AI claims to break a 30-year computing belief, Nvidia's Rubin ships slower than promised, and a rocket shortage, plus 3 India stories

Before, India often wasted extra daytime solar power. Grid batteries move it to the evening. More big battery auctions are open.

A 500 megawatt hour battery in Rajasthan just started work. IndiGrid's 250 megawatt, 500 megawatt hour battery in Rajasthan began paid, everyday use on October 27. First, it charges when solar power is cheap at midday. Second, it discharges in the evening peak under a 12 year contract with N V V N. Cells are mostly imported, and contract income is fixed. Our judgment: promising, start small.

Before, small electric water cleaners wasted most energy heating water. This one heats only a thin layer. It is a lab sample. Nobody knows yet if it works for a village or city.

A palm-sized solar gadget made drinking water from salty, oily and acidic water. Researchers in Nature Water showed a solar evaporator that made 1.64 litres of fresh water in 8 hours. First, sunlight powers a heated mesh layer that warms only a thin sheet of water. Second, the vapour is collected as clean water, leaving salt and dirt behind. It is a small lab device, not yet tested at scale. Our judgment: too early.

Before, growers walked the rows looking for pests. Sensors and cameras can catch problems earlier. It is a small funding round. Many rivals do similar things.

A Berlin startup raised 4.8 million euros to spot greenhouse pests with A I. hexafarms raised 4.8 million euros for wireless crop sensors and AI that scouts greenhouses for pests and disease. First, sensors and cameras collect data across the greenhouse. Second, AI flags pests early and forecasts the harvest. It is a small seed round in a crowded field. Our judgment: too early.

Before, no glasses with Android XR were on sale. Now Google's software has real glasses, up against Meta's. The price and the wire mean only keen early buyers will get them.

The first glasses running Google's Android XR are on sale. Xreal's Aura glasses went on order October 28 from 1,099 dollars and ship from mid-November. First, a small computer in your pocket runs Android apps. Second, the glasses show them on a sharp screen that seems to float in front of you. High price and the wired puck limit buyers. Our judgment: too early.

Before, most carbon-removal deals were small, one-off buys. Long deals let projects borrow money and grow. Fire or land changes can undo tree projects. So checking matters.

Google and McKinsey will buy 320 thousand tonnes of carbon removal over 10 years. Google and McKinsey signed 10-year deals for 320 thousand tonnes of carbon removal from trees on Brazilian pasture. First, farmers plant trees among cattle pasture. Second, the carbon the trees store is measured and sold as credits. Trees can burn, and the price is not public. Our judgment: too early.

Before, this drug was tested in only one animal type. Now there are two, but with tiny groups. The results come from the company. Most anti-ageing drugs that work in animals fail in people.

A company says its anti-ageing drug kept very old rats moving well. Eos SENOLYTIX reports that very old rats on PTC-2105 kept their balance-beam scores near where they started. First, the peptide targets the cell's power units, called mitochondria. Second, it is meant to clear out worn-out cells that cause inflammation. It is company-reported and not checked by other scientists. Our judgment: too early.

Before, quantum-safe codes were only advice. Now there are firm dates, and big cloud firms offer them. The danger: thieves store data today and read it once quantum computers arrive.

The US just fixed the dates for switching to quantum-safe encryption. NIST published its final roadmap, NISTIR 8547, with deadlines of 2030 and 2031 for the most sensitive US systems. First, organisations list every place old encryption is used. Second, they swap in the new NIST methods, like the one AWS just added. Audits show agencies are behind, so spending may be slow. Our judgment: promising, start small.

Before, drones were bought in small rush orders. Now they are part of the big, planned budget. This first yes is only step one. Contracts can take years.

India just approved over two lakh crore rupees of defence buying, with drones at the centre. On October 26 India's Defence Acquisition Council approved 2.12 lakh crore rupees of proposals, about 82 percent from Indian industry. First, each item goes out to tender. Second, the winning firms sign contracts and start deliveries. This is approval in principle only, and contracts can take years. Our judgment: promising, start small.

Before, fusion firms bought ready-made power parts. Helion needs switches for huge, fast bursts, so it will make its own. This is a first agreement, not a final deal. No fusion power has been sold to the grid yet.

A fusion startup may get up to 175 million dollars of US chip money. Helion signed a letter of intent for up to 175 million dollars from the US CHIPS office, plus 50 million private. First, the money funds chip and packaging research in Washington state. Second, the chips switch huge pulses of power in Helion's fusion machines. It is a letter of intent, not a final deal, and no fusion firm sells power yet. Our judgment: too early.

Before, driverless rides were only in small areas. Bigger areas and more cars mean robotaxis start to compete with normal taxi apps. Waymo still loses money. Fleet goals are from reports, not promises.

Waymo's driverless taxis now cover about 250 more square miles in five US cities. Waymo is expanding robotaxi service across Los Angeles, Nashville, Orlando, San Diego and the Bay Area. First, cars are mapped and tested in each new area. Second, the zone opens to paying riders who book through the app. Waymo is still unprofitable, and fleet numbers are press reports. Our judgment: promising, start small.

Before, human-shaped robots were mostly demos. If sales start, factories could buy all-round robots instead of one-job machines. This is the boss's forecast. Tesla's past dates have often slipped.

Tesla says its humanoid robot could go on sale to other companies by the end of 2027. On its October 21 earnings💡 call, Tesla said Optimus could ship to outside companies by the end of 2027. First, the robot learns tasks by watching and practice in Tesla's own plants. Second, outside buyers would pay for it to do repetitive jobs like moving parts. No price, customer or volume was given, and Tesla timelines often slip. Our judgment: too early.

Before, everyone treated this guess as true. If the claim holds, many arguments need rechecking. But the new method may only help on enormous problems. OpenAI also pulled three other AI maths papers after a mistake. AI proofs can be wrong.

OpenAI and Anthropic each say their AI disproved a 30-year-old computing belief; outside experts have started checking. Like a climber saying they reached a peak the guidebooks called unclimbable: the GPS log (the Lean check) shows a real route, but experts still have to confirm it is the same peak the guidebook meant. Our judgment: wait. Sources on our page.

AI often waits on memory, not on maths. So memory speed matters for real speed. ServeTheHome thinks the new memory couldn't reach Nvidia's very high target. Even so, it is still faster per pin than AMD's MI455X. Rubin is shipping on time.

Nvidia's Rubin AI chip ships with 11% less memory speed and 17% less chip-to-chip speed than first promised. Like a new highway opening with all its promised lanes but a lower speed limit on the ramps to the warehouse: the trucks (compute) are as big as advertised, but goods move to and from storage (memory) a bit slower. Our judgment: watch. Sources on our page.

Faster steps mean more work before the part forgets its data. That helps fix errors. Before, going faster caused more errors. This was for single parts. Many parts on one chip still disturb each other.

German researchers run quantum logic steps in 6.8 billionths of a second with 99.98% accuracy. Like a pianist playing a fast passage without brushing the neighbouring keys: speed usually causes wrong notes, and the new finger technique (pulse shape) keeps every stroke on the right key. Our judgment: watch. Sources on our page.

Before, speech decoding needed a button press and worked in one app only. Neuralink says only users can judge the designs. All features are early tests, with no measured results yet.

Neuralink's speech-implant user makes iPhone calls and types in any app in a 72-hour home test. Like fitting a hearing aid with Bluetooth: the core device already works, but it becomes truly useful once it connects directly to the phone and apps people use every day. Our judgment: wait. Sources on our page.

Solid batteries swap the liquid inside for a solid. That promises more range and more safety. The hard part is keeping the solid layers touching while charging. Output will be small until about 2030.

CATL starts 20 Ah trial cells of its all-solid-state EV battery, aiming for 500 Wh/kg and small-scale output in 2027. Like a shipyard moving from scale models to its first full-size test hulls: the design floats in the lab, but a working fleet (mass production) is still years away. Our judgment: watch. Sources on our page.

Before, Falcon 9 carried most Western business satellites. Now buyers must wait for new rockets. Quilty says if you haven't booked a launch, you're late. Makers of working rockets could see strong demand (our judgment).

Analysts see a shortage of about 150 rocket launches outside China through 2030. Like an airport where the busiest airline retires its main jet just as bookings boom: new planes are late, so seats sell out for years. Our judgment: watch. Sources on our page.

UPI keeps growing fast. That strengthens India's lead in phone payments. But most UPI payments are free. So payment apps earn little from them. The fee decision matters for their income (our judgment).

UPI moves over ₹1 lakh crore a day on average for the first time, in a record October. Like a national railway on its busiest festival day, except it happened every day of the month: 821 million trips (payments), each one tiny, adding up to a record load. Our judgment: watch. Sources on our page.

Before, most AI was built for English first. BharatGen says Param2 learned Indian languages from the start. No outside test results were given. The insurance time savings come from BharatGen's own talk.

India's BharatGen Param2 model, built from scratch for 22 Indian languages, is now running in government and insurance apps. Like a large orchestra where only the few sections needed for each tune play: the full band (17 billion parameters) is on stage, but only about 2.4 billion play each note, so the concert costs less to run. Our judgment: watch. Sources on our page.

India wants AI computers and data kept inside the country. New AI chips need liquid cooling and lots of power. So ready power and cooling are a real plus (our judgment). Data centres are a small part of L&T's business.

L&T's Vyoma.AI breaks ground on a 52 MW liquid-cooled AI data centre in Chennai. Like building a bigger power-hungry factory next to an existing one, with coolant pipes running straight to each machine: AI servers run so hot that air alone cannot cool them. Our judgment: watch. Sources on our page.

Before, surgery was the main choice. The virus stayed, so many needed surgery every year. Fast-track approval is based on fewer surgeries. Inovio must still run more trials to confirm it. Most trial patients still needed some surgery. Common side effects included sore arms, headache, tiredness and high blood pressure.

FDA approves Voceryq, the first plasmid-DNA immunotherapy in the US, to cut repeat airway surgeries. Like handing a building's security guards a wanted poster: the DNA makes the body's cells show the virus's 'face', so the immune system learns to find infected cells, instead of calling a repair crew again and again to scrape off the damage. Our judgment: watch. Sources on our page.