tokens

$200 a month gives access to $14,000 in tokens

by

Most of those who pay to use AI models do so with subscriptions. They take advantage of platforms like ChatGPT Plus either Claude Pro and with them they access an all-you-can-eat buffet that at first seems quite generous. Those who are in this modality surely have divided opinions: some will say that they immediately cross the limits and others that they almost never cross them. They are both right. Bargain subscriptions. The prestigious SemiAnalysis has carried out an investigation most curious. They signed up for Anthropic and OpenAI subscription plans and then tried to get the most out of them. The question they wanted to answer was simple: are they really profitable, or is it better to pay per use with the API? His conclusion is forceful: these subscriptions are an absolute bargain… if used well. You pay 200 dollars in tokens, they give you 14,000. SemiAnalysis research took advantage of these subscription plans to execute complex programming tasks that also extended over time to exhaust the weekly usage limits of each account. Popular belief is that these plans have a consumption limit of about $2,000 (costing $200). However, their tests showed that Anthropic’s plan allowed them to consume $8,000 per month in API tokens. In the case of OpenAI, things were even better: they managed to consume the equivalent of $14,000 worth of tokens per month. The savings are simply amazing and make one thing clear: if you make the most of your subscription, it is almost a free gift. The all-you-can-eat buffet trap. The figures make it clear that OpenAI and Anthropic are using the same model of business than that of gyms or free food buffets: users who do not go or eat little finance those who do not stop going and put on their boots. The technology companies are taking huge losses with the power users of AI that do not stop using these plans with autonomous agents that squeeze them. Risk. But that group is balanced by many subscribers who pay the flat rate but only ask a handful of fairly simple questions each day. According to analyst Ed Zitron, that’s dangerous: it is enough for 25% of users to decide to squeeze those usage rates for the profit margins of these companies to be negative. Price drop in sight. Coinciding with the study, in The Wall Street Journal they indicated this week how OpenAI is considering entering a price war by lowering the prices of its subscriptions. It could thus anticipate Anthropic—which could do the same—but for experts that could end badly. Gary Marcus explained that OpenAI is already surviving by creating hypeand a decision like that could go very badly for them. The Ghost of DeepSeek. This hypothetical price war could also be motivated by the costs of Chinese models such as DeepSeek, which offers much of the capacity of GPT or Claude, but at a much lower cost. Opposite that, of course, are the APIs that allow payment per use and that both Anthropic and OpenAI are increasingly forcing. These APIs impose a surcharge of between 40x and 70x the price of the subscription tokens according to some experts. Agents against subscription plans. What is threatening the future of these subscription plans are AI agents that are capable of completing complex tasks autonomously and in long sessions. These agents “burn” millions of tokens quickly, which is why both OpenAI and Anthropic limit (or ban) the use of their subscriptions to use them for example in OpenClaw. Amjad Masad, CEO of Replit, believe that the subscription tap will soon be closed in the face of these costs triggered by agentic AI. “Intelligence” keeps getting cheaper. But in the face of all those realities that seem to threaten the end of AI subscriptions, there is a factor that can contribute to their survival. As they point out in SemiAnalysisthe market laws themselves are confirming that generating tokens is increasingly cheaper. Efficiency improves and the costs of producing tokens decrease, and they could do so at such a pace that in the end access to AI remains profitable for these free buffets. Companies are subsidizing AI for us. All this leads to think that in many cases AI companies are subsidizing the use of their models. They do it by taking advantage of that scheme that only a few really take advantage of the subscription plans. The question is, of course, whether this situation has an expiration date. In Xataka | Anthropic is at the most important moment in its history and has a warning: we must lift the AI ​​accelerator

The problem is not spending a lot of tokens, it’s that most of them are being wasted

by

A year ago, Sam Altman did a striking prediction: as the production of data centers becomes automated, the cost of intelligence (AI) should at some point converge with the cost of electricity.” Or what is the same: access to AI would be very, very cheap. That has not happened by any means, but in addition to spending a lot of tokens, we are wasting them. So much AI for what?. He phenomenon of tokenmaxxing -he rampant token consumption more like fashion than something useful—has begun to set off alarm bells, because companies have realized that they are spending small fortunes for their employees to try to get the most out of AI. AI dismissal. A study by the startup EntelligenceAI affirms that for every dollar invested in AI, only 18 cents end up reaching production. The remaining 82% ends up being invested in correcting errors, rewriting code and executing review processes that do not generate direct value. This is what they call “unproductive spending,” and it is a warning sign because the success of this technology does not depend on us using AI non-stop, but on using it to improve productivity. Uber warns. Andrew Macdonald, COO of Uber, I questioned openly whether this massive spending by companies like yours on AI is really justified when it is not linked to improvements in productivity. The company has been one of those that has decided to cut spending on Anthropic models because the available annual budget had already been “vented” to use them. Investing in tokens ends up being unprofitable: the “useful part” is less than a fifth of what is invested, according to EIntelligence AI. The uncertainty is there. Other experts They warn just the opposite: This is just the beginning of what is to come, so taking action against AI consumption may be counterproductive. The problem is not so much that AI is being used, but rather that it is being wasted: this obsession with consuming tokens caused the CFO at Amazon, for example, to tell his employees “Don’t use AI just for the sake of using it”. The company rewarded those who used AI the most, so many ended up using it for trivial, redundant or useless tasks. Use AI appropriately. Matan Gringberg, CEO of the AI ​​startup Factory, told in WSJ how a manager at a major financial institution had told him that his employees were spending hundreds of thousands of dollars a month on tokens. The problem was that some were using the most powerful models to answer simple questions or just to chat. The message here is clear: these models must be used appropriately to avoid wasting them: “If your daughter needs private algebra classes, you can probably find someone cheaper than Albert Einstein to give them to her,” he concluded. We are consuming tokens beyond our means. At the Google I/O event Sundar Pichai, CEO of Alphabet, explained that the company currently processes more than 3.2 trillion tokens per month, seven times more than a year ago. Faced with this demand, both it and other companies are “punishing” the trivial use of AI models. AI agents consume tokens like there’s no tomorrow. What has also happened is that the arrival and popularization of agentic programming tools, such as Claude Code, Codex or Antigravity, causes many more tokens to be consumed because with them it is possible to automate the execution of programming tasks (or other areas) on a continuous basis. The AI ​​model prepares a plan, executes it, and at each step thinks and evaluates its responses before continuing with the plan. This process is intensive in token consumption, and is the main reason why token consumption has skyrocketed. Flat rates, nothing. Monthly plans like ChatGPT Plus or Claude Pro offered leeway for developers to consume huge amounts of tokens with hardly any limitations. However, both OpenAI and Anthropic and other companies have begun to change their strategies, limiting the cases in which these flat rates can be used so that users cannot abuse them. If they want to consume more they can, but always through a pay-per-use philosophy: the more you use, the more you pay for something that at least helps users be aware that they cannot use super-powerful models for useless conversations with their chatbots. Image | Xataka with Magnific In Xataka | If the question is whether using ChatGPT or Claude in English is more efficient and saves tokens, the answer is: yes

Anthropic does not offer its services in China. So China has invented a black market for Claude tokens

by

Claude has become in the most desired model by the most demanding developers and engineers, but it is not available in mainland China for regulatory and safety reasons. The demand there remains notable, and to satisfy it, an underground token economy has emerged that allows local developers to access models such as Claude Opus 4.7, avoiding all the measures imposed by the blockade. No paying with Alipay. One of the measures that Anthropic imposes to prevent the use of its models in China is to only accept international credit cards such as Visa or Mastercard. Their payment gateways reject local payment methods like Alipay or Wechat Pay, giving Chinese users a first and important hurdle. One that they have already overcome. Virtual cards. What they are doing in China to overcome this problem is using virtual credit cards (VCC) like DuPay or WildCard. With these services it is possible to obtain Hong Kong or US credit cards financed with cryptocurrencies or through local transfers. This makes it possible to deceive the billing systems of Anthropic and other companies that offer banned services to Chinese users. SMS verifications They are also solved through “SMS farms” that also avoid this problem and even others such as identity verification that also have implemented in Anthropic. The “Transfer Stations” arrive (中转站). Another problem is that even overcoming that first barrier, latency and micro-cuts mean that the use of Claude in China is affected by continuous connection problems. To avoid them, so-called “Transfer Stations” have emerged, which are nothing more than servers that act as a bridge between foreign servers and Chinese users. These gateways receive requests from China and forward them to Anthropic servers as if they were coming from an authorized location. The latencies are also relatively low, which means that for Chinese users the experience is basically identical to that of a user in the US or Spain, for example. These stations are publicly known and do not only appear in listings on GitHub: there is a ranking with the best. Claude is almost free in China. The surprising thing about these methods is that they don’t just give Claude access in China: they do with ridiculous prices which can be 10 and even 5% of (growing) original price of the service thanks to those transfer stations. The question, of course, is how it is possible to access Claude at those prices. The almond tree trick. Thanks to the transfer stations, developers can access Claude at a price of 1 yuan for every dollar of tokens, or in other words, up to a 90% reduction in the official price. It is something that is discussed publicly and that makes it clear that several methods are used to achieve this: Mass purchase of capacity, Use of accounts created with stolen or fraudulent cards, Use of promotional credits, and A simple hook: providers lose money with Claude, but they manage to attract developers to whom they then sell more profitable local models like DeepSek. Am I really using Claude? One of the growing risks in the cheap token market is direct fraud. Some Chinese resellers have been caught red-handed offering what they call the “Claude API” when in reality what they were providing were much cheaper and mediocre models. For a user to detect this type of deception it’s very difficult unless you are working with complex tasks or you have already used models and know more or less what to expect from them. For victims, the effect is clear: they believe they are paying for the intelligence of Opus 4.7 when in reality they are receiving answers from a low-end AI model. Goodbye to privacy. When a user purchases tokens at one of these transfer stations, they completely give up the confidentiality of their data. All queries and responses end up passing through the intermediary’s servers, which can and apparently does use them to sell them to AI companies that use them to post-train their models. So everything they do and say when using these models is filtered and used as training data without the user knowing. A double business. For these providers, this business of reselling conversations is especially interesting in the face of the famous “distillations” of US models that take advantage of this data to “copy” the capabilities of those models and apply them to Chinese models. Anthropic can read us, but (theoretically) it doesn’t. It is true that the conversations we have with Claude (from Spain, for example) are also stored on Anthropic’s servers, but the company makes it clear in your privacy policy that does not use that data. In fact, we can even explicitly prohibit the company from using them in the privacy settings of Claude’s account. The game of cat and mouse. At Anthropic they know very well what is happening and they are trying to prevent it. For example, they have begun to intensively block IP ranges associated with VPN services or data centers known to be used in these transfer stations. Even so, Chinese providers usually respond with an “elastic” architecture that allows IPs of domestic residences to rotate, making the traffic appear completely normal. Image | Xataka with Magnific In Xataka | There is a thing called “Ornn price index”, it is out of control and it is bad news for everyone

If the question is whether using ChatGPT or Claude in English is more efficient and saves tokens, the answer is: yes

by

You may not have stopped to think about it, but there is a striking reality in the world of chatbots: It is more expensive to speak in Spanish with AI than to do so in English. The reason is simple: AI does not understand words, it understands tokens. And when you talk to GPT, Gemini, Claude or any other LLM, you talk to him in a language, but to understand you he first “translates” what you are telling him and converts it into tokens. And the problem is precisely that: that not all languages ​​”cost” the same in terms of tokens. There is a very simple example that we can analyze thanks to tools like ClaudeTokenizer: the word “developer”, which in English is “developer” costs few or many tokens depending on the language in which we write it and also (importantly) the version of the AI ​​model used. In the image it is clearly seen, but just in case, we summarize: For ChatGPT (GPT-4o and GPT-5) the word “developer” has three tokens (des-developer-ador), but the word “developer” only costs one. For Claude (Opus 4.7) the word “developer” costs no less than 9 tokens (2 in Opus 4.6), but “developer” costs “only” 6 (1 in Opus 4.6). What is happening here? Well then each language model uses its own “tokenizer”your “translator” from a conventional language to the token language that the language model understands. And those tokenizers favor precisely the languages ​​in which these models are created. This is how AI understands how we speak. Each word is divided into tokens, and English is understood much better. “developer” only costs one token in GPT-5, but “developer” breaks down into three. Bad news for Spanish speakers. In fact, English has become the official language of artificial intelligence, whether we want it or not. The reason is not cultural, but architectural: 95% of the training data of the frontier models (GPT-5, Gemini 3.1, Claude Opus/Sonnet 4.7…) are in that language. That makes the rest of the languages ​​”foreign languages”, and that makes it necessary to pay extra when using them, an almost invisible toll on every interaction. In practical terms, what happens when we use Spanish to talk to an AI model is simple: we use more tokens, and therefore using Spanish is simply more expensive than using English when working with a large language model. If you want to save tokens, better use English The question, of course, is how much more does it cost us to speak in Spanish than in English with ChatGPT (GPT 5.x) or with Claude Opus 4.7? It is difficult to say because each word and each phrase is a world, but the truth is that English is almost always the most “economical”. We have used one of the first sentences in this article to compare that token consumption, and by translating the sentence into different languages ​​and querying that token consumption for different models, the data is clear. It is important to highlight that these results are not conclusive, but they do make the trend clear: English is the most efficient language in terms of token consumption, but be careful, because Spanish is not that bad, and is usually the second most efficient. It is even more efficient than English in Gemini, at least according to the tool consulted. But on average, it is normal that there is a significant extra cost when using different models. A conversation with Claude Opus 4.7 is already “expensive” because it is one of the most expensive models currently, but in Spanish it is almost 30% more expensive, not to mention in Arabic, 76.3% more expensive. In fact, according to this example, the difference between Claude or GPT-4o in terms of efficiency is clear: OpenAI tokenizer is “cheaper”and although there may be differences with GPT-5.x, what seems clear is that Anthropic has preferred to “spend more” to obtain better results (or that is the objective). Gemini is even more thrifty according to these tests, and that may also have a lot to do with the quality of the answers, although that question is for another topic. We have used one of the paragraphs of this article in Spanish and translated it with Deepl into English, Arabic, Norwegian, French and Chinese to find out how many tokens the phrase “cost” in each language. English is undoubtedly the most efficient Tokenizers advance and evolve. Sometimes they do it to save us tokens, as happened with the GPT-4o tokenizer: at that time OpenAI explained how that tool used 1.1 times fewer tokens when speaking to her in Spanish but up to 2.9 times fewer in Hindi or 3.5 times fewer in Telugu. With Claude Opus 4.7, just the opposite has happened: the tokenizer has been redesigned and consumes more tokens (up to 1.35 times more, they admitted) with the aim of better processing and understanding the text. Your chatbot thinks (and programs) in English Here we must also highlight something important: although we can talk to our favorite chatbot in any language and it will answer us in that language (unless we ask otherwise), AI models “think in English”. That is to say: when you talk to them what they do is translate what you tell them and then reason in English and finally they translate their response into the language in which you were speaking to them. This consumes additional reasoning tokens, but also has some impact on latency (how long it takes to start thinking or answer the model). In complex tasks, this can clearly influence response times for the simple reason that the AI ​​model does not stop translating from “its official language” (English) to our language. This preference for English is also noticeable in the benchmarks: in the Humanity’s Last Examin which the models are asked all kinds of general knowledge questions with several options to answer, it is reasonable to think that the models They answer better in English because that exam is designed in that language. … Read more

Within Meta there is a race to see which employee consumes the most AI tokens. It’s the ‘Tokenmaxxing’ of Silicon Valley

by

There is a battle within Meta: see who spends the most AI tokens. This is the basic unit that AI uses to understand the language with which we order actions. It is like the “bridge” between our words and the numbers that the machine can process and, therefore, when ChatGPT either Google They present a model, they brag about the millions of tokens they can process. But tokens are also becoming a ‘spending’ unit in AI companies. Silicon Valleyso much so that they may be generating a toxic work culture. And Meta is an example of a company where employees compete to see how many tokens they can consume to become a Token Legend. Tokenmaxxing. It is not the first time that we talked about this. A few days ago, Jensen Huang -CEO of NVIDIA and one of the main instigators of this phenomenon- commented that he would be worried if an engineer who earns $500,000 did not spend at least $250,000 a year on tokens. Because tokens cost money and NVIDIA is already considering offering tokens as part of the signing bonuses for its artificial intelligence engineers. Goals. As it could not be otherwise, Meta does not want to miss this party. The company, which changed its name when the metaverse was going to be the big thing and, after the swerveis defined as a “native AI company”, is one of those that promotes its artificial intelligence engineers to keep a count of the tokens spent during their day. There is no official data, but there are reports revealed to media such as Business Insider and The Information which point out that some of these teams have very specific objectives related to the use of tokens. For example, the company expects 65% of its engineers to write more than 75% of code using AI tools by the middle of this year. The Scalable Machine Learning division has another objective, and so on in each of the code-related departments within Meta. Legend Token. In The Information, they directly point out that there is an internal classification table created by the employees themselves to gamify the work. It shows the 250 most intensive AI users in their tasks with an easy premise: the more tokens you spend, the more you climb in the ranking. The winner of this particular competition takes the title of ‘Token Legend’, or ‘Legend of Tokens’. It is turning an expectation into a kind of internal sport. The first paragraph of this article converted to tokens crazy spending. If we put the first paragraph of 542 words in the tool ‘tokenizer‘ from OpenAI, we see that that simple phrase has already consumed 121 tokens. Well: according to The Information, in the last 30 days the total token panel usage of that internal table was more than 60 billion (of ours) tokens And even if they want to dress it for sports and competition, it is still obligatory. In late 2025, Meta launched the ‘Level Up’ program where employees who complete the most tasks using AI earn badges. And more important than this: it made the use of AI a central criterion in its employee performance evaluations. This, obviously, sets salary and promotion objectives. Doubts. But of course, beyond paying to work, there are other underlying issues. One of the criticisms of this tokenmaxxing system is that AI companies like Meta or NVIDIA encourage spending more on tokens because, in this way, their own employees become consumers of the product they are creating. An easy example that software engineering analyst Gergely Orosz exposed which is as if Tim Cook, CEO of Apple, said that if one of his employees who earns $500,000 a year did not spend $50,000 on purchases in the App Store, he would be worried. Orosz continuous stating that productivity should not be measured in tokens spent, but in the results obtained. Industry issue. In any case, Meta and NVIDIA are not the only ones that measure their employees by their consumption of AI at work. It is something that is soaking in other AI majors, turning the tokens into an extra work benefit incorporated into the engineers’ remuneration wheel along with the base salary, performance bonuses and shares. HE esteem that an OpenAI engineer can process 210 billion tokens in a week and there are Claude Code engineers who accumulate more than $150,000 in tokens in one month. Basically it is merging part of your salary into the company that pays you. And… have they said anything from Meta? Yes, it’s not about volume, but about quality, pointing that performance rewards are based on the impact of the work and not the raw use of AI. Image | ‘Wolf of Wall Street’, Meta Logo. Edited In Xataka | Google Earth shows the world. The Spanish Xoople wants AI to understand it

Jensen Huang believes he has found the perfect new bonus for software engineers. Not Stocks: AI Tokens

by

The CEO of Nvidia has been converting the AI tokens at the center of all their public conversations. Jensen Huang’s latest idea links these tokens to the efficiency of engineers and how the best engineers in the world are recruited: in addition to a generous salary, offer them an amount equivalent to half their annual salary in AI tokens as part of the hiring package.​ Huang verbalized his proposal during the inaugural speech of the GTC 2026 conferenceNVIDIA’s largest annual event for developers. In a later interview, the Nvidia CEO detailed that engineers would earn “a few hundred thousand dollars a year as a base salary,” and the intention would be to give them “probably half of that, also, in tokens, so they can multiply your productivity times ten”.​ What Huang proposes already has a name: Tokenmaxxing. In one podcast appearance ‘All-InHuang said he would be on “high alert” if an engineer earning $500,000 didn’t spend at least $250,000 a year on tokens. “If that person said (that he has used tokens worth) $5,000, I would go completely crazy,” Huang stated. When asked if NVIDIA planned to spend $2 billion on tokens for its engineering team, as proposed, Huang responded: “We’re trying.”​ As and how they counted in The New York Timesthat has generated a phenomenon called “Tokenmaxxing“, with which engineers brag about the number of tokens they consume to try to improve the perception of their productivity: the more tokens you consume, the more productive you are. Tokens as bonuses are a trend in Silicon Valley. The CEO of NVIDIA is not the only one who thinks this way, and the use of tokens as an extra work benefit it’s soaking among investors in the sector. Tomasz Tunguz of Theory Ventures stated to Business Insider that “companies are incorporating AI inference as a fourth component of engineer compensation: salary, bonus, stock and tokens.” The interest of whoever sells the chips. The NVIDIA CEO encouraging everyone to spend more on tokens is not disinterested advice. Gergely Orosz, analyst at software engineeringhe pointed it out bluntly in a publication from AND he added an analogy that sums it up accurately: “It’s almost as if the CEO of Apple said, ‘If someone who makes $500,000 a year doesn’t spend at least $50,000 a year on in-app purchases on iOS, I’d be deeply alarmed.’ And yes, you would be, because that would reduce the revenue you generate.” Huang is the head of the company that manufactures the chips for AI on which most of the world’s artificial intelligence runs. Huang himself made it clear to his investors: “Without computing, there is no way to generate tokens. Without tokens, there is no way to grow revenue,” he declared, describing his data centers as “token factories” whose demand will only grow as AI agents proliferate.​ Do not confuse value with price. However, Huang has incurred a bias when arguing his idea: confusing value with price. Orosz formulated it clearly in a message in X : “The advice that engineers should use tools that make them more productive IS correct… except that the cost of tools should NOT be what we focus on. Some of the most useful tools are very cheap. Of course, vendors will focus on selling the most expensive and most profitable tools.” Productivity is not measured in tokens spent, but in results achieved. The right question for companies should not be whether their employees use more AI, but whether increased use of AI is rewarded. with greater productivity. In Xataka | Customers demand that a human solve their problem. The surprising thing is that if humans serve them they think they are an AI Image | NVIDIA, Unsplash (Arif Riyanto)

China makes tokens cheaper than anyone else

by

Last month, Chinese AI models surpassed American ones in use on OpenRouter, an AI platform that allows you to detect interesting trends. And this in fact was just confirmed this month and has accelerated, because what we are seeing is that despite the obstacles that the US has tried to put in place to prevent China from competing in this market, the Asian giant has found a key tactic to do it: the so-called “token export”. Useless tariffs. The US government the era of globalization burst with the trade war with China and more recently with its aggressive tariff policy. That has had a clear effect on Chinese exportsbut the Asian country has found a way to avoid tariffs: with AI. Its artificial intelligence models can be used around the world without being affected by tariffs. Although they are inferior in performance and quality, they are much cheaper to use, so China is managing to convince the world with its old recipe: if the product or service is good enough and is also cheap, it wins. On the OpenRouter platform we have been seeing for two months how Chinese models are used more than those from the US for a simple reason: they are much cheaper and perform reasonably well. Token export. When we use energy we consume kilowatts. When we use AI we consume tokens from AI models. And that is where China is winning with the phenomenon called “token export”because the tokens of their AI models are extremely competitive and for many tasks those models are good enough. Minimax M2.5, Step 3.5 Flash or DeepSeek V3.2 clearly outperformed Gemini 3 Flash Preview, Claude Sonnet 4.6 or Claude Opus 4.6 in use in the last two months on the OpenRouter platform, for example. Developers from all over the world take advantage of these models and do so without being affected by tariffs: tokens do not pay those fees that, for example, apply to mobile phones, cars and many other products. Devastating price difference. While an American premium model like Claude Opus 4.6 costs 5 dollars per million entry tokens (Sonnet 4.6 costs 3), Chinese models like the MiniMax M2.5 cost as little as $0.25, 20 times less, and the Step 3.5 Flash, also very popular, costs just $0.10, 50 times less. AI agents ask for cheap models. That price gap is especially relevant now that AI agents —and especially, OpenClaw— begin to demonstrate their ability. These types of systems are capable of completing tasks for us and even controlling the machines to which we give them access, but to achieve this they use a huge amount of tokens. Using the best models guarantees better results, but it is also very expensive, but in many “simple” tasks, very cheap models like the Chinese ones They can perfectly solve the problem. The subscription trap. In recent weeks, the rise of OpenClaw and similar platforms has provoked a curious response from companies like Anthropic or Google. To these companies they don’t like it that subscription plans for your AI models be used for these types of AI agents because they argue that those plans are abused, and there are certain restrictions to that type of uses. This has caused many users to opt for AI models from Chinese companies, which are precisely positioning themselves as the cheap and trouble-free alternative to be able to take advantage of these agents. Why Chinese tokens are so cheap. There are several factors that favor the low cost of AI models in China. The first of them, cheap energy: the costs of industrial energy They are 40% smaller than in the United States. The second, its efficient architecture: as DeepSeek demonstrated, it is possible to achieve great results with techniques like Mixture of Experts (MoE). With it, the model is divided into multiple “experts” and only activates those that are necessary according to the request. The irony of tariffs. Curiously, the US restrictions on the export of advanced chips may have ended up being the great catalyst for this situation. By not having access to the most advanced NVIDIA chips, Chinese companies were forced to perfect the efficiency of their models to the maximum, and that has now caused be more competitive in the AI ​​inference market (that of the use of models in practice), which is where this new economic battle is being fought. Challenges. Although the “token export” is currently profitable for China, it faces significant challenges. The data sovereignty is one of them: for a company or a government, sending sensitive data to data centers in China is a red line. There is also the problem of latency: the response of China’s AI models is affected by the enormous distances that these data packets have to travel. It remains to be seen if Washington ends up applying some kind of measure to also restrict the use of AI models from Chinese companies, although that seems more complicated. In Xataka | NVIDIA already has a monopoly on AI hardware: now it wants to conquer software through agents

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.