The Capital Cycle Framework Missed Nvidia: He Audited His Own Miscalculation
Looking at AI hardware through a capital-return framework, Dai Bin admits he missed the main line. In 2023 he valued Nvidia at over $3 trillion based on server penetration; CAPEX later went from 2000 to 9000, and he never tracked the change in magnitude.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Research is only two things: how money is made, and how it is priced
Dai Bin splits investment research into two dimensions. The first is studying how money is generated inside a company — it arrives as operating cash flow after customer delivery, gets reinvested into capex, and finally returns to shareholders. The second is studying the stock pricing mechanism, which is the domain of the capital cycle. He says people who came up through research are very familiar with the first and relatively weak on pricing and cycles — that's determined by where they came from. Together the two dimensions form an assembly line: capital cycle first, then the company.
— Dai BinThe difference from Buffett is that asset managers can't control cash flow
Dai Bin points to the one important difference between asset management institutions and Buffett: Buffett can control a company's cash flow and force distributions; asset managers cannot, and can only earn returns through other people's quotes, dividends and buybacks. He cites his own loss on cigar-butt stocks — some companies hold a lot of cash on the books, but the controlling shareholder can choose not to pay a dividend or buy back for ten years. Anyone who is a minority shareholder must take the mechanism for ‘how does this money get back to me’ very seriously.
— Dai BinFour stages of the capital cycle: from tight supply-demand to PB arbitrage
Dai Bin breaks Marathon's capital cycle into four stages. Stage one: the industry enters a boom, demand exceeds supply, the income statement and cash flow improve. Stage two: the boom broadens, second- and third-tier companies' financials start improving too, and PB goes from low to high, because there is an arbitrage mechanism between financial markets and the real economy — building a factory is 1x PB, virtual equity is 4-5x PB, so entrepreneurs sell equity to arbitrage into the real economy. Stage three: the CAPEX plans land, everyone extrapolates demand too linearly, and the stock market falls first. Stage four: bad signals appear in the fundamentals and PB declines. But not every industry gets to enter a new cycle.
— Dai BinCapex intensity: CAPEX divided by depreciation tells you where you are in the cycle
Dai Bin offers a rule of thumb: capex intensity equals capex divided by that year's depreciation. For a mature industry, 0.8 to 1.2 counts as high, 0.5 to 0.8 counts as low — that applies to very mature industries like cement. For a somewhat newer industry like lithium carbonate, a bit over 1x to 2x is the bottom and 4-5x is the high. For an industry just going through its first wave of penetration, the low is 2-3x and the high is 10x. His example: in 2018 new energy vehicles were just starting, with a low of 2-3x and a high of 10x; by 2024 they already count as mature, and CAPEX can't possibly be that high again.
— Dai BinPigs versus monkeys: the longer the supply-side contraction, the greater the elasticity
Dai Bin contrasts pig farming with monkey farming: the industrialised share of hog farming is very high, and a pig grows out in a dozen-plus months; a rhesus monkey takes four or five years to reach a usable stage. Looking at it through the capital cycle, hog farming is unlikely to re-enter the boom everyone wants, while monkey prices may still have elasticity. He adds that chickens are faster than pigs — 30 to 45 days to market — so capacity supply ramps up extremely fast and prices get smoothed out easily. As long as demand is still there, the bigger the supply-side conflict and the longer the mismatch persists, the better the stock's elasticity.
— Dai BinThe capital cycle framework is not effective enough on new industries
Dai Bin admits the regret of this methodology: it is not effective enough on new industries. In this cycle he missed AI hardware, and the companies mentioned earlier were still found with the old methodology — the boss kept buying more, the stock was relatively cheap — without a detailed breakdown of how overseas giants' CAPEX transmits step by step into each industry to affect demand. He says people who think in capital cycles naturally think a bit more about risk, and that is where he missed.
— Dai BinValuing Nvidia in 2023: over $3 trillion from server penetration
Dai Bin walks through how he analysed Nvidia's valuation in 2023: the world sells roughly 18 million servers a year, a mature industry, and he used the new energy vehicle analogy, calculating with 30% or 50% penetration, assuming 30% market share, a few GPUs per unit, and roughly what a GPU costs, arriving at a market cap of about $3 trillion-plus. At the time Nvidia was already at $2 trillion-plus, and it eventually reached the $4-5 trillion level. He says this was the first stage, and that people who think in capital cycles naturally think a bit more about risk — that is where he relatively missed.
— Dai BinCAPEX went from 2000 to 9000 and he wasn't tracking the magnitude
Dai Bin says the second stage was a lack of close tracking of the magnitude of CAPEX. The demand side must not be strongly prejudged; the supply side you basically have a handle on, while the demand side requires intense tracking. When he valued Nvidia he felt the industry carried some risk, so he didn't track it intensely — but seeing CAPEX go from 2000 to 9000, a three-to-four-fold change in magnitude, plus the changes in memory in between, he could actually have analysed it with the capital-return framework. His reflection: Nvidia wasn't necessarily buyable, but he could have bought the memory chain or other companies in traditional chains that were improving.
— Dai BinDoing the math: semis are up 13 trillion, and the payers are also around ten trillion
Dai Bin does the math: this cycle semiconductors are up roughly ten-plus trillion, or 13 trillion in market cap if you include ChangXin; but the customers who can pay — China Mobile, Tencent and the like — add up to around ten trillion in market cap too. Overseas, memory plus Nvidia is also ten-plus trillion US dollars, and the CSPs adding money is also ten-plus trillion US dollars. From a math standpoint, hardware has reached a place where the odds aren't great, while the odds on applications may be better. Using the mobile internet experience, he judges that what follows will play out as using this hardware to make money.
— Dai BinIn their own words · checked verbatim
Buffett doesn't need a quote — he can force distribution of his cash flow. That may be an important difference between us.
巴菲特他可以不需要报价，他可以强制分配他的现金流，这可能是我们的一个重要的区别。
Dai Bin17:17
That is, if it's a mature industry, from my tally, between 0.8 and 1.2 counts as high, relatively high, and 0.5 to 0.8 probably counts as low.
就是说如果它是一个成熟产业，我统计下来，就是说0.8到1.2之间，它就算高了，算比较高了，0.5到0.8可能就算低。
Dai Bin33:28
The regret is that for some new industries, its effectiveness is insufficient. And indeed, looking at this cycle, on some AI hardware, I essentially missed that direction.
遗憾的地方就是对一些新型产业，它的有效性是不足的。那确实从这一轮周期来看，对一些AI硬件，我是相当于是错过的这个方向的。
Dai Bin50:39
There's no cure-all in this industry; it keeps changing. Look at the major inflection points we just shared — why did they change? Did it stop working? It really changed.
这个行业没有万金油，一直在变，你看我们刚才分享的重大时间点的变化，为什么有变化，失效了吗，实在变化了。
Dai Bin1:13:56
Figures
| Nvidia's estimated market cap in 2023 | Over $3 trillion | 51:40 |
| Nvidia's actual market cap level | $4-5 trillion | 51:40 |
| Change in CAPEX magnitude | From 2000 to 9000 | 53:43 |
| Market cap gain of the semiconductor sector | 13 trillion (including ChangXin) | 59:45 |
| Reasonable capex intensity range for a mature industry | 0.8 to 1.2 counts as high, 0.5 to 0.8 counts as low | 33:28 |
| Domestic annual sales of large-displacement motorcycles (2019) | About 100,000 units | 41:31 |
| Current monthly sales of large-displacement motorcycles | Approaching the 100,000-unit level | 42:31 |
| Total material in the Knowledge Planet | 3.8 GB | 1:20:01 |
Glossary
- MA200 / 200-day line
- The moving average Paul Tudor Jones used to flag opportunity and risk, and also the name of Dai Bin's WeChat account.
- Capex intensity / CAPEX-to-depreciation
- The ratio of capex to that year's depreciation, a rule of thumb for judging where an industry sits in its cycle.
- Paired research / paired research
- Dai Bin's method: once repeatable companies are in order, go study growth companies to reduce volatility.
- Act2 / second act
- One of Dai Bin's tools: a company doing a new business, or the same skill applied in a new scenario.
How to listen
Fund managers and analysts doing public-market investing and focused on capital cycles and supply-side analytical frameworks; anyone who wants to understand why value investors missed AI hardware.
The 1:14:56 to 1:15:56 chat about a century of Hong Kong real estate — low information density.