Closed-source LLM profits peak as margins flow to open-source models and applications
Last phase: nearly all profits went to Anthropic and OpenAI. Now domestic models have crossed the usability threshold, open-source models are eroding closed-source margins, and profits are flowing to open-source and application layers.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI's stage hinges on how you define penetration rate's denominator
Numerator is straightforward—Anthropic plus OpenAI's year-end ARR may hit $200+ billion. The hard part is the denominator. If AI equals coding, set against $550 billion in US programmer wages or $1.2–1.3 trillion globally, penetration is already not low, putting AI in mid-to-late stage. But if AI equals all tasks computers can do, the denominator becomes $50 trillion in global white-collar wages, making penetration quite low. How you define the denominator is itself a judgment about AI's stage—and this disagreement on framing is the root of industry disagreement.
— Tang Zong (Powell)Open-source and applications are taking away closed-source profits
Last phase: Anthropic and OpenAI captured nearly all value; open-source and application companies had no leverage—Cursor, for example, paid Anthropic for nearly everything, leaving zero margin. Now open-source models are approaching frontier capability and will compress closed-source 90% margins; profit flows to open-source layers and downstream applications. In parallel, domestic models have crossed the usability threshold—from elementary-school level to university level. Frontier models are PhD-level, unaffordable; but university-level is sufficient, and domestic demand will explode.
— Tang Zong (Powell)Whether downstream ROI is real demand or just eliminating redundancy
Hao raises a sharp question: when big tech firms use AI to replace workers, are those workers redundant organizational-friction positions to begin with? Is cutting them real ROI, or just eliminating fake headcount? He even questions whether giant internet companies could cut 60–70% workforce with no issue. Tang counters by examining supply: domestic coding plans sold out instantly and sold out again after price hikes, proving demand exceeds supply. Person-to-person communication always has friction costs; AI agent-to-agent context passing is genuine organizational efficiency gain, not mere headcount reduction.
— Hao Ge/Tang Zong (Powell)Storage bulls face a bear case that cannot be disproven
Storage stocks have surged this year, yet bear narratives always existed—just ignored. On supply: domestic expansion will grab share. More threatening: on demand, if algorithms improve context-window efficiency at all, it collapses huge capex spending. This is completely unfalsifiable and likely to occur. Micron beat earnings across the board; stock rose one day then fell. Sentiment has shifted. Hao is emphatic: storage only for shorts, never for long holds. Unlike overseas compute capacity, storage technology paths can flip without warning.
— Hao Ge/Tang Zong (Powell)AI is infrastructure like electrification, not internet—builder phase obscures use cases
Tang spent roughly 10 hours working through electrification history with AI. AI capital spending is about 2% of global GDP; that scale only matches infrastructure-tier revolutions like electrification. Internet (Cisco, dark fiber) was under 1% of GDP—too small to serve as analogue. Electrification went through about 20 years of competing technology routes, then about 20 years of infrastructure buildout. During buildout, manufacturing productivity growth was only about 1%—use cases were completely invisible until the late stage arrived. On this timeline, today's AI is roughly electrification around 1919, infrastructure is barely showing effects.
— Tang Zong (Powell)AI tracking feedback uncovers information sources you'd never think to request
Tang breaks research into six components: idea capture, prioritization screening, quick research, deep research, decision, and tracking feedback. AI handles three: idea sparking, information collection, and tracking feedback. The striking case: he asked AI to track one Nordic data-center firm. AI auto-decomposed the task into monitoring US domestic construction protests, EU new-build announcements, even scraping municipal websites for land-transaction records. It unprompted reported who signed with city today, acreage purchased, price paid—none of which he explicitly requested. The tool surfaced information sources he never would have thought to ask for.
— Tang Zong (Powell)Domestic compute has solid odds and favorable payoff ratios right now
Tang is most bullish on the domestic compute chain. Domestic models just crossed the usability threshold (on par with Opus 4.6+ when agents ubiquitously deployed); layered with Chinese big tech lowering adoption barriers, explosive adoption is a timing question, not an if. Once domestic vendors saw training and inference were profitable, capex intensity rose; yet supply remains constrained, much like the EV industry during its own lucrative window. Win rate exceeds 50%, payoff profile is decent; choosing between large-cap vendors, small-cap vendors, and upstream-versus-downstream is judgment-call territory. Application layer involves deeper specialization and wider disagreement.
— Tang Zong (Powell)Quantitative factor models are down to minute-level granularity; retail day traders always lose
Tang studied quantitative trading and concluded: daily-timeframe factors have degraded to too-low-frequency, zero alpha. Now it's all minute-timeframe factor machine-learning wars. No human can win at that resolution. Worse: as quant scale grows, they cannibalize each other, not just retail. This year, quant long-only alpha has evaporated. Tang's adaptation: abandon short-term timing, stretch to 3–6 month windows, bet only on major shifts like domestic model breakthroughs, then enter at an acceptable price. He will not race quantitative reaction time.
— Tang Zong (Powell)In their own words · checked verbatim
When I'm not in the trade, industry trends keep chugging along; the moment I board, we get a value reset. Really that correlated?
别十个点开始重新理解现金流 怎么我不带的时候 天天产业趋势不肯腻 我一上车就价值回归了 有这么强吗
Hao Ge1:00
Solar penetration is where now? New-energy vehicle penetration where? Generally speaking, for technology revolutions, the sweet spot is around 5–10% penetration, right? At that point the total addressable market is big enough.
光伏的渗透率到多少了 新电源车的渗透率 到多少了 然后一般来讲 技术革命的渗透率 最舒服的时候 大概是在5%到10的时候 对吧 那个时候总量足够大
Tang Zong (Powell)6:04
It was elementary-school level before; now it's college level, right? You're PhD level. I can't afford you. I'll get the college student. That elementary-school level? Definitely won't use it anymore.
它之前是小学生 它现在变成大学生了 对吧 你是博士生 我顾不起你 我就顾个大学生 那你小学生 我肯定不用你了
Tang Zong (Powell)15:12
So this supply-expansion narrative lands on storage, right—domestic makers expand to beat it. But I think the stronger narrative—completely unfalsifiable and I think it'll actually happen—is that algorithms will cut into it.
那所以这个叙事 就是说供给扩产的这个叙事 就到了长心仓存身上 对吧 国产去扩产去击败它 但是我觉得更强的一个叙事 就是完全无法证伪的 而且我觉得它会成为事实的 就是它会被算法砍一刀呀
Tang Zong (Powell)35:17
I need to be very clear here: storage, never go long. If you absolutely must go long, only short-term. You cannot hold for the long run.
我这里要特别 就表达一下我的观点 就我认为存储 千万不要超 就你要超的话只能短期超 就是你不能长期拿
Hao Ge37:20
Once you have this tool, the essential mindset is: have AI do everything first. Don't do it yourself. That's critical.
当你有了这个工具之后 一定要有的一个想法 就是说什么事情 你都要让AI先做 不要你自己去做 这个很重要
Tang Zong (Powell)1:03:27
Every single day it just drops straight down, shoots straight back up, drops straight down, shoots back up—no smooth curves, nothing but 30-degree angles the whole way.
他一天就是这样直接下去 再直接上来 再直接下去 再直接上来 他都没有那种柔和的曲线 全是那种30度角
Hao Ge1:08:28
Back then it was already $90 billion and making over $10 billion. People thought that was the peak, gave it a 10x to $150 billion. Five years from $90 to $150 billion? Only 60% upside, not much point. Instead it went from $90 billion to $500 billion, then crashed back to $100 billion.
然后那个时候 它已经九百亿了 它能赚一百多亿 然后觉得那个时候 可能pick了 给它十倍 那它一千五百亿 五年 大概从九百亿 涨到一千五百亿 才60%空间 没啥意思 结果它从九百亿 涨到五千亿 完了又从五千亿 跌回了一千亿
Tang Zong (Powell)1:16:33
Figures
| AI capital expenditure as share of global GDP | 2% | 5:04 |
| Comfortable penetration-rate range for technology revolutions | 5%-10% | 6:04 |
| Coding as share of Anthropic usage | 35% | 24:16 |
| Anthropic ARR | over $70 billion | 26:16 |
| OpenAI ARR | four to five hundred billion dollars | 26:16 |
| GPU rental (per 1GW) | approximately $50 billion per year | 27:16 |
| US AI-related capital expenditure (annualized) | eight to nine hundred billion dollars | 28:16 |
| DeepSeek payback cycle | 10 months | 19:13 |
Glossary
- Jevons Paradox
- Efficiency improvements and cost declines paradoxically increase total demand, causing resource consumption to rise rather than fall.
- AI for Science
- Using AI to accelerate scientific discovery such as drug development; considered AI's next high-value application domain.
- World Model
- Research direction aiming to let AI understand how the physical world operates; enables embodied AI such as robotics.
- AR
- Annualized run-rate; projected annual revenue based on current operating pace; standard metric for measuring AI company growth.
- Agent
- AI program capable of autonomously decomposing and executing multi-step tasks, distinct from single-question-single-answer chat models.
How to listen
Secondary-market investors and analysts tracking AI supply chains, and retail investors seeking to build personal research workflows with AI tools.
Opening 0:00–2:00 has low-density chat about market volatility; skip to 6:04 for the penetration-rate framework.