BlockBeats News, July 10th - SemiAnalysis founder Dylan Patel responded systematically to the AI investment return rate controversy in the latest podcast. He revealed that Anthropic achieved positive free cash flow in the second quarter of this year, with an annualized recurring revenue surpassing $50 billion and a gross margin exceeding 70%. "Profit in April, profit in May, and June looks the same." OpenAI's revenue has rapidly grown as the adoption rate of Codex increases. Patel used SemiAnalysis' own data as evidence— the company's 90-person team's annualized AI expenditure has surged from less than $100,000 in November last year to the current $11 million, with AI costs now surpassing one-third of labor costs. "Every time the model upgrades from 4.6 Opus to 4.7 Opus, expenditure first drops for a week, and then surges again—because what was impossible before can now be done." He believes that Anthropic's core advantage lies in higher Token efficiency and lower user comprehensive costs.

On the hardware front, 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." Inference model-driven KV cache explosive growth, while memory capacity only grows by 20% to 30% per year, the supply-demand gap will continue to widen. Memory has quadrupled, with still 2 to 3 times room for growth; consumer electronics will be the first to bear the pressure, with shipments from mid- to low-end phone manufacturers already down 40%. At the same time, his narrative on CPUs gives a clear warning—current demand includes a large number of historical catch-up effects, once the backlog is cleared, CPUs are far behind AI acceleration chips in absolute terms. "Memory and AI acceleration chips are the main part, CPUs are the reassessment after being underestimated, but will not grow at a speed exceeding AI chips indefinitely."

Regarding the market's enthusiasm for CPOs, he clearly stated that true large-scale production will have to wait until the end of 2028 to 2029— manufacturing yield, chip design, and supply chain maturity are not up to standard. NVIDIA's Rubin and its subsequent architecture Feynman will still use a full copper solution, unexpectedly extending the copper cable connector bonus period.

On the power side, he predicts this year will see an additional 20 gigawatts of power consumption in data centers, 30 gigawatts next year, and 50 gigawatts the year after, with half of it coming from enterprise self-built "behind-the-meter power supply." Currently, combined-cycle gas turbine units are mainstream, but within two years, the integrated cost of solar energy plus energy storage is expected to be lower than gas-fired power generation.