GPUS

NVIDIA has lost hope in China, which is why it has started manufacturing its own next-generation GPUs for AI

by

NVIDIA faces this 2026 a crucial year. They have become one of the largest strategic investors in the AI ​​ecosystem with dozens of billion-dollar investments in other companies, models, infrastructure and robotics. But, in the end, they are a company that supplies chips and, so far, the H200 They set the tone. According to a report by Financial Timesthat’s over. NVIDIA just ordered TSMC to start mass manufacturing Vera Rubinits next-generation hardware for AI. The reason? They have lost all faith in China. In short. With the entire AI industry looking to the future, and NVIDIA that has its Vera Rubin on the starting grid, it was strange that the company continued to invest so much in keeping TSMC working on a chip as old as the H200. Although it has been around for a while, it has positioned itself as unbeatable in the industry due to its price/power ratio, so these are the chips on which it has been built. the AI ​​empire. However, time passes and NVIDIA needs to move. Data centers need more power, new models are more demanding and the spearhead of the software sector – such as OpenAI either Google– have demanded new solutions. According to two sources consulted by the financial media, and close to NVIDIA’s plans, the company has grown tired of “waiting in limbo” and has begun to accelerate the delivery and deployment of Vera Rubin. Yoncomparable. As it could not be otherwise, TSMC is going to be in charge. The Taiwanese foundry would have already been asked to begin diversifying the production line to begin manufacturing the new chips. And if you’re wondering why it’s not enough for Google or OpenAI to simply buy more H200, the answer is because the chips have nothing to do with it. H200 is a more classic GPU for a data center. It is the configuration that AI and computing companies on these servers have been working with for years. Vera Rubin, however, is a paradigm shift made up of new CPUs, new GPUs and designed so that everything works as a single rack-scale accelerator. It has not only more power, but also the latest software and hardware additions from NVIDIA and something very important: incredible bandwidth. The higher the bandwidth on such a system, the more simultaneous data it can handle. This implies greater efficiency when training, but also a lower cost in inference. It is not an update, it is a platform change designed for models with trillions of parameters. Qgoose faith in China. To put it more simply, if the H200 is like a “super powerful graphics card”, Vera Rubin is like a mini data center in itself. And if you’re wondering why they didn’t start production sooner, the reason is… China. Jensen Huang, CEO of NVIDIA, has been ‘fighting’ with Washington for months to open their arms in the trade and technology war maintained by the US and China. Trump ended up agreeing and Huang commented earlier this year that they had returned to “turn on” all production lines to supply the very high Chinese demand. The problem is that that demand did not arrive. At least, It was not as high as Huang expected. In the presentation of results, NVIDIA’s financial director commented a few days ago that “although small quantities of H200 for Chinese customers were approved by the US government, we have not yet generated any income. And we do not know if imports to China will be allowed.” We already told the problem: The US was leaving for NVIDIA to sell its graphics, butThe Chinese government did not seem so convinced. Your main Big Tech They were demanding NVIDIA solutionsarguing that they need them to keep up with what their American rivals are doing, but the ball was in the court of the Government and Customs. China is promoting AI that is different from that of the US, more focused on low costs and rapid acceptance by the client, and at the same time want to build your own hardware network with companies like SMIC or a Huawei that you already have your supercomputer for AI. complicated swerve. From the Financial Times they point out that the president of China, Xi Jinping, and the president of the United States will meet at the end of March to discuss export controls. The problem is that, according to their sources, even if the barrier is lifted completely and not just for certain companies and China can buy H200s en masse, turning TSMC’s ship around so that it starts producing H200s again would be complicated. It is not as simple as pressing a button and going from producing one thing to another. If this situation occurs, “NVIDIA would take up to three months to reallocate or add capacity to the supply chain to produce H200.” One of Vera Rubin’s PCBs Rebound winner. What is clear here is that NVIDIA is not going to lose from the operation. Huang already argued that the United States could not miss the opportunity to take a slice of a multi-billion dollar market (because the US let the cards be sold… with a 25% tariff), but whether it is the Chinese or the Western industry, it is from NVIDIA that they continue to buy the H200 and, ‘shortly’, the Vera Rubin. And the rebound winner in this operation is Samsung. Of the three companies that manufacture memory (and that have catapulted the RAM and SSD crisis we are in), Samsung is the one that has completed its new generation HBM4 memory. It is the one that has passed the high standards of NVIDIA and the one that is already being mass manufactured to be able to integrate into Vera Rubin systems. Everyone attentive. As we said, NVIDIA has to the entire industry at his feet. Google, xAI and Meta are working on their own chips, but together with Microsoft, Amazon Web Services, OpenAI, Mistral and Anthropic they are some of the companies that they … Read more

A Singapore company has purchased 136,000 AI GPUs from NVIDIA. What is not clear is what he has done with them.

by

In the last three years, an unknown Singapore company has become the largest buyer of NVIDIA chips in Southeast Asia. This singular activity has caused alarms to go off, especially now that the trade war between the US and China means that the “illegal trafficking” of these components is extremely monitored. The suspicion. The company, called Megaspeed, is being investigated by the US government. The objective is to find out exactly if there are ties that unite this company with the Chinese government and if the NVIDIA chips that the company has purchased have ended up in China despite the veto and prohibition that said cards can end up there. The Singapore government is also checking whether Megaspeed has violated local laws, they say. on Bloomberg. Megaspeed denies the major. In a statement sent by mail to that newspaper, those responsible for Megaspeed declare that the company “is based in Singapore and operates fully in accordance with applicable laws, including United States export control regulations.” At the moment there is no evidence. An NVIDIA spokesperson indicates that its request for information from Megaspeed shows no evidence that there was a violation of the terms of those transactions. In their visits to Megaspeed’s data centers they confirmed that “the GPUs are where they are supposed to be.” Furthermore, according to its data, Megaspeed has owners and operates entirely outside of China, and there is no Chinese shareholder. But it does serve Chinese tech giants. Megaspeed has a “neocloud”, cloud infrastructure dedicated to offering computing capacity for AI projects. It has several data centers in Southeast Asia, and the company rents NVIDIA chips to Alibaba. This is an option that the US government does continue to allow: no buying chips, but access to those from suppliers from “non-vetoed” countries. Delicate situation. The question is whether Megaspeed has really done things right or whether it has ended up serving as an intermediary for NVIDIA chips to end up in Chinese technology companies. It would also be disturbing if in the end Megaspeed did have ties to companies or the Chinese government. This discovery comes just as President Donald Trump has stated that he would approve the sale of certain NVIDIA chips to China, something that until now was prohibited. Confusing data. Although Bloomberg admits that they have found no evidence that Megaspeed’s NVIDIA chips have ended up being sent to China, doubts remain. They have analyzed documents with records of commercial transactions, appointments and job offers from both Megaspeed and some of its collaborating companies, and have detected “inconsistencies” between the inventory of chips and those that should really be installed in their data centers. Megaspeed has thousands of NVIDIA GPUs. And the problem is that this company has a huge number of company chips. Since it was founded in 2023 and until November 2025, Megaspeed has imported at least 136,000 NVIDIA GPUs according to Malaysian and Indonesian customs records. More than half are Blackwell chips, which Trump said I would not approve of them being exported to China. Most of those newer GPUs were purchased six months ago, but NVIDIA employees who visited the data centers did not definitively clarify that those that were exported actually ended up where they were supposed to be. The suspicion: a mysterious data center in China. On the Megaspeed website it says that they have three data centers in Malaysia and Indonesia. There is also mention of a room under construction in an unspecified “specific area.” The problem is that Megaspeed showed an image of a render with a data center in Shanghai financed in part by Megaspeed’s original parent company, a Chinese company. Not only that: Megaspeed has a kind of corporate twin in China with an identical website that shows that in reality the employees of the Singapore company are its employees. All of this raises clear questions that remain unresolved and that raise even more suspicions. In Xataka | The US believed it had dealt a mortal blow to China when it deprived it of NVIDIA. He only accelerated one plan: ‘Delete America’

AMD’s problem is not that it does not make good gpus for ia. Is that it is not even close to Nvidia

by

AMD is doing things well, but even doing them still unable to compete with Nvidia. The company has just raised its renewed road map with promising models, but that is not a guarantee of anything to a NVIDIA that will not let its absolute leadership position escape. The problem for AMD is not to be, but get others to take note. IDC consultancy data indicate that Nvidia dominates the AI ​​chips market with 85.2% market dick, for 14.3% AMD. Other analysts like Jon Pedie Research go beyond and According to your data The NVIDIA quota in this segment is 92%. AMD instinct mi350 are just the beginning. The GPUS for IA, which AMD calls “accelerators”, follow its evolution. During the event they presented their family or Instinct Mi350 series with two variants, MI350X and MI355X. According to the manufacturer, these chips are four times higher in general performance with respect to the previous generation, but are up to 35 times more powerful in the field of inference AI (that is, in the practical use of models such as Chatgpt, which “infers” “their responses from our prompts). They have 288 GB of HBM3E memory and a memory bandwidth of 8 TB/s. Its yield is 18.45 pflops in FP4 precision and 9.2 pflops in precision FP8. Instinct Mi400 in 2026. Next year the new family of AMD’s accelerators will arrive. It’s about future MI400 instinctwhich will arrive with up to 432 GB of HBM4 memory, 19.6 TB/s of bandwidth of that memory, and a performance of 40 pflops in precision FP4 and 20 Pflops in precision FP8. These monsters will be sold in future racks with infrastructure “Helios“, that You can house Up to 72 Mi400 with up to 260 TB/s total bandwidth thanks to its interconnection technology, Ultra Accelerator Link. EPYC VENICE. AMD not only talked about GPUS: it also has its future processors for servers in data centers in full development. The Epyc Venice will arrive in 2026 and will be based on Zen 6 architecture. Among the variants, an especially spectacular with 256 cores that will offer up to 70% more performance compared to the previous generation. These processors will be built with future MI400 instinct. They are expected to be manufactured with the N2P (2 Nm) node of TSMC. Helios against Oberon. The aforementioned Rack Helios will compete with not already with the current Nvidia AI server, the GB200 NVL72 which connects 36 CPUS Grace and 72 Gpus Blackwell. He is destined to compete with his successor, which has Oberon’s code name and will use IA B300 GPUS with Vera Rubin architecture. The yields and benefits of these future racks are absolutely dizzy, and for example their Precision Power FP8 is 1.4 Exaflops. The same in some things, better in others. AMD promises to match NVIDIA in several sections, but also ensures that it will exceed it remarkably (50% more) in memory quantity and width, something crucial for training and inference AI. Be careful, because at the end of 2027 NVIDIA prepares the Rubin Ultra architecture, which promises racks with up to 5 Exaflops in FP8 precision, three times more than Helios or Oberon. In 2027 we will have another “summer”. The AMD roadmap goes further, and they have already prepared the development of their new generation of chips for summer Epyc servers, which will replace the Epyc Venice. These CPUS will be paired with the future MI500X instinct, and it is expected – although it is not safe – that both types of chip take advantage of the one already announced TSMC A16 node (1.6 Nm), which will begin to be used at the end of 2026. There are no specifications for these developments, surely because they will depend on the manufacturing node that AMD ends up using to produce them. Frantic race. All these ads show that AMD does not want to be left behind in that race to place their solutions in data centers worldwide. The Crusoe company, which is dedicated to the construction of large AI data centers, advertisement A few days ago I would spend 400 million dollars in AMD’s chips, and even Sam Altman, CEO of OpenAi, made a surprise appearance During the inaugural talk of the Lisa Su, CEO of AMD event. Altman said they will also use AMD chips in the data centers they use, and highlighted that the new AMD ia gpus “will be somewhat amazing.” AMD presumes to be more efficient (and cheap). AMD’s message was clear during the event: its MI355 offer much more efficiency and are cheaper than NVIDIA B200 and GB200 with comparable yields. The sales prices of those GPUS are not known, but we do know that at the beginning of 2024 the MI300x of AMD They cost a maximum of $ 15,000 for the more than $ 40,000 that cost The NVIDIA H100. The biggest challenge is still CUDA. The benefits of AMD AI chips are not in fact the problem of this company. Detailed studies revealed months ago that MI300X are clearly higher than NVIDIA H100 and H200 on performance and power. However, Nvidia has a Cudathe de facto standard in the industry for services of services and applications of AI. Using AMD native software is feasible, yes, but software experience, They assured in SEMIANALYSIS“Software is full of errors that make training (AI models) with AMD it is impossible.” AMD’s hope is Rocm. In that AMD event also presented Rocm 7, the latest version from your own Open Source programming platform for your GPUS. In AMD they indicated that this version is 3.5 times more powerful than Rocm 6, and even claim that it is 30% more powerful than CUDA in the B200 when serving the model Deepseek R1. Even so, they indicate In another report of semi -health, it is still lower in some sections. Getting that component allows developers to take advantage of all the potential of AMD’s chips is precisely key to the future of those efforts. Even … Read more

If Nvidia lived only from PC GPUs, she would be about to die of success. And US tariffs don’t help

by

During the last weeks take a family graphics card GeForce RTX 50 Nvidia without falling into the hands of speculators is a real odyssey. At the time I am writing this article the only solution of this family that is easy to find at a price close to the recommended By nvidia is the GeForce RTX 5070. As is understandable, this situation generates frustration among the players who have decided to get one of these graphics cards, as you have noticed in the comments of our articles. To understand what is happening, we are interested in taking a look at “the photograph” of the PC graphics card market. According to the consultant Jon Pedie Research During the last 2024 quarter, more than 78 million GPUs were distributed for PC, which represents a 6.2% growth compared to the third quarter of last year. Nvidia leads this market with a quota of 65%, while Intel and AMD are formed with 16% and 18% respectively. This is our starting point. Difficult times arrive for the PC graphics card market This statement of Jon Pediethe head of the consultant who I have mentioned a few lines above, describes very precisely what the current situation is: “Nvidia, with the highest market share, had difficulty satisfying the demand. And as a result of its size and influence, it prevented the GPU market Tariffs will include profits for the majority, even for everyone, for 2025 “. TSMC is responsible for producing these chips for Jensen Huang’s company Nvidia has not officially revealed how many GPUs belonging to the GeForce RTX 50 family has delivered to graphics card assemblies, but it is evident that not necessary to meet demand. The Taiwanese manufacturer of semiconductors TSMC is responsible to produce these chips for the company of Jensen Huang, but it is important that we do not overlook that it not only manufactures the GPUs for Nvidia games; also produces its chips for artificial intelligence (AI). And the production capacity of TSMC is limited. From one thing we can be sure: in the current circumstances the GPU market for AI is a priority for Nvidia. After all, it is Its greatest source of income. The Jon Peddie Research report expects that from 2024 to 2028 the general market of the GPUs will be contracted annually, and in 2030 only 15% of the computers will have a dedicated graphics card. In these circumstances it seems reasonable to assume that Nvidia and AMD will pay more and more attention to their chips for AI, and perhaps a little less to their GPU for games. Anyway, there is another fact in Jon Peddie’s statement that is worth not overlooking: The tariffs which is deploying the administration led by Donald Trump. At the moment TSMC produces Nvidia chips in Taiwan. Perhaps in the medium term diverts the manufacture of these integrated circuits to Your new US plantsbut for the moment neither Nvidia, nor AMD, nor any other client who buys TSMC his chips produced in Taiwan will be fought from the tariffs. And in these circumstances, as Peddie states, the profits of these companies will be affected. If prices do not raise their benefits will be reduced. And if they are presumably submitted less, so their earnings will also resent. More information | Jon Pedie Research In Xataka | The Singapore government has revealed which companies are involved in the delivery of GPU from Nvidia A Deepseek

GPT-4.5 It is the demonstration that using more GPUS and more data is no longer useful

by

In the last two years we have seen how companies that develop AI models have not stopped showing voracity almost without limits. They bet on climbing and using more data and more GPUS to improve those models. However, there has been a surprise: it turns out that this strategy no longer works. GPT-4.5 will be the last of your lineage. We have always associated with chatgpt with the traditional models “that do not reason”, although in recent times it also gives Access to reasoning modes. Even so, its current base is GPT-4Oand that model will have a last successor. It will be GPT-4.5, which will not be renewed. That is precisely the interesting thing. Climbing no longer serves much. As they point out experts like Gary MarcusGPT-4.5 It seems to be the finding that spending more and more money on climbing, using more and more GPUS and data to train models no longer makes no sense. OpenAi’s hope was Orionwhich aimed to be GPT-5, but it is not: it is (probably) GPT-4.5. Shock against a wall. The jump in performance and capacity It was never the expectedwhich resulted in the deceleration of AI. At least, of the generative AI that does not reason. That of course seems to have collided with a wall, and can no longer improve. We are, in the face of a change of total focus towards reasoning models. It is happening to all. GPT-4.5 is the acceptance of this new reality by OpenAI, but there are many other AI companies that are in the same situation. The new versions of the models “that do not reason” do not just arrive. Grok 3 does not arrive and Xai is staying behindbut we have also not seen Claude 3.5 successor and we don’t know what Anthropic is working. Google just Present Gemini 2.0but the leap in capabilities with respect to Gemini 1.5 is not spectacular, at least if we do not take into account its reasoning version, Flash Thinking. I told you. Experts like Yann Lecun, head of goal, since warning that this strategy of “more data and more GPUS” had an expiration date. Ilya Sutskever, Openai co -leaflet and now with her own startup of AI, It also made it clear months ago. For him the massive training of an AI model using a large set of data without labeling so that the model detects patterns and structures no more than itself, and even trying to do it more and larger, also did not offer too many advantages. So, why spend so much money? If traditional models can no longer advance with that climbing, the question is obvious: why are companies investing billions of dollars in data centers? The answer is diverse. First, the climb is still useful to improve the models and make them behave better and comment less errors. Data centers make sense. But it is also the section of inference: that gigantic infrastructure in which companies are investing is not so much to train models with the traditional approach, but so that hundreds or even billions of people end up using AI in their day continuously. That is the current bet. That live the models that reason. The deceleration of the AI ​​that takes time speaking is not “of the whole AI”, but as we say of the traditional generative models that did not reason. The new models such as O1, Deepseek R1 or Gemini 2.0 Flash Thinking are clearly the trend: increasingly precise and with answers that have more and more quality and really help us to trust them. To do work for us almost “blind.” We have advances in AI for a while. The AI ​​still has a long way forward. That the climbing approach (more Gpus, more data, this is war) does not make much sense, because there are other paths. Many. And that of reasoning models is just one of them. Image | Amazon In Xataka | OpenAi wants to be the new Google with GPT-5: You will ask and the AI ​​will already decide how it answers you

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.