MIT Technology Review’s Climate Op-Ed Turns a 49-Actor AI-Assisted Roster Into a Verdict on a 2035 Projection, but That Projection Has No 2026 Checkpoint, Its Power Lever Is a 20% Load-Factor Assumption Retrieved From a Blog, and Several Examples Drift From Their Sources

On November 9, MIT Technology Review published “The climate case for AI has not been made,” an opinion piece by Mary Nichols, a former chair of the California Air Resources Board, and Jeremy Tamanini, founder of the sustainability advisory Dual Citizen. It opens with the UN climate conference in Antalya, where the authors expect AI to be a main topic, and it argues in three steps. AI leaders say the technology will fix the climate. A peer-reviewed 2025 study from the London School of Economics’ Grantham Research Institute and Systemiq put the possible savings at 3.2 to 5.4 billion tons of carbon dioxide equivalent a year by 2035 in three sectors. And an AI-assisted review of 49 market actors in those sectors found that few are doing what the study’s numbers need.

The review is unusually easy to check. Tamanini published it on September 14 with a methodology and the full roster, and the page embeds the data behind its tables. I read the essay, the study and its Technical Annex, the roster page and the data inside it, and the sources behind nine of the essay’s specific details, and I wrote a short standard-library script that recounts the roster and redoes the study’s arithmetic. The review was run with Anthropic’s Claude Code, and this article was researched with a Claude model too, so I checked each figure against its own source rather than against what I expected a model to say.

Here is what I found:

• The count reproduces. The roster’s own data hold 49 actors and put 6 at Rank 1, deployed and quantified: 1 of 27 grid operators, 3 of 14 alternative-protein firms and 2 of 8 vehicle firms. The page also says only 13 of about 26 EU grid operators on its original list were assessed, and that a “not found” should not be read as “confirmed absent.”
• The projection has no 2026 checkpoint. The study defines its AI scenario by “assessing changes in enabling conditions for both scenarios in 2035.” Its annex says “the time dimension of the S-curve is normalised” and never states what should be visible in 2026. A 2026 census is a gap analysis, not a test of whether the path is on track.
• The power lever is a flat assumption. The study’s 1.8 billion tons from power is 56 percent of its low end. The annex gets there by assuming a 20 percent load-factor gain across the whole solar and wind fleet and cites NREL in a 2023 reference. The only retrieval link is an AI company’s blog post about wind turbines, and I could not find an NREL publication that states the figure.
• The top end needs a different scenario. 5.4 is reached only by adding the highly ambitious alternative-protein case, 3.0 billion tons, which assumes cost parity with meat and dairy in all categories. That one lever is 56 percent of the top end, and the essay’s own funding data show the cultivated-meat part of that sector shrinking.
• The one quantified grid result counts a different thing. Open Climate Fix says its tool at NESO cuts about 300,000 tonnes a year by reducing reliance on fossil-fuel peakers. The study’s arithmetic prices more output from the same panels and turbines. Both are real effects of better forecasting; they are not the same quantity.
• Some illustrations drift. The Didi study compares pooled rides with solo rides and uses machine learning as the researchers’ tool, not as Didi’s routing. The CATL release the essay links has no screening-time figure. The Kraft Heinz venture with NotCo was announced in February 2022. And the study sets AI’s footprint beside its reductions instead of subtracting it.
• A sister paper points the same way by a different method. A model paper in the same journal, published August 4, finds net emissions rising by 0.47 to 1.8 billion tons under parallel adoption. The essay cites it for one clause about oil and gas.

None of this overturns the essay’s direction. The cultivated-meat funding series holds, the roster’s count reproduces, and the sister paper points the same way. What does not hold up is the confidence of the design: a census of 2026 deployments is presented as a test of a 2035 number that has no 2026 checkpoint, and several of the examples that make the essay vivid say less, or something else, at their sources.

The roster is public, and its count reproduces

Dual Citizen’s page, “Backcasting 2035 AI-Induced Emission Reduction Projections (with Claude Code),” states the method plainly. For each of the study’s three sectors, Tamanini asked Claude to identify the “most significant market actors” in China, the United States and the European Union whose AI use the 2035 numbers would require. Each actor then got a rank: 1 for “AI deployed & quantified,” 2 for “real but partial or adjacent,” and 3 for “no evidence found, or deprioritized.” The page embeds the full tables as JSON, so the count can be redone in a few lines. Table 1 is that recount, and the page’s own sentence, “only 6 of the 49 total market actors ranked 1,” matches it.

Table 1. The roster recount, from the data embedded in the page (Rank 1 means deployed and quantified).

What the page says about its own limits

The methodology section is franker than the essay. It says “Roster coverage is not exhaustive in any sector,” that “~13 lower-priority EU TSOs, Panasonic and Lyft were only lightly checked,” and that nearly all of its deployed findings rest on “self-reported claims by the organizations involved, not independent audits.” It adds that “not found” is not “confirmed absent,” and calls that “a materially weaker claim for entities with limited English-language press coverage,” naming several EU grid operators and US federal power authorities. The coverage figures are on the page too: both Chinese grid companies, 12 of 12 in-scope US operators, and 13 of about 26 originally listed EU transmission operators. One of the 14 alternative-protein rows is a sector-wide line, “Cultivated meat sector generally,” not a company.

The essay’s version of all this is one sentence: “It is possible that some companies and utilities in these industries are pursuing AI-assisted applications that would fit into the initial study’s scenarios but haven’t yet been disclosed publicly.” The essay’s reader gets “only one” grid operator and that hedge, and does not get the denominators.

Deployed is not the same as quantified

Rank 1 means deployed and quantified, and only one grid operator clears it. But eight of the 27 power rows carry a status that begins “Deployed,” and seven of those are ranked 2. They include State Grid, marked “Deployed (nuanced),” and China Southern Power Grid, marked as a recent deployment of an AI grid-planning suite unveiled in July 2026. Those are the two entries with the largest electricity volumes stated in terawatt-hours on the page, about 5,600 and about 1,000 to 1,100 a year of sales, both marked approximate, against about 300 for NESO. The volume column mixes units across rows (terawatt-hours, gigawatts of peak load, customers, population), so I did not weight by it. The point is smaller: what the roster records for the largest systems is deployment without published results, which is a different finding from an absence of deployment.

The two texts by the same author also weigh the same three alternative-protein Rank 1 rows differently. The roster page says the sector “exhibits several strong, commercially deployed AI integrations,” while the essay says “few other alternative-protein companies had such a deeply integrated way of using AI in product development.” Both can be true of 3 in 14; they read differently. (The page also prints the study’s range as 3.2 to 5.2 in its opening paragraph and 3.2 to 5.4 in its methodology; the paper says 5.4.)

What the 3.2 to 5.4 range is made of

The study’s main text prints one number for power, a range for each of the other two sectors, and a headline range for the total. Table 2 puts them together. The sums reproduce the printed endpoints exactly, which is how I know which scenario each end belongs to. The shares are my arithmetic.

Table 2. The headline range rebuilt from the study’s printed components (billion tons of CO2e a year in 2035).

With the AI scenario’s own top for meat and dairy, 1.6, the same sum is 4.0. The paper’s headline spans both scenarios, and the essay’s “3.2 to 5.4” does not say so.

The top end needs alternative-protein cost parity

The study, “Green and intelligent: the role of AI in the climate transition,” published in npj Climate Action on June 23, 2025, estimates adoption of alternative proteins at roughly 8 to 14 percent in the business-as-usual case, 18 to 33 percent in an ambitious AI scenario and 27 to 50 percent in a highly ambitious one. The paper assumes “AI improves attractiveness (taste and texture) through identifying proteins with suitable properties.” Only in the highly ambitious case does it “also significantly” reduce “production costs to the point of achieving cost parity with traditional products in all categories and products.” That case supplies the 3.0, and the 3.0 supplies more than half of the 5.4.

This is the lever the essay says is stalling, and the essay’s evidence for that holds. AgFunderNews, on June 22, 2026, wrote that cultivated-meat funding “peaked at $989 million in 2021,” fell to $55 million in 2024, and “edged up to $82.6 million in 2025,” and that Believer Meats and Meatable ceased operations in December. The essay’s “around $83 million last year” matches; the low point was 2024, and 2025 was an uptick. One scope note: the study’s alternative proteins are a broad category, and NotCo, the roster’s strongest example, is plant-based, so a cultivated-meat funding collapse bears on the lever without being the whole of it.

The power lever is a flat 20 percent assumption

The study’s main text says AI can improve renewable systems “by optimising grid management and increasing the load factor of solar PV and wind by up to 20%,” which would cut 1.8 billion tons a year by 2035. The essay restates it as AI-driven forecasting raising the amount of renewable electricity that reaches the grid from already-built plants “by up to 20%.” The Technical Annex shows how the number is used, and “up to” is gone. The annex assumes the load factor of solar PV and wind improves between 2022 and 2035 in the AI case, so that “more power is being generated from the same capacity of solar PV and wind,” “by 20% (NREL, 2023).” It holds load factors flat in the business-as-usual case, sends all AI-driven efficiency gains to renewables (they “are assumed to be directed towards renewables”), removes the resulting oversupply and recalculates emissions. The 20 is a fixed input applied to the whole solar and wind fleet.

The citation is thin. The annex’s reference reads “Paper on wind power efficiency from AI,” attributed to NREL and dated 2023, and its only retrieval link is a January 2024 blog post by the AI company Datategy on wind turbine optimization. That post discusses turbine settings such as blade pitch and yaw angles and says “AI can improve the efficiency of wind turbines by up to 20%,” followed by “Study by the National Renewable Energy Laboratory (NREL),” with no title or link. So the figure is a wind-turbine efficiency claim from an unnamed study, and the annex applies it to solar PV as well, as a load-factor gain for the whole fleet. I searched for an NREL source and did not find one; the closest item the search surfaced, a May 2024 NREL study on wake steering (as the search summary describes it), reports land-use savings of 18 percent on average, not a 20 percent efficiency gain. nrel.gov did not resolve from this machine, so that part rests on the search summary.

For scale, 1.8 billion tons is 11.8 percent of the 15.3 billion tons of 2023 power-sector emissions the study prints, and 16.5 percent of the roughly 10.9 billion tons of 2035 power-sector emissions that the annex says its business-as-usual case tracks (my arithmetic). One input produces a cut of that size, and the annex’s uncertainty ranges come from the S-curve ceiling for adoption; I found no variation of the 20 percent. The study’s data are “not openly available due to reasons of sensitivity,” so I could not rerun its model to see how the 1.8 moves with the 20.

“Even after subtracting” is not what the paper prints

The essay says the 3.2 to 5.4 billion tons hold “even after subtracting AI’s own energy footprint.” The paper sets the two side by side. Its abstract says the reduction estimates are “compared with the increase in data centre-related emissions generated by all AI-related activities,” and its results say “these estimated emissions reductions outweigh” a separate increase of 0.4 to 1.6 billion tons from data centres and AI. No net figure is printed, and the discussion says macroeconomic outcomes “require further research which must consider the compounding, systemic and intersectoral net impact of AI.” The two sides also use different boundaries: the reductions cover three sectors, and the paper says they “may be subjected to underestimation since the analysis is only based on three sectors,” while the increase covers all AI activity, “not just those related to decarbonisation.”

If you do subtract, the extremes are 3.2 minus 1.6, or 1.6, and 5.4 minus 0.4, or 5.0 (my arithmetic, in the script below). That changes the comparison the essay draws with the US figure. EIA puts US energy-related emissions in 2025 at “about 4,904 million metric tons.” The gross 5.4 is 110 percent of that, the AI scenario’s top of 4.0 is 82 percent, and the best-case net of 5.0 is 102 percent. The essay’s phrase, roughly on par on the high end, is fair for the net figure and about 10 percent low for the gross one.

A 2026 census cannot test a 2035 end state

The study states 2035 end points, not a dated path for the AI levers. It says it assumes “technology adoption follows a Sigmoid-curve (S-curve),” and it defines each scenario by “assessing changes in enabling conditions for both scenarios in 2035.” The annex builds the curves from a maximum adoption level, a growth rate and an inflection point, and says “the time dimension of the S-curve is normalised,” so a technology’s maturity is a number between 0 and 1, not a date. For the power sector the growth rate is calibrated on historical adoption of technologies such as “liquid biofuels and nuclear for the power sector,” and for the power lever the AI effect is not on an S-curve at all: it is a single load-factor input between 2022 and 2035.

I searched the annex for calendar years. 2025 and 2027 to 2029 never appear. The lone 2026 is in a reference title, the IEA’s “Electricity 2024” forecast to 2026. The 2030 mentions concern food-consumption data years and the data-centre growth rate. There is no interim target, no date by which any AI lever should be visible. The one dated path I found is for the baseline build-out: the annex assumes the adoption rate of solar PV and wind “follows a linear trajectory between the 2022 and 2035 values,” which describes how much capacity exists, not what AI adds to it.

That matters for how to read the roster. The page itself says its value is “to demonstrate the current state of deployed and quantified AI integration among market actors,” and that is what it does. But 12 percent of actors at Rank 1 cannot say whether the world is early, on time or late against a curve with no published dates, and the same limit applies in the other direction: nobody can use the census to show the projection is on track. The clearest trend the essay uses is cultivated-meat funding, and it is also the least ambiguous evidence in the piece, because funding is a flow that can be measured year by year.

The one Rank 1 grid result counts a different thing

The sole Rank 1 power actor is NESO, the National Energy System Operator, which the roster lists as “National Grid ESO / NESO.” The roster cites “Open Climate Fix case study page (primary vendor source)” and a NESO document from February 2026 that I did not read. Open Climate Fix’s post of November 6, 2025 says its forecasting tool is “helping avoid an estimated £30 million in imbalance costs every year” and that “Enhanced short-term accuracy reduces reliance on fossil-fuel peakers, cutting around 300,000 tonnes” of CO2 annually. It adds: “Quartz Solar has halved large errors, and is 2.8 times more accurate than previous solar generation forecasting tools used by NESO.”

Three things follow. The essay says the partnership “halved the grid operator’s solar-forecast error”; the vendor says it halved large errors, and the roster note says “Halved solar forecast errors,” so both dropped a word. The tonnage is the vendor’s estimate; I found no NESO-published method for it. And the channel is fewer fossil peakers, which the roster describes as feeding “a Dynamic Reserve Setting model,” not more output per panel. The roster page defines its power test as AI “for increasing the load factor of solar PV and wind,” while its column header reads “forecasting/DER management”; NESO’s evidence fits the header, not obviously the sentence. The study’s annex prices the load-factor version.

For scale only, 300,000 tonnes is about one six-thousandth of 1.8 billion tons (my arithmetic). One national operator is not the world, and the point is not that the result is small. It is that the roster’s best evidence and the study’s lever are measured in different units, so a Rank 1 count cannot be added up toward the 1.8.

Where the essay’s illustrations drift from their sources

Table 3 checks the essay’s specific statements against the sources it links, or against the closest primary source I could read. Quotation marks around the essay’s words are exact; the right-hand column is my judgment.

Table 3. Claim-by-claim check.

Didi: pooled rides, not routing

The essay links a paper in Travel Behaviour and Society. ScienceDirect blocks scripts, so I read the authors’ arXiv version, “Revealing the CO2 emission reduction of ridesplitting and its determinants based on real-world data,” whose title differs from the published one. OpenAlex lists the published paper’s authors at four Chinese universities and none at Didi, and the ride and GPS data are “provided by the GAIA program” from Didi’s Chengdu service. The study computes the CO2 of each shared ride and of the “substituted single rides” that riders would otherwise have taken, and compares them. Its Table 3 gives a mean saved distance of 2.11 kilometers and a mean emission reduction of 16.49 percent per trip, with a standard deviation of 15.58 and a minimum of minus 27.26. The abstract says “not all ridesplitting trips reduce emissions from ridesourcing in the real world,” and the highlights say “15% of ridesplitting trips even increase CO2 emissions compared with regular ridesourcing.”

So 16 percent is a property of pooling rides, and the machine learning is the authors’ way of explaining which trips gain, using “interpretable machine learning models, gradient boosting machines.” Overlap and detour rates matter most. “AI-optimized routing to minimize detours” turns a finding about detour rate into something Didi did. The roster’s own Didi row says the result is not company-disclosed and names the tools as academic models applied to Didi’s public trip data. It also cites a separate Beijing study that, by the roster’s note, concerns AI-driven order matching; the essay does not use it, and I did not check it.

CATL: a figure the linked page does not contain

The essay links CATL’s January 19, 2026 release on a World Economic Forum award. It matches the essay on training data: “Trained on more than 100,000 battery design cases, 600TB of test data.” On speed it says the platform “generates design recommendations in seconds and virtual cell designs in minutes” with “prediction accuracy of up to 95 percent.” It has no screening-time figure and no orders of magnitude. The roster’s CATL row cites that release and a May 2026 interview with CATL’s chief scientist, and its note reads “7 target materials identified in 3 months with improved ionic conductivity.” I did not find “two orders of magnitude” in the release, the roster, a September 2024 Caixin summary of CATL’s AI materials work, or a web search, and I could not find the interview. The figure may exist in a source the essay does not link; the essay attributes it to “the company,” so a reader should be able to check it, and from the essay’s link cannot.

The sister paper counts the other half of the ledger

The essay links a second npj Climate Action paper, published August 4, 2026 by Will Alpine and colleagues, for the sentence that oil and gas companies are also interested in these tools. Its title says AI-driven productivity gains enable more CO2 emissions than they avoid in a global energy-economy model, and the abstract gives the sizes: under parallel adoption scenarios, net annual emissions rise by 0.47 to 1.8 billion tons, or 1.2 to 4.8 percent of 2024 global energy-related emissions. The results section calls parallel adoption “a neutral benchmark representing all pathways at the same adoption tier.” “Enabled emissions exceed avoided emissions whenever fossil-sector gains are nonzero,” and breaking even needs renewables gains four to five times larger than fossil gains. The model is a computable general equilibrium model that treats AI as productivity shocks on competing energy pathways. Its authors say the results “describe the direction and approximate scale of economic incentives under current conditions, not time-resolved forecasts,” and that real-world outcomes are “likely to differ in magnitude while maintaining directionality.”

The two papers answer different questions, and each states its boundary. The 2025 study says “Since we are not assessing the impact of rebound effects, we have also not considered any effects of AI on emission-intensive activities,” and its annex sends all AI efficiency gains to renewables. The 2026 paper models gains on both sides and reports data-centre emissions separately, for reference. Both are scenarios built from assumed productivity gains, not measurements.

This is the essay’s strongest support, and it is cited in passing. It also exposes the asymmetry of the census: the roster asks only whether clean-sector actors deploy AI on the study’s routes. There is no counterpart census of oil and gas operators, although the essay itself says they are interested. The four-to-five-times threshold is the number to watch, because it states what has to be true for the sign to be negative. Our earlier posts on Microsoft’s emissions jump and on Google’s nuclear uprate deal cover the footprint and power-supply side of the same ledger.

What a fair test of the projection would need

• Interim checkpoints. Dates derived from the curves, such as the load-factor or curtailment trend a 20 percent gain implies by 2030, or adoption and cost-parity milestones for alternative proteins, so that a 2026 census has something to be compared with.
• The lever’s own unit. If the lever prices output from the same capacity, count output (availability, curtailment, load factor), not forecast error or reserve held.
• Both sides. A census of fossil-side AI deployment next to the clean-side one, since the sister paper says the sign depends on the ratio.
• One footprint basis. Reductions and AI’s own emissions on the same boundary and the same grid-intensity assumption, with a printed net.
• Operator data over vendor estimates. A tonnage from the vendor of the tool is an estimate; operator-published forecast and outturn data, where they exist, are a measurement.

If you cite these numbers

If you write sustainability or IT-strategy documents that lean on AI-climate figures, a few habits would have caught most of what is above. Call 3.2 to 5.4 a scenario-conditional, gross figure for three sectors in 2035, not a forecast and not a net, and name the scenario when you quote the top end. When a claim says AI cuts emissions through a grid or a vehicle fleet, name the channel (fewer peakers, higher output, shared rides) and check that the source measures that channel. When a sentence says only N of M, ask for M and for the rule that chose it. And treat a vendor’s tonnage as an estimate until an operator publishes the method. We applied the same approach to other opinion pieces in the refusal essay, the agentic-web essay and the Apollo productivity note.

Reproduce the recount

The script below needs Python 3.13 and nothing else. It fetches the roster page, pulls out the three embedded tables, recounts the ranks and then redoes the arithmetic from the figures printed in the study and from EIA’s 2025 figure. It needs the network; the page could change, and the output below is from November 9, 2026.

# roster_check.py
"""Recount the roster behind the MIT Technology Review climate op-ed and redo the
arithmetic on the headline range of the npj Climate Action study it tests.
Standard library only. Written for Python 3.13."""
import json
import urllib.request
from collections import Counter

PAGE = ("https://dualcitizeninc.com/backcasting-projections-for-2035-ai-induced-"
        "emission-reductions-in-partnership-with-claude-code/")
SECTORS = ["Power (grid operators)", "Meat and dairy (alt-protein)", "Light vehicles"]


def fetch(url):
    request = urllib.request.Request(url, headers={"User-Agent": "Mozilla/5.0"})
    with urllib.request.urlopen(request, timeout=60) as response:
        return response.read().decode("utf-8", "replace")


def embedded_tables(html):
    """Every JSON array of objects embedded in the page, in page order."""
    decoder = json.JSONDecoder()
    found, pos = [], 0
    while (start := html.find('[{"', pos)) >= 0:
        try:
            value, length = decoder.raw_decode(html[start:])
        except ValueError:
            pos = start + 3
            continue
        if isinstance(value, list) and len(value) >= 5 and all(isinstance(r, dict) for r in value):
            found.append(value)
        pos = start + length
    return found


def column(row, prefix):
    return next(v for k, v in row.items() if k.startswith(prefix))


def name(row):
    return row.get("Operator") or row.get("Company")


def main():
    tables = [t for t in embedded_tables(fetch(PAGE)) if "Operator" in t[0] or "Company" in t[0]]
    assert len(tables) == 3, len(tables)

    print("=== 1. Roster recount (rank 1 = deployed and quantified)")
    total = Counter()
    for label, rows in zip(SECTORS, tables):
        ranks = Counter(column(r, "AI Integration Rank") for r in rows)
        total.update(ranks)
        top = ", ".join(name(r) for r in rows if column(r, "AI Integration Rank") == 1)
        print(f"{label:<30} actors={len(rows):>2}  rank1={ranks[1]}  rank2={ranks[2]:>2}  "
              f"rank3={ranks[3]:>2}  rank 1: {top}")
    count = sum(total.values())
    print(f"{'All sectors':<30} actors={count:>2}  rank1={total[1]}  rank2={total[2]:>2}  "
          f"rank3={total[3]:>2}  rank-1 share={total[1] / count:.1%}")

    power = tables[0]
    deployed = [r for r in power if column(r, "AI adoption status").startswith("Deployed")]
    print(f"\nPower rows whose status text starts with Deployed: {len(deployed)} of {len(power)}")
    for r in deployed:
        print(f"  rank {column(r, 'AI Integration Rank')}  {name(r)}")

    print("\n=== 2. The study's headline range rebuilt from its printed components (Gt CO2e a year, 2035)")
    low = {"power": 1.8, "meat and dairy": 0.9, "light vehicles": 0.5}
    high = {"power": 1.8, "meat and dairy": 3.0, "light vehicles": 0.6}
    for label, parts in (("low end ", low), ("high end", high)):
        whole = sum(parts.values())
        shares = ", ".join(f"{k} {v / whole:.0%}" for k, v in parts.items())
        print(f"{label}: {' + '.join(str(v) for v in parts.values())} = {whole:.1f}   shares: {shares}")
    print(f"AI scenario only (meat and dairy 1.6 at most): 1.8 + 1.6 + 0.6 = {1.8 + 1.6 + 0.6:.1f}")

    print("\n=== 3. Reductions against the 0.4 to 1.6 Gt increase the paper prints separately")
    print(f"worst pairing: 3.2 - 1.6 = {3.2 - 1.6:.1f}")
    print(f"best pairing:  5.4 - 0.4 = {5.4 - 0.4:.1f}")
    us_2025 = 4.904  # EIA: U.S. energy-related CO2 in 2025, billion metric tons
    for label, value in (("5.4 gross", 5.4), ("4.0 AI-scenario top", 4.0), ("5.0 best net", 5.0), ("1.6 worst net", 1.6)):
        print(f"{label:<20} = {value / us_2025:>4.0%} of the 2025 U.S. figure (4.904)")

    print("\n=== 4. Size of the power lever against the paper's own bases")
    print(f"1.8 / 15.3 (2023 power-sector emissions)        = {1.8 / 15.3:.1%}")
    print(f"1.8 / 10.9 (2035 business as usual, IEA STEPS)  = {1.8 / 10.9:.1%}")


if __name__ == "__main__":
    main()

The output:

=== 1. Roster recount (rank 1 = deployed and quantified)
Power (grid operators)         actors=27  rank1=1  rank2=11  rank3=15  rank 1: National Grid ESO / NESO
Meat and dairy (alt-protein)   actors=14  rank1=3  rank2= 3  rank3= 8  rank 1: NotCo, Climax Foods, Shiru
Light vehicles                 actors= 8  rank1=2  rank2= 4  rank3= 2  rank 1: CATL, Waymo
All sectors                    actors=49  rank1=6  rank2=18  rank3=25  rank-1 share=12.2%

Power rows whose status text starts with Deployed: 8 of 27
  rank 2  State Grid Corporation of China (SGCC)
  rank 2  China Southern Power Grid (CSG)
  rank 2  ISO-NE (ISO New England)
  rank 2  RTE (Reseau de Transport d'Electricite)
  rank 2  Terna
  rank 1  National Grid ESO / NESO
  rank 2  APG (Austrian Power Grid)
  rank 2  CEPS

=== 2. The study's headline range rebuilt from its printed components (Gt CO2e a year, 2035)
low end : 1.8 + 0.9 + 0.5 = 3.2   shares: power 56%, meat and dairy 28%, light vehicles 16%
high end: 1.8 + 3.0 + 0.6 = 5.4   shares: power 33%, meat and dairy 56%, light vehicles 11%
AI scenario only (meat and dairy 1.6 at most): 1.8 + 1.6 + 0.6 = 4.0

=== 3. Reductions against the 0.4 to 1.6 Gt increase the paper prints separately
worst pairing: 3.2 - 1.6 = 1.6
best pairing:  5.4 - 0.4 = 5.0
5.4 gross            = 110% of the 2025 U.S. figure (4.904)
4.0 AI-scenario top  =  82% of the 2025 U.S. figure (4.904)
5.0 best net         = 102% of the 2025 U.S. figure (4.904)
1.6 worst net        =  33% of the 2025 U.S. figure (4.904)

=== 4. Size of the power lever against the paper's own bases
1.8 / 15.3 (2023 power-sector emissions)        = 11.8%
1.8 / 10.9 (2035 business as usual, IEA STEPS)  = 16.5%

Section 1 is Table 1. Sections 2 to 4 are Table 2, the net-of-footprint comparison and the size of the power lever; the study’s figures are typed in as printed, and the script adds nothing to them but sums and ratios.

What I could not check

• ScienceDirect, the IEA site and NREL’s site were not readable from this machine (a bot block for the first two, a DNS failure for NREL), so the published Didi paper, the IEA’s 2035 figure and NREL’s own material rest on the arXiv preprint, secondary coverage and search summaries.
• I did not find the CnEVPost interview the roster cites for CATL, and I did not read the NESO document it cites for NESO.
• The study’s data are not open, so I could not rerun its model or vary the 20 percent input.
• I read the rows I name here, not all 49. The roster is a snapshot, and its authors may update it.
• The shares, the net pairings, the ratios to the US figure and the six-thousandth comparison are my arithmetic on printed figures.

Sources

• Mary Nichols and Jeremy Tamanini, The climate case for AI has not been made, MIT Technology Review, November 9, 2026.
• Jeremy Tamanini, Backcasting 2035 AI-Induced Emission Reduction Projections (with Claude Code), Dual Citizen, September 14, 2026, with the roster data embedded in the page.
• Green and intelligent: the role of AI in the climate transition, npj Climate Action 4, article 56, June 23, 2025, and its Technical Annex.
• AI-driven productivity gains enable more CO2 emissions than they avoid in a global energy-economy model, npj Climate Action 5, article 71, August 4, 2026.
• Open Climate Fix, National Energy System Operator adopts AI solar forecasting, November 6, 2025.
• CATL, CATL wins World Economic Forum’s MINDS Award, January 19, 2026, and Caixin, CATL is using AI to discover next-generation battery materials, September 20, 2024.
• Li and colleagues, Revealing the CO2 emission reduction of ridesplitting and its determinants based on real-world data, arXiv 2204.00777 (published in Travel Behaviour and Society, DOI 10.1016/j.tbs.2023.100734).
• AgFunderNews, After the cultivated meat crash, who’s still standing?, June 22, 2026, and FoodNavigator, Kraft Heinz creates joint venture with AI-powered foodtech startup NotCo, February 22, 2022.
• Kate Brandt and Rich Lesser, AI could accelerate progress toward the world’s climate goals, Fortune, December 18, 2023.
• U.S. Energy Information Administration, What are U.S. energy-related carbon dioxide emissions by source and sector? and the IEA’s Energy supply for AI page the essay cites (not read directly).
• Datategy, Wind turbine optimization with AI, January 17, 2024.