During this National Day holiday, there has been another significant change in the field of AI.

"One day in the AI world is like ten years in the human world." After the National Day holiday in 2025, we will have an increasingly profound perception of the concept of the "super-accelerated society." As we all know, each industrial revolution brings exponential social changes, and AI is no exception. Although we were mentally prepared, we didn't expect that the impact of AI would be so rapid, like time being compressed and the social rhythm suddenly accelerating.

During this National Day holiday, from the increased investment in data center infrastructure to the continuation of the valuation myth of OpenAI, it is no longer a simple technological leap but a comprehensive game of capital, computing power, and geopolitical competition. A social and economic version of the "scaling law" in the AI era seems to be gradually taking shape.

What's even more chilling is that while humans are "exiled" during the holiday, AI is quietly "charging" - literally in terms of power consumption. It reminds us that, with humans as the reference frame, the essence of AI progress is to accelerate inequality. Regarding the future, it's "unimaginable."

Below, we will analyze several landmark events during the holiday one by one. Each event is like a chess piece, reconstructing the global chess game of AI and should not be overlooked.

01 Computing Power Monopoly 2.0: The Chip Alliance between OpenAI and AMD

After the "century handshake" between Nvidia and Intel, OpenAI is not willing to lag behind and starts to regard AMD as its backup option.

On October 6, OpenAI announced at its annual developer conference (DevDay) that it had reached a multi-year agreement with AMD. The latter will provide AI chips for OpenAI and grant OpenAI an option to purchase approximately 10% of AMD's shares.

This deal is worth tens of billions of dollars. It not only strengthens OpenAI's leading position in the field of generative AI but also highlights the industry's extreme thirst for computing resources.

Logically, this stems from the exponential demand for AI model training and inference. It is estimated that the training of a GPT-5 level model uses tens of thousands of GPUs, but it obviously hasn't reached the expected milestone. After GPT-5, more crazy exploration of computing power is needed to discover new limits. AMD's involvement aims to break Nvidia's monopoly and provide the more cost-effective MI300X series of chips.

Professional analysis shows that this move will accelerate OpenAI's "Stargate" super data center project, with the goal of deploying a computing power corresponding to 6 gigawatts of electricity demand by 2030 - equivalent to the annual electricity consumption of a medium-sized city in the United States.

The harsh reality is that this exacerbates the energy crisis: Data centers already account for 2 - 3% of global power consumption, and the expansion of AI may push it up to 8%. The paradox is that while these tech giants claim that AI will "save the world," they are building a power - consuming monster, as if saying, "Green AI? Let me get enough to eat first."

This phenomenon is touching on the core of technological ethics: Computing power is a double - edged sword. It can drive medical breakthroughs (such as personalized drug design), but it may also magnify environmental injustice - developed countries seize resources while developing countries can only watch.

02 Bubble or Value: OpenAI's Valuation Soars to $500 Billion

Also on October 6, OpenAI completed an employee share sale transaction, surpassing SpaceX with a valuation of $500 billion to become the world's largest startup.

This milestone is due to its revenue in the first half of the year reaching $4.3 billion (a 16% year - on - year increase), but it also exposes high - cost spending: The computing expenditure in the first half of the year was as high as $2.5 billion.

Logically, this reflects investors' enthusiasm for generative AI: ChatGPT has a user base of over 200 million, and its potential commercialization paths (such as enterprise integration) have made it a favorite on Wall Street.

Professionally speaking, this valuation depends on model iteration. For example, the newly launched Agent Kit toolkit allows developers to build, test, and deploy AI agents. In addition, ChatGPT has added an instant checkout function and third - party application integration (such as Spotify, Zillow), transforming the chatbot into an all - around platform.

Of course, it must be pointed out that if innovation slows down, the bubble may burst at any time - the burn rate has reached 58% of the revenue.

03 Infrastructure Investment Craze: The $14 Billion Agreement between Meta and CoreWeave

Just before the National Day, Meta signed a $14.2 billion computing agreement (until 2031) with CoreWeave, involving NVIDIA GB300 chips for AI infrastructure expansion. At the same time, Oracle announced that it would purchase $40 billion worth of Nvidia chips for OpenAI's US data centers.

These transactions mark the collective bet of big tech companies on AI hardware: It is estimated that in 2025, the capital expenditures of Amazon, Meta, Microsoft, and Google will reach $320 billion, mainly for data centers.

From a professional perspective, this reflects the transformation from experimentation to industrialization: CoreWeave, as a GPU cloud provider, focuses on efficient inference. The agreement includes outsourcing 10.5 gigawatts of power to partners such as Brookfield. The underlying logic is that the scale effect dominates everything - whoever controls computing power dominates AI.

But as the "old saying goes," the energy bottleneck is emerging: AI data centers may push up the global power demand by 40%.

These giants are like hungry dinosaurs, scrambling for the last piece of "silicon meat" while ignoring the climate "meteorite." It echoes the paradox of capitalism: Innovation drives growth but at the cost of sustainability. If we don't switch to green energy, this "change of the sky" will spread from technology to an ecological crisis.

04 Model Iteration Wave: Anthropic Claude 4.5 and DeepSeek V3.2

Within a week, Anthropic released Claude 4.5 (Sonnet 4.5), emphasizing long - context reasoning and tool integration. At the same time, DeepSeek launched V3.2, claiming to reduce the long - context computing cost by more than 50%.

These iterations reflect the evolution of AI from general models to more efficient and specialized ones: Claude 4.5 won in a coding competition, and DeepSeek is optimized for enterprise - level applications - it's all about cost.

According to the product guide, the "Agent SDK" customization function of Claude 4.5 allows users to build personalized agents to improve productivity. The efficiency improvement of DeepSeek comes from architecture optimization, which is suitable for fields such as biological computing.

To some extent, this reflects the overall trend of the current AI competition between China and the United States: The United States leads in the number of leading models, but China is constantly catching up in terms of quality.

Of course, although the competition is fierce, the process often includes a lot of "funny realities." As some pungent third - party comments say: These models are like "top students" in academia, copying each other but pretending to be original.

05 New Milestone in AI Video: Sora 2, Moving from Research Prototype to Consumer - Grade Product

Even earlier, on September 30, OpenAI officially released Sora 2, a major upgrade of its text - to - video generation model, marking the transformation of AI in the field of video creation from a research prototype to a consumer - grade product.

Sora 2 improves the physical authenticity and controllability of videos. It also integrates synchronous audio generation for the first time and is accompanied by the launch of an iOS social app called "Sora," which is similar to TikTok and allows users to generate, share, and remix videos.

OpenAI did not disclose the complete training details but emphasized enhancing the training data through re - captioning and learning from research such as Step - Video - T2V and DiTraj to improve temporal consistency and motion trajectories. Early user feedback shows that Sora 2 has a higher accuracy rate in the first - time generation, although it is slower.

These improvements enable Sora 2 to surpass competitors such as Google Veo 3 or Runway Gen - 2, with more advantages in multi - shot narrative and world state maintenance.

Some believe that Sora 2 marks AI's step towards a general world simulator, which may empower robots and more complex simulations.

OpenAI plans to further accelerate the commercialization process of Sora 2 through APIs and global promotion, reshaping social media (TikTok is trembling). However, high computing costs and ethical reviews remain challenges - "deep fakes," privacy infringement, and copyright disputes are always there. Sam Altman also joked about the copyright risks of the Sora 2 video tool at the press conference, saying, "I hope we won't be sued."

In a sense, while we are creating a utopia, we are also building our own Pandora's box.

Above, the capital frenzy, model leaps, and infrastructure expansion are intertwined - the concrete - and - steel process of AI accelerated during our National Day holiday. We can even say that our holiday is just a gap in its acceleration.

This also reminds humans that technology cannot be isolated to the point of being out of control and must dance with humanity. Otherwise, the whole society will change during the next National Day.

This article is written based on publicly available information and is only for information exchange, not constituting any investment advice.

This article is from the WeChat official account "Jinduan Research Institute," author: Mu Yang. It is published by 36Kr with authorization.