Tech, in brief

OpenAI's ChatGPT, Anthropic's Claude and xAI's Grok all started failing within about ninety minutes of each other on Thursday morning US time. [1][3]

SpaceX, which owns xAI, said the Grok failure came from an outage at its computer centre in Memphis, Tennessee, and apologised to its "impacted compute partners". [2]

Anthropic and xAI announced a deal to buy computing from SpaceX in May. Anthropic declined to comment on Thursday. [2]

OpenAI blamed a "routing error", meaning requests were sent down the wrong path inside its own systems. It began at 7:43am Pacific time and a fix went in 34 minutes later. [2][3]

Cloudflare, whose network sits in front of all three companies, said flatly that its services were running normally. The status pages of Amazon, Google and Microsoft's clouds showed nothing either. [3][2]

Anthropic's outage ran three hours and six minutes and covered Claude.ai, Claude Code, Claude Cowork and the programming interface other companies build on. [3]

OpenAI's took out search, file uploads, agents, voice mode, image generation, ChatGPT Work and its Codex coding tools at the same time. [5]

Complaints about Grok on Downdetector, a site where users report faults, went from under ten before 9am Eastern to 1,365 by 9:45am. [1]

Google never confirmed a fault, but Gemini complaints rose from 23 to 412 and a monitoring service recorded a likely outage of Google's AI programming interface between 10:45 and 11:15am Eastern. [1][2]

This is rare. Anthropic reports 99.4% uptime for Claude over 90 days and OpenAI reports 99.63% for ChatGPT; four frontier services faltering in one morning is close to unheard of. [1]

The Verge asked OpenAI, xAI and Anthropic what had happened while it was still happening. None replied immediately. [4]

Two days on, no company has said whether the three failures were one event. [2][3]

Why it matters. Watch whether OpenAI or Anthropic publishes a full account of Thursday morning. Until one of them does, the only thing linking the three failures is that they happened together.

Microsoft's Exchange Online started failing at 11:55am UTC on Monday 31 August, and by Tuesday afternoon the company still would not call it fixed. [6][7] Exchange Online is what actually delivers business email behind Outlook, and the fault spread from it to SharePoint, OneDrive, Teams search, Copilot, Defender and the Microsoft 365 admin console. [6][7] Microsoft named the cause as an issue in a core authentication configuration used by several services at once - the shared component that checks who a user is. [6] Downdetector logged more than 5,000 Outlook complaints by 2pm Eastern on Monday. [8]

Why it matters. Mail flow came back first and search stayed broken for another day, so many companies spent Tuesday able to send email and unable to find any. [7] OpenAI's ChatGPT Work went down the same Monday afternoon, unconnected as far as either company has said. [9]

Judge Leonie Brinkema's refusal on Wednesday to make Google sell its ad exchange leaves the European Commission the only regulator still demanding one, Politico reported on Friday. [10] The Commission fined Google 2.95bn euros a year ago over digital advertising and said then that selling part of the business looked like the only fix. [10] Its own competition director general, Anthony Whelan, has since called structural remedies very difficult. [10] Lawyers for the companies that brought the complaint say privately they no longer expect a sale. [10]

Why it matters. Brinkema found Google liable in April 2025, then accepted mostly behavioural changes and issued the opinion under seal. [11][12] It is the third time in a row that a US judge has rejected an attempt by antitrust enforcers to break up a big technology company. [48][13] Semafor argued the same week that a breakup would simply strengthen a different monopoly - spin off Android and cheap phones get dearer, which helps Apple. [14]

Nvidia confirmed on Thursday it will buy Hugging Face for $12.93bn, its second-largest deal after the $20bn Groq purchase in December. [15][16] Hugging Face is where AI models are published and downloaded: Nvidia says 18 million developers use it and it holds more than three million models. [16] Two weeks earlier Stripe agreed to buy OpenRouter, which routes developer traffic to over 400 models and handles more than 10 trillion tokens a day - tokens being the chunks of text that AI systems are billed by. [16] That is roughly $21bn in a fortnight on two firms that make no models of their own. [16]

Why it matters. Both sit between the people who build models and the people who use them, which is exactly where developers went to avoid depending on any single AI company. [16] Hugging Face's chief executive Clement Delangue approached Nvidia himself, Jensen Huang wrote. [15]

AT&T, the US telecoms company, has shifted 40% of its AI work onto open models, up from 20% in May. Open models are the kind you download and run on your own machines. AT&T says it now pays up to 80% less for AI than it did earlier this year. [17] Airbnb and Deloitte have made the same move. [17] Open models were 58% of US traffic on the marketplace OpenRouter last month, against 10% a year earlier. [17] The word open is doing loose work: most of these models release their weights but not their training data, which falls short of the Open Source Initiative's definition. [16][18]

Why it matters. This is the demand that made Hugging Face worth $12.93bn to Nvidia, whose chips those models run on. [15][17] A company that switched to open models to escape one supplier now downloads them from a library the chipmaker owns.

Moonshot AI, the Beijing lab behind the Kimi models, has filed confidentially for a Hong Kong listing and is weighing a raise of up to $5bn, Bloomberg reported on Friday. [19] Bank of America, Goldman Sachs and Deutsche Bank are reported to be arranging it alongside China International Capital Corporation. [19] Two other Chinese labs, Zhipu and MiniMax, listed in Hong Kong in January, before OpenAI or Anthropic had filed anywhere. [19] US export rules govern which chips American firms may sell to China. They say nothing about which listings American banks may underwrite. [19]

Why it matters. Moonshot trained its Kimi K3 model on a 20,000-chip cluster rented through Alibaba's cloud, so the hardware limit did not bind either. [19] The reported size moved from about $3bn on Thursday to $5bn on Friday, and Moonshot has announced neither. [19]

Cognition, an American AI coding startup, is closing a funding round of around $1bn that values it at about $47bn, Bloomberg reported on Wednesday, citing people who asked not to be identified. [20] The round drew nearly $10bn of investor interest, one of those people said, and the final size may end up larger than $1bn. [20] Talks are still going on and the details could change. [20] Its main product is a coding agent, software given a programming task that then writes and tests the code itself. [20]

Why it matters. Investors offered close to ten times what the company is actually raising. [20] Cognition itself has announced nothing, and every figure here comes from people who would not be named. [20]

Andreessen Horowitz, one of Silicon Valley's largest venture capital firms, has grown its fifth growth fund to $8.5bn, adding $1.75bn since it launched in January at $6.75bn. [21] Days earlier it raised $1.1bn for a separate fund aimed at AI hardware startups working on chips, memory, networking and storage. [21] The growth fund backs companies already selling and scaling rather than new ones, and has invested in more than 100 over seven years. [21] The firm announced $15bn of fresh funding in January, when it managed $90bn. [21]

Why it matters. Money for the dull layers - memory, networking, storage - is now being raised in a fund of its own, which is a fair guide to where the shortages are. [21]

AIR, an Israeli startup founded by two veterans of Israel's Unit 8200 intelligence corps, where they worked on offensive cybersecurity, raised $50m in two rounds weeks apart: $10m led by Sequoia and $40m led by Greenoaks. [22] Its software finds AI agents running inside a company and vets the tools, skills and add-ons those agents install, blocking any that fail its checks. [22] Co-founder Yair Saban compares the gap to device drivers, which nobody signed in the early 2000s and everybody signs now. [22] Agents currently install add-ons with no equivalent check. [22]

Why it matters. Agents are getting the run of company systems faster than anyone has built the permissions to govern them, and that gap is now a funded market. [22]

Thyme Care, an American company that guides cancer patients through treatment by video and phone, raised $125m in a Series E round led by Morgan Health. [23] The health insurer Humana and CVS Health Ventures both took part, alongside AlleyCorp, HealthQuest Capital and a16z Bio + Health. [23] The round values the company above $2bn, roughly double what it was worth less than a year ago. [23] It is now forming a parent company to build further businesses around how cancer care is delivered and paid for. [23]

Why it matters. The money came largely from the insurers who pay the bills for the patients the service guides, which is a bet that a large share of the cost of cancer care is disorganisation rather than medicine. [23]

South Korean semiconductor exports rose 209% from a year earlier to a record $46.65bn in August, the country's trade ministry reported this week. [24] That was 47.5% of the $98.25bn of goods the country sold abroad that month, and one bank strategist put chips at nearly 80% of all export growth. [24] South Korea's trade ministry credited AI spending by large cloud companies including Google and Amazon. [24] Car exports fell 29.8% over the same month, which it blamed partly on holiday timing and strikes. [24]

Why it matters. An economist at Moody's Analytics said a gradual slowdown would be manageable and an abrupt stall would not, because the sectors that would have to take up the slack are the ones already struggling. [24] South Korea's central bank raised its main rate to 3% in August, its second rise in a row. [24]

A South Korean presidential official told reporters on Friday that semiconductor investment in the United States is part of wider talks between the two countries. [25] US Commerce Secretary Howard Lutnick said this week his government is planning a targeted, thoughtful tariff policy on imported chips. "If you don't build here, expect to pay to enter the greatest market in the world," he said. [25] Under a deal struck last year, South Korea committed $350bn of investment in US manufacturing in return for chip tariff terms no worse than any comparable competitor's. [25]

Why it matters. South Korea is home to Samsung and SK Hynix, the world's two largest makers of memory chips, so a US chip tariff would land on the country whose exports have just tripled on American AI demand. [25][24]

A Chinese court has frozen up to 2.14bn yuan, about $300m, of assets belonging to Nexperia, a Dutch chipmaker that supplies the basic chips inside cars and consumer electronics. [26] The case was brought by Wingtech, the Chinese electronics group that owned Nexperia until Dutch authorities stripped it of control last year over fears technology and money would be moved abroad. [26] Wingtech is seeking 8bn yuan in damages, and stakes in four Nexperia units in China are covered. [26] Nexperia says day-to-day operations are unaffected and the case has not been heard on its merits. [26]

Why it matters. The last round of this fight saw China impose export controls on Nexperia's operations there, disrupting shipments and threatening global supplies of the chips used in cars. [26] This move gives Wingtech leverage without changing who runs the company.

Taiwan's Ministry of Justice Investigation Bureau has examined 166 companies in six years for concealing Chinese ownership. It completed 67 further trade-secret investigations in the technology sector over the same period. Both figures were given to Rest of World and have not been public before. [27] On 5 August it raided another 18 technology firms. [27] One engineer described taking a $120,000 job - double the local average - at a Hsinchu company presented as American, then finding every colleague was in Nanjing and his chip designs went there. [27] The Taiwanese arm was shut down after a 2021 raid. [27]

Why it matters. Taiwan makes more than 60% of the world's chips and 90% of the most advanced ones, and Chinese companies need government approval to operate there at all. [27] Hiring the engineers is cheaper than building the knowledge, and far harder to see than shipping hardware.

A Heatmap Pro and Embold Research poll found 75% of Americans would oppose a data centre being built in their community, including 66% of Republicans, 80% of independents and 83% of Democrats. [28] A year ago opinion was almost evenly split; net approval has swung 62 percentage points since. [28] Some $130bn of projects were delayed or cancelled in the first quarter of this year alone. [28] Five hundred jurisdictions have passed bans or pauses, 150 of those local bans in July. [28]

Why it matters. Pennsylvania's governor Josh Shapiro, who signed an executive order on data centres on 18 August, told CNBC on Tuesday that too many developers do not care about the communities they build in. [29] President Trump had said the day before that the only reason a community should not want one is if it wants to end up backwards and poor. [29]

The suppliers of electrical gear are the quiet winners of the AI build, Reuters reported on Tuesday. [30] South Korea's HD Hyundai Electric, which makes transformers - the boxes that step high-voltage grid power down to something servers can use - grew its order backlog 23% to $8.5bn in six months. [30] It expects data centres to be 16% of its power division's new orders next year, up from 6.3% this year. [30] China's Hainan Jinpan more than quadrupled its data-centre orders in the first half. [30]

Why it matters. Buyers want sites delivered within six months, while grid connections take up to 24 months in some emerging markets and more than eight years in major developed ones. [30] Bank of America estimates power use per AI rack could pass 1.5 megawatts by the end of 2030, nearly 100 times a conventional rack. [30]

Global X, which runs a range of AI investment funds, has filed a prospectus with the SEC, the US financial regulator, to launch one tracking makers of capacitors. [31] Capacitors are the small components that smooth out the violent swings in electricity demand inside a data centre. [31] It is due within a few weeks and the company is in a quiet period until then. [31] Similar narrow funds already exist for memory chips and for the optical parts that shuttle data between machines. [31]

Why it matters. Investors are moving from the chips everyone can name down to the parts inside the building, which is where the shortages now are. [31][30]

Australia's federal energy minister Chris Bowen said on Friday that Australia's government still intends to legislate national standards requiring new data centres to run on 100% renewable electricity, backed up by gas when needed. [32] Prime Minister Anthony Albanese wants a bring-your-own-power rule. The Queensland and Northern Territory governments want to sell them electricity from their own state-owned gas and coal stations instead. [32] A national cabinet meeting last week produced a statement worded so every leader could claim a win, and no clarity at all about the rule. [32]

Why it matters. Bowen said exceptions would be considered only where state-owned generation is cheaper. [32] Electricity is about 10% of a data centre's total costs, so the argument is mostly over who gets to sell the power. [32]

At the G20 innovation ministers' meeting in Chapel Hill, North Carolina, a Trump White House official said the United States would press other members to take a hands-off approach to regulating AI. [33] Elon Musk, speaking by video on Tuesday, told the meeting governments should make things default legal, not default illegal, and predicted AI would add 20% to 30% to the global economy. [33] Former White House AI adviser David Sacks said local communities can benefit from data centre construction. [33]

Why it matters. In the same week, three in four Americans told pollsters they would oppose a data centre near them and 500 US jurisdictions had already restricted them. [28] The argument the US is making abroad is the one it is losing at home.

Google released Gemini 3.8 Flash on Wednesday, its third Flash model in six weeks, aimed at coding and multi-step tasks, alongside a cybersecurity model for government and enterprise customers. [34] Flash models are the smaller, cheaper ones: this one costs 75 cents per million tokens of input and $3.75 per million of output, tokens being the chunks of text AI systems are billed by. [34] Tulsee Doshi of Google DeepMind said the recent Flash models had surprised the company in positive ways. [34] Analyst Gil Luria of D.A. Davidson said it keeps Google in the race but probably will not shift its distant third place among business customers. [34]

Why it matters. It landed the same day a judge refused to break up Google's advertising business, after a summer of talent departures and a restructuring inside DeepMind. [34][11] Alphabet's shares rose 0.6% on the day and are still down slightly for the month. [34]

OpenClaw, the open-source AI coding agent, released version 2.0 at the end of August, under the name v2026.8.1, with a rebuilt browser interface putting conversations, files, approvals, configuration and live agent activity in one shared workspace. [35] It adds shared cloud sessions, multi-user collaboration, stronger sandboxing, role-based permissions, approval controls, secrets handling and auditing. [35] Those are the controls a company needs before it lets staff use a tool at all. [35] Peter Steinberger wrote on X on 31 August that the team had spent two months building OpenClaw with OpenClaw, moving off individual local setups onto a shared environment. [35]

Why it matters. The security and isolation gaps in the older version are what prompted rivals such as NanoClaw; version 2.0 closes them, though not always by default. [35]

Broadcom announced TrueSource, a service promising clean and secure versions of the Spring Java framework and RabbitMQ, and later of wider Java, Python and Node.js libraries. [36] Spring is the framework a large share of business Java software is built on; Broadcom acquired it through VMware and already employs many of the people who maintain it. [36] Purnima Padmanabhan, who runs Broadcom's Tanzu division, said the company chooses and builds every Spring library and database component itself. [36]

Why it matters. Companies use open-source libraries partly so they do not depend on one supplier. Broadcom is now selling the assurance that those libraries are safe. [36]

The first laptop built on Nvidia's RTX Spark chip was shown at IFA in Berlin: Lenovo's Yoga 9n, a 16-inch machine 0.69 inches thick, which is 0.03 inches off a 16-inch MacBook Pro. [37] The chip puts an Arm processor of up to 20 cores, a graphics part of up to 6,144 cores and the memory onto one piece of silicon. [37] Nvidia announced small desktop machines using it too. [37] The pitch is running AI agents on your own computer instead of somebody else's servers. [37]

Why it matters. Every AI task run on your own machine is one that does not depend on a service that can stop answering at 6:23 in the morning. [37][1]

Lenovo demonstrated Project AeroBlade at IFA, a concept laptop cooled without fans, heatsinks or large vents. [38] It uses AirJet modules from Frore Systems, which build about 2,000 pascals of back pressure - air moving at roughly 124 miles an hour. That air is blasted across a copper plate to strip away the still layer of hot air that makes chips slow themselves down. [38] Lenovo is the first computer maker to put the technology in a laptop, and says the space saved leaves room for a bigger battery. [38]

Why it matters. Fans set the thickness, the noise and the battery size of every thin laptop, so removing them changes all three at once. [38]

Acer showed Project DualPlay Mini at IFA, a concept 8-inch gaming handheld that splits in the middle. The front flips up on a hinge to reveal a keyboard, and the screen folds back over so the thing works as a very small laptop. [39] Acer stressed it is a concept only and gave almost no specifications beyond an Intel chip and the same twin-fan cooling as its Predator Atlas 8. [39] It also launched the Predator Atlas 7, a 7-inch handheld with an Intel Arc G3 Extreme chip and a 1080p screen. [39]

Why it matters. Handhelds have converged on one shape over three years, and Acer is testing whether buyers want one that also types. [39]

Minisforum launched two machines at IFA built on AMD's Ryzen AI Max+ Pro 495 chip, which the company rates at up to 131 trillion operations a second and pairs with up to 192GB of shared memory. [40] The AI Agent NAS N5 is a storage box holding up to 200TB with the OpenClaw agent pre-installed; the MS-S1 is a small desktop for the same work. [40] The idea is running agents where the data already sits, so nothing has to leave the machine. [40] Minisforum says agent software can use a thousand times more text than an ordinary AI request. [40]

Why it matters. Every one of those words is billed when the work runs on somebody else's computer, and free when it runs on yours. [40]