first glance, Gemini seems the greenest per query (with footprints barely visible in the chart below), while ChatGPT has a moderate impact, and DeepSeek is the least efficient. But real-world AI use involves billions of queries daily. So, even small differences matter.

As AI scales, overall energy and CO₂ use skyrocket unless systems are optimized for efficiency.
Data Centers or Carbon Centers? The Stakes for Climate
The environmental stakes are real. Experts estimate global data center use could hit 945 terawatt-hours (TWh) by 2030, with AI responsible for 652 TWh—an 80× jump from today.
Generative AI alone may cause 18–246 million tons of CO₂ emissions per year by 2035, similar to entire industries like aviation or shipping.
Without green design, AI growth could claw back efforts to reduce climate impact. Companies need to think beyond speed and accuracy—AI must grow sustainably, too.
AI Growth Meets Climate Responsibility: What Comes Next
The AI competition is intensifying—with ChatGPT, Gemini, and Grok pushing each other forward. Users benefit from better tools, but rising usage means rising environmental costs. To move forward responsibly, analysts suggest these actions:
- Developers should optimize AI models for energy efficiency, just like Gemini’s leap.
- Companies should track and reveal full lifecycle impacts—not just inference costs.
- Cloud providers and AI firms need policies favoring renewable energy and efficient data center cooling.
- Public policy could reward low-carbon AI, possibly with incentives or carbon pricing.
The a16z report shows that generative AI has entered a new phase—competition among equals, not a single leader. ChatGPT, Gemini, and Grok are all driving innovation in AI. But with growing usage comes growing environmental responsibility.
As the field speeds up, AI’s impact on climate can’t be ignored. Models that combine high performance with low energy use will define the future. If innovators balance progress with sustainability, AI’s value could be even greater—and greener.
The post ChatGPT, Gemini, and DeepSeek Are on an AI Race – But at What Climate Cost? A Comparison appeared first on Carbon Credits.
Carbon Footprint
Insetting vs Offsetting: Which Actually Counts Toward Your Scope 3 Targets
The accounting differences that decide whether your nature investment shows up in inventory, in BVCM, or nowhere at all.
The question reaches a procurement team about three weeks before the next sustainability committee meeting. Someone has read about insetting. Someone else has just signed off on an offset purchase. The CSO wants to know if the two are interchangeable. The answer is no, and the GHG Protocol Land Sector and Removals Standard is the reason why.
This article walks through what each term means at audit-grade specificity, what the standards actually say about how each gets counted, and how to decide which tool fits which target. The insetting vs offsetting question is one of the most-searched in corporate climate strategy, and one of the most poorly answered. By the end of this piece, you should be able to brief a committee on the difference without notes.
The two definitions, in plain English
Offsetting means buying carbon credits generated outside your value chain and retiring them against your residual emissions. The reduction happens somewhere else, financed by you, and the credit is the receipt.
Insetting means investing in emission reductions or removals inside your own value chain, typically with suppliers, where the reduction is directly linked to the products and services you buy. The reduction happens inside the boundary of your Scope 3 inventory, and the accounting treatment is fundamentally different.
The shorthand from the University of Oxford’s Nature-based Insetting Initiative is useful: insetting is what you do with the supply chain you have; offsetting is what you do with the supply chain you do not have.
What the GHG Protocol Land Sector Standard actually says
The GHG Protocol Land Sector and Removals Standard, finalised in 2024 after a multi-year pilot, sets the rules for how land-based emission reductions and removals enter corporate inventories. The Standard distinguishes between inventory accounting (Scope 1, 2, and 3) and project or intervention accounting (a separate methodology for crediting).
For insetting, the practical implication is that supplier-level interventions, when properly measured and attributed, can reduce your Scope 3 category 1 (purchased goods and services) emissions in your inventory. The reduction is not a credit retired against the inventory; it is a lower inventory number, period.
For offsetting, the credit is retired separately. It can be reported as a contribution toward a net-zero claim under the SBTi Beyond Value Chain Mitigation framework or as part of a VCMI Carbon Integrity claim, but it does not lower the inventory number.
A practical consequence: if your Science Based Target requires a 50% absolute reduction in Scope 3 emissions by 2030, insetting moves you toward the target. Offsetting does not. This single point of difference reshapes the procurement decision.
When insetting counts toward Scope 3 (and when it does not)
Insetting counts toward Scope 3 only when several conditions are met:
- The intervention must occur with an entity in your value chain.
- The emissions reduction or removal must be measured against a defensible baseline.
- The reduction must be attributed to your share of that supplier’s output, not double-counted with other buyers.
- It must follow the inventory accounting rules in the GHG Protocol Land Sector Standard, not the project accounting rules used to generate credits.
The most common failure mode is double counting. If your supplier sells the same reduction as a credit on the voluntary market and also reports it to you as a Scope 3 reduction, the math breaks. The Standard requires you to address this risk, typically by purchasing and retiring the supplier-issued credit as part of your inventory or by contractual provisions that prevent the supplier from selling the reduction twice.
When insetting does not count toward Scope 3: when the intervention sits with a supplier you do not buy from, when the baseline is not defensible, when the attribution is unclear, or when the documentation does not survive audit. Those cases default to Beyond Value Chain Mitigation, which is still useful but operates on a different ledger.
The procurement and supplier engagement question
Insetting is harder than offsetting. That is the unfashionable truth most buyers eventually confront. Offsetting is a transaction; insetting is a relationship.
To run an insetting program, you need supplier mapping precise enough to know which farms or facilities sit at which Scope 3 boundary. You need an engagement model that gets suppliers to participate, which usually requires multi-year commitments and shared economics. You need an MRV architecture that measures the right things and produces audit-ready documentation. And you need a contractual structure that prevents double counting and protects both sides.
The trade-off you receive in return is significant. Reductions count against your inventory rather than your residual. Supplier relationships deepen, which protects sourcing continuity. Yield and quality improvements often follow regenerative interventions, which reduces your input cost over time. And the regulatory file, under CSRD, CSDDD, EUDR, and the SBTi FLAG Guidance, is materially stronger.
Choosing the right tool for the right target
A practical decision rule. If your target is a science-based Scope 3 reduction and you operate in a FLAG sector or source FLAG commodities, insetting is the structurally correct tool. If your target is a net-zero claim that includes neutralising hard-to-abate residual emissions outside your value chain, BVCM via high-integrity offsets is the structurally correct tool. Most companies with material Scope 3 exposure need both, in different proportions, sequenced over time.
The sequencing matters. Insetting takes longer to stand up but produces a permanent reduction in the inventory. Offsetting can be transacted faster but does not change the inventory and now sits under tighter claim restrictions. Treat them as complementary tools with different jobs, not as substitutes. The Accountability Framework Initiative and the IUCN Global Standard for Nature-based Solutions both provide useful guardrails for the insetting side, with biodiversity, human rights, and benefit-sharing requirements that go beyond carbon math.
If you are mapping a Scope 3 reduction roadmap and need to scope which interventions count toward your inventory versus which sit in Beyond Value Chain Mitigation, the carbon and sustainability experts at Carbon Credit Capital can help you structure a nature-based supply chain investment program that fits your FLAG exposure, your target architecture, and your audit horizon. Schedule a consultation.
Carbon Footprint
Net zero needs nature: a carbon credit guide
Net zero is often described as a balancing act: cut what you can, account for the rest, and reach zero on the ledger. That framing is useful, but it leaves something out. It treats every tonne of carbon as interchangeable and every route to zero as equally sound, while the science tells a more specific story.
![]()
Carbon Footprint
Deforestation in Malawi: causes and solutions
Malawi has lost a striking share of its forests over the past three decades. Woodlands that once covered well over a third of the country now cover less than a quarter, and the pressure on what remains is increasing. Behind those figures sit two practical questions: what is driving the loss, and what reverses it?
![]()
-
Climate Change1 year ago
Guest post: Why China is still building new coal – and when it might stop
-
Greenhouse Gases1 year ago
Guest post: Why China is still building new coal – and when it might stop
-
Greenhouse Gases2 years ago嘉宾来稿:满足中国增长的用电需求 光伏加储能“比新建煤电更实惠”
-
Climate Change2 years ago嘉宾来稿:满足中国增长的用电需求 光伏加储能“比新建煤电更实惠”
-
Climate Change2 years ago
Bill Discounting Climate Change in Florida’s Energy Policy Awaits DeSantis’ Approval
-
Renewable Energy10 months agoSending Progressive Philanthropist George Soros to Prison?
-
Greenhouse Gases1 year ago
嘉宾来稿:探究火山喷发如何影响气候预测
-
Carbon Footprint2 years agoUS SEC’s Climate Disclosure Rules Spur Renewed Interest in Carbon Credits

