Quantum Secrecy, Space Lasers and an AI Talent Flip: The Week in Research

pv magazine reported on 28 September that researchers at China's Hunan University of Technology modelled how different air injection strategies affect the performance of compressed air energy storage (CAES) caverns. A linearly increasing mass flow rate raised the energy storage rate from 85.5% to 87.5% compared with constant-rate injection, across eight-hour cycles injecting the same total mass of air. The study, published in Case Studies in Thermal Engineering, validated its numerical model against operating data from Germany's Huntorf CAES plant. Maximum differences between simulated and measured values were 2.66 K for temperature and 0.30 MPa for pressure, which the researchers deemed acceptable. They then tested four injection patterns in a spherical cavern of 268,000 m3 connected to a 30-meter wellbore.

These details matter for project economics.

The linearly increasing mode raised average temperature by 1 K and average pressure by 0.13 MPa. Work done on the gas increased by roughly 12.7%, while wall heat loss to the surrounding rock dropped by roughly 11% compared with constant injection. Wellbore diameter also played a role. Increasing it from 0.5 meters to 2.0 meters cut heat loss through the wellbore wall from 8.5 x 10^8 J to 2.5 x 10^8 J, while lifting the average air temperature in the cavern by about 1 K. Surrounding rock geology shifted conductive heat loss by approximately 7% between basalt at 2.17 W/(m K) and quartzite at 5.53 W/(m K). The team also found that an adiabatic model overestimates average temperature and pressure by 1.7 K and 0.05 MPa at the end of injection, and by 2.2 K and 0.06 MPa at the end of discharge. None of this is a commercial breakthrough on its own. It is the kind of modelling work that determines how operators size caverns, wells and injection schedules before signing an offtake agreement.

Storage auctions and EU infringement letters

On the same day, pv magazine also reported that Italy's energy regulator Arera raised the MACSE auction premium cap to 27,000 euros per qualified MWh per year, up from the 22,000 euros proposed in July. This was set in resolution 332/2026 dated 24 September. The magazine spoke to independent researcher Mario Iovino, who explained that the cap represents the cost of new entry. It consists of the annualized cost of a benchmark investment plus fixed operating costs, rounded to the nearest thousand. The benchmark investment cost rose from 125,000 to 150,500 euros/MWh, and WACC moved from 8.0% to 8.2%, based on data as of 31 August. Compared with the first auction, the cap is 27% lower.

Iovino's numbers specify what the contract actually pays. Consider a reference asset of 75 MW and a nominal 600 MWh with eight-hour duration. Qualification factors leave it with 552 MWh, 8% below nameplate, because the premium is paid on qualified MWh while costs are incurred on nominal MWh. At the cap, the contract provides 24,840 euros per nominal MWh per year, the maximum fixed revenue this plant can receive. The pre-auction guarantee of 10%, payable in cash by 20 October, comes to 2,700 euros per qualified MWh. Peak guarantee requirements, including the post-auction guarantee and the fund, reach 12,150 euros per qualified MWh, or 6.71 million euros for this asset.

Also on 28 September, pv magazine reported that the European Commission sent letters of formal notice to 18 EU member states for failing to report full transposition of energy sharing and related consumer rules under Directive (EU) 2024/1711. The deadline was 17 July 2026. The countries named include Belgium, Ireland, Greece, Spain, France, the Netherlands and Poland. The other nine member states, among them Germany and Italy, were not included, though the Commission's announcement does not state whether each has fully transposed the directive. The letters concern failure to notify transposition, not a finding on the substance of national rules.

Elite AI talent: China pulls ahead

A Carnegie China study reported by The Next Web on 28 September provides the next major update. In 2025, 41% of the top AI researchers tracked by the think tank worked in China and 34% in the US. Three years earlier the US led: in 2022, 46% worked in the US and 27% in China. The authors, Damien Ma and Binyi Yang, examined researchers with papers accepted at NeurIPS 2025, a group the study calls the Navy Seals of the field. There were 25,677 authors on 5,823 papers, and the study traced more than 10,000 of them with full career histories. Data came from OpenReview, the platform NeurIPS uses for submissions, with AI coding tools helping process the data and human checks on results.

The pipeline numbers reinforce the shift. In 2025, 57% of the group took their first degree in China, up from 46% in 2022. More of them are staying in China to work, the study found, pointing to job openings in China's AI industry and tighter US visa rules, especially for Chinese graduate students in science. Among AI researchers working in the US, the share with a Chinese undergraduate degree rose by four percentage points. Peking University and Tsinghua University now rank first and second as workplaces, with South Korea's KAIST third. Google held first place in 2022. The National University of Singapore and Nanyang Technological University both moved into the top 20. Europe lost ground: ETH Zurich and the University of Oxford dropped out of the top rankings, and Europe kept 60% of its own talent against 89% for the US.

Policy is part of the picture, though the study does not quantify its effect. In Washington, the Trump administration has proposed a $103,265 fee on new H-1B visas. Chinese AI and chip firms, meanwhile, are handing out shares to retain their engineers. Last week, Xi Jinping told Donald Trump that the two countries can jointly prevent AI misuse.

Zero-knowledge proofs and a failed secret

Scientific American reported on 27 September on a different kind of research result: one published without disclosure. In March 2026, Google shared a finding that quantum computers could crack common encryption methods by the end of the decade. For the first time in scientific history, they used a zero-knowledge proof to demonstrate it, revealing nothing about the methods used. The claim was that a specific form of elliptic curve cryptography could be broken by a quantum algorithm operating with just 1,175 logical qubits and more than two million Toffoli gates. Until then, experts had estimated at least 20 times that number of qubits would be needed. As of July 2026, the record was just under 100 logical qubits. The proof was flawed. A team led by computer scientist Keegan Ryan found it could also prove false claims, including that the encryption could be cracked without Toffoli gates. This is obviously false because the algorithm would then be unable to perform any logical operations. In mid-April 2026, Google's developers corrected the proof, and the original claim was upheld. The error had been in the implementation, not the finding. But secrecy had already invited reproduction. After 63 days, cryptologist Andre Schrottenloher of the Inria Center at the University of Rennes in France matched the result. He surpassed it with an improved quantum algorithm that is now public. Google scientist Craig Gidney, a co-author, responded on his personal blog: "My congratulations to Andre on being the first to match our circuits."

Lasers, robots and the cost of dreaming

Wired reported on 28 September that Star Catcher, a Jacksonville startup, is preparing to launch a prototype aboard a SpaceX rocket this week. It will beam energy to another satellite in orbit. If it works, it would be the first time lasers have beamed energy between two untethered objects in space. The company announced a $65 million funding round earlier this year and won a $30 million award from the US Space Force. It has 40 letters of intent and 10 power purchase agreements. Chief executive Andrew Rush said the company is still in the crawl phase. Hanieh Fattahi of the Max Planck Institute for the Science of Light noted that tracking systems, heat management and long equipment lifetimes in vacuum still need work. The Naval Research Laboratory ran a space laser experiment in 2023 that operated for 100 days, but the beam travelled less than five feet and efficiency was 11%.

On the robotics side, The Decoder reported on 27 September that researchers from Stanford and Caltech built HomeBody. This system lets a Unitree G1 robot navigate an unfamiliar kitchen, tidy up and fetch items from drawers using GPT-6 Astra directly, with no trained control layer in between. The robot explores the room, builds a digital twin in Nvidia's Isaac Sim, and logs objects in spatial memory. Limitations include latency, overheating finger servos and high compute costs. The same outlet reported the same day that Boris Power, OpenAI's head of applied research, said 80 to 90 percent of the company's research already targets GPT 7, GPT 8 and beyond.

ScienceDaily reported on 26 September on a Tohoku University study of REM sleep. It found that brain blood volume starts rising about 50 seconds before REM begins, while neuronal ATP then drops. Professor Ko Matsui framed the puzzle bluntly: "Ever felt exhausted after a vivid dream?" The work suggests the dreaming brain may run unusually high energy demands even as fuel supply increases.

One older item rounds out the picture. Canary Media reported on 25 September that 87% of US clean energy generation happened outside urban centres as of last year, per a Rural Climate Partnership analysis. More than 60% of counties with renewable restrictions are rural. This is background to the newer research above, but it reflects the same argument every energy technology eventually faces: the physics is only half the story.