Many trains in the U.S. are vulnerable to a hack that can remotely lock a train's brakes, according to the U.S. Cybersecurity and Infrastructure Security Agency (CISA) and the researcher who discovered the vulnerability. From a report:The railroad industry has known about the vulnerability for more than a decade but only recently began to fix it. Independent researcher Neil Smith first discovered the vulnerability, which can be exploited over radio frequencies, in 2012. "All of the knowledge to generate the exploit already exists on the internet. AI could even build it for you," Smith told 404 Media. "The physical aspect really only means that you could not exploit this over the internet from another country, you would need to be some physical distance from the train [so] that your signal is still received."

Read more of this story at Slashdot.

Perplexity CEO Aravind Srinivas warned young entrepreneurs that tech giants will "copy anything that's good" during a talk at Y Combinator's AI Startup School, telling founders they must "live with that fear." Srinivas said that companies raising tens of billions need to justify capital expenditures and search for new revenue streams. Perplexity pioneered web-crawling chatbots when it launched its answer engine in December 2022, but Google's Bard added internet-crawling three months later, followed by ChatGPT in May 2023 and Anthropic's Claude in March 2025. The competition has extended to browsers, with Perplexity launching its Comet browser on July 9 and Reuters reporting that OpenAI is developing a web browser to challenge Google Chrome. Perplexity's communications head Jesse Dwyer said larger companies will "drown your voice."

Read more of this story at Slashdot.

NIST: There's a new record holder for the most accurate clock in the world. Researchers at the National Institute of Standards and Technology (NIST) have improved their atomic clock based on a trapped aluminum ion. Part of the latest wave of optical atomic clocks, it can perform timekeeping with 19 decimal places of accuracy. Optical clocks are typically evaluated on two levels -- accuracy (how close a clock comes to measuring the ideal "true" time, also known as systematic uncertainty) and stability (how efficiently a clock can measure time, related to statistical uncertainty). This new record in accuracy comes out of 20 years of continuous improvement of the aluminum ion clock. Beyond its world-best accuracy, 41% greater than the previous record, this new clock is also 2.6 times more stable than any other ion clock. Reaching these levels has meant carefully improving every aspect of the clock, from the laser to the trap and the vacuum chamber. The team published its results in Physical Review Letters. "It's exciting to work on the most accurate clock ever," said Mason Marshall, NIST researcher and first author on the paper. "At NIST we get to carry out these long-term plans in precision measurement that can push the field of physics and our understanding of the world around us."

Read more of this story at Slashdot.

A survey of 500 IT asset managers in organizations that use Oracle Java has found that 73% have been audited in the last three years. From a report: At the same time, nearly eight out of 10 Oracle Java users said they had migrated, or planned to shift, to open source Java to try to avoid the risk and high costs of the dominant vendor's development and runtime environments. Oracle introduced a paid subscription for Java in September 2018, and in January 2023, it decided to switch its pricing model to per employee rather than per user, creating a steep price hike for many users. In July 2023, Gartner recorded users experiencing price increases of between two and five times when they switched to the new licensing model. Two years later, the survey conducted by market research firm Dimensional Research showed only 14% of Oracle Java users intended to stick with the vendor's subscription model.

Read more of this story at Slashdot.

Nearly half of young Americans feel unprepared for future jobs as AI reshapes the workforce faster than career guidance can adapt, according to a new study from the Schultz Family Foundation and HarrisX. The survey of thousands of workers aged 16-24, along with parents, counselors and employers, revealed differences between generations about job availability and requirements. While 71% of employers say sufficient opportunities exist, only 43% of young people agree. Parents rely on outdated personal experiences when advising children, with 79% drawing from their own career paths despite 66% believing their children should pursue different directions. Employers require at least one year of experience for 77% of entry-level positions while offering internships for just 38% of roles.

Read more of this story at Slashdot.

An anonymous reader shares a report: Hugging Face, a company with a multi-billion dollar valuation and one of the most commonly used platforms for sharing AI tools and resources, is hosting over 5,000 AI image generation models that are designed to recreate the likeness of real people. These models were all previously hosted on Civitai, an AI model sharing platform 404 Media reporting has shown was used for creating nonconsensual pornography, until Civitai banned them due to pressure from payment processors. Users downloaded the models from Civitai and reuploaded them to Hugging Face as part of a concerted community effort to archive the models after Civitai announced in May it will ban them. In that announcement, Civitai said it will give the people who originally uploaded them "a short period of time" before they were removed. Civitai users began organizing an archiving effort on Discord earlier in May after Civitai indicated it had to make content policy changes due to pressure from payment processors, and the effort kicked into high gear when Civitai announced the new "real people" model policy.

Read more of this story at Slashdot.

Candy Crush-maker King is cutting approximately 200 employees, with many positions filled by AI tools the departing workers helped develop, according to multiple sources who spoke anonymously to industry publication MobileGamer.biz. The layoffs heavily target level designers, user research staff, and UX and narrative writers across King's London, Barcelona, Stockholm, and Berlin studios. The London-based Farm Heroes Saga team faces cuts of roughly 50 people, including key leadership positions. "Most of level design has been wiped, which is crazy since they've spent months building tools to craft levels quicker," one staffer said. "Now those AI tools are basically replacing the teams."

Read more of this story at Slashdot.

Microsoft employs engineers in China to help maintain Defense Department computer systems, with U.S. citizens serving as "digital escorts" to oversee the foreign workers, according to a ProPublica investigation. The escorts often lack advanced technical expertise to police engineers with far more sophisticated skills, and some are former military personnel paid barely above minimum wage. "We're trusting that what they're doing isn't malicious, but we really can't tell," one current escort told the publication. The arrangement, critical to Microsoft winning federal cloud computing contracts a decade ago, handles sensitive but unclassified government data including materials that directly support military operations. Former CIA and NSA executive Harry Coker called the system a natural opportunity for spies, saying "If I were an operative, I would look at that as an avenue for extremely valuable access."

Read more of this story at Slashdot.

An anonymous reader quotes a report from Bloomberg: CoreWeave is expanding a data center that is projected to double the electricity needs of a city near Dallas, another example of the strains that artificial intelligence workloads are placing on the US power supply. Local officials have grappled with how to handle the increased stress on the electricity grid from the project, according to a late 2024 presentation and emails seen by Bloomberg. The site is being developed by Core Scientific and will be used by OpenAI in Denton, Texas. Last week, CoreWeave announced it would acquire Core Scientific for about $9 billion, in part, to gain direct control of its data centers aimed at supplying AI work. Denton, about 50 miles northwest of Dallas, has almost doubled its population in the last 25 years to about 166,000 residents. To meet the spike in AI-related power demand, the city is passing on any extra costs to the data center operator and constructing additional grid infrastructure, Antonio Puente, general manager of local utility Denton Municipal Electric, said in an interview. "To serve the entire load from Core Scientific, we do have some transmission challenges," Puente said. "We will have to make some additional transmission investments." [...] Like some other large AI data center projects, the site in Denton was focused on cryptocurrency mining before pivoting to AI workloads in December. This transition means unrelenting power consumption -- the site will no longer curtail operations when power prices are high -- which will increase grid strain. "Now you're talking about a facility that has to have energy 24 hours a day, 365 days a year," Puente said. That challenge will be mitigated by the addition of backup generators and batteries, he added. Unlike many large projects, the Denton data center didn't receive local tax exemptions. Officials expect more than $600 million in property and sales tax from the data center expansion, more than double the costs it plans to incur, according to an analysis document seen by Bloomberg. It also anticipates that 135 new jobs will be created, according to the document. The Denton site, which is already being rented by CoreWeave, is Core Scientific's largest planned project at about 390 megawatts of power. It's "utilizing the majority of extra system capacity" in the city, wrote a utility executive in a January email seen by Bloomberg. Any additional large power users will exacerbate overloads on the grid, the executive added. "When fully built out, it will host one of the largest GPU clusters in North America," Core Scientific Chief Executive Officer Adam Sullivan said of the site during a May call. "Denton is a flagship facility." The report notes that Texas could face electricity shortages as soon as 2026 due to surging power demand from data centers, oil and gas operations, and crypto mining.

Read more of this story at Slashdot.

A team of Japanese researchers has set a new world record for internet speed, transmitting data at 125,000 gigabytes per second over 1,120 miles using a new type of 19-core optical fiber. "That's about 4 million times the average internet speed in the U.S. and would allow you to download the entire Internet Archive in less than four minutes," notes Live Science. It's also "more than twice the previous world record of 50,250 Gbps, previously set by a different team of scientists in 2024." From the report: To achieve this new speed -- which has not been independently verified -- the team developed a new form of optical fiber to send information at groundbreaking speeds over roughly the distance between New York and Florida. Details about this achievement were presented April 3 at the 48th Optical Fiber Communication Conference in San Francisco, according to a statement from Japan's National Institute of Information and Communications Technology. The new type of optical fiber is equivalent to 19 standard optical fibers in its data transmission capacity. The new optical fiber is better suited to long-haul transmission than existing cables because the centers of all 19 fibers interact with light in the same way, so they encounter less light fluctuation, which results in less data loss. The new cable squeezes 19 separate fibers into a diameter of five-thousandths of an inch (0.127 millimeters), which is the same thickness as most existing single-fiber cables already in use. This effort means the new cable can transmit more data using existing infrastructure. [...] For this demonstration, the data ran through a transmission system 21 times, finally reaching a data receiver after traveling the equivalent of 1,120 miles.

Read more of this story at Slashdot.

The LIGO-Virgo-KAGRA Collaboration has detected the most massive black hole merger to date, forming a final black hole around 225 times the Sun's mass. Caltech reports: Before now, the most massive black hole merger -- produced by an event that took place in 2021 called GW190521 -- had a total mass of 140 times that of the Sun. In the more recent GW231123 event, the 225-solar-mass black hole was created by the coalescence of black holes each approximately 100 and 140 times the mass of the Sun. In addition to their high masses, the black holes are also rapidly spinning. "The black holes appear to be spinning very rapidly -- near the limit allowed by Einstein's theory of general relativity," explains Charlie Hoy of the University of Portsmouth and a member of the LVK. "That makes the signal difficult to model and interpret. It's an excellent case study for pushing forward the development of our theoretical tools." Researchers are continuing to refine their analysis and improve the models used to interpret such extreme events. "It will take years for the community to fully unravel this intricate signal pattern and all its implications," says Gregorio Carullo of the University of Birmingham and a member of the LVK. "Despite the most likely explanation remaining a black hole merger, more complex scenarios could be the key to deciphering its unexpected features. Exciting times ahead!"

Read more of this story at Slashdot.

An anonymous reader quotes a report from MIT Technology Review: California's statewide power grid operator is poised to become the first in North America to deploy artificial intelligence to manage outages, MIT Technology Review has learned. "We wanted to modernize our grid operations. This fits in perfectly with that," says Gopakumar Gopinathan, a senior advisor on power system technologies at the California Independent System Operator -- known as the CAISO and pronounced KAI-so. "AI is already transforming different industries. But we haven't seen many examples of it being used in our industry." At the DTECH Midwest utility industry summit in Minneapolis on July 15, CAISO is set to announce a deal to run a pilot program using new AI software called Genie, from the energy-services giant OATI. The software uses generative AI to analyze and carry out real-time analyses for grid operators and comes with the potential to autonomously make decisions about key functions on the grid, a switch that might resemble going from uniformed traffic officers to sensor-equipped stoplights. But while CAISO may deliver electrons to cutting-edge Silicon Valley companies and laboratories, the actual task of managing the state's electrical system is surprisingly analog. Today, CAISO engineers scan outage reports for keywords about maintenance that's planned or in the works, read through the notes, and then load each item into the grid software system to run calculations on how a downed line or transformer might affect power supply. "Even if it takes you less than a minute to scan one on average, when you amplify that over 200 or 300 outages, it adds up," says Abhimanyu Thakur, OATI's vice president of platforms, visualization, and analytics. "Then different departments are doing it for their own respective keywords. Now we consolidate all of that into a single dictionary of keywords and AI can do this scan and generate a report proactively." If CAISO finds that Genie produces reliable, more efficient data analyses for managing outages, Gopinathan says, the operator may consider automating more functions on the grid. "After a few rounds of testing, I think we'll have an idea about what is the right time to call it successful or not," he says.

Read more of this story at Slashdot.

Saudi Arabia has asked consultants to reassess the feasibility of The Line, its ambitious 170km linear city project and centerpiece of the Neom initiative, as rising costs and falling oil prices force the kingdom to scale back its megaprojects. Middle East Eye reports: In April, The Financial Times reported that the CEO of Neom had launched a "comprehensive review" of the kingdom's megaproject. Neom, along with luxury Red Sea hotels and a ski resort, is the flagship project of Crown Prince Mohammed bin Salman's Vision 2030 plan to transform the kingdom's economy and reduce its dependence on oil revenue. Bloomberg reported in 2024 that Saudi Arabia was cutting back plans for The Line. Instead of 1.5 million people living there by 2030, Saudi officials were said to anticipate fewer than 300,000 residents. Meanwhile, only 2.4km of the city is expected to be completed by 2030. In April, Goldman Sachs painted a bleak picture for Saudi Arabia's projects in a note to clients, projecting "pretty significant" budget deficits and more scaling back of megaprojects. Neom has already faced internal challenges. Nadhmi al-Nasr, who managed Neom's construction from 2018 to 2024, departed from his post in November. Nasr earned a chilling reputation managing Neom. He bragged that he put everyone to work "like a slave," adding, "When they drop down dead, I celebrate. That's how I do my projects." Two other foreign executives also left Neom at the end of 2024, according to The Wall Street Journal.

Read more of this story at Slashdot.

Microsoft is testing a new adaptive energy saver mode in Windows 11 that automatically turns energy saver on or off based on system workload instead of battery percentage, aiming to extend laptop battery life without dimming screen brightness. The feature is currently available to Windows Insider testers and expected to roll out later this year. The Verge reports: The energy saver mode in Windows 11 typically dims a display brightness by 30 percent, disables transparency effects, and stop apps running in the background. Non-critical Windows update downloads are also paused, and certain apps like OneDrive, OneNote, and Phone Link may not sync fully while energy saver is enabled. This new adaptive energy saver mode, which will only be available on devices with a battery, will automatically enable or disable without affecting screen brightness. That will make it less noticeable on devices like laptops, tablets, and handhelds. "Adaptive energy saver is an opt-in feature that automatically enables and disables energy saver, without changing screen brightness, based on the power state of the device and the current system load," explains Microsoft's Windows Insider team.

Read more of this story at Slashdot.

The U.S. Department of Defense has awarded contracts worth up to $200 million each to OpenAI, Google, Anthropic, and xAI to scale adoption of advanced AI. "The contracts will enable the DoD to develop agentic AI workflows and use them to address critical national security challenges," reports Reuters, citing the department's Chief Digital and Artificial Intelligence Office. From the report: Separately on Monday, xAI announced a suite of its products called "Grok for Government", making its advanced AI models -- including its latest flagship Grok 4 -- available to federal, local, state and national security customers. The Pentagon announced last month that OpenAI was awarded a $200 million contract, saying the ChatGPT maker would "develop prototype frontier AI capabilities to address critical national security challenges in both warfighting and enterprise domains." The contracts announced on Monday deepen the ties between companies leading the AI race and U.S. government operations, while addressing concerns around the need for competitive contracts for AI use in federal agencies. "The adoption of AI is transforming the (DoD's) ability to support our warfighters and maintain strategic advantage over our adversaries," Chief Digital and AI Officer Doug Matty said.

Read more of this story at Slashdot.

Aylo (ex Mindgeek), éditeur de trois parmi les plus importants carrefours d’audience au monde en matière de contenus pornographiques, a à nouveau mis en place le blocage de ses sites pour les internautes français.

Depuis la France, Youporn, Pornhub et Redtube sont remplacés par un nouveau message dans lequel l’éditeur s’insurge contre la façon dont la France a choisi de mettre en œuvre l’obligation de vérification de l’âge des internautes.

La mise en ligne de cet écran et le blocage associé font suite à la publication, mardi, d’une décision du Conseil d’État qui rétablit la possibilité de bloquer les sites pornographiques si ces derniers n’ont pas répondu à l’exigence de mise en place d’un dispositif de vérification d’âge.

« Les contestations juridiques, les décisions et les revirements en cours mettent clairement en évidence une chose : l’absence de direction, l’absence de solution globale et, pour dire les choses simplement, le dysfonctionnement législatif », affirme l’éditeur de Pornhub, qui une nouvelle fois cherche à prendre à partie les utilisateurs de ses services :

« Votre gouvernement propose que nous vérifiions votre âge à chaque fois que vous visitez notre site. La dure réalité : Cette mesure ne protège pas les mineurs, elle met en péril la vie privée de chacun et met les enfants en danger, car elle conduit le trafic vers des milliers de sites qui contournent délibérément les réglementations, ne vérifient pas l’âge des participants dans les contenus et encouragent activement les utilisateurs à contourner la loi ».

Aylo avait déjà bloqué l’accès à ses sites en France début juin, avant de faire marche arrière quand le tribunal administratif de Paris avait ordonné la suspension de l’obligation de contrôle.

« Seuls des adultes identifiés devraient être autorisés à déverrouiller l’accès à des contenus inappropriés pour l’âge. La technologie nécessaire à cette fin existe aujourd’hui. Nous sommes déterminés à faire partie de cette solution et souhaitons collaborer avec les gouvernements, la société civile et les partenaires technologiques pour parvenir à une solution efficace de vérification de l’âge basée sur les appareils », affirme encore l’éditeur.

Ce nouveau rebondissement intervient alors que la Commission européenne vient de publier ses lignes directrices en matière de protection des mineurs sur Internet, et d’annoncer le développement d’un module de vérification destiné aux éditeurs, qui sera notamment testé par la France.

Aylo (ex Mindgeek), éditeur de trois parmi les plus importants carrefours d’audience au monde en matière de contenus pornographiques, a à nouveau mis en place le blocage de ses sites pour les internautes français.

Depuis la France, Youporn, Pornhub et Redtube sont remplacés par un nouveau message dans lequel l’éditeur s’insurge contre la façon dont la France a choisi de mettre en œuvre l’obligation de vérification de l’âge des internautes.

La mise en ligne de cet écran et le blocage associé font suite à la publication, mardi, d’une décision du Conseil d’État qui rétablit la possibilité de bloquer les sites pornographiques si ces derniers n’ont pas répondu à l’exigence de mise en place d’un dispositif de vérification d’âge.

« Les contestations juridiques, les décisions et les revirements en cours mettent clairement en évidence une chose : l’absence de direction, l’absence de solution globale et, pour dire les choses simplement, le dysfonctionnement législatif », affirme l’éditeur de Pornhub, qui une nouvelle fois cherche à prendre à partie les utilisateurs de ses services :

« Votre gouvernement propose que nous vérifiions votre âge à chaque fois que vous visitez notre site. La dure réalité : Cette mesure ne protège pas les mineurs, elle met en péril la vie privée de chacun et met les enfants en danger, car elle conduit le trafic vers des milliers de sites qui contournent délibérément les réglementations, ne vérifient pas l’âge des participants dans les contenus et encouragent activement les utilisateurs à contourner la loi ».

Aylo avait déjà bloqué l’accès à ses sites en France début juin, avant de faire marche arrière quand le tribunal administratif de Paris avait ordonné la suspension de l’obligation de contrôle.

« Seuls des adultes identifiés devraient être autorisés à déverrouiller l’accès à des contenus inappropriés pour l’âge. La technologie nécessaire à cette fin existe aujourd’hui. Nous sommes déterminés à faire partie de cette solution et souhaitons collaborer avec les gouvernements, la société civile et les partenaires technologiques pour parvenir à une solution efficace de vérification de l’âge basée sur les appareils », affirme encore l’éditeur.

Ce nouveau rebondissement intervient alors que la Commission européenne vient de publier ses lignes directrices en matière de protection des mineurs sur Internet, et d’annoncer le développement d’un module de vérification destiné aux éditeurs, qui sera notamment testé par la France.

Titans sans Tartare

Dans une série de messages sur Threads, Mark Zuckerberg a annoncé « investir des centaines de milliards de dollars ». Meta veut accélérer l’agrandissement de son parc de datacenters : actuellement, l’entreprise utilise même des tentes pour étendre ses centres. Pourtant, leur installation a des conséquences pour le voisinage.

Sur son réseau social Threads, Mark Zuckerberg a annoncé, lundi 14 juillet, que son entreprise, Meta, va investir « des centaines de milliards de dollars » dans de nouvelles infrastructures de calcul pour l’intelligence artificielle. Cette annonce arrive après le débauchage par son entreprise de plusieurs spécialistes chez Apple, Anthropic, OpenAI ou Scale AI.

Meta a bien développé une famille de modèles de langage, Llama, qu’elle présente comme open source, même si bien des spécialistes critiquent cette appellation. Mais l’entreprise, qui avait parié sur le metaverse, s’est montrée moins à la pointe dans le développement de l’IA générative que ses concurrents comme OpenAI, Google ou Microsoft. C’était pourtant chez Facebook que l’un des pionniers du deep learning, Yann Lecun, avait posé ses pénates en 2013, donnant un boost au domaine.

Des « titans » demandant jusqu’à 5 GW de puissance

« Nous sommes en train de construire des clusters de plusieurs gigawatts », a expliqué de manière offensive Mark Zuckerberg sur Thread. « Nous avons baptisé le premier Prometheus et il sera mis en service en 2026 », a-t-il ajouté, « nous construisons également Hyperion, qui aura une puissance qui pourra aller jusqu’à 5 GW dans plusieurs années. Nous construisons également de nombreuses autres clusters de titans. L’un d’entre eux couvre une grande partie de l’empreinte de Manhattan ». Le CEO de Meta parle de datacenters « de titans » et les baptise de noms inspirés de ces divinités de la mythologie grecque, donnant une impression de gigantisme à ses centres de données.

Vendredi 11 juillet, SemiAnalysis expliquait que Meta mettait en place des parties de ses datacenters sous tentes pour accélérer leur expansion. L’entreprise a confirmé auprès de Business Insider, tout en précisant que ces tentes permettent seulement d’augmenter la capacité de datacenters existants. SemiAnalysis donnait aussi plus d’informations sur le projet Prometheus. « Lorsque le réseau électrique local n’a pu suivre, Meta s’est mis en mode Elon [Musk]. Avec l’aide de Williams [énergéticien étasunien, ndlr], ils construisent deux centrales au gaz naturel de 200 MW sur site », affirmait SemiAnalysis.

Comme nous l’expliquions récemment, ces gros datacenters ne cherchent plus la proximité des fibres, mais celle des grosses lignes électriques pour alimenter des serveurs toujours plus gourmands et denses. Jusqu’à, comme ici, imaginer des centrales dédiées.

• Pourquoi l’IA oblige les datacenters à se réinventer

Pour mettre en perspective cette annonce de Mark Zuckerberg, rappelons qu’Olivier de Nomazy, responsable analyste chez Data4, nous expliquait qu’ « il y avait à peu près 6,5 GW en 2023 de datacenters déployés ». L’annonce d’un datacenter de 5 GW par Meta implique donc un besoin d’énergie colossal par rapport à ce que l’on connait actuellement.

Des problèmes pour les riverains

En parallèle, lundi, le New York Times a publié un article racontant les problèmes d’une famille du Comté de Newton, dont la maison se situe juste à côté d’un des datacenters de Meta. Beverly et Jeff Morris habitent à 300 mètres de ce centre de données, et l’eau de leur maison vient directement de la nappe d’eau souterraine de leur terrain.

Le journal étasunien explique que quelques mois après le début de la construction du data center de Meta, en 2018, le lave-vaisselle, la machine à glaçons, le lave-linge et les toilettes des Morris ont tous cessé de fonctionner. « En l’espace d’un an, la pression de l’eau s’est réduite à un filet d’eau ».

Le média syndical américain More Perfect Union avait déjà publié, il y a trois mois, sur YouTube une vidéo sur les problèmes de la famille :

Titans sans Tartare

Dans une série de messages sur Threads, Mark Zuckerberg a annoncé « investir des centaines de milliards de dollars ». Meta veut accélérer l’agrandissement de son parc de datacenters : actuellement, l’entreprise utilise même des tentes pour étendre ses centres. Pourtant, leur installation a des conséquences pour le voisinage.

Sur son réseau social Threads, Mark Zuckerberg a annoncé, lundi 14 juillet, que son entreprise, Meta, va investir « des centaines de milliards de dollars » dans de nouvelles infrastructures de calcul pour l’intelligence artificielle. Cette annonce arrive après le débauchage par son entreprise de plusieurs spécialistes chez Apple, Anthropic, OpenAI ou Scale AI.

Meta a bien développé une famille de modèles de langage, Llama, qu’elle présente comme open source, même si bien des spécialistes critiquent cette appellation. Mais l’entreprise, qui avait parié sur le metaverse, s’est montrée moins à la pointe dans le développement de l’IA générative que ses concurrents comme OpenAI, Google ou Microsoft. C’était pourtant chez Facebook que l’un des pionniers du deep learning, Yann Lecun, avait posé ses pénates en 2013, donnant un boost au domaine.

Des « titans » demandant jusqu’à 5 GW de puissance

« Nous sommes en train de construire des clusters de plusieurs gigawatts », a expliqué de manière offensive Mark Zuckerberg sur Thread. « Nous avons baptisé le premier Prometheus et il sera mis en service en 2026 », a-t-il ajouté, « nous construisons également Hyperion, qui aura une puissance qui pourra aller jusqu’à 5 GW dans plusieurs années. Nous construisons également de nombreuses autres clusters de titans. L’un d’entre eux couvre une grande partie de l’empreinte de Manhattan ». Le CEO de Meta parle de datacenters « de titans » et les baptise de noms inspirés de ces divinités de la mythologie grecque, donnant une impression de gigantisme à ses centres de données.

Vendredi 11 juillet, SemiAnalysis expliquait que Meta mettait en place des parties de ses datacenters sous tentes pour accélérer leur expansion. L’entreprise a confirmé auprès de Business Insider, tout en précisant que ces tentes permettent seulement d’augmenter la capacité de datacenters existants. SemiAnalysis donnait aussi plus d’informations sur le projet Prometheus. « Lorsque le réseau électrique local n’a pu suivre, Meta s’est mis en mode Elon [Musk]. Avec l’aide de Williams [énergéticien étasunien, ndlr], ils construisent deux centrales au gaz naturel de 200 MW sur site », affirmait SemiAnalysis.

Comme nous l’expliquions récemment, ces gros datacenters ne cherchent plus la proximité des fibres, mais celle des grosses lignes électriques pour alimenter des serveurs toujours plus gourmands et denses. Jusqu’à, comme ici, imaginer des centrales dédiées.

• Pourquoi l’IA oblige les datacenters à se réinventer

Pour mettre en perspective cette annonce de Mark Zuckerberg, rappelons qu’Olivier de Nomazy, responsable analyste chez Data4, nous expliquait qu’ « il y avait à peu près 6,5 GW en 2023 de datacenters déployés ». L’annonce d’un datacenter de 5 GW par Meta implique donc un besoin d’énergie colossal par rapport à ce que l’on connait actuellement.

Des problèmes pour les riverains

En parallèle, lundi, le New York Times a publié un article racontant les problèmes d’une famille du Comté de Newton, dont la maison se situe juste à côté d’un des datacenters de Meta. Beverly et Jeff Morris habitent à 300 mètres de ce centre de données, et l’eau de leur maison vient directement de la nappe d’eau souterraine de leur terrain.

Le journal étasunien explique que quelques mois après le début de la construction du data center de Meta, en 2018, le lave-vaisselle, la machine à glaçons, le lave-linge et les toilettes des Morris ont tous cessé de fonctionner. « En l’espace d’un an, la pression de l’eau s’est réduite à un filet d’eau ».

Le média syndical américain More Perfect Union avait déjà publié, il y a trois mois, sur YouTube une vidéo sur les problèmes de la famille :

Blender 4.5 is now officially available today as the newest feature release and one that is a Long Term Support (LTS) for this popular, cross-platform 3D modeling software...