Security

FCC To Rescind Ruling That Said ISPs Are Required To Secure Their Networks (arstechnica.com) 47

The FCC plans to repeal a Biden-era ruling that required ISPs to secure their networks under the Communications Assistance for Law Enforcement Act, instead relying on voluntary cybersecurity commitments from telecom providers. FCC Chairman Brendan Carr said the ruling "exceeded the agency's authority and did not present an effective or agile response to the relevant cybersecurity threats." Carr said the vote scheduled for November 20 comes after "extensive FCC engagement with carriers" who have taken "substantial steps... to strengthen their cybersecurity defenses." Ars Technica reports: The FCC's January 2025 declaratory ruling came in response to attacks by China, including the Salt Typhoon infiltration of major telecom providers such as Verizon and AT&T. The Biden-era FCC found that the Communications Assistance for Law Enforcement Act (CALEA), a 1994 law, "affirmatively requires telecommunications carriers to secure their networks from unlawful access or interception of communications."

"The Commission has previously found that section 105 of CALEA creates an affirmative obligation for a telecommunications carrier to avoid the risk that suppliers of untrusted equipment will "illegally activate interceptions or other forms of surveillance within the carrier's switching premises without its knowledge,'" the January order said. "With this Declaratory Ruling, we clarify that telecommunications carriers' duties under section 105 of CALEA extend not only to the equipment they choose to use in their networks, but also to how they manage their networks."
A draft of the order that will be voted on in November can be found here (PDF).
Power

The World's Secret Electricity Superusers Revealed (bloomberg.com) 35

An anonymous reader shares a report: The rush to secure electricity has intensified as tech companies look to spend trillions of dollars building data centers. There's an industry that consumes even more power than many tech giants, and it has largely escaped the same scrutiny: suppliers of industrial gases.

Everyday items like toothpaste and life-saving treatments like MRIs are among the countless parts of modern life that hinge on access to gases such as nitrogen, oxygen and helium. Producing and transporting these gases to industrial facilities and hospitals is a highly energy-intensive process. Three companies -- Linde, Air Liquide and Air Products and Chemicals -- control 70% of the $120 billion global market for industrial gases. Their initiatives to rein in electricity use or switch to renewables aren't enough to rapidly cut carbon emissions, according to a new report from the campaign group Action Speaks Louder.

"The scale of the sector's greenhouse gas emissions and electricity use is staggering," said George Harding-Rolls, the group's head of campaigns and one of the authors of the report. Linde's electricity use in 2024 exceeded that of Alphabet's Google and Samsung Electronics as well as oil giant TotalEnergies, while the power use of Air Liquide and Air Products was comparable to that of Shell and Microsoft. Yet unlike fossil fuel and tech companies, these industrial gas companies are far from household names because their customers are the world's largest chemicals, steel and oil companies rather than average consumers.

The industry relies on air-separation units, which use giant compressors to turn air into liquid and then distill it into its many components. These machines are responsible for much of the industry's electricity demand, and their use alone is responsible for 2% of carbon dioxide emissions in China and the US, the world's two largest polluters.

EU

EU Carmakers 'Days Away' From Halting Work as Chip War With China Escalates (theguardian.com) 116

Carmakers in the EU are "days away" from closing production lines, the industry has warned, as a crisis over computer chip supplies from China escalates. From a report: The European Automobile Manufacturers' Association (ACEA) issued an urgent warning on Wednesday saying its members, which include BMW, Fiat, Peugeot and Volkswagen, were now working on "reserve stocks but supplies are dwindling."

"Assembly line stoppages might only be days away. We urge all involved to redouble their efforts to find a diplomatic way out of this critical situation," said its director general, Sigrid de Vries. Another ACEA member, Mercedes, is now searching globally for alternative sources of the crucial semiconductors, according to its chief executive, Ola Kallenius. The chip shortage is also causing problems in Japan, where Nissan's chief performance officer, Guillaume Cartier, told reporters at a car show in Tokyo that the company was only "OK to the first week of November" in terms of supply.

United States

US Agencies Back Banning Top-Selling Home Routers on Security Grounds (msn.com) 89

More than a half dozen federal departments and agencies have backed a proposal to ban future sales of the most popular home routers in the United States on the grounds that the vendor's ties to mainland China make them a national security risk, Washington Post reported Thursday, citing people briefed on the matter. From the report: The proposal, which arose from a months-long risk assessment, calls for blocking sales of networking devices from TP-Link Systems of Irvine, California, which was spun off from a China-based company, TP-Link Technologies, but owns some of that company's former assets in China.

The ban was proposed by the Commerce Department and supported this summer by an interagency process that includes the Departments of Homeland Security, Justice and Defense, the people said. "TP-Link vigorously disputes any allegation that its products present national security risks to the United States," Ricca Silverio, a spokeswoman for TP-Link Systems, said in a statement. "TP-Link is a U.S. company committed to supplying high-quality and secure products to the U.S. market and beyond."

If imposed, the ban would be among the largest in consumer history and a possible sign that the East-West divide over tech independence is still deepening amid reports of accelerated Chinese government-supported hacking. Only the legislated ban of Chinese-owned TikTok, which President Donald Trump has averted with executive orders and a pending sale, would impact more U.S. consumers.

China

New China Law Fines Influencers If They Discuss 'Serious' Topics Without a Degree (iol.co.za) 74

schwit1 shares a report from IOL: China has enacted a new law regulating social media influencers, requiring them to hold verified professional qualifications before posting content on sensitive topics such as medicine, law, education, and finance, IOL reported. The new law went into effect on Saturday. The regulation was introduced by the Cyberspace Administration of China (CAC) as part of its broader effort to curb misinformation online.

Under the new rules, influencers must prove their expertise through recognized degrees, certifications, or licenses before discussing regulated subjects. Major platforms such as Douyin (China's TikTok), Bilibili, and Weibo are now responsible for verifying influencer credentials and ensuring that content includes clear citations, disclaimers, and transparency about sources.
A separate report notes that if influencers are caught talking about the "serious" topics, they will face a fine of up to 100,000 yuan ($14,000).
United States

US Needs 'Finesse' to Stay Ahead of China, Nvidia Boss Says (bloomberg.com) 31

Nvidia chief executive Jensen Huang said that maintaining the US edge in AI will require a steady approach that ensures China remains hooked on American technology. From a report: The chipmaker is in an "awkward place" as President Donald Trump prepares to meet with his Chinese counterpart Xi Jinping later this week, Huang told reporters Tuesday at a company conference in Washington. The Nvidia chief praised Trump's commitment to winning but urged careful engagement with China because of the country's massive software developer base and its growing technology capabilities.

During the meeting, Trump and Xi are expected to finalize an agreement to ease trade tensions between the world's two largest economies. When it comes to those negotiations, Huang said he has "no idea" if GPUs -- the chips central to artificial intelligence capabilities -- will be a topic between Trump and Xi.

Huang was careful to leave the negotiating to Trump but encouraged US leadership to think longer term on its overall AI strategy. "A policy that causes America to lose half of the world's developers is not beneficial long-term," Huang said, warning that it was still possible for the US to cede the AI race to China. Keeping US technology in front requires finesse," he said. "It requires balance. It requires long-term thinking."

China

China Bars Influencers From Discussing Professional Topics Without Relevant Degrees (iol.co.za) 196

schwit1 writes: China has enacted a new law regulating social media influencers, requiring them to hold verified professional qualifications before posting content on sensitive topics such as medicine, law, education, and finance, IOL reported. The new law went into effect on Saturday.

The regulation was introduced by the Cyberspace Administration of China (CAC) as part of its broader effort to curb misinformation online. Under the new rules, influencers must prove their expertise through recognized degrees, certifications, or licenses before discussing regulated subjects. Major platforms such as Douyin (China's TikTok), Bilibili, and Weibo are now responsible for verifying influencer credentials and ensuring that content includes clear citations, disclaimers, and transparency about sources.

Audiences expect influencers to be both creative and credible. Yet when they blur the line between opinion and expertise, the impact can be severe. A single misleading financial tip could wipe out someone's savings. A viral health trend could cause real harm. That's why many believe it's time for creators to acknowledge the weight of their influence. However, China's new law raises deeper questions: Who defines "expertise"? What happens to independent creators who challenge official narratives but lack formal credentials? And how far can regulation go before it suppresses free thought?

AI

Nvidia Becomes World's First $5 Trillion Company 31

Nvidia became the world's first $5 trillion company on Wednesday after its stock climbed 5% in early Wall Street trading to push its market capitalization to $5.13 trillion. The Silicon Valley chipmaker reached the milestone three months after hitting $4 trillion and three years after it was valued at roughly $400 billion before the debut of ChatGPT.

Nvidia chief executive Jensen Huang said Tuesday that Nvidia had secured half a trillion dollars in orders for its AI chips over the next five quarters. The stock had already gained 5% on Tuesday and added more than $200 billion to its market value. President Donald Trump said Wednesday he planned to discuss Nvidia's Blackwell chip with China's President Xi Jinping when the two leaders meet later this week. Nvidia's latest generation of graphics processing units is not currently available in China because of US export controls. The company's shares have risen more than 85% in the past six months.
Biotech

China Pushes Boundaries With Animal Testing to Win Global Biotech Race (bloomberg.com) 36

China is accelerating its biotech ambitions by pushing the limits of animal testing and gene editing (source paywalled; alternative source) while Western countries tighten ethical restrictions. "Editing the genes of large animals such as pigs, monkeys and dogs faces scant regulation in China," reports Bloomberg. "Meanwhile, regulators in the US and Europe demand layers of ethical reviews, rendering similar research involving large animals almost impossible." From the report: Backing the work of China's scientists is not only permissiveness but state money. In 2023 alone, the Chinese government funneled an estimated $3 billion into biotech. Its sales of cell and gene therapies are projected to reach $2 billion by 2033 from $300 million last year. On the Chinese researchers' side are government-supported breeding and research centers for gene-edited animals and a public largely in approval of pushing the boundaries of animal testing.

The country should become "a global scientific and technology power," Xi said, declaring biotechnology and gene editing a strategic priority. For decades, the country's pharmaceutical companies specialized in generics, reproducing drugs already pioneered elsewhere. Delving head first into gene editing research may be key to China's plan to develop innovative drugs as well as reduce its dependence on foreign pharmaceutical companies.

The result is a country that now dominates headlines with stories of large, genetically modified animals being produced for science -- and the catalog is startling. Its scientists have created monkeys with schizophrenia, autism and sleep disorders. They were the first to clone primates. They've engineered dogs with metabolic and neurological diseases, and even cloned a gene-edited beagle with a blood-clotting disorder.

AI

China's DeepSeek and Qwen AI Beat US Rivals In Crypto Trading Contest (yahoo.com) 31

hackingbear shares a report from Crypto News: Two Chinese artificial intelligence (AI) models, DeepSeek V3.1 and Alibaba's Qwen3-Max, have taken a commanding lead over their US counterparts in a live real-world real-money cryptocurrency trading competition, posting triple-digit gains in less than two weeks. According to Alpha Arena, a real-market trading challenge launched by US research firm Nof1, DeepSeek's Chat V3.1 turned an initial $10,000 into $22,900 by Monday, a 126% increase since trading began on October 18, while Qwen 3 Max followed closely with a 108% return.

In stark contrast, US models lagged far behind. OpenAI's GPT-5 posted the worst performance, losing nearly 60% of its portfolio, while Google DeepMind's Gemini 2.5 Pro showed a similar 57% decline. xAI's Grok 4 and Anthropic's Claude 4.5 Sonnet fared slightly better, returning 14% and 23% respectively. "Our goal with Alpha Arena is to make benchmarks more like the real world -- and markets are perfect for this," Nof1 said on its website.

Technology

China Dives in on the World's First Wind-Powered Undersea Data Center (wired.com) 33

China has completed the first phase of what it claims is the world's first underwater data center in Shanghai's Lingang Special Area. The facility cost roughly 1.6 billion yuan ($226 million) and operates on twenty-four megawatts of power drawn entirely from wind energy.

Seawater acts as a natural cooling system for the submerged servers. Traditional land-based data centers devote up to 50% of their energy consumption to air conditioning. The underwater design reduces cooling energy demand to less than 10%. The first phase is designed to achieve a power usage effectiveness rating of no more than 1.15. More than 95% of the facility's electricity comes from offshore wind turbines in the East China Sea. The project reduces land usage by more than 90% and eliminates the need for fresh water. The main contractors signed an agreement to launch another offshore wind-powered underwater data center with a capacity of 500 megawatts.
Crime

North Korea Has Stolen Billions in Cryptocurrency and Tech Firm Salaries, Report Says (apnews.com) 21

The Associated Press reports that "North Korean hackers have pilfered billions of dollars" by breaking into cryptocurrency exchanges and by creating fake identities to get remote tech jobs at foreign companies — all orchestrated by the North Korean government to finance R&D on nuclear arms.

That's according to a new the 138-page report by a group watching North Korea's compliance with U.N. sanctions (including officials from the U.S., Australia, Canada, France, Germany, Italy, Japan, the Netherlands, New Zealand, South Korea and the United Kingdom). From the Associated Press: North Korea also has used cryptocurrency to launder money and make military purchases to evade international sanctions tied to its nuclear program, the report said. It detailed how hackers working for North Korea have targeted foreign businesses and organizations with malware designed to disrupt networks and steal sensitive data...

Unlike China, Russia and Iran, North Korea has focused much of its cyber capabilities to fund its government, using cyberattacks and fake workers to steal and defraud companies and organizations elsewhere in the world... Earlier this year, hackers linked to North Korea carried out one of the largest crypto heists ever, stealing $1.5 billion worth of ethereum from Bybit. The FBI later linked the theft to a group of hackers working for the North Korean intelligence service.

Federal authorities also have alleged that thousands of IT workers employed by U.S. companies were actually North Koreans using assumed identities to land remote work. The workers gained access to internal systems and funneled their salaries back to North Korea's government. In some cases, the workers held several remote jobs at the same time.

IT

Some Startups Are Demanding 12-Hour Days, Six Days a Week from Workers (msn.com) 151

The Washington Post reports on 996, "a term popularized in China that refers to a rigid work schedule in which people work from 9 a.m. to 9 p.m., six days a week..." As the artificial intelligence race heats up, many start-ups in Silicon Valley and New York are promoting hardcore culture as a way of life, pushing the limits of work hours, demanding that workers move fast to be first in the market. Some are even promoting 996 as a virtue in the hiring process and keeping "grind scores" of companies... Whoever builds first in AI will capture the market, and the window of opportunity is two to three years, "so you better run faster than everyone else," said Inaki Berenguer, managing partner of venture-capital firm LifeX Ventures.

At San Francisco-based AI start-up Sonatic, the grind culture also allows for meal, gym and pickleball time, said Kinjal Nandy, its CEO. Nandy recently posted a job opening on X that requires in-person work seven days a week. He said working 10-hour days sounds like a lot but the company also offers its first hires perks such as free housing in a hacker house, food delivery credits and a free subscription to the dating service Raya... Mercor, a San Francisco-based start-up that uses AI to match people to jobs, recently posted an opening for a customer success engineer, saying that candidates should have a willingness to work six days a week, and it's not negotiable. "We know this isn't for everyone, so we want to put it up top," the listing reads.

Being in-person rather than remote is a requirement at some start-ups. AI start-up StarSling had two engineering job descriptions that required six days a week of in-person work. In a job description for an engineer, Rilla, an AI company in New York, said candidates should not work at the company if they're not excited about working about 70 hours a week in person. One venture capitalist even started tracking "grind scores." Jared Sleeper, a partner at New York-based venture capital firm Avenir, recently ranked public software companies' "grind score" in a post on X, which went viral. Using data from Glassdoor, it ranks the percentage of employees who have a positive outlook for the company compared with their views on work-life balance.

"At Google's AI division, cofounder Sergey Brin views 60 hours per week as the 'sweet spot' for productivity," notes the Independent: Working more than 55 hours a week, compared with a standard 35-40-hour week, is linked to a 35 percent higher risk of stroke and a 17 percent higher risk of death from heart disease, according to the World Health Organization. Productivity also suffers. A British study shows that working beyond 60 hours a week can reduce overall output, slow cognitive performance, and impair tasks ranging from call handling to problem-solving.

Shorter workweeks, in contrast, appear to boost productivity. Microsoft Japan saw a roughly 40% increase in output after adopting a four-day work week. In a UK trial, 61 companies that tested a four-day schedule reported revenue gains, with 92 percent choosing to keep the policy, according to Bloomberg.

China

China's Zhuque-3 Reusable Rocket Passes Key Milestone (universetoday.com) 42

China's private space company LandSpace has completed a key static fire test of its Zhuque-3 (ZQ-3) reusable rocket -- a stainless-steel, methane-fueled launcher modeled after SpaceX's Starship. Universe Today reports: The latest milestone took place on Monday, Oct. 22nd at the Dongfeng commercial space innovation pilot zone (where the JSLC is located). It involved another static fire test, where the rocket was fully-fueled but remained fixed to the launch pad while the engines were fired. This kind of testing is a crucial prelaunch trial (what NASA refers to as a "wet dress rehearsal"), and places the company and China another step closer to making an inaugural flight test, which is expected to happen by the fourth quarter of 2025.

In traditional Chinese, Zhuque is the name of the Vermillion Bird that represents fire, the south, and summer, and is one of the four Symbols of the Chinese constellations. Like the Starship, the Zhuque-3 is composed of stainless steel and relies on a combination of liquid methane (LCH4) and liquid oxygen (LOX) propellant. The rocket will be powered by nine Tianque-12A (TQ-12A) engines and will measure 65.9 m (216 ft) tall and weigh 550,000 kg (1,210,000 lb). It's payload capacity will be significantly less than the Starship: 11,800 kg (26,000 lbs) in its expendable mode, and 8,000 kg (18,000 lbs) for the recoverable version. This is closer in payload capacity to the Falcon 9, which is capable of delivering 22,800 kg (50,265 lbs) to Low Earth Orbit (LEO).

In time, the company hopes to transition to the larger Zhuque-3E, which will be 76.2 m (250 ft) tall and powered by nine TQ-12B engines, and will be capable of delivering to 21,000 kg (46,000 lb) in its expandable mode and 18,300 kg (40,300 lb) recoverable. The long term goal is to create a reusable system that can rival the Falcon rocket family, bringing the country closer to its goal of achieving parity with NASA.

The Internet

Browser Promising Privacy Protection Contains Malware-Like Features, Routes Traffic Through China (arstechnica.com) 16

A web browser linked to Chinese online gambling websites and downloaded millions of times routes all internet traffic through servers in China and covertly installs programs that run in the background, according to findings published by network security company Infoblox. The researchers said the Universe Browser, which advertises itself as offering privacy protection, includes features similar to malware such as key logging and surreptitious connections.

Infoblox collaborated with the United Nations Office on Drugs and Crime on the research. The investigators found links between the browser and Southeast Asia's cybercrime ecosystem, which has connections to money laundering, illegal online gambling, human trafficking and scam operations using forced labor. The browser is directly linked to BBIN, a major online gambling company that has existed since 1999. Infoblox researchers examined the Windows version of the browser and found that it checks users' locations and languages when launched, installs two browser extensions, and disables security features including sandboxing.
China

China's New Five-Year Plan Sharpens Industry, Tech Focus (reuters.com) 30

An anonymous reader shares a report: China's Communist Party elite vowed on Thursday to build a modern industrial system and make more efforts to achieve technological self-reliance, moves it sees as key to bolstering its position in its intensifying rivalry with the United States. As expected, the Party's Central Committee also promised more efforts to expand domestic demand and improve people's livelihoods - long-standing goals that in recent years have been little more than an afterthought as China prioritised manufacturing and investment - without giving many details.

[...] The full five-year plan will only be released at a parliamentary meeting in March, but the post-plenum outline from state news agency Xinhua hinted at policy continuity, which concerns economists who have been calling for a shift towards aâgrowth model that relies more on household demand. Building "a modern industrial system with advanced manufacturing as the backbone" and accelerating "high-level scientific and technological self-reliance" were listed ahead of the development of "a strong domestic market," the communique showed.

Google

Google's Quantum Computer Makes a Big Technical Leap (nytimes.com) 30

Google announced Wednesday that its quantum computer achieved the first verifiable quantum advantage, running a new algorithm 13,000 times faster than a top supercomputer. The algorithm, called Quantum Echoes, was published in the journal Nature. The results can be replicated on another quantum computer of similar quality, something Google had not demonstrated before. The quantum computer uses a chip called Willow, which was announced in December 2024. Hartmut Neven, head of Google's Quantum AI research lab, called the work a demonstration of the first algorithm with verifiable quantum advantage and a milestone on the software track.

Michel H. Devoret, who won this year's Nobel Prize in Physics and joined Google in 2023, said future quantum computers will run calculations impossible with classical algorithms. Google stopped short of claiming the work would have practical uses on its own. Instead, the company said Quantum Echoes demonstrated a technique that could be applied to other algorithms in drug discovery and materials science.

A second paper published Wednesday on arXiv showed how the method could be applied to nuclear magnetic resonance. The experiment involved a relatively small quantum system that fell short of full practical quantum advantage because it was not able to work faster than a traditional computer. Google exhaustively red-teamed the research, putting some researchers to work trying to disprove its own results.

Prineha Narang, a professor at UCLA, called the advance meaningful. The quantum computer tested two molecules, one with 15 atoms and another with 28 atoms. Results on the quantum computer matched traditional NMR and revealed information not usually available from NMR. Google's research competes against Microsoft, IBM, universities and efforts in China. The Chinese government has committed more than $15.2 billion to quantum research. Previous claims of quantum advantage have been met with skepticism.
NASA

NASA Opens SpaceX's Moon Lander Contract To Rivals Over Starship Delays (reuters.com) 61

NASA has reopened SpaceX's $4.4 billion moon lander contract to new bidders like Blue Origin and Lockheed Martin after delays in Starship's development threatened the 2027 Artemis 3 mission. Reuters reports: The move paves the way for rivals such as Jeff Bezos' Blue Origin to snatch a high-profile mission to land the first astronauts on the moon in half a century. "I'm in the process of opening that contract up. I think we'll see companies like Blue get involved, and maybe others," the U.S. space agency's acting chief Sean Duffy, who also serves as U.S. Transportation Secretary, told Fox News' "Fox & Friends" program.

Duffy's comments follow months of mounting pressure within NASA to speed up its Artemis lunar program and push SpaceX to make greater progress on its Starship lunar lander, while China progresses toward its own goal of sending humans to the moon by 2030. It represents a major shift in NASA's lunar strategy, starting a new competitive juncture in the program for a crewed moon lander just two years before the scheduled landing date. Blue Origin is widely expected to compete for the mission, while Lockheed Martin has indicated it would convene an industry team to heed NASA's call.

Starship, picked by NASA in 2021 under a contract now worth $4.4 billion, faces a 2027 moon landing deadline that agency advisers estimate could slip years behind schedule, citing competing priorities. Musk sees Starship as crucial to launching larger batches of Starlink satellites to space and eventually ferrying humans to Mars, among other missions. "They do remarkable things, but they're behind schedule," Duffy said of SpaceX's lunar lander work, adding President Donald Trump wants to see the mission take place before his White House term ends in January 2029.

United Kingdom

London Became a Global Hub for Phone Theft. Now We Know Why. (nytimes.com) 133

London police finally understand why 80,000 phones disappeared from the city's streets last year. The answer involves budget cuts [non-paywalled source] that hollowed out British policing in the 2010s, the arrival of electric bikes that made theft easy, and a lucrative black market in China where stolen British phones retain full functionality. The Metropolitan Police discovered an industrial-scale operation in December when officers traced a woman's iPhone to a Heathrow warehouse on Christmas Eve. Boxes labeled as batteries and bound for Hong Kong contained almost 1,000 stolen iPhones. The police arrested two men in their thirties in September as suspected ringleaders of a group that sent up to 40,000 stolen phones to China.

The epidemic took root after Conservative-led austerity measures reduced police numbers and budgets. In 2017 the Metropolitan Police announced it would stop investigating low-level crimes to focus resources on serious violence and sexual offenses. Thieves on rented electric bikes began mounting sidewalks to snatch phones at high speed while wearing balaclavas and hoods. Police data shows only 495 people were charged out of 106,000 phones reported stolen between March 2024 and February 2025. Thieves earn up to $401 per device. The phones sell for up to $5,000 in China because Chinese network providers do not subscribe to the international blacklist for stolen devices.
Cloud

Alibaba Cloud Says It Cut Nvidia AI GPU Use By 82% With New Pooling System (tomshardware.com) 27

Alibaba Cloud claims its new Aegaeon GPU pooling system cuts Nvidia GPU use by 82%, letting 213 H20 accelerators handle workloads that previously required 1,192. The advancements have been detailed in a paper (PDF) at the 2025 ACM Symposium on Operating Systems (SOSP) in Seoul. Tom's Hardware reports: Unlike training-time breakthroughs that chase model quality or speed, Aegaeon is an inference-time scheduler designed to maximize GPU utilization across many models with bursty or unpredictable demand. Instead of pinning one accelerator to one model, Aegaeon virtualizes GPU access at the token level, allowing it to schedule tiny slices of work across a shared pool. This means one H20 could serve several different models simultaneously, with system-wide "goodput" -- a measure of effective output -- rising by as much as nine times compared to older serverless systems.

The system was tested in production over several months, according to the paper, which lists authors from both Peking University and Alibaba's infrastructure division, including CTO Jingren Zhou. During that window, the number of GPUs needed to support dozens of different LLMs -- ranging in size up to 72 billion parameters -- fell from 1,192 to just 213. While the paper does not break down which models contributed most to the savings, reporting by the South China Morning Post says the tests were conducted using Nvidia's H20, one of the few accelerators still legally available to Chinese buyers under current U.S. export controls.

Slashdot Top Deals