I just finished watching the latest All In Podcast and talked about Anthropic's 2 trillion yuan IPO, Zuckerberg's AI manifesto, Nvidia's 500 billion yuan financialization of computing power, and Grok's comeback. Here's a summary Anthropic is about to sprint for a $2 trillion IPO According to the Financial Times, Anthropic plans to advance its IPO at a valuation of $2 trillion around October, with an expected annualized operating revenue (ARR) of $100 billion to $120 billion by the end of the year, achieving nearly tenfold exponential expansion for three consecutive years. According to this ARR, the market to sales ratio is only about 16 to 20 times, which is cheaper than some overvalued listed technology companies before. Its moat mainly comes from the programming vertical track. I bet on programming models early on and seized the first mover advantage by observing high-intensity calls from ecosystems like cursor. It is rumored that I am still in talks to acquire DART, an AI energy efficiency optimization startup, for $6 billion. However, OpenAI is also accelerating its catch-up, with GPT-5.6 fully leveraging its coding capabilities, and the month on month growth rate has rebounded to over 20%. The guests also discussed whether Anthropic could move from billions of ARR to trillions of ARR. Although the market potential for enterprise knowledge work is as high as tens of trillions of dollars, the shortage of computing chips, power infrastructure, and grid access restrictions will be the ceiling in reality. The general consensus is that ARR can reach 400 billion to 500 billion dollars by the end of 2027, which is a relatively pragmatic expectation. 2. Zuckerberg's AI Manifesto Zuckerberg has released a 6500 word long manifesto titled 'The Future is for Everyone', advocating for open source AI, empowering every individual with super intelligence, and decentralized checks and balances. Several speakers strongly criticized the "AI doomsday/savior mentality" of Anthropic and Effective Altruism (EA), believing that likening AI to nuclear weapons and attempting to create industry cartels through government approval mechanisms such as the AI version of FAA or FDA would not only erase the leading advantage of enterprises, but also bind themselves in the geopolitical competition of AI between China and the United States. They also talked about the economics of coexistence between open source and closed source models. Open source models such as Llama and China's GLM are 90% cheaper than closed source frontier models and undertake approximately 80% of conventional inference tuning. And closed source frontier models such as Claude and GPT play the role of highly intelligent "central scheduling brains", occupying 65% to 85% of the highest economic value allocation, forming a relatively healthy hierarchical ecology. Nvidia has launched a $500 billion financialization campaign for computing power Huang Renxun collaborates with Blackstone, Goldman Sachs KKR、 Apollo, BlackRock, and other top Wall Street asset managers have established a $500 billion independent funding platform for computing power. The operation method is somewhat similar to aircraft leasing or equipment mortgage loans. Downstream customers borrow money to purchase GPU clusters and earn money from renting out computing power to repay their loans. Nvidia acts as a matchmaker and provides residual value guarantees, while enjoying excess profits beyond the guaranteed returns. This model essentially transforms Nvidia from a simple hardware seller to a role similar to that of a central bank in the AI field, using Wall Street's private equity credit funds to break through the financing bottleneck of huge capital expenditures and directly support trillion level computing power deployment. Of course, some people also raised the risk. The core concern is whether there will be an absolute surplus of computing power in the future (a bit like the dark optical fiber in the Internet foam period). However, the current two bottlenecks of power and approval objectively play a role of speed limiter to prevent the instantaneous surplus of computing power. The energy bottleneck of data centers: electricity The construction of data centers is constrained by physical limitations such as heavy gas turbine production capacity and power grid transformer delivery cycles. Capital is rushing to build computing power nodes through dismantling and renovating old aircraft engines and building their own natural gas power generation. The program also refuted some media exaggeration about data centers depleting civilian water resources and significantly increasing electricity prices, pointing out that data centers that operate in compliance, build their own power plants, and feed back electricity to the grid can significantly improve the tax and infrastructure levels of remote towns. The "energy audit" implemented by Texas Governor Abbott aims to standardize grid connection safety, not to halt computing power construction. Grok counterattacks, PE funds also start to refocus on SaaS The evaluation shows that the capability of Grok 4.6 has approached the forefront, and the inference cost has significantly decreased, breaking the previous market expectation that Anthropic and OpenAI would have a long-term duopoly. Musk has played a clever strategy on SpaceX's computing power cluster, using large models for personal use and subleasing them at high prices (including the previously signed agreement with Anthropic for computing power sublease and a 90 day adjustable clause), which can be considered a two-way option gameplay. In addition, private equity giant Silver Lake is rumored to be acquiring Workday at a huge cost, causing its stock price to skyrocket as soon as the news came out. This also reflects a new trend. After a long period of valuation decline, open source big models combined with AI brought cost reduction and efficiency improvement (layoffs, AI priority operation model) are providing space for software companies with mature cash flow to restructure and privatize arbitrage. PE funds are starting to refocus on the enterprise software track. Finally, a labor dispute at Amazon was also mentioned. New York and New Jersey sued Amazon for using third-party outsourced fleets (DSPs) to evade labor benefits responsibilities, and guests debated the efficiency of free contracts and labor compliance costs.