Tech, in brief

Taiwan's Keelung prosecutors indicted nine people on 24 August over the illegal export of Nvidia B300 AI servers to China. [1][2][3]

One of the nine is a sales manager at Nvidia Taiwan. Two work for Super Micro's Taiwan arm. [1][3][4]

The order was for 130 servers. Seventy-four reached China: 16 went direct, 50 through Indonesia, and eight via Japan and Hong Kong. [1][3]

Every control existed. Super Micro sells these systems only to approved buyers, who sign a promise not to re-export, and an order of eight or more triggers an on-site inspection. [1]

The inspection happened in September 2025 and it passed. Inspectors found a rented site with no racks, no power and no bandwidth to run 130 servers. Nobody said so. [1]

The buyer, a Taiwanese trading firm called Flying Tiger Tech, had shown a price quotation for the space instead of an actual lease. [1]

The people checking were the people paid on the sale. An Nvidia sales manager pushed the order through and emailed headquarters that the inspection was done. [1]

A Super Micro manager coached the buyer through the review. A second manager worked out where the servers were really going and was cut into the commissions instead. [1]

The shipments that got through earned about $21.2 million. [1]

What stopped the rest was outside the chain. Customs flagged 56 units declared to Japan, and the defendants answered with a fake mock-up of Super Micro's website. [1]

Super Micro's own outside investigation cleared its senior management and found no evidence of banned sales. Taiwan's prosecutors found this channel, not that audit. [1][4]

Washington is widening the net. The US is investigating Singapore freight forwarder Apex Logistics, which would be the first case against a transport company in the chip trade. [5]

It is also drafting a rule aimed at Chinese firms renting Nvidia servers remotely in Thailand and Singapore, and could show it to industry as early as September. [6]

Why it matters. Watch whether Commerce actually circulates the remote-access rule next month, and whether it survives the objection that it can only police goods that physically cross a border. [6]

Z.ai says it used 100,000 China-made chips to serve every request to GLM-5.3-Flash, the model it released anonymously as Ox Alpha. Its shares rose 8%. CNBC could not verify the claim, and the company would not name its suppliers. [8][12]

Why it matters. It is the first public claim of a top-ranked model served end to end on Chinese hardware, which is the outcome the export rules were meant to delay.

Tencent published a preview of Hy4 on Friday, a free-to-download model for software engineering, research and financial analysis. It has 770 billion parameters in total but uses only about 49 billion for any one request. [13]

Why it matters. Chinese labs keep releasing capable models for nothing, which is a hard price for paid American models to answer.

Nvidia said on Wednesday it is optimising its hardware for leading open Chinese models, including DeepSeek and Alibaba's. The same week it warned in a filing that White House restrictions on Chinese AI are a business risk. [7]

Why it matters. Nvidia is hedging against its own government, because whoever tunes the hardware for the popular model keeps the customers.

At the Hot Chips conference, OpenAI published benchmarks for Jalapeno, the inference chip it built with Broadcom. It reports more work per kilowatt and faster replies than Nvidia's Blackwell. Volume deployment comes in 2027. [14][15]

Why it matters. Inference is the part of AI that runs every day, and the biggest buyers are now building the silicon for it themselves.

Nvidia agreed a $6 billion deal for model builder Poolside, Stripe bought OpenRouter for more than $7 billion, and Hugging Face is reported in talks with Nvidia. [16]

Why it matters. A great deal of money is buying companies whose product is given away, because the value sits in the developers, not the file.

Micron told Hot Chips that high-bandwidth memory needs roughly three times the silicon of ordinary DDR5 for the same capacity, and that the gap widens every generation. Ordinary memory contract prices rose 90% to 95% in a single quarter. [9]

Why it matters. The same factories make both, so every wafer moved to AI memory deletes several ordinary chips. That is why laptop memory costs what it does.

Memory now accounts for about 30% of what the big cloud firms spend on kit, up from around 8% in 2023. Meta is running recovered DDR4 from decommissioned machines across millions of servers, cutting server counts by up to 25% for some jobs. [10]

Why it matters. Recycling old memory was a sustainability idea two years ago. The shortage has turned it into a cost measure.

SK hynix broke ground this week on the first American plant to package high-bandwidth memory, in Indiana. It says production starts in 2029. [11]

Why it matters. The fix for the shortage is three years out, which is roughly the shape of every supply answer in this industry.

Chinese customs are delaying shipments of germanium, quartz and permanent magnets to Taiwanese manufacturers, according to Nikkei. Some firms are losing orders while they wait for clearance. [49]

Why it matters. Beijing controls inputs Taiwan's optics and robotics makers need, and a slow customs queue is a squeeze that never has to be announced.

From 30 August, Google will stop applying manual demotions to searchers in the 27 EU countries plus Iceland, Norway and Liechtenstein. The change settles a Commission investigation that could have ended in a fine. The policy is unchanged everywhere else. [17][18]

Why it matters. The same page can now rank differently depending on where the reader is sitting, which makes a global ranking rule visibly a policy choice.

Google agreed to settle a mass London claim brought for app developers for 260 million pounds, about $353 million, with no admission of liability. The tribunal still has to approve it. [19]

Why it matters. It is the second UK collective claim Google has settled rather than fought to judgment this year.

Poland's digital affairs minister asked the European Commission to fine Meta 250 million euros over fraudulent advertising. Poland cannot impose the fine itself; only the Commission can act against very large platforms. [20][21]

Why it matters. The Digital Services Act allows up to 6% of global turnover, so 250 million euros looks like a figure chosen to be plausible rather than painful.

Brazil's data protection agency fined ByteDance 153.7 million reais, about $29.9 million, over children's data. The ruling covers people browsing without an account, not just logged-in users. [22]

Why it matters. Most child-safety enforcement asks what age a user typed in. This one applies to people who never signed up at all.

Zillow and Redfin settled with the US Federal Trade Commission and five states hours before trial. The case concerned a 2025 deal under which Redfin showed Zillow's rental listings instead of competing for the same advertisers. [23]

Why it matters. The arrangement could have kept Redfin out of rental advertising for up to nine years.

A San Francisco judge said on Thursday he was likely to deny California's request to block Amazon from what the state calls price fixing. The attorney general says Amazon pressed brands to push rivals' prices up. [24]

Why it matters. The state can still take the case to trial, but it loses the fast remedy it wanted.

KKR agreed to pay $250 million over failures to file at least 16 deals with US antitrust regulators before completing them. It is the largest such penalty ever. KKR says outside law firms will reimburse it in full. [25]

Why it matters. A record fine that costs the company nothing is a lesson about who actually bears a penalty.

Ireland still bans funding a lawsuit you are not party to, under a statute from 1634. Only one collective action has been filed there since the EU created the right in 2020. [26]

Why it matters. Many large technology firms have their European base in Ireland, so the venue where they can be sued together barely functions.

Emerald AI raised $150 million at a $1.05 billion valuation to make data centres adjust their power draw when the grid is strained. It says the approach could free more than 100 gigawatts on the existing US grid. Silicon Valley Power, which serves about 55 data centres in 20 square miles, is a customer. [27][28][29]

Why it matters. Flexibility is cheaper than new wires, and new wires are the thing that takes a decade.

Ofgem wants developers to pay a large non-refundable deposit, line up customers in advance and prove funding before joining the grid queue. There are 315 projects in the queue representing 73 gigawatts of demand. [30][31]

Why it matters. A queue that costs nothing to join stops telling anyone what demand is real.

Analysis by the group Foxglove says the planned Wapseys Wood and Quest Park sites will use 1.3 gigawatts and produce more than 4.5 million tonnes of carbon a year. ExxonMobil's UK emissions were 3.9 million tonnes in 2023. [31]

Why it matters. The UK's climate targets are legally binding, and two buildings would move the national number.

Australia's grid operator says data centres will draw 13% of the national electricity market by 2035-36, up from 3% today. It arrives as 13 gigawatts of coal retires and more than a third of the project pipeline has already been cancelled. [32]

Why it matters. Demand is forecast to rise as the supply that would have met it is switched off.

Under a draft Spanish decree, data centre operators must be established in the EU, and the data and metadata they handle must stay in the EU. Efficiency and water requirements would match the top tier of a coming EU label. [33]

Why it matters. Spain has attracted tens of billions in projects, and is now choosing among them rather than competing for them.

Gas-fired power in development specifically for US data centres has almost doubled in under a year, according to Global Energy Monitor. The group counted 97 gigawatts of such projects at the end of 2025. [34]

Why it matters. The companies with the loudest clean-energy commitments are building private fossil plants to meet their own demand.

A review of multiple US polls found opposition to data centres is driven by local electricity price rises and noise, not by views on AI or big technology firms. Political affiliation had little effect. [35]

Why it matters. The industry has been answering an argument about AI that the neighbours were never having.

Indiana Michigan Power proposed cutting residential bills by about $59 million by 2027, roughly $100 a year for a typical home, plus a three-year rate freeze. It says large customers including data centres are driving the extra revenue. [36]

Why it matters. It is the first sizeable case pointing the other way, against a year of local bills going up.

Data centre demand has pushed solid-state transformers out of the lab. A one-megawatt system from NC State is now being field tested on a live distribution line. [37]

Why it matters. Transformers are dull, made slowly, and currently one of the longest waits in connecting anything to a grid.

AWS agreed to acquire DuckLabs, the Amsterdam company behind DuckDB, and its roughly 30 engineers. DuckDB itself stays free under the MIT licence, held by an independent foundation. AWS has not said what it wants from the team. [38][39][40]

Why it matters. Cloud firms could always use open-source code for free. Buying the people who write it is a different move.

An Anthropic paper reports automated research systems improved a model's score on all ten tests of misbehaviour it was given, without making the model worse overall. Each system reads the literature, proposes a method and trains for 30 minutes. [41]

Why it matters. If a model can improve its own training, the researcher becomes the slowest part of AI research.

OpenAI, Anthropic and Meta each disclosed that a model had broken into outside systems during testing. All three tests were run by the same Israeli start-up, Irregular, which made a mistake. [42]

Why it matters. The safety evaluations are themselves a shared dependency, and one error propagated to three frontier labs.

The medical device maker said a cyberattack is causing a global disruption to its operations. [43]

Why it matters. The devices are implanted; the disruption is to the company that supplies and supports them.

The FBI seized two hacking platforms and a network of compromised devices it attributes to a Chinese group called QTFY. Court filings say a private Nanjing company was paid by China's Ministry of State Security. Targets included NASA, the Energy Department and the US Senate. [44]

Why it matters. The filings describe state hacking as a contracting market, with former military staff winning the work.

OpenAI, Anthropic, Google, Microsoft, AWS, IBM, Cloudflare, CrowdStrike and more than a hundred others signed an open letter warning about increasingly capable AI attacks. [45][46]

Why it matters. The signatories include both the firms building the capability and the firms selling protection from it.

India's TCS agreed to buy Porsche's IT division MHP for 320 million euros, about $373 million. Porsche committed 1.25 billion euros of work to TCS over five years as part of the deal. [47]

Why it matters. The commitment is worth more than three times the price, which is what is actually being bought.

Xiaomi's new Xring O3 is reportedly made at TSMC on a 3-nanometre process, for a folding phone expected to ship 200,000 to 300,000 units. Xiaomi ships tens of millions of handsets a quarter. [48]

Why it matters. At that volume it cannot be about cost. It is about proving the company can design a leading-edge chip at all.

GitHub's build and test service went down on Wednesday afternoon. The company failed a database over to a replica, which did not fix it, then throttled traffic while it worked on the problem. [50]

Why it matters. It came days after GitHub publicly promised developers it was scaling up.

Instinct raised $250 million at a $2.5 billion valuation, General Intuition reached $6 billion weeks after a $2.3 billion round, and robotics firm Generalist hit $3 billion. Stability AI raised $76 million, backed by three major music groups and EA. [51][52][53][54]

Why it matters. Stability's investors are the companies it needs licensing deals with, which is a different kind of round from the others.

Korea's Wrtn raised about 100 billion won at an $870 million valuation. Mexico's Primero took $12 million in seed funding to connect legacy systems at large Latin American firms. Keenable raised $26 million to index the web for AI agents rather than people. [55][56][57]

Why it matters. The clearest returns are outside the frontier: joining up software that already exists.

Nvidia is discussing an investment in Perplexity in a round valuing the search company at more than $30 billion, according to The Information. Reuters could not verify the report. [59]

Why it matters. Nvidia keeps investing in the customers who buy its chips, which makes its demand harder to read.

Governments outside the US and China have spent years trying to build alternatives to American cloud providers, and the dependency has not moved much. [58]

Why it matters. A 2024 update from a single security vendor took down 8.5 million machines, airports and banks among them.

SiFive launched what it calls the first enterprise-grade 2U rack server built on RISC-V, the open chip instruction set. SiFive designs the chips and does not make them. [60]

Why it matters. An open instruction set is the one part of the chip stack no government can restrict.

The Register worked through the filings behind the reported $120 billion Google and Marvell chip agreement. The 8-K says very little, Google said nothing publicly, and the detail is in a share warrant. [62]

Why it matters. Large round numbers in this industry are usually a headline built on a multi-year option.

Cursor described how it rebuilt Git on object storage to handle its own scale, after finding the standard tool struggled. Its engineers say AI agents made it worse: more code, more pull requests, more test runs. [63]

Why it matters. Agents write faster than the tools underneath them were designed to absorb.

A consultant reviewing three years of corporate AI projects concludes the failures come from poor data, unclear processes and missing governance, not from the models. [64]

Why it matters. It is the least fashionable diagnosis and the most common one.

Models went from solving 18% of New York Times Connections puzzles in late 2024 to near-perfect within months. Small changes to familiar riddles still trip them up. [65]

Why it matters. What a model fails at tells you more about how it works than what it passes.

Rest of World's publisher travelled with a delegation through China's AI and robotics industry, expecting a different future, and found the same debates in a different register. [66]

Why it matters. The two-competing-futures framing is doing a lot of work in Washington that the ground does not support.

Thirty-five years ago Linus Torvalds posted a note about a hobby operating system he said would not be big. He nearly called it Freax. [61]

Why it matters. It now runs most of the servers every story above depends on.