Net Zero, Counted Twice

The UK Conservative party's "Right Way" policy booklet says the Climate Change Committee (CCC) estimated in 2020 that its route to net-zero would cost £957bn. Carbon Brief's Simon Evans finds that the CCC's figure was exactly half that, £478bn. The larger number comes from a spreadsheet error that counted every value twice.

How the double-count happened

The CCC's 2020 spreadsheet, which supports figure 5.3 of its sixth carbon budget advice, lists the extra investment needed in each sector for 2020-50, covering low-carbon power, heat pumps, EVs and more. Over the period, these CAPEX lines add up to £1.38tn. They are set against OPEX savings of £0.90tn, mainly from lower spending on fossil fuels. The net result is £478bn.

The sheet also has a separate line giving the combined total for each year. If you add up every row, including that totals line, each cost and each saving is counted twice. The result is £957bn. In 2020 the CCC said the £478bn net cost was under 1% of GDP over 30 years. It also said the investment would raise GDP by about 2% overall.

The figure has since fallen. In 2025 the CCC revised its estimates to £670bn of investment and £562bn of operating savings for 2025-50. That gives a net cost of £108bn, under 0.2% of GDP. The drop is mostly due to cheaper EVs. Earlier this year the committee said reaching net-zero would cost less than a single fossil-fuel price shock and would bring benefits worth £110bn a year.

Wood Mackenzie published an analysis Thursday finding that the levelized cost of electricity (LCOE) for four-hour battery storage is now lower than that of gas peakers in markets around the world. The firm did not release exact prices. A spokesperson told Utility Dive that for US projects starting commercial operation in 2026, four-hour storage is 65% to 75% cheaper than new open-cycle gas turbine peakers. Where the figure falls in that range depends on whether state-level carbon pricing applies.

Why the gap is widening

The two costs are moving in opposite directions. On the gas side, WoodMac said in April that turbine prices would reach $600/kW by the end of 2027, up 195% since 2019. GE Vernova, Siemens Energy and Mitsubishi each have backlogs of 35 GW to 116 GW. WoodMac expects North American gas capacity investment to stay in "a supply deficit cycle through the late 2030s," driven by data center load growth, which keeps thermal capital costs high. Principal analyst Ahmed Jameel Abdullah also cites rising fuel volatility. On the battery side, growing manufacturing capacity keeps pushing storage costs down.

Manufacturing scale shows up clearly in the regional numbers. Storage in China costs more than 55% less than the rest of the Asia Pacific average. In the Middle East and Africa, WoodMac forecasts four-hour storage will fall 33% to $80/MWh by 2035, beating gas peaking on cost in every gas market in the region.

Baseload and the US policy picture

A bipartisan group of four senators released the Bipartisan American Affordability and Jobs Act (BAAJA) just over a week ago. Heatmap's Robinson Meyer says the bill produces a messy set of winners and losers:

• Losers: utilities, data centers and the Trump administration.
• Winners: electricity ratepayers, clean energy companies, long-distance transmission and natural gas pipeline builders. Geothermal startups and virtual power plant providers also gain.

Who backs it

The building-trades unions mostly support the bill. The International Brotherhood of Electrical Workers, which often sides with utility executives, has said nothing. Clean energy and fossil fuel trade groups both back it, including the American Clean Power Association and the American Petroleum Institute.

Environmental groups are split. The Sierra Club published a "hard look," and the NRDC posed "hard questions." Both are critical, neither rejects the bill outright, and both signal that more statements are coming.

The Sierra Club sorts the bill into "green flags" and "red flags."

• Green flags: easier planning for long-distance power lines, protection for ratepayers from utility and data center "freeloading," and clearer rules on which federal agency can approve lines.
• Red flags (four in total): the "hollowing out" of court authority over some permits, removal of a Clean Water Act provision that lets governors block pipelines and power lines, and the option to hand partial Endangered Species Act enforcement to states.

Constellation Energy, the largest U.S. nuclear operator, will add 890 MW of capacity at 11 existing plants in the PJM Interconnection under a 20-year deal with Google. PJM is the country's biggest electricity market and its most short of capacity. The work is a set of power uprates: replacing turbines, steam generators, digital control systems and other non-nuclear equipment so the same reactors produce more electricity.

The first Google-backed uprate should be finished by 2028, with the rest done in the early 2030s. The largest single project adds 348 MW to the 2,317-MW Limerick plant near Philadelphia. The 2,327-MW Salem plant in New Jersey gets 97 MW. Constellation says the scope "varies plant to plant."

The cost math

The program is budgeted at more than $4.3 billion, about $4,800 per kW. That is well above utility-scale solar with storage, a new combined-cycle gas plant, or Constellation's own $1.6 billion restart of the Three Mile Island unit (now Crane). It is less than half the per-kW cost of Vogtle Units 3 and 4, which came in seven years late and at more than double their original $14 billion budget.

What the money buys is reliability. DOE puts the average U.S. nuclear capacity factor at 92.3%, against 59.9% for combined-cycle gas, 34.3% for wind, 23.4% for solar and 17.2% for gas peakers. PJM expects demand to grow more than 5% a year for the next decade and retail prices to rise as much as 60% by 2030, and says it welcomes the capacity even at a premium. Renewables trade group MAREC Action also backed the deal, pointing to the long-term contract model its members already use.

Water may be the data center fight where both sides talk past each other most, Shayle Kann argues on Latitude Media's Catalyst podcast. He offers some scale: a 300 MW data center with evaporative cooling uses about 150 million gallons a year. A combined-cycle gas plant of the same capacity uses roughly three times that, and 150 million gallons is about what 550 acres of irrigated soybeans need. His guest, thermal engineer Tim Shedd of Ebullient Insights, explains why older consumption figures say little about new builds.

Why air cooling ran out

Racks used to average about 9 kW, which air handlers at the edges of the room could cool. In-row coolers and rear-door heat exchangers brought cooling closer to the servers, but they top out around 80 kW per rack. Nvidia's NVL72 is rated at 220 kW. Shedd recalls one supercomputer rack at Lawrence Berkeley at about 280 kW, though he isn't sure of the figure, and engineers have mapped designs to 400–500 kW. Racks are packed this densely to shorten the expensive, high-speed links between GPUs.

The dominant answer is single-phase direct-to-chip cooling. Water mixed with 25% propylene glycol, plus a 1–2% corrosion-inhibitor package, runs through copper cold plates. The plates have fins about 100 microns thick and 100 microns apart. This setup cools a 1,400–1,800 W processor with 45°C water, which wasn't possible three years ago. Two-phase systems boil a refrigerant off the chip. They perform about the same but use smaller plumbing, and they aren't yet available at scale. Oil immersion transfers heat poorly, and two-phase immersion is limited to roughly 1,000 W chips by vapor-pressure limits. Cold plate output has gone from about 100,000 a year in 2022 to hundreds of thousands a month, and could approach a million a month next year.