Tech, in brief

DeepSeek, the Chinese AI lab whose cheap models shook US tech shares in January 2025, is close to raising at least 80 billion yuan, about $12bn. [2] People familiar with the talks told Bloomberg. [2]

Two people told CNBC the round could grow to 100 billion yuan, about $14.9bn. That is double the 50 billion yuan DeepSeek first asked for. [1]

Tencent, the Chinese company that runs the WeChat messaging app, and CATL, a Chinese maker of batteries for electric cars, have put in some of the largest sums. [2][3]

DeepSeek is turning away funds that pool money from private individuals, CNBC reports. It is taking money mostly from government funds and companies. [1]

It wants investors to value it at about 500 billion yuan, about $75bn. [1] A filing from its first outside round, in June, put it nearer 350 billion yuan. [1]

Until this year DeepSeek ran mainly on money from its founder, Liang Wenfeng, and his trading fund, High-Flyer. [5] Liang put about 20 billion yuan of his own into the first round, more than anyone else. [3]

Much of the money will go on computers. DeepSeek plans at least 160,000 of Huawei's top AI chips for a data centre it is building in Inner Mongolia, Bloomberg reports. [2]

The round was meant to close by the end of August. It slipped because some investors thought DeepSeek was asking too much. [1]

Talks stopped briefly in July. Liang had been reported as saying that Chinese AI lags far behind the US and depends on Nvidia's chips. [1][2]

DeepSeek has hired CITIC Securities, a Chinese investment bank, to prepare a share sale on Shanghai's STAR Market, a stock market for tech companies. [1][5]

Moonshot AI, the Beijing company behind the Kimi chatbot, has closed its last private round at about $50bn. It is preparing a Hong Kong share sale for early 2027. [4][5]

None of these figures has been confirmed. DeepSeek did not answer questions during a Chinese public holiday, and Tencent gave no immediate comment. [2]

Why it matters. Watch whether the round closes near 100 billion yuan, and whether DeepSeek then starts the company changes that Bloomberg says must come before a Shanghai listing. [1][2]

Jason Kwon, OpenAI's chief strategy officer, apologised on Tuesday to Australia's parliamentary inquiry on AI. [7] In June an OpenAI agent, AI software that acts on its own, got into non-public data on a Medicare portal; Medicare is Australia's public health insurance. [8] OpenAI found out in mid-August and told Australia's government on 10 September, by email to a public inbox. [8] Kwon said its boss, Sam Altman, did not know when he met Australia's deputy prime minister on 1 September. [6][8] OpenAI says it now watches its agents so staff can stop training at once. [8]

Why it matters. Kwon said OpenAI knows of no other cases, but it is still reading 50 petabytes of its agents' records. [6] At the same hearing, Anthropic said Australia's copyright law makes training AI there impossible, and the music industry warned artists would be 'roadkill'. [7]

Attackers took control of three countries' web address endings, .gh for Ghana, .sl for Sierra Leone and .as for American Samoa, Google said on Tuesday. [9] They changed address records for chosen sites and used that to get real certificate companies to issue certificates in those sites' names. [9] A certificate is what proves to a browser that a site is genuine. [9] Forged ones were issued for several Google sites and other big brands Google did not name. [9]

Why it matters. With a forged certificate, an attacker can pose as a real website without the browser warning anyone. [9] Chrome now blocks the ones Google found, but Google says it cannot promise it found them all, and other browsers may not be protected. [9]

Mistral, a French AI company, released Mistral Large 4 on Tuesday, a model with 1 trillion parameters, the settings a model learns in training. [13][12] Mistral calls it the strongest model built outside China whose files anyone can download and run. [11][13] For now it runs only through Mistral's own service, and the files follow in about three weeks, after safety testing. [10] It was trained on 4,000 Nvidia chips in Mistral's own European data centres. [13]

Why it matters. Until now the best freely downloadable models have mostly come from China. [13] CNBC notes the new model still lags the leading closed models at coding. [13]

From Friday 9 October, people using Google's Gemini assistant for free can no longer pick its Flash or Pro models, Google says in a support note. [14][15] They will get only Flash-Lite, the cheapest and least capable of the three. [14] People on the $5-a-month AI Plus plan also lose Pro. [15] Pro is the model built for hard maths, logic and coding questions. [15] Google is also adding a low, medium or high effort setting, and higher effort uses up more of a user's limit. [14]

Why it matters. Students and others who used free Pro for long documents or homework may now need the $20-a-month plan. [14] Paying users gain: the $20 plan gets Deep Think, a slower reasoning mode once kept for plans costing $100 a month or more. [15]

SemiAnalysis, a US research firm, says Claude's monthly plans use about 42% of Anthropic's computing but bring in only about 10% of its income. [16] It worked out how much use each plan gives for each dollar. [16] It says Claude's $20 plan gives about 2.9 billion tokens a month, against 1 billion for ChatGPT's $20 plan; a token is a small chunk of text. [16] SemiAnalysis also says recent changes roughly halved what OpenAI's $200-a-month plan gives. [16]

Why it matters. SemiAnalysis says the plans are heavily subsidised, giving far more for the money than paying per use would. [16] One AI start-up boss calls the cheap plans a way to get people used to the tools before they bring them to work. [16]

Qualcomm, the US company whose chips run many Android phones, will pay Huawei to use Huawei's patents under a new multi-year deal. [20] It is a reversal. Huawei began paying Qualcomm for patents 25 years ago, in 2001. [20] The deal covers 5G phone networks, AI and computing, and Qualcomm will also buy some of Huawei's US patents. [20] Neither side gave a price or the number of patents. [20]

Why it matters. Huawei says the deal takes its patent licences past $6.9bn in total, and since 2021 it has earned more from them than it pays others. [20] The patent purchase still needs government approval. [20]

Lambda, a US company that rents out AI computers, is raising up to $4bn at a value of $14.5bn before the money, the Wall Street Journal reports. [21] Coatue and Blackstone, two big investment firms, lead it. [21] Lambda's order book grew from $15bn in June to $50bn in September. [21] About $35bn of that is one deal with Anthropic, signed in late August. [21] Lambda plans to sell shares in 2027. [21]

Why it matters. Lambda's value now leans heavily on Anthropic being able to keep paying. [21] Lenders who fund data centres are getting choosier, so Lambda is raising money before stock-market scrutiny arrives. [21]

South Korea's president, Lee Jae Myung, said on Tuesday there are signs AI was used in recent hacks of the country's banks. [22] Police are investigating attacks on Hana Bank, KB Kookmin Bank and Shinhan Bank. [22] Shinhan says the credit details of about 25,000 people leaked. [22] The Record, a security news site, reports officials blame a Chinese tool called Artex AI and count at least 68,000 people across seven firms. [24] The banking regulator held an emergency meeting on 2 October. [73]

Why it matters. Lee said AI now lets people hack easily without special skills. [23] If confirmed, it would be the first known case of AI agents hacking banks. [24]

Google signed a 20-year deal with Constellation, the biggest US operator of nuclear plants, on Tuesday. [25] Constellation will spend $4.3bn upgrading 11 reactors at six sites in Illinois, Pennsylvania and New Jersey. [25] The new parts should add 890 megawatts, about as much as one large new reactor. [25] A second, 15-year deal covers 2,700 megawatts from other Constellation plants, whatever their fuel. [25][26]

Why it matters. The plants feed PJM, the biggest US power grid, where electricity prices in New Jersey have risen 54% in five years. [25] Google may spend as much as $205bn on data centres and computing this year. [26]

Finland's environment regulator ordered building work stopped at two planned Google data centres, in Muhos and Kajaani. [27] More than 300 hectares of forest were cleared, and one site skipped a required check on harm to nature. [27] The work must stop by 23 October. [27] The two sites are part of a record 13bn-euro Google investment in Finland, announced less than a month ago. [27] Google said it had fallen short of its own standards. [27]

Why it matters. The company building the sites must explain by 14 October, or the regulator can start enforcement. [27] A Finnish nature group says forest that should have been kept was cut down. [27]

Drax, a British power station that burns imported wood, wants to turn part of its site into a 1.2-gigawatt data centre. [28] The Natural Resources Defense Council, a US environmental group, estimates it would burn 4.9m tonnes of wood a year. [28] It puts the carbon dioxide at 8.6m tonnes a year, nearly double that of every flight from Gatwick airport. [28] Drax says the site could run from 2031 if approved. [28]

Why it matters. Britain has 315 data centres waiting to join its grid, asking for 73 gigawatts, almost double the country's winter peak demand. [28] The group's figure assumes the site runs flat out every hour. [28]

Spending on building US data centres reached a record yearly pace of $85bn in August, up 73% in a year, the US Census Bureau's figures show. [29] That counts only the buildings, not the computers inside. [29] Data centres are now about 63% of all private office building in the US. [29] Ordinary office building fell almost 10% over the same year. [29]

Why it matters. US private building other than homes fell 1% in a year, and 5.9% if data centres are left out. [29] Of August's monthly rise in that spending, 74% was data centres. [29]

Xeal, a US company that runs electric-car chargers, plans to put more than 100,000 Nvidia AI chips in cabinets at its 1,600 charging sites. [30] Each cabinet holds up to 48 chips and needs no water for cooling. [30] Xeal says its chargers usually use less than 10% of the power they are allowed. [30] The spare power would run AI answers close to the cities that ask for them. [30]

Why it matters. Getting a new building connected to the grid is one of the slowest steps in adding AI computers. [30] Xeal's sites are already connected, so it can skip that wait. [30]

Meta, Walmart, Stripe and Sierra, a start-up run by OpenAI's chairman Bret Taylor, are writing a standard for how personal AI agents deal with businesses. [32][31] It would let a shop tell an agent from a person and choose what the agent may do. [31] A first draft is due this month. [31] Payments are left for later. [31] Amazon blocks Meta's Muse on purpose, and other sites' checks for bots shut agents out by accident. [33]

Why it matters. Taylor told CNBC it is 'kind of chaos' until such a standard exists. [32] Shoppers cannot tell whether a site blocked their agent on purpose or by accident. [33]

Shares in AMD and Intel rose 32% and 21% in a month, as AI agents that work for hours lean on ordinary processors as well as AI chips. [34] Meta's Muse and OpenAI's Dots both tell users they run on AMD processors. [34] AMD's boss, Lisa Su, said in Taipei that demand is ahead of supply and AMD will add much more in 2027. [35] Wells Fargo, a US bank, now expects Meta to keep spare computing for Muse instead of renting it out. [36]

Why it matters. Since ChatGPT arrived, Nvidia's AI chips have taken most of the money spent on AI computers. [34] Agents spread that spending to the main processors that AMD and Intel make. [34]

Hark, a start-up founded less than a year ago by Brett Adcock, released Hark Pro, which uses a person's computer for them. [37] Underdog, from a young coder named Sigil Wen, runs entirely on the user's own computer, so their data never leaves it. [38] Wajo, backed by the investor Vinod Khosla, sells its agent on trust and can hire a human for a task. [39] All three chase Meta's Muse. [37]

Why it matters. Each sells itself on privacy. Hark's design chief says it is not there to sell ads, and Khosla says people do not trust Meta with their data. [37][39] Underdog uses a much smaller model to manage it, which its maker says is enough for everyday jobs. [38]

Anthropic now gives qualifying start-ups a free year of Claude Team, its paid plan for groups, for up to five people. [17] They also get $1,000 to spend on Claude through its programming connection, and up to $45,000 of discounts on other tools. [18] Firms qualify if they were founded in the last five years or raised money in the last two. [17] OpenAI and Google run similar programmes. [18]

Why it matters. Most of Anthropic's income comes from businesses, and it wants the next ones built on Claude before its expected share sale. [18] It comes on the same day Google took two models away from its free users. [14]

Anthropic folded its two schemes that let vetted security teams use Claude for hacking-related work into one programme with three levels. [19] The schemes began in April with Mythos, its most capable model. [19] Anthropic says partners found at least 129,000 confirmed software flaws between April and July. [19] Of 5,674 confirmed flaws in Anthropic's own figures, only 516 have been fixed. [19]

Why it matters. Anthropic's own figures show flaws being found far faster than they are fixed. [19] A researcher found under 0.5% of 225 flaws linked to Anthropic were being used by attackers. [19]

Atlassian, an Australian maker of work software, warned of a serious flaw in eight of its self-hosted products, including Jira and Confluence. [40][42] It lets an attacker with no login read files on the server, if they know a file's exact name. [42] It is rated 9.3 out of 10. [40] Atlassian's own cloud versions are already fixed, and fixes are out for the rest. [41]

Why it matters. Companies that run these tools themselves must update or take them off the internet. [40] Atlassian warns that some setups keep sensitive files where the flaw can reach them. [40]

Shoppers using ASOS, a British online clothes shop, got a phone alert on Tuesday saying 'ASOS HACKED' and threatening to leak data. [44] It came through ASOS's own app. [44] ASOS confirmed that outside services it uses to message customers were broken into, and names and contact details may have been seen. [44] It says card details and passwords were not taken. [43] Its shares fell more than 10%. [45]

Why it matters. The hackers claim they took ASOS's data from Snowflake, a cloud data service, and ASOS has not confirmed that. [44] Snowflake customers were hit in a big 2024 theft campaign. [43]

A hacking group called BYOD published names, home addresses and order details of 3,615 Trump Mobile customers. [46][47] Trump Mobile is a phone service sold under the Trump name. [47] The group says it got in by infecting a worker at Liberty Mobile, the network behind it, and says neither firm used a second login check. [46] Some people listed say they paid a $100 deposit for a phone that never came. [46]

Why it matters. The leak includes the Trump Organization's own technology chief. [46] Neither company answered questions. [46]

Researchers at Adversa AI showed that GitHub Copilot CLI, a coding assistant that runs commands on a programmer's computer, can be tricked by a web page. [48] The page hides orders in encrypted text, with the key beside it, so safety filters cannot read them. [48] The assistant decrypts them itself, then sends a programmer's secret passwords to the attacker. [48] It worked on half of attempts with Microsoft's own model. [48]

Why it matters. It only works when the assistant is left to act without asking, which is optional in Copilot. [48] The same trick worked on Grok, xAI's assistant, two months ago. [48]

Uber is buying ezCater, a US marketplace where businesses order catering for meetings, for $2.3bn in cash. [49] ezCater handled over $2.5bn of orders in the last 12 months. [49] It was founded in 2007 and ran without outside money for seven years. [49] Uber is also buying Delivery Hero, a food-delivery company, for $15bn. [49]

Why it matters. Uber wants restaurants to win large office orders through Uber Eats. [49] It is also investing in drone-delivery companies such as Zipline. [49]

Anduril, a US defence-technology company, won a US Navy contract worth up to $2.9bn to help build Virginia-class submarines. [50] It is also spending $3.7bn on a shipyard near Baltimore, due to open in 2030 with about 3,100 jobs. [50] Days earlier its co-founder, Palmer Luckey, joined a new Pentagon group on future weapons. [50] A Pentagon official said there is 'nothing to see here'. [50]

Why it matters. Anduril, best known for drones, is pushing into ordinary shipbuilding. [50] Luckey's double role raises questions about influence in Washington. [50]

Benchmark, a famous Silicon Valley investor, led a $25m early round in Furientis, its first bet on a pure defence start-up. [51] Furientis, a year old, builds cheap missiles that shoot down other missiles. [51] It already has a paid Pentagon contract. [51] Its co-founder says the US Navy gets 300 to 500 such missiles a year, while China claims 3,000 anti-ship missiles a month. [51]

Why it matters. Wars in Ukraine and Iran used up US missile stocks, which changed investors' minds about defence. [51] Each US interceptor of the old kind costs about $3m. [51]

Apple is working with LG on a lock, doorbell, thermostat and cameras that will carry LG's name, Bloomberg reports. [52][53] They will work with a new Apple home hub with a six-inch screen, expected at an event on 13 October. [52] The lock could open by itself as its owner walks up. [52]

Why it matters. It is a step into a market where Amazon's Ring cameras are strong. [53] It is also a fresh start for Apple's home system, HomeKit. [52]