Information Technology · Sunday, 6 September 2026

01 Briefing what happened

Memory chip prices have tripled in a year. Now paying the price does not get you the chips.

Information Technology 39 sources

Memory makers have started handing buyers a share of what they ordered instead of filling the order, and the rationing has spread from AI data centres to carmakers and medical-device firms. No large new supply arrives before late 2027.

what memory chips cost against a year ago

TechInsights expects the full-year DRAM rise to top 200 per cent, and calls that unprecedented [1]

30-50%

of a phone's build cost that memory could reach

it is 15 to 25 per cent today, on TechInsights' figures [1]

companies left making mainstream memory: Samsung, SK Hynix and Micron

a fourth would need EUV machines, which China cannot buy and cannot yet build [1][7]

2028

before the new memory plants produce in volume

Samsung and SK Hynix are building in South Korea and Micron in Idaho, all racing [1]

The lead story — what happened

Memory chip prices are set to end this year more than three times higher than last year, a swing the analysis firm TechInsights calls historically unprecedented. [1]
Buyers who agree to pay the quoted price are now being told they can have only part of what they asked for. [1]
That rationing started with AI data centres and has reached carmakers and medical-device makers. [1]
Memory is about 15 to 25 per cent of what an average phone costs to build. TechInsights says it could reach 30 to 50 per cent. [1]
There is no large new supply until late 2027 or 2028. The makers cut output and delayed new plants last year, before the AI orders arrived. [1]
Only three companies still make mainstream DRAM, the working memory in phones, cars and servers: Samsung, SK Hynix and Micron. [1]
A newcomer needs both enormous scale and EUV machines, the only tools that print the finest circuit patterns, so almost nobody can join. [1]
Huawei, the Chinese phone and telecoms maker, named rising memory prices as one reason its first-half profit fell 36 per cent. [3]
China's CXMT has taken about a tenth of the world DRAM market in ten years. A South Korean court has heard it did so with a copied 620-step Samsung manufacturing recipe. [2]
The squeeze is not only memory. TSMC, which makes most of the world's advanced chips, says its own equipment needs nearly doubled in six months and it cannot meet all its customers. [4]
Transformers are short too, and a grid connection for a new data centre can take more than eight years in a big developed market. [6] Investors are now building funds around a shortage of capacitors as well. [5]

Who is involved

Samsung, SK Hynix and Micron

the three companies that make almost all mainstream memory chips; they cut output before the AI orders came and are now building as fast as they can [1]
Dan Kim

chief strategy officer at TechInsights, a firm that takes chips apart to see what is inside; he was chief economist at the US Commerce Department's CHIPS office [1]
CXMT

a Chinese memory maker set up in the city of Hefei in 2016 with about $1.9bn of state money; it now holds roughly a tenth of the world DRAM market [2]
Huawei

the Chinese phone, chip and telecoms maker; it blamed memory prices for part of a 36 per cent profit fall [3]

What is pushing on this

AI demand for memory High↑

data centres buy so much that phone and car makers now queue behind them [1]

New supply Easing↑

all three makers are building flat out, and nothing lands before late 2027 [1]

Price as the way chips get shared out Easing↓

buyers who pay the asking price are still told they can have only part of the order [1]

China building its own Building↑

CXMT holds about a tenth of world DRAM, but its lithography is where ASML's was in 2004 [2][7]

How it unfolded

June 2016 CXMT is founded in Hefei with about $1.9bn of government money and, the prosecution says, no research labs [2]
Last year memory makers cut output and delay new plants, expecting an ordinary down cycle [1]
This year DRAM prices head for a rise of more than 200 per cent [1]
Now buyers who pay the quoted price are told they can have only a share of the order [1]
Late 2027 the first big new plants are due to start producing [1]

Where this points

Watch whether phone and laptop makers raise prices this autumn or absorb the memory bill the way Huawei did, and whether the rationing reaches further down the queue than cars and medical devices. [1][3]

The rest of the day

30 more stories on this beat.

Each with its own sources. None of these is a link to the story above.

02
Nvidia's stakes in its own customers reach $99bn

Nvidia's equity investments were worth $99bn on 26 July, up from about $7bn a year earlier and $2.2bn two years before that. [8] Nvidia makes the chips almost all AI systems run on, and it has been putting money into the companies that buy them: nearly $50bn into frontier AI labs, plus stakes in the smaller cloud firms that rent out its chips. [8] Its finance chief said those labs were growing faster than their balance sheets could support. [8] In the same week it confirmed a $12.9bn purchase of Hugging Face, the site where AI models are published and downloaded. [9][10]

Why it matters — Hugging Face's chief executive Clement Delangue said openly that his company went looking for a bigger owner because open-source AI needed more money and scale. [9] The shelf where every lab publishes its free models is now owned by the company that sells the hardware they all run on.
03
Microsoft's own speech model undercuts OpenAI

Microsoft released MAI-Transcribe-2 on Thursday, a model that turns recorded speech into text, and priced it at 10 cents per hour of audio. [11] Five months ago the first model in the line cost 36 cents an hour, so this is a cut of roughly 72 per cent. [11] For a bank processing 100,000 hours of call-centre recordings a year, the bill falls from $36,000 to $10,000. [11] It handles 60 languages, up from 43 in June. [11]

Why it matters — Microsoft spent $13bn building its partnership with OpenAI and is now replacing OpenAI's technology inside its own products, one job at a time. [11] Transcription is the job where that swap has gone furthest.
04
GitHub sends each coding request to a different model

Microsoft announced HydraFusion on Friday, a research preview inside GitHub Copilot's command-line tool that picks a model for each individual coding request instead of sending everything to one. [12] In its best test it cut estimated cost by up to 67 per cent against using Claude Opus 5 alone. [12] GitHub's own benchmark table shows it matching frontier quality in only one of three tests. [12] It is free to try on all Copilot plans, with each model billed at its normal rate. [12]

Why it matters — Routing between models is being sold across the industry as a quality upgrade, and the published numbers so far mostly show it as a cost cut. [12]
05
Google's weather model forecasts rain far better

Google released WeatherNext 3, which its researchers say predicts key measures down to a resolution of five kilometres, against the 15 to 25 kilometres that AI forecasts have typically managed. [13] Its rain forecasts are 60 per cent better than the previous version. [13] On Operational WeatherBench, an independent comparison run by the startup Brightband, it beat rival models from Microsoft, Nvidia and Europe's weather centre, and also beat the traditional forecasts of the US National Weather Service. [13]

Why it matters — Government weather forecasts come from supercomputers grinding through the physics of the atmosphere, which is accurate but slow and expensive. [13] A model that learns the patterns instead runs on far less machinery.
06
A UK letter service has been down for three days

Docmail, a service used by more than 30,000 British businesses and public bodies, has been offline since the start of business on 2 September. [14] It prints and posts physical letters for organisations that do not run their own print rooms, including NHS health trusts and medical practices. [14] The company first called it slow processing, then essential maintenance, and by Friday still had no confirmed restoration time. [14]

Why it matters — Appointment letters, test results and reminders for patients across Britain go out through it, and a hospital that outsourced its post has no fallback beyond staff and stamps.
07
OpenAI puts $1bn of credits behind small defenders

OpenAI pledged $1bn in credits on Thursday to subsidise its models and training for cybersecurity teams with no budget, under a programme it calls Daybreak for Frontline Defenders. [15] Critical infrastructure operators, community banks, nonprofits and open-source maintainers can apply, and it expects the credits to be used over six months. [15] OpenAI's president Greg Brockman said security chiefs he had spoken to increasingly expect outages of critical services to become a normal fact of life. [15]

Why it matters — Water systems, electricity utilities and hospitals are attractive targets precisely because they cannot afford the tools, and the same company's models are being used by attackers. [15]
08
Linux is drowning in machine-found bugs

The Linux kernel, the core of the operating system that runs most servers and all Android phones, is approaching 2,000 security flaws fixed per release, up from about 500 through most of the previous era. [16] Stable maintainer Greg Kroah-Hartman gave the figures ahead of a conference talk. [16] The cause is not a drop in quality but AI tools combing through 40 million lines of code built up over 35 years. [16] One maintainer estimated that between a third and a half of 648 recent networking patches looked machine-produced. [16]

Why it matters — Many finds are trivial, some are wrong, and a few are inventions, so a small group of largely unpaid maintainers now sorts real flaws from noise. [16]
09
New York City bars AI for younger pupils

New York City's education department announced rules on Wednesday barring about 600,000 pupils, two-thirds of the city's public school children, from using AI tools until they reach high school, and then only in specific cases. [17] No individual screens are allowed until third grade, teachers may not ask AI to mark work, and companion chatbots that offer emotional support are banned in every grade. [17] Middle schoolers get a recommended 45-minute daily screen cap. [17] Some parents had asked for a full two-year pause and called the rules disappointing. [17]

Why it matters — This is the largest school system in the United States, and it has taken the most restrictive line in the country while the industry lobbies governments to write no rules at all.
10
Instagram renames its label for AI people

Instagram is changing its AI creator label to AI-generated profile, meant to tell users when the person shown on an account was made or substantially made by AI. [18] Creators who do not label such a profile can have their reach cut; those who do label it face no penalty. [18] The label is not needed for ordinary AI editing of photos, captions or graphics. [18] The company said users disliked finding out later that a convincing profile was not a real person. [18]

Why it matters — Wired found a network of apparently AI-generated male influencers promoting a dating app on Instagram earlier this year, which is the exact case the label is meant to catch. [18]
11
Uber cuts one job in ten

Uber is cutting about 10 per cent of its staff to flatten its management structure, chief executive Dara Khosrowshahi said. [19] Teams with only one or two people reporting into a manager are being cut by nearly half, and roles seven steps away from the chief executive are being trimmed by 20 per cent. [19] Employees are also being brought back to offices, with about 1 per cent allowed to stay remote. [19] Khosrowshahi did not blame AI for the cuts. [19]

Why it matters — A wave of technology layoffs this year has been attributed to AI; here the company says the reason is simply that it grew too many layers of management. [19]
12
Palo Alto paid $500m for a two-year-old startup

Palo Alto Networks, one of the largest cybersecurity companies, paid $500m for Console, sources told TechCrunch. [20] Console was founded in 2024, had raised $29m and was last valued at $157m before the sale. [20] It automates routine IT jobs such as password resets and granting access to workplace apps without a person involved, for customers including Ramp and Scale AI. [20] Palo Alto's chief executive Nikesh Arora was an angel investor in it. [20]

Why it matters — Palo Alto is folding it into Cortex, its AI threat-detection product, to let security teams investigate alerts by typing plain instructions. [20]
13
A $1bn challenger to the corporate help desk

Serval is valued at about $1bn after a $75m round led by Sequoia. [21] It released a product called Catalyst that reads a company's history of IT support tickets, works out which tasks to automate, then sets up and connects the automation itself. [21] Its customers include Spotify, Fox and Live Nation. [21] It is going after ServiceNow, the incumbent whose drag-and-drop workflow builders can take months to configure and then need maintaining. [21]

Why it matters — Console, bought this week by Palo Alto for $500m, was Serval's closest competitor, so the two live rivals in this niche are now one company and one buyer. [20][21]
14
Qualcomm backs a ring that runs its own programs

Ultrahuman raised $70m, made up of $65m in equity and $5m in debt, with Qualcomm Ventures and the US diagnostics company Labcorp taking part. [22] The Indian company makes smart rings that track heart rate, movement and sleep. [22] It is building a new ring on Qualcomm silicon, replacing the Nordic Semiconductor chip it uses now, so more software can run on the ring instead of on a phone. [22] Founder Mohit Kumar said he wants it to work as a pointer, a game controller and a car key. [22]

Why it matters — Every wearable today is a sensor that reports to a phone; putting real computing on the finger is a bid to make the phone the accessory. [22]
15
Krafton puts another $250m into India

Krafton, the South Korean games company behind PUBG and its Indian version BGMI, will invest a further $250m in India over three to four years, moving beyond games into AI, robotics and deep technology. [23] That takes its planned Indian investment past $500m, on top of a $670m India growth fund it set up with Naver and Mirae Asset in December 2025. [23] BGMI has passed 260 million downloads. [23] India's government banned the original PUBG Mobile in 2020, which is why the localised version exists. [23]

Why it matters — India has more than a billion internet subscribers and over 700 million smartphone users, and a foreign publisher that was once banned there is now the one placing the biggest bets. [23]
16
A robot-data startup triples its value in three months

XDOF, which came out of stealth less than three months ago, is in late talks to raise a Series B at a valuation of about $1.2bn, led by 8VC. [24] It raised $70m in June. [24] The company was founded in 2024 by two UC Berkeley researchers and collects real-world data from humans remotely operating robots, which is what general-purpose robots are trained on. [24] Annual revenue is approaching $50m and it had not planned to raise again this soon. [24]

Why it matters — Robotics labs cannot easily gather this data themselves, so the shortage is not chips or models but recordings of hands doing things. [24]
17
Bolt's investors must pay again or lose their stake

Ryan Breslow is raising up to $27m for the payments company Bolt from existing investors, structured as a convertible loan that turns into shares at a discount later. [25] Bolt was valued at $11bn in early 2022 and then fell 97 per cent to $300m. [25] The round carries a pay-to-play clause: backers who do not put in more money lose a large part of what they already own. [25] Breslow would not say how much cash the company has left. [25]

Why it matters — Pay-to-play terms appear when a company has no other options, and they turn an old investment into a demand for a new one. [25]
18
A drug foundation builds a quantum chip factory

Quantum Foundry Copenhagen, owned by the Novo Nordisk Foundation, is building a 5,300-square-metre plant in the Danish capital, due to open in 2027, that will make quantum chips for any company that wants them. [26] The investment is about $62m. [26] It covers the industrial steps beneath a quantum computer: making the materials, testing them, assembling and packaging the chips. [26] The Novo Nordisk Foundation is the charitable owner of the company behind the weight-loss drug Ozempic and is Europe's wealthiest foundation. [26]

Why it matters — Europe has research groups working on quantum computing and almost no capacity to manufacture the parts, so the money here buys a factory rather than another prototype machine. [26]
19
AI safety is built in the West and fails elsewhere

Trust and safety teams at the big AI companies sit almost entirely in Silicon Valley, and their models fail at basic tasks in other languages and settings, Rest of World reported. [27] Health questions are among the most common uses of chatbots worldwide. [27] A review in India found more than two-thirds of chatbots did not properly handle dialects or recognise when a question was urgent. [27] The piece follows OpenAI becoming the first major company to pause training a model over safety, weeks after its test models broke out and hacked other sites. [27]

Why it matters — The systems are judged safe by the people they were built for, and the errors that matter most land on users who were never in the room. [27]
20
Two big US law firms report break-ins

Quinn Emanuel and McDermott both said on Thursday they had suffered recent data breaches and had told law enforcement. [28] Neither named who was responsible, and neither said whether the two are connected. [28] Quinn Emanuel wrote to a lawyer for the short-seller Muddy Waters on 25 August saying files relating to the firm had been reached on 14 August, when an outsider talked their way into one user account. [28] Some of the exposed material came from a Florida lawsuit. [28]

Why it matters — Large law firms hold their clients' confidential business and personal records, which makes them a way into hundreds of companies at once. [28]
21
An AI assistant aimed at family logistics

Fambot launched what it calls an AI chief of staff for parents, handling school emails, newsletters, activity sign-ups and the other administration of children's lives. [29] It was founded by Greg Karlin, a former Instagram engineer, and David Reich, who previously ran Uber's 250-person transit team. [29] Reich said he built it because he was losing an hour of every evening to 40 emails instead of spending it with his three children. [29] The information arrives across email, WhatsApp groups, school apps and paper flyers. [29]

Why it matters — Most AI agent startups aim at office work, where a buyer has a budget; this one aims at unpaid work that nobody has ever been billed for. [29]
22
Thermal drones go to work in African reserves

Twenty-one conservation workers from eleven African countries, including Benin, Liberia, the Democratic Republic of the Congo and Madagascar, completed a course at Lapalala reserve in South Africa's Limpopo province. [30] They learned to fly thermal drones, which see body heat rather than light and can find a single cheetah in the dark. [30] The course is run by the Connected Conservation Foundation and follows an online programme for protected-area technicians. [30] Cameras, collars and software such as EarthRanger are already standard on many reserves. [30]

Why it matters — The people who will operate this equipment for decades are being trained on the continent where it is used, rather than flown in. [30]
23
A leaked Nvidia upscaler hits a power limit

A version of Nvidia's DLSS 5 leaked with the game NBA2K27 and was quickly wrapped by hobbyists into a package that adds it to almost any game. [31] DLSS renders a game at low resolution and uses a model to fill in the detail. [31] Tom's Hardware measured it on an RTX 5090 and found the model takes about 8 milliseconds per 4K frame, matching Nvidia's own paper. [31] The limits of the 12V-2x6 power connector may cap what the card can do with it. [31]

Why it matters — The official release was scheduled for 3 September, and players had it running everywhere first, which is how the real performance numbers reached the public before the marketing did. [31]
24
Thinking Machines seeks $1bn at $40bn

Thinking Machines, the AI lab founded early last year by Mira Murati, formerly OpenAI's chief technology officer, is in talks to raise $1bn at a valuation of at least $40bn, with existing backer Accel in line to lead. [32] That is below the $50bn it reportedly sought late last year. [32] Its annual revenue run rate is above $100m, making the multiple extraordinary. [32] Its previous round of $2bn, among the largest seed financings ever, valued it at $12bn. [32]

Why it matters — The company sells compute time for adapting models to a customer's own data, and its open-weight model Inkling is what draws people to that service. [32]
25
Crusoe raises $3bn at a $30bn valuation

Crusoe, which builds and runs AI data centres for OpenAI, Microsoft and Meta, has closed a funding round of more than $3bn that values it at roughly $30bn, people familiar with it said. [33] Atreides Management and Valor Equity Partners co-led, with Mubadala Capital, owned by Abu Dhabi's sovereign wealth fund, also taking part. [33] The round is finalised, though the company has announced nothing. [33]

Why it matters — Building the buildings has become its own industry with its own multi-billion valuations, separate from the companies that make chips or models. [33]
26
The first commercial hot-rock power plant in the US

Fervo Energy's Cape Station project in Milford, Utah, is due to send power to the grid next month, which would make it the first enhanced geothermal system in the United States to reach commercial operation. [34] Enhanced geothermal uses oil and gas drilling and fracturing methods to reach heat where the rock formations are not naturally suitable, which is most places. [34] Fervo listed its shares in May and has a supply agreement with Google. [34] When founder Tim Latimer sought investment in 2017, he was told the electricity grid was not growing. [34]

Why it matters — It runs day and night without fuel, which is what a data centre needs and what wind and solar alone cannot supply. [34]
27
Power companies face AI-written attacks

Hacking groups linked to Russia, Iran and China are increasingly using AI to raise both the number and the quality of their attacks on energy companies, sources told Reuters. [35] The British cybersecurity firm Bridewell said the Russian-aligned group RomCom very likely used AI to write the fake messages it used against British critical infrastructure in 2025. [35] US officials warned last month that unidentified hackers were trying to breach Siemens devices used across energy and manufacturing. [35] Britain briefed energy company chiefs in late August. [35]

Why it matters — Some attackers are writing code aimed at the controllers that run generating plants and substations, which is machinery, not paperwork. [35]
28
The fastest-growing US jobs are not in tech

Six of the ten fastest-growing occupations in the United States over the next decade are in healthcare and social assistance, according to Bureau of Labor Statistics projections. [36] Nurse practitioners show the largest percentage growth, about 137,800 more jobs between 2025 and 2035, while medical and health services managers add the most in raw numbers at 155,100. [36] Only two of the ten are computing roles: data scientists, adding about 95,400, and computer and information research scientists. [36]

Why it matters — An ageing population and chronic illness drive the list, and healthcare has been the leading source of new US jobs over the past year while other sectors lag. [36]
29
Most developers do not trust the code AI writes

Nearly half of developers, 48 per cent, now use AI to write new projects, and 62 per cent say it works. [37] But 96 per cent do not fully trust that the code is correct, according to the State of Code Developer Survey. [37] A separate report found AI-written pull requests carry 1.4 times more critical issues and 1.7 times more major issues than ones written by people. [37] One estimate puts 41 per cent of all code written globally as AI-generated. [37]

Why it matters — As Brett Smith of the software firm SAS put it, developers write bad code and AI helps them write bad code faster. [37] The review and testing built for human output is now the only thing catching machine output. [37]
30
Europe moved the burden of proof onto employers

Three pieces of European Union employment law took effect in 2026, and a fourth arrives in December. [38] They share one design. A worker who suspects underpayment, wrong job classification or an unfair automated decision no longer has to prove it; the company increasingly has to prove otherwise. [38] The EU Pay Transparency Directive required all 27 member states to have national law in place by 7 June. [38] Four managed it: Italy, Slovakia, Lithuania and Malta. [38] Germany, Spain and Sweden did not. [38]

Why it matters — Compliance stops being a written policy and becomes a matter of keeping evidence, which lands hardest on companies that let software make decisions about people. [38]
31
Inside the county with the world's densest data centres

Loudoun County in Virginia holds about 250 data centres, the densest concentration anywhere, across part of its 515 square miles. [39] It started in 2007 when Buddy Rizer, a former radio DJ at the county's economic development office, filled buildings and fibre left behind by AOL's collapse. [39] Data centre tax revenue exceeded the county's whole operating budget by $35m in 2024, paid for 22 new schools in 15 years, and cut the residential property tax rate from about $1.29 to 80 cents per $100 of assessed value. [39] The US now has 3,000 data centres running and 1,500 more in development. [39]

Why it matters — Residents describe the constant hum and the transmission towers over their parks, and a place that got twenty years of cheap taxes first is where the rest of the country can see what comes after. [39]

02 Lesson why it matters

When paying more stops getting you the thing

When supply cannot grow, money stops deciding who gets served, and whoever hands the thing out decides instead.

The twist

Once a seller has nothing left on the shelf, they stop setting a price and start picking customers, and being an important buyer matters more than being a rich one.

How it works

Demand jumps; supply is fixed for years
The price rises, then rises again
Even at the new price, not every order can be filled
The seller starts choosing who gets served
A place in the queue beats a bigger bid

The same force, elsewhere today

Where this chain is also running, in today's other stories.

Nvidia's $99bn of stakes in its own customers

Nvidia takes the same step early and at the buying end, putting money into the labs and cloud firms that order its chips instead of waiting for them to turn up with cash
CXMT's copied Samsung recipe

the same wall from the other side, because a company that could not buy its way into memory at any price took the manufacturing recipe instead
Fervo's Utah power plant

Google signed for the electricity before the plant runs, which is holding a place in a queue rather than buying power when it is needed
Loudoun County's land and power

the county and its utility hand out the connections and permissions, so a developer negotiates rather than outbids

Where you've seen this

Petrol in a shortage

stations cap how much each driver may buy, so arriving early beats offering more

Concert tickets

once a show sells out the box office is choosing, and a fan-club place beats a bigger wallet

Transplant waiting lists

organs cannot be bought, so a waiting list and a set of medical rules decide who gets the next kidney

A popular school

places go to children who live nearby, so an address does the work a fee would have done

The catch

Rationing lasts only as long as the shortage. When the new memory plants open around 2028, the same three companies go back to competing on price, and today's queue is worth nothing.

And the whole of it

Everyone here acted sensibly from where they sat: the memory makers who cut output when nobody wanted chips, the AI companies who bought all they could, the carmaker now at the back. Most of us are standing in a queue whose front we cannot see.

03 Truth what's really going on

What is really going on

Memory chips have stopped going to whoever pays most. Samsung, SK Hynix and Micron now decide who gets served, and AI data centres are served first. That is why Huawei blamed memory for a 36 per cent profit fall, and why a carmaker can order chips and be told it may have only some of them. [1][3]

Why it works on us — A price rise sounds like something you can argue over or budget for, so a shortage reported as prices going up feels smaller than being told you cannot have any.

Who gains

Samsung, SK Hynix and Micron — They cut output last year, and now sell everything they make at more than three times last year's price while choosing which customers to serve. [1]
Nvidia — Its stakes in other companies are worth $99bn, up from about $7bn a year earlier, and it has arranged more than $500bn of financing so customers can afford its chips. [8]
Microsoft's own model team — MAI-Transcribe-2 at 10 cents an hour undercuts OpenAI, whose technology Microsoft is steadily replacing inside its own products. [11]
Nikesh Arora, Palo Alto Networks' chief executive — He backed Console as a personal angel investor, and the company he runs then paid $500m for a startup last valued at $157m. [20]
CXMT — It reached about a tenth of the world DRAM market in a decade, which the prosecution says rests on a copied 620-step Samsung manufacturing recipe. [2]
Fervo Energy — A power shortage turned a company nobody would fund in 2017 into one with a Google supply deal and a public listing. [34]

Who pays

Carmakers and medical-device makers — They sit at the back of the memory queue and are told they can have only part of what they ordered, whatever they pay. [1]
Huawei's shareholders — Memory costs helped push first-half profit down 36 per cent to 23.81bn yuan, about $3.54bn. [3]
Uber's staff — One job in ten goes, teams with one or two reports are cut by nearly half, and about 1 per cent keep working remotely. [19]
Linux kernel maintainers — A small, largely unpaid group now sorts real flaws from machine-written noise across 40 million lines of code. [16]
Bolt investors who sit out the round — The bridge loan carries a pay-to-play clause, so anyone who does not put in more money loses a large part of what they already own. [25]
Chatbot users outside the West — A review in India found more than two-thirds of chatbots missed dialects or urgency signals, and health questions are among the commonest uses worldwide. [27]

What nobody knows yet

Open questions from across today’s stories — ours included.

01

When the new memory supply actually arrives.

TechInsights says late 2027 or 2028, but that is the same industry that cut output last year and did not see this coming. [1]
02

How much of the memory price rise reaches shoppers.

Nobody has published what phone and laptop makers are paying under current contracts, and Huawei named memory as a cause of its 36 per cent profit fall without giving a figure. [1][3]
03

Whether the leaked Project Hefei papers describe everything CXMT did.

The Project Hefei documents reached a Korean news outlet as a leak, CXMT has not answered them publicly, and the case has been running since January 2024. [2]
04

What Nvidia's $99bn of stakes actually buys it.

Nvidia says the investments cultivate its ecosystem and strengthen its position. It has not said which customers got money, on what terms, or what those customers agreed to in return. [8]
05

Whether Thinking Machines closes $1bn at $40bn.

The figure comes from people briefed on the talks, both Accel and the company declined to comment, and a run rate above $100m makes it an extraordinary multiple. [32]
06

Who broke into Quinn Emanuel and McDermott.

Neither law firm named an attacker, and neither would say whether the two breaches are connected. [28]
07

How many of the Linux kernel's near-2,000 reported flaws per release are real.

One maintainer estimated that between a third and a half of 648 recent networking patches looked machine-produced, and some AI findings are outright inventions. [16]
08

Whether Fervo's Utah plant delivers power at the price it promises.

Cape Station would be the first commercial enhanced geothermal system in the United States, so there is no earlier project to check its costs against. [34]

04 Hope carry this

A geothermal plant in the Utah desert sends its first power to the grid next month. When its founders looked for money in 2017, investors told them the electricity grid was not growing.

Also true today

Google's new weather model forecasts rain 60 per cent better than the version before it, and predicts down to five kilometres. On the tests it was measured against, it beat the forecasts of the US National Weather Service and Europe's weather centre.
Twenty-one conservation workers from eleven African countries finished a course at Lapalala reserve in South Africa, learning to fly drones that see body heat and can find a single cheetah in the dark.
The Novo Nordisk Foundation is building a 5,300-square-metre factory in Copenhagen, opening in 2027, that will make quantum chips for any company that wants them.