This article on finance is worth reading carefully and summarizing a few key points: 1. The infrastructure construction period of AI is longer than that of the Internet. The core is not penetration, but "the multiplication of two indexes": number of users and per capita token consumption: Internet infrastructure has only one dimension: user penetration. Penetrating into the middle of sigmoid, hardware requirements will shift. AI has raised the upper limit by one level: the number of users multiplied by the per capita token usage, both of which are converted into tokens. The penetration rate has exceeded about 50%, and the first slope is starting to slow down; The second paragraph is still early - the median per capita AI spending of enterprises is about $12 per month, and if it reaches the 10% level of white-collar wages (about $1000) in the long run, there are nearly two orders of magnitude in between. Two S-curve relays are rare in history, so this round of semiconductor super cycle will be much longer than the Internet infrastructure construction period. From a personal perspective, the current competition in the market is not about whether there is demand, but whether to measure this demand in a few years or ten years. The growth point after coding is not about changing tracks, but about "broad coding eating up knowledge work". Narrow coding is only the first type of programmer to run through the "observation execution verification" loop. Fin's big judgment is that the next move is still coding, but it has spread from programmers to non programmers using the same infrastructure for non coding tasks (finance, legal, sales, pharmaceuticals, research modeling, Excel/PPT backend, etc.). OpenAI's enterprise side Codex has accounted for the majority of output tokens, with legal/sales departments growing much faster than engineering departments; Anthropic's revenue is about 40% for narrow software companies, and there is already a considerable volume in vertical industries such as finance and insurance. The slowdown in growth from June to July is mainly due to the competition for computing power, anti distillation measures, and market share, rather than a peak in demand. From an investment perspective, the second growth engine does not need to imagine a new "AI for Science super scenario". Let's first look at the speed at which Co work/computer use migrates workflows. Previously, it happened to be here at https://(x.com)/qinbufark/status/2091420164109836623? S=46&t=k6rimWSEbo2D2TXolYcM-A also mentioned: "Vibe Coding is becoming the paving technology for non coding workflows. In the future, many non-technical employees will not think they are 'writing code', but will generate financial reconciliation tools, sales dashboards, inventory warning systems, internal approval pages, customer data cleaning program automatic reports, contract inspection processes, and department specific agents through natural language. After the Coding Agent reduced the cost of customized software, a large number of non coding processes became economically viable for software and agent based processes for the first time. ” 3. The unit economy of O+A can be calculated, and Anthropic has approximately 2GW of effective computing power corresponding to approximately 62B ARR. 1GW is constructed at 50B and amortized over six years, with an annual cost of approximately 10B. The estimated revenue is approximately 30B/GW, and the multiples and ROIC are not bad; Pure reasoning is better. Approximately 6-7 GW by mid year, 11 GW by the end of the year, and 20-24 GW planned by the end of 2027- as long as the ARR is roughly synchronized (with a monthly increase of about 10% in the second half of the year and another doubling by 2027), the account can be balanced. I started guessing after 2028. Borrowing and FCF turning negative is indeed gambling, but in the face of this ROIC, the risk of short selling is greater than the risk of FOMO. This is the same yardstick as watching CSP financial reports and whether AI revenue continues to exceed capital expenditures. Overbuild is almost certain to occur, but it will not be visible for at least two years; The real change in profit distribution is that after the construction, the Inference is too profitable: closed source gross profit is 70% -85%, open source token factory can also reach more than 60%, 1GW rent is moving towards 20B, and short-term rental cluster gross profit is also high. Those with channels and technology will flock to Neocloud, and there is still a 1-2 year delay in construction, so it will inevitably be overbuilt in the end. But with current demand and planning, any paradigm shift (reasoning → agent, or application layer innovation) will push the time of over construction back. Once over construction occurs, it is more favorable for the top model factories and application layers - the gross profit of data centers is suppressed, and profits are concentrated in upstream models. SpaceX has been named as a latecomer with strong execution, cost, and planning capabilities, and is also an important contributor to future development. From a personal perspective, the market is now afraid of "having built too many", while people in the industry chain are afraid of "not enough electricity and cards" - the opposite feeling is the typical structure in the middle of the cycle. 5. Open source will generate profits, but the impact in the past year has been smaller than the market narrative; The closed source uses Qualcomm style defense, and the difference in open source and closed source capabilities is maintained at around six months; Distillation is more about reducing costs, not smoothing out barriers. The three barriers of computing power, data, and self iteration are narrowing, but they have not disappeared yet. Using ARR instead of OpenRouter sampling: North American open source accounts for about 4-5B, while the control closed source accounts for about 120B, accounting for approximately 3%. The North American open source token factory grew approximately three times in the first half of the year, not significantly faster than closed source; Expansion is still limited by computing power and financing scale. Closed source defense is very similar to Qualcomm's collaboration with MediaTek in the past: flagship defense of brand and upper limit (Sol), active coverage of low-priced products (Luna price reduction), and not leaving profit pools for low-cost competitors. Value tier does not automatically eliminate the flag demand. Open source changes profit distribution, not the direction of the entire AI infrastructure. Is it worth re evaluating the value of CSP in mid July? Long article https://(x.com)/qinbafrank/status/2076839550119444702? S=46&t=k6rimWSEbo2D2TXolYcM-A has also been discussed: the open source model brings about a huge change in the cost structure of CSP, greatly improving its gross profit margin. At the same time, it will also drive more enterprises to accelerate the process of AI adoption, resulting in higher overall token consumption. 6. What really panics the market is the mismatch of "highly visible expenditure and highly invisible demand". Capex numbers are hard, and token demand, per capita usage, non coding scale, and whether ROIC can cover the cost of new capital all rely on subsequent financial reports to prove. Therefore, the foam debate is more intense than in the Internet era, and financial risks are also leveraged. The base case of Fin is that even without a new paradigm, the infrastructure of 2027 – 2028 is not a foam. People nearby see shortages, while those far away see surpluses - the debate is often not about whether the numbers are correct, but about what time units should be used to look at this curve.