SemiAnalysis: Storage space is still two to three times larger, CPO mass production postponed until the end of 2028 According to BlockBeats, on July 10th, Dylan Patel, founder of SemiAnalysis, responded systematically to the controversy over AI investment returns in his latest podcast. He revealed that Anthropic achieved a positive free cash flow in the second quarter of this year, with annual recurring revenue exceeding $50 billion and a gross profit margin of over 70%. OpenAI made a profit in April, a profit in May, and it looks the same in June. Its revenue grew rapidly with the increase in Codex adoption rate. Patel cites SemiAnalysis's own data as evidence - the company's 90 person team's annualized AI spending has skyrocketed from less than $100000 in November last year to $11 million currently, with AI costs exceeding one-third of labor costs. Every time the model is upgraded from 4.6 Opus to 4.7 Opus, the expenditure will first decrease for a week and then skyrocket again - because things that couldn't be done before can now be done. It believes that Anthropic's core advantages are higher token efficiency and lower overall user costs. On the hardware level, Patel's judgment on memory is the most firm: 'This is not a short-term shortage, but a structural shortage that will last for many years.' The inference model drives the explosive growth of KV cache, while memory capacity only grows by 20% to 30% annually, and the supply-demand gap will continue to widen. The memory has increased fourfold, but there is still 2 to 3 times more upstream space; Consumer electronics will be the first to come under pressure, with mid to low end mobile phone manufacturers experiencing a 40% decline in shipments. At the same time, a clear warning is given to the narrative of CPUs - current demand contains a large number of historical fill in effects, and once the outstanding debts are filled in, in absolute terms, CPUs are far inferior to AI acceleration chips. Memory and AI acceleration chips are the main focus, and CPU is a reassessment after being underestimated, but it will not grow indefinitely at a rate faster than AI chips. Regarding the highly sought after CPO in the market, its clear judgment is that true large-scale production will not be achieved until the end of 2028 to 2029- manufacturing yield, chip design, and supply chain maturity have not yet met the standards. NVIDIA Rubin and its subsequent architecture Feynman will still use an all copper solution, which unexpectedly extends the dividend period of copper cable connectors. In terms of electricity, it is predicted that 20 gigawatts of new data center electricity will be added this year, 30 gigawatts next year, and 50 gigawatts the following year, half of which will come from self built "meter power sources" by enterprises. Combined cycle gas turbines are currently the mainstream, but the comprehensive cost of solar energy and energy storage is expected to be lower than that of gas-fired power generation within two years.