AI chip demand forecasts are everywhere, but the hardest data is in trucks and power grids

The freshest numbers point away from the semiconductor earnings calendar. On 3 November, pv magazine Global published an Ember analysis of hybrid wind, solar and battery plants in Kazakhstan. The plants could deliver electricity at an average of $56/MWh, against $70 to $170/MWh for new coal, while covering up to 90% of typical demand.

That is a power story, not a chip story. It matters because the AI demand narrative now runs through electricity, freight and grid capacity as much as through wafer fabs. The most dateable evidence sits in those adjacent markets, not in a single chipmaker's guidance.

Ember's numbers are unusually specific. The think tank simulated 15 sites across Kazakhstan's three power zones, according to pv magazine. It tested 125 wind, solar and battery combinations at each, for 1,875 simulations run hour by hour over a 30-year project life at a 100 MW peak. Demand in the country rose 16% in the last five years, and coal still supplied 62% of generation in 2025. Those units average over 50 years old, and there were 891 incidents of various severity between January and July this year.

"Solar is really cheap and very fast to build here," Kostantsa Rangelova, senior battery analyst at Ember, told pv magazine. "Shifting some of it to the evening with a battery unlocks its full value, making its potential contribution to electricity supply larger than with standalone solar."

The Kazakhstan work is the newest item, and the material is thinner on semiconductors than the topic implies. Only two of the six sources are AI research papers, both posted to arXiv in the past week, and neither forecasts chip demand. The rest cover electric trucks, European car registrations, football governance and the Kazakh grid.

What the freight data says

Electrek reported on 2 November that 33,100 electric big rigs were sold in China in September, up 37% year on year. The overall heavy truck market shrank 21%. Through the first nine months of 2026, China sold 226,400 electric heavy trucks, up 64%, for a 38.1% share, already close to the whole of 2025. In tractors the September share hit 64.5%.

Beijing's target, set this year by 11 government bodies, is 40% new energy heavy truck sales by 2030. The market cleared it on a monthly basis four years early. CATL now says new energy trucks will be half of China's annual truck sales by 2028, according to Electrek. The cost case is doing the work: CATL claims savings of about 0.6 yuan per km. A study by the ICCT and China's CATARC found battery electric trucks about 19% cheaper per tonne-km than diesel in good conditions. That edge disappears at low annual mileage or when battery weight eats payload.

Charging is answered with swapping. CATL's Qiji brand ran about 300 heavy truck swap stations at the end of 2025. It plans 900 by the end of this year and 2,000 in 2027. In August it agreed to pay 2.56 billion yuan, about $370 million, for nearly 25% of Qiyuan Green Power. Its standardized packs hold 432 kWh and 513 kWh and swap in about five minutes. XCMG led the segment in 2025 with a 15.4% share ahead of SANY and FAW Jiefang, the top three held 42.9%, and CATL supplied 60% of the batteries.

Electrek also cites Kpler on the fuel shift. LNG trucks held 37% of heavy duty registrations in March against 23% for electric. By July electric was at 47% while LNG had fallen to 11%, with monthly LNG registrations dropping from more than 30,000 to about 6,200. Those figures come from Kpler via Electrek and were not independently checked for this piece.

Europe's record month

CleanTechnica reported on 31 October that 422,000 fully electric passenger cars were registered in Europe in September. That is a 63% year on year increase and a record 31% share. Plug-in hybrids added 167,000 units, up 23%, for a combined 590,000 and a 43% share. Year to date the battery electric share is 25%, above the 20% full-year result for 2025.

The same piece flags a possible turn in plugless hybrids, which grew 5% against a market up 10%. Their share slipped to 34% from 36% a year earlier. Petrol fell 21% and diesel 28%. In the midsize category, 59% of Mercedes GLC sales came from its battery electric version or plug-in hybrids, leaving combustion versions at 41%.

None of these outlets attaches a chip demand number to its findings. The connection is inferred, not reported: more electric drivetrains and more grid storage mean more power electronics, and that is where the reporting stops. Treat any figure beyond what is quoted above as unverified.

The forecasting research

The two arXiv papers published in the last week are the only AI material. Yuxuan Wang, Yishu Wang and Jiaqi Deng submitted FABench on 29 October. It is a backtesting environment for forecasting agents whose first release poses 287 questions across nine domains. Testing agents built on eight recent models, the authors found retrieval improved the Brier score of all eight in every retrieval setting. It also lifted the leading agents to market consensus on the 54 questions linked to Polymarket, with the best three slightly above it.

A separate paper submitted on 30 October by Haofu Yang, Hongyang Yin and five co-authors describes Ask on Demand. It is a communication framework that lets parallel agents request targeted help from peers. On AppWorld and WebShop the authors report consistent gains over independent parallel execution at lower inference cost than existing sharing baselines.

Whether either method helps anyone produce a better semiconductor forecast is untested here. FABench is a benchmark for the forecasting process, not a forecast. Its authors describe recurring failures in how agents weigh evidence rather than a solved problem.

Background, and one unrelated item

For completeness, the oldest source is a SportsPro report from 21 October. In it, LaLiga president Javier Tebas called for "justice" after Manchester City were found guilty by an independent commission of 114 of 115 charges. The investigation found revenues artificially inflated and costs reduced by more than UK£900 million between 2009/10 and 2017/18. It has no bearing on chip demand and is included only because it appears in the reporting.

Read together, the material supports a narrow conclusion. The measurable demand growth here is in electric freight and power capacity. The AI research is about how agents forecast rather than what they forecast. No source here gives a defensible number for AI semiconductor demand in 2027.