The real price of AI never shows up in the demo

0.142 against 70

The UK government put the CO2 emissions expected from new AI datacenters at 0.142 megatonnes per year. Carbon Brief redid the math in March 2026 and found realistic scenarios ranging from 2 to 70 megatonnes. That is a factor of ten to five hundred. Shortly after the analysis appeared, the UK Department for Innovation quietly removed its original figure from the official site.

The same day, May 18, 2026, The Register documented an Australian developer who received a $10,000 Google Cloud bill despite setting his budget cap at $250. Another user got hit with $127,000. An AWS Bedrock customer who expected a few hundred dollars saw $38,000 instead.

Two facts, same mechanic. What does not show up in the demo always shows up somewhere else.

Why the UK is firing up the gas

More than 100 UK datacenter projects have asked to connect to the gas network so they can generate their own power on site. Total request: roughly 15 TWh per year.

The reason is operational. Grid connection delays on the UK electricity network now run to five to ten years. Operators cannot wait, so they put small gas plants next to the building.

The IEA published its global projections in 2025: datacenter consumption will more than double by 2030, reaching about 945 TWh. That is slightly more than Japan's entire current electricity use. In the United States, these facilities are expected to absorb close to half of the growth in electricity demand over the same period. The mix being built to feed them is a blend of renewables and natural gas, because those two sources are available and competitive in the near term.

That is where the 0.142 versus 70 gap starts to mean something. The government calculation assumed a near-fully decarbonized UK grid by 2035. Carbon Brief applied a different scenario, closer to what is actually being observed: 5 to 15 percent of electricity generated from gas.

With just 5 percent gas, AI datacenter emissions reach about 2 megatonnes, ten times the official estimate. With heavier reliance, the curve climbs to 30 megatonnes, the annual footprint of Denmark. In the worst-case scenario at 20 GW of capacity, it tops out at 70 megatonnes, roughly Sweden's emissions.

Meanwhile, your cloud bill

The compute crunch does not just show up in planning applications. It shows up on the bank statements of companies running AI agents in production. The Register compiled a series of cases where the monthly bill tripled or grew tenfold with no warning.

On Google Cloud, the mechanism is documented. A policy change in March 2026 allowed accounts with $1,000 in lifetime spend to see their budget cap raised automatically from $250 to $100,000. The new usage-priced models (Nano Banana, Veo 3) are expensive enough that a leaked API key can drain a runway in days. The lag between actual consumption and the charge appearing in the billing console is 28 days.

On AWS, the mechanic is different but the outcome is identical. AWS Bedrock bills through AWS Marketplace, which makes it incompatible with the platform's own Cost Anomaly Detection tool. Users get an anomaly detector active by default, and that detector does not see the very charges most likely to explode.

O'Ryan Johnson, quoted by The Register, summed it up: "If you're using Cost Anomaly Detection, it should stop you running up a massive invoice. In this case, it didn't. It was completely silent."

The thread running through both cases is the opacity of token pricing. An AI agent burns tokens in the background, sometimes for hours, without any human alert firing. The delayed billing does the rest.

France is not off the curve

RTE published its forward outlook in December 2025. The trajectory is clear: French datacenters use around 10 TWh per year today, or 2 percent of national electricity consumption. The 2035 projection sits between 23 and 28 TWh. Cloud, generative AI and agentic workloads account for most of that growth. ADEME estimates that AI multiplies energy needs by four to five compared with a classical datacenter.

None of these numbers says France is heading for disaster. They say the trajectory is being set now, in the infrastructure choices of the next five years, not in the 2050 carbon-neutrality pledges.

The "AI is efficient" story hides a transfer

AI produces useful results. That is the honest baseline of the debate, and ignoring it disqualifies everything else. But the sentence "AI is efficient" is half a sentence. Efficient for whom, measured how, paid for by what.

The productivity gain displayed to the end user is financed by the cloud provider. The cloud provider finances that cost through token pricing. Token pricing is itself backed by datacenter electricity consumption. Datacenter consumption lands on the grid, then on the energy mix, then on the gas-versus-renewable decisions of operators like the ones asking for 15 TWh in the UK today. At every stage, a cost is transferred to the next link.

As long as you only look at the demo, the chain stays invisible. The prompt answers in two seconds, and the output serves a measurable business goal. It is only when you look at next month's cloud bill, or the ten-year demand curves from RTE and NESO, that the chain reappears.

The calculation or the belief

The UK government published 0.142. The independent analysis published 70. The gap measures what happens when the material costs are left out of the frame. An ROI calculated without the cloud bill, the electricity consumption and the emissions trajectory is a sales projection, not a calculation.

The honest debate is not "pro or anti AI." It is: what are we willing to pay, where, and why. Until somebody adds up that bill, the assessment remains an act of faith.