NVIDIA

OpenAI teamed up with NVIDIA and made circular financing fashionable. Anthropic has returned the ball with a surprise girlfriend: Google

by

Let’s see if we were going to believe that OpenAI was going to be the only one to look for powerful allies. Nothing of that: Anthropic just did the same and has announced an eye-catching agreement with Google. The AI ​​startup will have access to up to one million Google TPUs in a pact that is worth “tens of billions of dollars.” Less noise, but a lot of nuts. The figures of the agreement are modest if we compare them with those that OpenAI has managed in its circular financing agreements with NVIDIA, amd either Broadcombut here Anthropic seems to take a very different position. Compared to colossal projects like Stargate, Anthropic’s idea is focused on execution. Without making much noise, the company led by Dario Amodei has been gradually conquering the business sector. More than 1 GW of computing capacity. On CNBC indicate that this investment will allow the creation of a data center with a computing capacity greater than 1 GW and have it ready in 2026. It is estimated that a center of these characteristics would cost about 50,000 million dollars, of which about 35,000 million would be dedicated to AI chips. It may not be comparable to Stargate and the idea of ​​investing $500 billion in data centers, but the alliance between Anthropic and Google is significant. More than circular financing. The partnership certainly features elements of circular financing, but it is more of a symbiotic relationship with that cross-investment component. The dynamic is simple and is now completed with that commercial return. The agreement requires Anthropic to buy or rent infrastructure services from Google Cloud. Virtuous circle. With its original investment in Anthropic, Google helped that company grow, which in turn allows Anthropic not only the ability to grow, but the need for enormous computing power… provided by Google. In essence, some of the money Google invests in Anthropic returns to Google Cloud as revenue. The vicious (or virtuous, as they say in the US) circle is complete. Anthropic diversifies. Anthropic’s AI models are trained and used using infrastructure from various manufacturers. Thus, they use both Google TPUs and Amazon Trainium processors and NVIDIA GPUs: each platform is assigned to a specialized workload. In the case of Google’s TPUs, according to Anthropic the focus is “its strong price/performance ratio and its efficiency.” Promising successes, but… Anthropic’s growth is evident, and its annualized revenue rate (ARR) is now estimated to reach $7 billion. Claude Code, its developer assistant, managed to generate 500 million dollars after just two months on the market. But as always, that revenue can’t hide the fact that Anthropic, like other AI startups, you continue to spend much more money than you earn. Amazon is your other great ally. In fact, the company led by Andy Jassy has invested around $8 billion, when official data indicates that Google has invested $3 billion. AWS is still considered the largest infrastructure provider for Anthropic, and its supercomputer Project Rainierbased on the Trainium 2, allows you to have a large computing capacity for every dollar invested, they point out on Amazon. The company’s influence is not only financial: it is structural. Image | Wikimedia | Fortune Brainstorm Tech In Xataka | You thought you had an amazing connection on Tinder, but you were actually chatting with ChatGPT

AI is running out of power in this world. So Nvidia has opted for servers in space

by

The energy appetite of data centers is nothing new. Elon Musk predicts a shortage of transformers in two years. Sam Altman believes we will need an energy revolution, such as nuclear fusion, to keep pace. The planet was not prepared for so much energy demand. And that’s why Nvidia is funding a possible solution: deploy the servers outside of Earth. It’s not science fiction. It is the business model of several startups that propose building the next hyperdata centers in Earth orbit and even on the Moon. The idea, which until recently sounded far-fetched, is gaining traction driven mainly by two factors: the insatiable demand for AI and the low-cost launches that Starship promises. One of the companies leading this idea is Starcloud, supported by the NVIDIA Inception program. And he is so serious that he plans to launch his first satellite, the Starcloud-1in November. On board it will carry the first GPU for data centers launched into space: an NVIDIA H100. The difficult part will come later. Starcloud-1 is a test unit the size of a small refrigerator, but the company’s goal is to build a monster five-gigawatt orbital data center. Adding the solar panels and the enormous radiator, it would measure four kilometers wide. Its goal is the training of large AI models in orbit. Why in space? As detailed in an extensive white paperfuture models like GPT-6 or Llama 5 could require multi-gigawatt clusters, something “simply impossible with the current energy infrastructure” on Earth. In space, there is no such limitation. It’s more. According to Starcloud calculations, server energy costs are 10 times lower in space than on Earth. The value proposition of space data centers is based precisely on two pillars that are a problem on Earth: energy and cooling. Solar energy 24/7. On Earth, solar energy is intermittent. They depend on the day/night cycle, the weather and the atmosphere, which attenuates the radiation. In space, things change. By placing your data centers in a sun-synchronous “dawn-dusk” orbit, Satellites follow the line that divides day and night on Earth. With the panels illuminated by the sun almost continuously, the system increases its capacity to more than 95%. “Almost unlimited, low-cost renewable energy,” in the words of Starcloud. And the refrigeration? How would they dissipate all that heat? Land-based data centers consume millions of liters of fresh water to cool. There is no water in space, but they have something much better: an infinite heatsink at -270°C. The plan is not to ventilate the servers. The heat generated by GPUs (such as the H100) will be managed within sealed modules using liquid cooling (direct-to-chip or immersion), like high-performance systems on Earth. The difference is that this hot liquid does not go to an evaporation tower, but is pumped to gigantic radiator panels. These panels simply radiate waste heat into the vacuum of space in the form of infrared radiation. The Starcloud white paper details the calculations using the Stefan-Boltzmann law, estimating that a radiator at 20°C can cleanly dissipate more than 630 watts per square meter. Without using a single drop of water. Not everything that glitters in space is gold. The pillar that supports this entire concept is the launch of high-capacity reusable rockets, such as SpaceX’s Starship. Starcloud calculations are based on a long-term cost of $30 per kilo put into orbit. But Starship is not ready, and it is certainly far from achieving its full and rapid reusability capability. If that cost does not materialize, the economic viability of the system collapses. The other big problem is radiation. Commercial GPUs are not designed for space. Cosmic radiation and solar flares can fry electronics. The solution is shielding, which adds mass and therefore launch cost. Not to mention that maintenance is not possible with current technology.

This is your asset so as not to depend so much on Nvidia

by

OpenAI has announced A multiannual agreement with AMD so that the chips company supplies artificial intelligence processors that will feed part of its AI infrastructure. The pact includes the hardware deployment equivalent to 6 gigawatts of power and gives the ChatgPT creators the option to acquire up to 10% of participation in AMD. A colossal dimensions agreement. The plan contemplates that Openai begins to use the AMD Instinct Mi450 chips in the second half of 2026, with a first installation of a gigavatio. The New York Times Compare The magnitude of the total deployment (6 gigawatts) with the equivalent of electric consumption of all Massachusetts households. AMD ensures that the agreement could generate tens of billions of dollars in annual income and more than 100,000 million in four years, telling the drag effect on other customers. Beyond the economic. As part of the pact, AMD has issued a purchase option (Warrant) that allows Openai to acquire up to 160 million shares from a cent. This option is unlocked by sections as specific objectives are met, which include the first sending of MI450 chips and AMD contribution objectives that scale up to $ 600 per share. AMD’s actions shot More than 20% In operations prior to the opening of the market after the news is known, adding 80,000 million dollars to its capitalization. OpenAI diversifies suppliers. The movement arrives just weeks after Openai closed An agreement of 100,000 million dollars With Nvidia, the technological giant who has the domain of the chips of AI. With Nvidia, Openai promised to deploy hardware equivalent to 10 gigawatts. According to ReutersSam Altman has set expectations of reaching 250 gigawatts of total computing capacity by 2033, which explains this multiple suppliers strategy. The company also works with Broadcom in The development of their own processors (Xpus). AMD looks for its hole in front of Nvidia. For AMD they are great news, since the agreement represents the validation of its chips and software in a market where Nvidia prevails almost completely, controlling approximately 90% of the quota in processors for AI. “We consider this agreement certainly transformative, not only for AMD, but for the dynamics of the industry,” declared Forrest Norrod, Executive Vice President of AMD. The company has been collaborating with Openai for years, contributing ideas in the design of previous generations such as MI300X chips. Hunger of AI infrastructure. OpenAI and other great technological plan to spend More than 325,000 million dollars in data centers only this year. Unlike giants such as Amazon, Microsoft or Google, which finance these projects With your available operational cashOpenai, which according to the latest reports It has generated about 4.3 billion dollars In revenue in the first half of 2025 while burning 2.5 billion in cash, it needs to look for creative financing formulas. The agreement with AMD, like Nvidia, allows Openai to ensure more supply while aligning its strategic interests with its suppliers. In Xataka | Everything you ask the goal AI on WhatsApp or Instagram will be used to sell you things: this is the new mandatory clause

CUDA is the standard that grips the world and Nvidia is the only company with chips capable of running it. Until now

by

Goal will acquire rivos, a Californian startup specialized in the design of chips based on RISC-Vaccording to sources of Bloomberg. In addition to the capabilities of its chips, the operation is part of a broader strategy: free itself from the NVIDIA dependence and thus take control of its infrastructure for artificial intelligence without its chips. What is at stake. Throughout these last years, Nvidia has dominated the GPUS market For the thanks to CUDAits owner development platform that has become the de facto standard to train and execute artificial intelligence models. Today, we have reached the point that whoever wants to make a large scale needs Nvidia chips, and that gives the company a huge market power, since they put the necessary hardware for an industry in which everyone wants to enter. Goal, despite having some of the best open models in the sector with Callskeep spending billions annually in Nvidia hardware. The strategic movement. With rivos, goal not only buys a company, buy an alternative to the current technological stack. The startup Develop GPUS and RISC-V-based acceleratorsan open source architecture standard that threatens the traditional X86 (Intel and AMD) and ARM. Goal already works in its own internal chip, the goal Training and Inference Accelerator (Mtia), designed next to Broadcom and manufactured by TSMC, but the advances are not as fast as Zuckerberg would like. According to sources cited by Bloombergthe CEO would have been actively looking for market reinforcements to accelerate development. It is not the only one. Goal adds to a career in which their technological rivals already have an advantage. Google has His tpusAmazon has Trainium and Microsoft has developed Maia. The AI ​​war does not win only with the best models, but also With the chip that executes them And goal, despite being burning hundreds of billions of dollars in AI, it was staying behind in this front. The context. Rivas acquisition is not an isolated movement. Target there was already tried to buy furiosaaia South Korean startup specialized in chips to train AI systems, but the offer of 800 million dollars was rejected. In addition, the company has recently announced An investment of 29,000 million dollars To build a huge data center in Louisiana and plan to spend up to 72,000 million this year on infrastructure related to AI. The RISC-V challenge. Rivas represents an ambitious bet. Although RISC-V has not yet managed to penetrate massively into US data centers (its presence is mainly limited to microcontrollers and IoT devices), its potential is undeniable. China is already launching tablets and laptops with this architecture. If Meta manages to develop an AI accelerator based on RISC-V capable of replacing The NVIDIA H200 In its internal operations, it would be a considerable blow for the dominant standard. Cover image | Nvidia and Goal In Xataka | Openai has just presented Sora 2 with a Tiktok -style app. This is outlined a new wave of viral videos

“Circular financing” between Nvidia and Openai can be the genius of the century … or collapse

by

Nvidia has announced A “strategic investment” of up to 100,000 million dollars in Openai. But it is an investment with trap: Openai will use that money to buy Nvidia chips. The semiconductor manufacturer thus becomes the financier of its own most important client. Why is it important. This maneuver dangerously reminds the “circular financing” schemes that characterized the end of the 2000 Puntocom bubble. Companies like Lucent, Nortel and Cisco financed operators as Global Crossing to buy them equipment. We are not the first to see this simile At this stage of AI. When the bubble exploded, both suppliers and customers sank into a spiral of debts and overcapacity. The agreement will allow OpenAI to build data centers with a joint capacity of 10 gigawatts, equivalent to about 10 nuclear reactors. Jensen Huang, CEO of Nvidia, has acknowledged that this represents between 4 and 5 million GPUS: “double those we distributed last year.” Brutal scale In figures. The numbers are astronomical. According to Huang himself in August, creating a 1 Gigavatio data center costs between 50,000 and 60,000 million dollars, of which about 35,000 million are destined for Nvidia chips. With that logic, the 10 projected gigawatts would cost more than 500,000 million dollars. The bags have reacted with euphoria: Nvidia shares rose almost 4%, adding 170,000 million dollars to their stock market capitalization. Jensen Huang Broza’s company is already 4.5 billion dollars of valuation. Yes, but. Parallelism with the ‘Puntocom’ bubble is disturbing. These same schemes of ‘Financing vendor‘We already saw them in the final stage of the 2000 technological bubble. They did not end well for any of the parties. The difference is that current numbers are much larger, even adjusting for inflation. The key is whether the productivity profits of the generative AI will compensate for the spent money. Between bambalins. The agreement explains the current situation in the AI ​​ecosystem: OpenAi desperately needs computing capacity to maintain its competitive advantage over the 700 million weekly users of their products. But infrastructure costs are so high that it needs constant external financing. Nvidia, on the other hand, seeks to ensure the future demand of its most advanced chips. The agreement guarantees mass orders while consolidating its dominant position against competitors such as AMD and Intel. “It is a closed cycle: Nvidia gives OpenAi money, and OpenAi uses it to buy Nvidia products,” Summary Summary Javier Pastor. The threat. Anti -Ponopoopoly experts are already arched eyebrows. Andre Barlow, a lawyer specialized in competition, explained to Reuters that “the agreement could change the economic incentives of NVIDIA and OpenAI, potentially blocking the Nvidia chips monopoly with OpenAi software leadership.” The structure creates extra barriers so that competitors such as AMD in OpenAi chips or rivals in AI models can climb their operations. They paint basts. In perspective. The story is full of similar schemes that ended badly. Global Crossing, the telecommunications operator that broke in 2002it was funded precisely by the same suppliers that sold equipment. When it was discovered that the real demand was much lower than the projected, both Global Crossing and its financiers lost thousands. The key question is whether the demand for AI services will be sufficient to justify this billionaire investment, or if we are faced with the recreation of the same speculative pattern with even more exorbitant figures. As Stacy Rasgon concludesBernstein analyst: “On the one hand, Openai helps meet very ambitious infrastructure objectives. On the other hand, it will further feed concerns about ‘circular’ financing.” Outstanding image | In Xataka | Openai estimates that it will enter 200,000 million dollars in 2030. The figure, like everything in OpenAi, is extremely ambitious

Nvidia will invest 100,000 million dollars in OpenAI. Actually a single euro will not be spent

by

Openai has signed a “strategic agreement” with Nvidia. According to this agreementNvidia “intends to invest up to 100,000 million dollars” in OpenAI gradually, but the truth is that this investment is misleading. Especially since Openai will spend those 100,000 million dollars to buy GPUS to Nvidia. Everything remains at home. What happened. These two companies have initiated the procedures to complete an agreement with a clear objective: create and display AI data centers With a joint gigantic computing capacity: 10 GW. The investment will be made gradually and will be completed “as each gigawatt” of computing capacity is installed in those Data centers. Nvidia will thus become a “computing partner and strategic connectivity” for the development plans of new data centers, says Openai. Millions of Gpus. According to Jensen Huang statementsCEO of Nvidia, that represents between four and five million gpus. Or what is the same: it is the number of units of their GPUS of ia that they expect to distribute this year, and “twice the ones we distributed last year.” The strategy “seller finances buyer”. This agreement is not a simple investment, but a strategic association in which the hardware provider invests a massive amount of money in its main client. In return that client undertakes create a mass infrastructure With supplier technology. It is nothing more than a closed cycle: Nvidia gives OpenAi money, and OpenAi uses it to buy Nvidia products. This sounds like a bubble. There is Several analysts that They speak How this remembers once again The bubble of the Puntocomwhere companies lent money to buy products from the other. That raises suspicions and questions about the long -term sustainability of these agreements. Companies becoming stronger among them. The circular agreement serves in fact to strengthen both companies and solidify their positions as dominant and indispensable actors in the AI ​​industry. In fact, this strategic alliance makes rivals like AMD or Intel very difficult. Nvidia is worth 170,000 million dollars more. The announcement caused immediate reactions in the NVIDIA assessment, whose shares increased almost 4%. The stock market capitalization of the company of Jensen Huan grew by 170,000 million dollars in that session and already touch the 4.5 billion dollars, and manages to distance itself even more from Microsoft, Apple or Google, which already exceed three billion. Long live Hype. Here once again there is a reinforcement of the speech of expectations and Hype. The confidence of these companies in the future of AI is patent, but they are interested and for now Openai’s income – no rivals – are well below spending They are doing in these technologies. Energy challenge. The plans to create infrastructure with 10 GW capacity are also astronomical. According to Some estimatesthose 10 gigawatts They are equivalent to the production of about 10 nuclear reactors, which normally provide a capacity of 1 GW per plant. A colossal cost. The current data centers range between very modest capabilities of 10 MW and other extraordinary 1 GW. Openai’s plans would leave those facilities very behind in computing capacity. In August Huang told investors to create a 1 GW data center is a cost of between 50,000 and 60,000 million dollars, of which about 35,000 are dedicated to Nvidia chips. With those figures, the total cost of those 10 GW of joint computing power would amount to more than 500,000 million dollars, a figure that – one—curiously— It coincides with that of the Project Stargate. Image | Flikr (Techcrunch) | Nvidia In Xataka | 5,000 “tokens” of my blog are being used to train an AI. I have not given my permission

Nvidia has paid 900 million for one

by

Nvidia has signed a new talent from Ia. The news is not that, but the way in which he has signed it: to achieve “capture it” he has made an investment of more than 900 million dollars in the company he directed. It is a new hiring modality that allows “stealing talent” without current legislation being able to do much. What happened. According to CNBC dataNvidia has invested more than 900 million dollars in the AI ​​Hardware Startup Contractwhich manufactures connectivity components for AI servers. As part of that agreement, Nvidia has signed his CEO, Rochan Sancar, who will begin to work directly for the firm led by Jensen Huang. Clusters to power. Contractivity connectivity solutions – founded in 2019 – allow the company to connect more than 100,000 GPUS to put them to work within AI. These types of solutions can help Nvidia offer integrated systems that use their chips. Or what is the same: all this points to new “supernodos” of computing with thousands of Nvidia chips ready to be installed in large data centers. The firm already its GB200 NVL72 marketsfor example, but this agreement allows you to go more in that field. They already knew each other. In 2023 Nvidia already participated in a round of investment in which Undergraduate raised 125 million dollars. In 2024 another new 115 million round He made companies such as ARM, Samsung or Cisco participate. According to Pitchbook, after this round the contribute assessment was 600 million dollars: the investment made by NVIDIA is enormous considering that data. Big Tech invest fortunes to sign talent. This tactic is the same that used goal in June, when invested 14.3 billion dollars In the startup Scale AI and signed his new tsar of the Superintelligence Division, Alexandr Wang. Google did the same in July by announcing an agreement with Windsurf: would invest 2.4 billion dollars in itand incidentally, he would sign his CEO, Varun Mohan. Microsoft did the same With Inflection and Mustafa Suleyman In July 2024, and Amazon also He moved file At that time when signing managers of the STARTUP of the ADEPT. One way to avoid legal problems. Traditionally, these types of operations were carried out through the well -known “Acquihires”. A large company bought another and in many cases it did to get talent, not because of the product or service offered by the “victim.” These agreements have ended up suffering remarkable legal scrutiny, which has made large companies go to other forms of talent. These “pseudoinversiones” are nothing more than a mechanism to achieve that talent without being so exposed – at least, for the moment – to legal scrutiny. And a distortion of the startup market. These operations, however, pose an important problem for the global startup panorama. If a large company can go to methods like this to sign talent, change the dynamics and strategies of startup themselves. After the investment in Underground, the company should continue to be separated, but to what extent is this a undercover acquisition? More elements for the bubble of the AI. There is also a threat to the risk capital market. Big Tech are using their huge cash reserves to inflate startup assessments such as getting out. Not because of its market potential, but as a “hiring premium” covered by its founders. This can create a bubble and change the strategy of risk capital investors, which can now value talent more than long -term business viability. Image | Hillel Steinberg In Xataka | We still do not have the four -day week and there are already CEOS dreaming with the next level: work only three days

one million terabytes and 24,000 nvidia chips for a key mission

by

In an increasingly digitized world and where artificial intelligence (AI) is transforming the way we work, investigate and relate, the supercomputing has established itself as the rod of measure technological power. It is a strategic resource that allows us to accelerate advances in science, innovation and defense. Not all super -taders play in the same league. Frontierof the United States Department of Energy, marked a milestone in 2022 by becoming the first to officially overcome the exaescala barrier, with 1,102 Exaflops in the Benchmark HPL. To that achievement they joined later The Captain and Auroraalso on American soil, consolidating its leadership position on paper. In the case of China, the information remains opaque, With very few public data about the status of their projects. Europe, however, just moved. Your first superorous to exaescala is already underway: Jupiter. Installed in the Jülich Supercomputing Centerin Germany, one of the most important advanced research poles of the continent. Jupiter is driven by the platform Nvidia Grace Hopper And Evidan xh3000 Bullsequana architecture is based on a liquid -refrigerated system designed to squeeze efficiency and performance. It is expected to reach up to 90 exaflops in artificial intelligence loads. Their applications will be diverse: from climatic research to neuroscience and quantum simulation, placing Europe in a new calculation capacity league. An inauguration with historical air September 5 The official inauguration ceremony in Jülich took placewith the presence of German authorities, European and leaders of the technology industry. German Chancellor Friedrich Merz presented him as a Pioneer project for Europe. “With Jupiter, Germany now has the fastest supercomputer in Europe and the fastest room in the world. Open completely new possibilities, from the training of AI models to scientific simulations.” In the Top500 list, Jupiter already appears as the fourth Most powerful supercomputer in the world, only behind the Captain, Frontier and Aurora in the United States. The European Union stands outIn addition, what works entirely with renewable energyby hiring green supply on the German network, and that its Rack Jedi leads the Green500 energy efficiency classification. The figures behind Jupiter To understand its magnitude, just review some technical data: 24,000 superchips nvidia gh200 grace hopper 51,000 network connections with infiniband quantum-2 technology Storage capacity close to an exabyte Modular installation with 50 specialized containers Maximum consumption of 17 MW, equivalent to about 11,000 homes A rack called Jedi leads the World Energy Efficiency Classification Why is it relevant to Europe Europe had been behind in the supercomputing career for years, with a landscape dominated by the United States. JUPITER offers researchers, companies and academic centers direct access to a top -level machine without depending on external resources. This means forming their own talent, consolidating experience in the management of these systems and reinforcing technological sovereignty at a time when artificial intelligence and calculation capacity have become strategic issues. Concrete applications The first projects already selected show how far a supercomputer of this category can go: Climate: The ECMWF works with a kilometer scale simulations, capable of representing extreme storms and feeding the Destination Earth project, whose objective is to build digital twins of the planet European: The Trustllm consortium trains language models in multiple European languages ​​for industrial and scientific applications Neuroscience: With the arbor simulator, neurons behavior will be modeled at the subcellular level, key to developing therapies against diseases such as Alzheimer’s Quantum: JUPITER aims to exceed the 50 -QBITS record in simulation, a relevant step towards quantum practical computing Astrophysics: The Max Planck Institute will use it to study cosmic reion, the period in which the first stars and galaxies emerged Particle physics: The University of Wuppertal will increase the resolution of its calculations on the Mon, which could open the door to new discoveries Video models: The University of Munich explores compression and dissemination architectures to advance applications that go from medicine to autonomous driving Multimodal models: The University of Lisbon Scale open and multilingual models, integrating different fields of science and automatic learning Access and future Researchers may request access to the system in calls that will be held twice a year. At the moment, there are already 30 projects underway. The expected useful life is at least six years, which guarantees continuity and stability in a land where technological cycles are increasingly fast. A strategic movement Jupiter is not just a technological achievement. It is a strategic commitment to provide Europe on their own capacity in an area where part of the future of artificial science and intelligence is played. With him, the continent finally has a tool that allows him compete at the highest levelwith energy efficiency and technological independence. Images | Nvidia | Jülich Supercomputing Center In Xataka | Alibaba has just demonstrated that Openai spends 78 million to do the same as them for $ 500,000

We believed that Nvidia was the company that had benefited most from AI. Micron is ridiculous

by

Micron is at historical maximums in Nasdaq, and rightly. The American manufacturer is taking a lot of benefit from the FEVER through the AI ​​and the data centers. The demand for memory chips is growing extraordinarily, but that has two faces. A good for micron, and another bad for customers and consumers. They all love micron. Citigroup analysts They promoted these days Micron’s target price from $ 150 to 175. The reason: according to its data, the company will have financial results “much better than consensus” when these are presented on September 23. Micron is doing so well that it even exceeds the growth of Nvidia. Source: Bloomberg The chips devastate. Yeah A week ago The shares were around $ 125, yesterday they closed at $ 150 and before market openings that figure is $ 155. This year the value has already grown by 81%, exceeding 33%Nvidia growth, although it is also true that the company led by Jensen Huang grew especially in 2024 (approximately 170%). Other companies such as Broadcom (55%), SK Hynix (91.88%in the South Korean bag), or TSMC (31%) also show an outstanding growth in the bags. Micron’s “Compute Networking” division is the one corresponding to the data centers. As can be seen, sales in that segment are already more than half of all of the last quarter. Source: Paul/Note. The commitment to HBM memories goes well. Micron has dedicated many resources to boost the manufacture of HBM memories, used precisely in the accelerators (GPUS) that are used in data centers. Independent analysis confirm the increasing weight of both these memoirs and the AI ​​segment in the micron business Micron will raise prices. According to Citi analysts, workloads for the inference of AI need more DRAM and NAND memories, and demand is spectacularly. The problem is that this demand will overcome the supply, and Micron will take advantage of the occasion to do something logical (for her): upload prices. Up to 30%. This is what it indicates Trendforce And also Some media In China, according to which Micron has notified its distribution channel partners today that the prices of their storage products will rise between 20% and 30%. In fact, the quotes of the DDR4, DDR5, LPDDR4 and LPDDR5 memories have been suspended among others: “All prices agreed with customers will be canceled and the quotes will be suspended. All products are expected to stop quoting for a week. “That involves not only industrial and consumer memories, and the chips for the automotive industry will rise in price by 70%. Sandisk and TSMC have already announced up. Both TSMC and Sandisk announced Price increases For memory chips in the past days. That will affect its great clients –apple, Nvidia, among others – and as indicated In Techpowerup It is a clear confirmation that manufacturers want to maintain their gross margins. In Sandisk there have been 10% prices due to the “growing demand” of the AI ​​market, data centers and mobile devices. At the moment, they indicate In Trendforcethat climb has encountered resistance from customers. In Xataka | Intel’s recent history is that of a failure. Now he has found a niche from which to resurface: HBM memories

China is no longer made up of moving away from Nvidia. His next step is the heart of the AI ​​with a system that breaks molds

by

In 2017, the Paper “Attention is all you need”Google changed the technical basis of language generation: the Transformers They allowed to process long sequences in parallel and climb models to sizes that were previously unfeasible. That climbing route has driven architectures such as GPT and Bert and has converted self -how The central piece of generative AI Contemporary. But this new approach was accompanied by growing costs in memory and energy when the context lengthens, a limitation that has motivated research to develop alternatives. Spikingbrain-1.0 aims to break molds. Of the “attention is all you need” to the brain: the new commitment to break limits in the A team from the Chinese Academy of Sciences Automation Institute He has just presented Spikingbrain-1.0. We are talking about a family of spiky models aimed at reducing data and computation necessary for tasks with very long contexts. The experts propose two approaches: Spikingbrain-7B, of linear architecture focused on efficiency, and spikingbrain-76b, which combines linear attention with Mixture of Experts (MOE) mechanisms of greater capacity. The authors detail that much of the development and the tests were carried out in clusters of GPU Metax C550, with libraries and operators specifically designed for that platform. This makes the project not only a promising advance at the software level, but also a demonstration of own hardware capabilities. An especially relevant aspect if China’s effort is taken into account for reducing his dependence on Nvidia, A strategy that we already saw reflected with Depseek 3.1. Spikingbrain-1.0 is directly inspired by how our brain works. Instead of having neurons that are always “burning” by calculating numbers, uses spiky neurons: units that accumulate signals until they exceed a threshold and trigger a peak (spike). Between peak and peak they do nothing, which saves operations and, in theory, energy. The key is that not only does it matter how many peaks there are, but when they occur: the exact moment and the order of these peaks carry information, as in the brain. In order for this design to work with the current ecosystem, the team developed methods that convert traditional self -acting blocks into linear versions, easier to integrate into its spiky system, and created a kind of “virtual time” that simulates temporal processes without stopping the yield in GPU. In addition, the Spikingbrain-76B version includes Mixture of Experts (MOE), a system that “awakens” only to certain submodos when we are needed, which we have also seen in GPT-4O and GPT-5. The authors suggest applications where the context length is decisive: analysis of large legal files, complete medical records, DNA sequencing and massive experimental data sets in high energy physics, among others. That lace appears reasoned in the document: if the architecture maintains efficiency in contexts of millions of tokens, would reduce costs and open possibilities in domains today limited by access to very expensive computer infrastructure. But validation in real environments is pending outside the laboratory. The team The code of 7,000 million parameters has released in Github next to a detailed technical report. It also offers a web interface similar to chatgpt to interact with the modelwhich according to the authors are deployed entirely in national hardware. Access, however, is limited to Chinesewhich complicates its use outside that ecosystem. The proposal is ambitious, but its true scope will depend on the community to reproduce the results and make comparisons in homogeneous environments that evaluate precision, latencies and energy consumption in real conditions. Images | Xataka with Gemini 2.5 | ABODI VESAKARAN In Xataka | Openai believes having discovered why the IAS hallucinates: they don’t know how to say “I don’t know”

Log In

Forgot password?

Forgot password?

Enter your account data and we will send you a link to reset your password.

Your password reset link appears to be invalid or expired.

Log in

Privacy Policy

Add to Collection

No Collections

Here you'll find all collections you've created before.