Information Technology · Tuesday, 8 September 2026

01 Briefing what happened

China now builds its own chip printers. Its biggest chipmaker is testing them, and still prints the real chips on ASML's.

Information Technology 32 sources

Huawei is funding the lens and light-source makers China never had, and twelve home-made machines are due off the line this year. Inside SMIC's factories they run on simple chips while the work that matters stays on the Dutch machines.

Chinese-built chip printers due off Yuliangsheng's line this year

the company remains a long way from competing with ASML, the FT's sources said [1]

2004

where one analyst places China's chip printers against ASML's

UBS's Francois-Xavier Bouvignies, who expects Chinese mass production of the immersion type in two to five years [2]

37%

of the lens maker Fujian Zhiqi Photonics, bought by a Huawei-backed fund

Rmb100mn, about $15mn, paid in 2023; lenses are one of the two parts China cannot yet make well [1]

1,400%

quarterly revenue growth at Enflame, a Chinese AI chip designer heading for a listing

it is taking business off Nvidia by making the switch away from Nvidia's software easier [4]

The lead story — what happened

China's home-made chip printers have left the laboratory. Machines built by Yuliangsheng are being tested at SMIC, China's largest contract chipmaker, and on Huawei's own production lines. [1]
Yuliangsheng expects to build 12 of the machines by the end of this year, one person familiar with the plan told the Financial Times. [1]
A chip printer, or lithography machine, uses light to print circuit patterns onto discs of silicon. The Dutch company ASML makes the best ones, and export rules bar it from selling its most advanced models to China. [1]
Huawei is not building the machine. It is organising the people who do. Bernstein analyst Lin Qingyuan calls that a project management role, pulling thousands of vendors and customers in one direction. [1]
Huawei's investment arm Habo and a second Huawei-backed fund have bought into the makers of the two parts China lags on: the projection lenses and the light source. [1]
One of them, the lens maker Fujian Zhiqi Photonics, sold 37% of itself to a Huawei-backed fund for Rmb100mn, about $15mn, in 2023. [1]
Yuliangsheng's own machines still use projection lenses made by Germany's Zeiss, two people told the FT. Zeiss says it sells its lithography optics only to ASML. Industry insiders said the lenses can be bought second-hand in China. [1]
The machines are being run on less complex chips while engineers improve them, and the FT's sources say they will take several years to match ASML's output rate. [1]
SMIC has bought and is testing the Chinese machines. It still prints its real chips on ASML's, whose China sales are forecast at about a fifth of the Dutch firm's revenue this year. [1]
A UBS analyst, Francois-Xavier Bouvignies, put China's lithography ability at roughly where ASML stood in 2004, and expects mass production of the immersion type in two to five years. [2]
The chip designs are moving faster than the machines that print them. Huawei launched its Mate XT2 folding phone on Monday running its own Kirin 9050 Pro, which it says does more computing per unit of electricity. [3]
Lin also said the US restrictions could speed China up by accident. If they push Chinese factories to work with Chinese equipment makers, he said, those makers could improve much faster than expected. [1]

Who is involved

Yuliangsheng

the Chinese lithography start-up spun out of SiCarrier; its machines are the ones SMIC is testing [1]
Huawei

China's biggest phone and telecoms maker, cut off from Western chip technology by US export rules; it is coordinating the machine effort and funding its suppliers [1]
ASML

the Dutch company that makes the machines which print the finest circuit patterns; its best models cannot legally be sold to China [1]
SMIC

China's largest contract chipmaker; it is testing the domestic machines and producing on ASML's [1]
Lin Qingyuan

a semiconductor analyst at the research firm Bernstein; he says a prototype is not enough and the coordination is the hard part [1]

How it unfolded

2022 Fujian Zhiqi Photonics is spun out of the optics group Castech to make ultra-precision lenses [1]
2023 a Huawei-backed fund pays Rmb100mn for 37% of Zhiqi [1]
Earlier the FT reports SMIC testing a 28-nanometre Yuliangsheng machine, and printing 7-nanometre chips by passing wafers through it repeatedly [1]
Monday Huawei launches the Mate XT2 folding phone on its own Kirin 9050 Pro chip [3]
By year end Yuliangsheng expects 12 machines built [1]

Where this points

Watch whether a Chinese factory starts printing real production chips on a Chinese machine rather than only testing one. Bernstein's Lin says that switch, not the prototype, is what would speed China up. [1]

What is pushing on the whole day

The bar and the word are our reading of how hard each one is pushing today. The arrow is where it is heading. The evidence is in the stories below.

China's push to build the machines Building↑

12 machines due this year, tested at SMIC and inside Huawei [1]

Willingness of factories to use them Easing→

SMIC tests the Chinese tools and prints its real chips on ASML's [1]

Foreign parts inside the Chinese machines Steady↓

some prototypes use Zeiss lenses, others are testing Chinese ones [1]

Export controls High→

ASML's best machines cannot be sold to China; the memory makers CXMT and YMTC stocked up three years of supply [1]

The rest of the day

26 more stories on this beat.

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

02
Anthropic drops a $6bn purchase

Anthropic, the American company behind the Claude AI models, has walked away from buying the start-up Decart, Bloomberg reported on Tuesday. The two sides had been talking at about $6bn. Anthropic ran due diligence first, the standard trawl through a target's finances, contracts and technology, and then decided against it. Neither company has said what it found. People familiar with the talks said the two may still work together in other ways. [5]

Why it matters — A price agreed and then dropped is one of the few honest signals in a market where private valuations are set by whoever agrees to pay them. Anthropic itself is valued at close to $1 trillion by private investors, and is heading toward a public listing. [9]
03
Belgium holds a chip executive over alleged spying

Prosecutors in Belgium said on Monday that a 52-year-old man with Belgian and Chinese nationality has been held since May on suspicion of chip espionage. He had a senior job at Belgan, a Belgian maker of gallium nitride chips, used in electric cars, aircraft and weapons. Prosecutors suspect he was at the same time a director of a Chinese company making the same chips. He was arrested at Brussels airport boarding a flight to China, and a second suspect is still at large. [6]

Why it matters — Export rules stop machines crossing a border. They do not stop a person who already knows how the thing is made, which is the part of the chip industry no customs officer can inspect.
04
Apple shows its first folding phone on Wednesday

Apple holds its annual iPhone event on 9 September, the first under John Ternus, who took over as chief executive from Tim Cook. Reports say the company will show a folding iPhone with a 5.5-inch outer screen that opens to 7.8 inches, and that it will cost more than $2,000. To keep it thin enough to pocket, Apple has reportedly dropped the long-range zoom camera. AirPods, HomePods and a non-folding iPhone 18 line are also expected. [7]

Why it matters — Apple has made buyers wait more than half a decade for a folding phone. [7] It arrives two days after Huawei launched its own new folding model in China, where both are fighting for the expensive end of the market. [3]
05
A new model finished a video game on its own

An enthusiast using the name CozyBlaze had OpenAI's new GPT-6 Astra model play Valve's puzzle game Portal from beginning to end without help. The run took about 24 hours, 3,336 tool calls and a headline cost of $571.18 in tokens, which their existing $200 subscription covered. The game was paused while the model thought, then unpaused to execute each planned set of moves. CozyBlaze said plainly this is not a benchmark. [8]

Why it matters — It is one of the very few public tests of the new model run by somebody who does not sell it, and the whole recording is published for anyone to check.
06
Four AI labs shipped in one week

Anthropic updated its Fable and Mythos models, Meta and Google pushed out changes, and OpenAI released GPT-6 Astra, all inside seven days. Buyers are calling the result model fatigue: chief executives and IT managers are spending large amounts of time comparing prices and abilities instead of using any of them. OpenAI's Sam Altman told CNBC that everyone is moving to faster release schedules, and put some of it down to people getting back from summer holidays. [9]

Why it matters — The pace is set by the labs' need to look fast to each other and to investors, and the cost of keeping up lands on the customers who have to re-test everything each time.
07
UN rights chief calls AI an existential risk

Volker Turk is the United Nations High Commissioner for Human Rights. He told the UN Human Rights Council in Geneva on Monday that he shares the concern of people inside the industry that advanced AI could threaten humanity's survival. He called for an all-out effort to put cast-iron guarantees around AI safety and security before it is too late, and said he would press companies directly in the coming days. He was speaking ahead of his second term, which starts next month. [10]

Why it matters — The most senior human-rights official in the UN system is now using the language of the safety researchers inside the AI labs. He said it a week after the Trump administration pressed the G20 group of large economies to take a hands-off approach to AI rules. [12]
08
The US and China prepare their first AI safety talks

Reuters reported that the two governments are preparing a dedicated dialogue on AI safety for the middle of this month, led on the American side by Treasury Secretary Scott Bessent. An official in the Trump administration told Reuters there is currently no planned AI meeting in mid-September. Separately, Nikkei Asia reported that the US will raise AI safety when China's president Xi Jinping visits the United States on 24 September, with AI-directed cyberattacks top of the American list, while China wants the chip export controls reopened. [11]

Why it matters — A Reuters Breakingviews column argued the same week that any real international control of AI has to start with a deal between these two governments, because nothing else binds the labs that matter. [12]
09
OpenAI takes Malaysian computing capacity

Firmus, an Australian data-centre company backed by Nvidia, said on Tuesday it has signed a multi-year deal making OpenAI the anchor customer for two data centres in Malaysia. The deal takes Firmus's contracted capacity across all customers past 900 megawatts. It already runs sites in Australia and Singapore, with five more under development across Asia-Pacific, and plans to install Nvidia's next-generation Vera Rubin processors. Firmus declined to give the contract's value and OpenAI did not respond to Reuters. A stock market listing is rumoured for this year. [13]

Why it matters — Malaysia is now Southeast Asia's fastest-growing data-centre market, and the buildout there is already drawing questions about how much electricity and water it takes.
10
Data-centre money looks at Argentine Patagonia

Investors are scouting Argentina's Patagonia for very large data centres, drawn by cool air, renewable power, shale gas and empty land, Reuters reported on Monday. Pampa Energia, one of the country's big electricity generators, wants investors for a site of up to 500 megawatts. It would sit beside its Loma de la Lata power plant in Neuquen province, fed by the Vaca Muerta shale field. Mapuche Indigenous families live in that rural area and have clashed with energy companies before. Connectivity and grid infrastructure remain the main obstacles. [14]

Why it matters — Argentina has no data-centre backlash because it has no data centres yet, which is precisely what makes it attractive to developers being turned away elsewhere.
11
A fire showed nobody owns a data centre's safety

In June a fire broke out in an unfinished building at the $3.2bn Lake Mariner data centre in Somerset, New York. Firefighters found no working alarm, no sprinklers and three dead hydrants, and the safety sheets they are legally entitled to see had reportedly burned in the blaze. Fire chief Steve Matisz said his crew went in blind. TeraWulf owns the site and leases the land from a company owned by its own chief executive. The British firm Fluidstack runs it, and Google guarantees Fluidstack's lease. [15]

Why it matters — The town was promised 165 permanent jobs and $85m of investment in 2019 in return for a power discount; a 2024 planning presentation put the finished site at 35 to 40 jobs. [15]
12
Truckers are having a good year off the buildout

American freight has been squeezed by tariff rounds and by diesel hitting a record $5.85 a gallon, but the data-centre boom is pulling in the opposite direction. Every transformer, generator, cooling unit and load of concrete arrives by truck. Spot rates for flatbed and heavy haul, the trailers that carry oversized equipment, reached multi-year highs this summer, and capacity is tightest in construction-heavy markets. Because sites are chosen for cheap power and land, the freight is moving into places that were never freight hubs. [16]

Why it matters — It is one of the few parts of the AI buildout whose gains reach people who do not work in technology, and it is reshaping which American towns trucks drive through.
13
SoftBank's power arm files to go public

SB Energy, the SoftBank-owned company building power projects for data centres, has filed for a US stock market listing. Nvidia has committed to buy $1.5bn of shares at the listing price, and OpenAI holds warrants worth about $5.5bn; both companies had already put $500m each into it earlier this year as part of the Stargate data-centre plan. SB Energy reports a backlog of roughly $439bn and has no operational data centres of its own yet. [17]

Why it matters — It is a test of whether public investors will fund the electricity half of the AI buildout on the same terms private ones have, at a moment when the sustainability of that spending is being questioned.
14
Saudi Arabia pitches itself as neutral ground for AI

Tareq Amin, chief executive of the Saudi state-backed AI company HUMAIN, told Semafor he wants the company to be a Switzerland for open AI models, sitting outside the contest between the US and China. HUMAIN is building towards three gigawatts of data-centre capacity at an industry average of $11m to $12m a megawatt, which implies more than $30bn. Blackstone has committed $3bn. Amin wants only 30% of the money to come from its shareholder, Saudi Arabia's Public Investment Fund. It is opening offices in the US, France and the UK. [18]

Why it matters — Saudi Arabia is trying to turn cheap power and empty land into a position no export rule can take away. Its plan is to sell other people's open models rather than build a leading one of its own.
15
China's military is preparing humanoid robots for war

Two days after robots boxed and danced at the World Humanoid Robot Games, the People's Liberation Army's newspaper called for the technology to be moved from laboratories to military training grounds. Reuters reviewed more than 100 Chinese military procurement notices, patents and academic papers and found institutions testing humanoids against battlefield requirements. A 2025 paper from the National University of Defence Technology modelled six combat robots clearing an occupied building room by room, using equipment its authors expect within five to ten years. Chinese manufacturers made about 95% of the world's humanoid robots in 2025. [19]

Why it matters — Reuters found no armed humanoid with an operational PLA unit, and the machines are still power-hungry and unreliable outside demonstrations, so this is a procurement programme rather than a capability. [19]
16
Meta pays for your prompts with a discount

Most AI companies let customers opt out of having their usage read. Meta has put a price on the choice instead. For its new Muse Spark model, aimed at coding and other agent work, customers who agree to share their prompts and the model's answers pay about 95% less. A million input tokens costs $0.10 rather than $1.25, and a million output tokens $0.20 rather than $4.25. Meta has had a hard time getting training data, including a scheme earlier this year to record its own staff's computer use. [20]

Why it matters — Meta has put a price on a privacy setting. The developers most likely to take the discount are the small ones, who would otherwise pay twelve times as much to run the same model. [20]
17
Anthropic says copying Claude has gone underground

Jacob Klein, Anthropic's head of threat intelligence, told CNBC that an illicit ecosystem has grown up to reach its Claude models at scale and copy their answers into cheaper rival products. The technique is called distillation: you ask a strong model millions of questions and train a small one on its answers. Done with permission it is legal and ordinary. Klein says what he is seeing evades the company's controls by any means available, including opening accounts at extreme scale. [21]

Why it matters — Every safeguard Anthropic builds into Claude has to survive being asked a million questions by somebody who only wants the answers, which is a different problem from stopping a person misusing the model.
18
AI security funding doubles into a new market

HiddenLayer, which makes tools to protect AI models and agents from attack, raised $100m, three years after a $50m round raised when nobody could point to many real attacks on AI. Two other rounds closed in the same week for companies vetting the tools that agents install. Gartner estimates companies will spend $2.83bn this year on securing AI, 83% more than in 2025, and nearly $4.78bn next year. [22]

Why it matters — The money is arriving ahead of the incident record: there are still few public cases of agents being exploited, and buyers are paying for a risk they have mostly been told about rather than seen.
19
South Korea's chip exports set a record

South Korean semiconductor exports reached $46.65bn in August, up 209% on a year earlier and 47.5% of the country's total goods exports that month, its trade ministry reported. One economist estimated chips accounted for nearly 80% of August's export growth. Car exports fell 29.8% over the same period, which South Korea's trade ministry blamed partly on holiday timing and strikes. The Bank of Korea raised its base rate to 3% in August, its second rise running. [23]

Why it matters — An economy whose exports are half one product is exposed to how long the AI buildout lasts, and economists warn the sectors that would have to take up the slack are the ones already under pressure. [23]
20
Nvidia's stakes in other companies reach $99bn

Nvidia, whose chips nearly every AI model runs on, held equity investments worth $99bn as of 26 July, up from about $7bn a year earlier and $2.2bn two years before. It has committed more than $40bn during 2026. In August it announced arrangements with investment firms to mobilise more than $500bn of financing for its own chips, and up to $105bn of conditional credit for an OpenAI data centre in Ohio. The company says the investments cultivate its ecosystem and strengthen its position. [24]

Why it matters — The chip maker is increasingly financing the customers who buy its chips, which makes demand for those chips partly a product of its own balance sheet.
21
Investors are paying to walk around Chinese factories

A small industry has grown up flying foreign investors and executives to Chinese industrial cities to see the technology for themselves. Robert Wu of the Shanghai research firm Baiguan has run two tours for more than two dozen people at up to $15,000 for five days, about half of them from Southeast Asia. The Shanghai agency Glopen says inquiries rose 50% this year and it now runs more than 100 single-day company tours a month. The circuit covers Beijing, Shenzhen, Shanghai, Hangzhou and Hefei. [25]

Why it matters — Wu says visitors arrive with wrong assumptions, including that China leads on robotaxis when policy caution about job losses has kept it a step behind. [25]
22
Opendoor's founder targets the construction shortage

Eric Wu, who built and ran the American home-buying company Opendoor until 2022, has brought a new company out of stealth. It makes hands-free AI coaches for construction and other field workers, spoken instructions rather than a screen. The trade group Associated Builders and Contractors says the industry needs roughly 349,000 extra workers this year to keep up with demand. Where a large data-centre campus once needed about 750 workers at peak, the biggest now need far more. [26]

Why it matters — The buildings the AI industry needs are held up by a shortage of people who can pour concrete and pull cable, and the industry's answer is to sell those people software.
23
Gimlet valued at $3bn for spreading work across chips

Gimlet Labs, which makes software that splits AI work between different types of chip rather than running it all on one, raised $300m at a $3bn valuation. Andreessen Horowitz led the round, joined by Arm Holdings, which designs the chip blueprints in most phones, and M12, Microsoft's venture fund. Chief executive Zain Asgar said the company is also working with Arm. [28]

Why it matters — Its product exists because buyers want to use chips that are not Nvidia's, and the thing standing between them and that choice is software rather than silicon. [4]
24
ByteDance readies a model that generates worlds live

Zhang Yiming, ByteDance's founder, is personally overseeing a new AI model that generates spatial video in real time, Bloomberg reported, with a launch possible as soon as next month. It would be built on Seedance, the company's existing video system, and produce interactive worlds for live streams, short dramas and games at around 20 frames a second with about 0.05 seconds of delay. The generating happens in the cloud rather than inside the headset, so the headset itself can be simpler and cheaper to build. ByteDance owns the headset maker Pico. [30][31]

Why it matters — If the world is generated remotely, the contest over virtual reality shifts from headset specifications to models, cloud capacity and content, which is the ground ByteDance already holds. [31]
25
Kai-Fu Lee says China wins on reach, not frontier

Kai-Fu Lee, who ran Google's China business and has backed dozens of Chinese technology start-ups, told Bloomberg that Chinese AI companies are closing the gap despite years of restrictions on their access to advanced chips. The thing he says is most underrated is the speed of improvement and the fall in cost: a model that could handle a four-minute task a year ago can now handle a forty-minute one. He argues the adoption that follows will be faster than any previous technology. [29]

Why it matters — It is an argument that the race is not decided by who has the single best model, and it comes from someone who has run American technology businesses inside China.
26
Retroid teases a two-screen handheld

The handheld maker Retroid has released a teaser for the Pocket Duo, its version of the two-screen Nintendo DS shape, with no specifications, price or release date. It follows the Pocket 6 and the Pocket Flip 2, and unlike the company's existing clip-on second screen the Duo appears to be a single cable-free device. Rivals already on sale run from the $100 Anbernic RG DS to the Ayaneo Pocket DS at $579 or more. [27]

Why it matters — A shape Nintendo abandoned a decade ago is being rebuilt by small firms for people who want to play old games, on chips powerful enough to simulate consoles that were current until recently.
27
France and South Korea put money behind cinema

France gathered film industry figures from around the world on Monday to talk about streaming services and AI. President Emmanuel Macron, who chaired it with South Korean president Lee Jae Myung, announced plans to teach young people about cinema, preserve ageing film stock, save independent cinemas and fund domestic film industries. France and South Korea each pledged 500 million euros, about $580m, for their own filmmakers. Organisers said the threat needs a response across many countries at once. [32]

Why it matters — Two governments are trying to answer a technology shift with public money for the people it displaces, which is a different response from the one the US government took to the G20 last week.

02 Lesson why it matters

Somebody has to run the bad version until it gets good

A new machine only improves by doing real work, and the first factories to use one pay for that in ruined output, while everyone who buys it later gets it working.

The twist

A copy does not usually fail on the drawing board. It fails because every sensible customer keeps buying the original, and a machine nobody runs cannot get better.

How it works

A new machine starts out slower and less reliable than the one it copies
So every sensible buyer keeps ordering the proven one
But the new machine can only improve by doing real work, where its faults show up
So it improves only where a buyer has no alternative, or decides to eat the loss
And the proven maker stays ahead partly because it has customers, not only because it is better

The same force, elsewhere today

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

Saudi Arabia's HUMAIN and its Arabic model

Tareq Amin built Allam on about 4,000 chips against OpenAI's hundreds of thousands, and says what it really produced was a team that now knows how the technology works
Gimlet's $3bn valuation

the same step, sold as a product: its software exists to make chips that are not Nvidia's usable, which is what a buyer needs before they will risk the switch
China's military and its humanoid robots

the robots are unreliable outside demonstrations and the PLA is buying and testing them anyway, which is the buyer who accepts the losses so the machine can improve

Where you've seen this

Learner drivers

the only way to become safe on the road is to drive while you are not yet safe, and everybody else on the road carries part of that

A surgeon's first operations

they are slower and riskier, and those patients pay for every patient who comes after

A new bus route

it runs half empty for a year before anybody plans their day around it, and somebody has to fund the empty year

The catch

Forcing a buyer onto the worse tool is expensive, and it can simply buy a worse tool for longer. Being cut off does the forcing for free, which is why a ban sometimes helps the side it was aimed at.

And the whole of it

The phone in your pocket, the search box you typed into today and the plane you last flew on all work because earlier users put up with them not working. None of those people signed up for the job. Somebody is doing it for you right now, and you will never know who.

03 Truth what's really going on

What is really going on

China's answer to being locked out of ASML is not one machine. Huawei's investment funds are buying into the small Chinese firms that make the lenses and the light source a chip printer needs, and Yuliangsheng's machines are now being tested inside SMIC's factories. SMIC still prints its real chips on ASML's. [1]

Why it works on us — The word race turns every prototype into a lap time. The FT's sources describe something slower: years of tuning inside factories before a machine matches the one it copies. [1]

Who gains

Fujian Zhiqi Photonics and Beijing RSLaser — Two small Chinese suppliers of lenses and light sources now have Huawei money and a guaranteed customer, in the exact two parts where China had nothing. [1]
ASML — Being restricted made its machines worth stockpiling. Its China sales are forecast at about a fifth of revenue this year, and the memory makers CXMT and YMTC have bought enough to cover three years of expansion. [1]
American flatbed and heavy-haul trucking firms — Spot rates hit multi-year highs this summer hauling transformers, generators and concrete to data-centre sites, in a freight market otherwise squeezed by tariffs and record diesel prices. [16]
Firmus — OpenAI as an anchor customer pushes its contracted capacity past 900 megawatts, weeks before a rumoured Australian stock market listing. [13]
AI security start-ups — HiddenLayer raised $100m into a market Gartner puts at $2.83bn this year, 83% up on 2025, largely on risks buyers have been told about rather than seen. [22]
Pampa Energia — A country with no data centres yet, and so no backlash, lets it sell up to 500 megawatts from beside its own gas plant. [14]

Who pays

SMIC — It is running unproven Chinese machines on simpler chips to help improve them, which costs output, while everything that matters still depends on ASML. [1]
The people of Somerset, New York — The site's landlord promised 165 permanent jobs and $85m of investment in 2019 in return for a power discount. A 2024 planning presentation put the finished data centre at 35 to 40 jobs. [15]
Developers using Meta's new Muse Spark model — The cheap price is paid in prompts: $0.10 per million input tokens if you let Meta read your usage, $1.25 if you do not. [20]
Anthropic — Its head of threat intelligence says an illicit market now opens accounts at extreme scale to query Claude and train cheaper copies on its answers. [21]
South Korea — Chips were 47.5% of August exports and nearly 80% of export growth. Economists say the sectors that would have to take up the slack if the boom stalls are the ones already under pressure. [23]
IT managers buying AI — Four labs shipped models in one week, so the comparison work never finishes and the buying decision is never settled. [9]

What nobody knows yet

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

01

Whether any Chinese factory will print real production chips on a Chinese machine.

SMIC has bought and is testing them and still runs ASML's for the work that matters. The FT's sources give no date, and say the machines need several years of tuning to match ASML's output rate. [1]
02

Where the lenses in Yuliangsheng's machines actually came from.

Two people told the FT the machines use Zeiss projection lenses. Zeiss says it sells its lithography optics only to ASML. Industry insiders said they can be bought second-hand in China, and nobody has shown the route. [1]
03

What Anthropic found when it looked inside Decart.

Bloomberg's sources say Anthropic ran due diligence on a deal worth about $6bn and then walked away. Neither company has said why, and neither has commented publicly at all. [5]
04

Whether US and Chinese officials are meeting about AI this month.

Reuters reported a dedicated AI safety dialogue for mid-September led by Treasury Secretary Scott Bessent. An official in the Trump administration told Reuters there is currently no planned AI meeting in mid-September. Both accounts rest on anonymous sources. [11]
05

What the Belgian case will prove.

Prosecutors say a 52-year-old has been held since May over gallium nitride chip secrets and that a second suspect is at large. No trial has been held, the man has not been convicted of anything, and China's embassy in Belgium was not immediately available for comment. [6]
06

Who is responsible for safety at the Lake Mariner data centre.

TeraWulf owns it and leases the land from a company owned by its own chief executive, Fluidstack runs it, and Google guarantees the lease. After June's fire the local fire chief said in mid-August that the hydrants were still dry. [15]
07

Whether China's humanoid robots can do anything in a real fight.

Reuters found no armed humanoid with an operational PLA unit, and researchers say the machines are power-hungry and unreliable outside controlled demonstrations. China's defence ministry did not respond to questions. [19]
08

What OpenAI is paying Firmus for Malaysian computing.

Firmus declined to give the contract value and OpenAI did not respond to Reuters, so the only public number is the 900 megawatts of capacity it now has contracted across all customers. [13]
09

How many of the week's four model launches are real improvements.

Anthropic, Meta, Google and OpenAI all shipped inside seven days, and buyers say they cannot finish comparing one before the next arrives. The labs published their own numbers; nobody outside has tested them all. [9]
10

How much electricity a Patagonian data centre would need and where it would come from.

Pampa Energia is looking for investors for a site of up to 500 megawatts and says connectivity and grid infrastructure are still the obstacles. No permit or contract has been published. [14]

04 Hope carry this

American truck drivers are having their best summer in years. Spot rates for flatbed and heavy haul, the trailers that carry transformers and generators, reached multi-year highs, and small carriers are picking up new customers in towns that were never freight hubs.

Also true today

South Korea's semiconductor exports reached $46.65bn in August, the highest month it has ever recorded. Its non-chip exports rose 20% at the same time.