Nvidia Has No Monopoly, Only First-Mover Advantage — Take 90% of the Money and Someone Will Come Dig Up Your Foundation
Nvidia's moat is the CUDA software layer plus a hardware lead, but it has not formed a monopoly. The higher the gross margin, the stronger the incentive for model makers to build their own chips — swinging that shovel is partly about answering to the capital markets.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Taking 90% of the money means everyone wants to dig you up
Nvidia's moat is software plus hardware: the hardware is the GPU, the software is the CUDA middle layer that acts as a bridge between human programming and the GPU. But it has not formed a monopoly, only a leading first-mover advantage. The problem is that when you run especially fast and take 90% of the world's money, the price is that everyone wants to settle that money with you, with a crowd behind you chasing with guns. And the higher the gross margin, the greater everyone's incentive to chase. Model makers themselves have an interest in raising gross margins, whether through cost cuts or efficiency gains; squeeze to a certain point and the knife naturally swings at Nvidia's ecosystem — even if it is only a test swing.
Swinging this shovel is about answering to capital
Why do model makers insist on developing their own chips even knowing how hard it is? This is not just a technology-selection question, it is even a question of how to answer to the capital markets. I want to swing this shovel to tell the capital markets that I am trying to do this, that I have not raised my hands in surrender and been monopolized by a single ecosystem. Amazon and Google have already proven that building your own chips is feasible, and the bigger the scale, the more money you save. So OpenAI also announced it would make its own chips, with the evaluation score called Mexican chili pepper. But this takes too much money — OpenAI said it would spend 45 billion US dollars to rent compute; can 45 billion build its own? Probably not enough.
An eight-card server is 14 million RMB in China spot
Take the relatively advanced B300 server as an example: 4T of memory, eight cards per unit, spot price in China over 14 million RMB. The same machine in the US costs only just over 500,000 US dollars, about six to seven million RMB, three times the price. There are multiple reasons for the high cost: US government export controls — the so-called defense is about preventing you from shipping it out. You can order on Amazon in the US and have it delivered to a home in Silicon Valley, but it cannot leave the country — it needs KYC, it cannot be China, and it needs IDC proof. When the machine is routed through layer after layer of forgery to Southeast Asia, say Thailand, it basically doubles; then from Thailand it breaks through layer after layer of defense into the mainland, going through both sides, and it triples.
512 units delivered means plugging in and lighting up 512 times
Spot trading is like housing, generally sold in powers of two: two units, four, eight, 128, 256. Someone ran into 512 units quoted at over 7 billion. The transaction process: first you screenshot your bank account to prove you really have the money, and the other side has to shoot a video proving the goods really exist, saying the company name and holding that day's newspaper. On delivery it is one unit at a time: plug in and light up to show it is good, wire the money, haul one away, then the next, repeated 512 times, taking almost a day — because lighting up one server with cards pulled takes half a day. The 80 billion HKD Alibaba raised is only enough to buy 5,000 units.
Futures are all fake, and 70% of spot is brokers
In the flood of futures inquiries, an 8 million quote is cheaper than the 14 million spot price, but nobody dares do it; now it is full-payment letters of credit. A letter of credit is a bank guarantee, understandable as an old paper version of Alipay, used for bulk commodity trade, and it also needs a third-party warehouse: goods go into the third-party warehouse, the warehouse receipt goes to the bank, the bank pays on seeing the receipt, and the buyer takes the receipt to the warehouse to collect the goods. But the current play is: the buyer issues a full letter of credit, the seller takes the letter of credit and the buyer's documents to the bank as collateral, gets the money and goes to buy the goods. So futures are 100% all fake — as long as the goods arrive, there is no compensation risk, and you can sell spot directly, who wouldn't do it? Spot is also estimated to be 70% fake, with 70% of sellers being brokers.
What must a 14 million machine run to pay for itself
Do the math: 14 million buys eight cards; for training, one or two units is definitely not enough, you need at least several hundred; for inference, how much does buying tokens cost? When does it pay back? It used to be said ten months to pay back, now that has already fallen. Someone saw the news of DeepSeek's IPO and said model makers have pretty high margins, seemingly reaching 60-70% — but that does not count the investment, nor the servers; with hardware assets amortized, plus data center, power and maintenance costs, maybe you calculate a 60-70 gross margin. The key is whether there is a steady stream of writable content to keep the cards fully loaded. Many businesses do not make money because there is too much idle time, just like a restaurant: if it can be full 24 hours a day, the margin is also very high.
A domestic GPU runs a top model; Nvidia falls then rises
Zhipu released GLM 5.3 Flash; the score is not first, but it basically caught up with Opus 4.8, and it claims to have entered the top tier. The most important thing is not the score, but the line in the article: our compute runs on domestic GPUs, and after optimization, the computational efficiency of domestic GPUs has caught up with that of overseas GPU products. That sentence went viral. The timing of the news was also bad, landing during the evening development window. Nvidia fell only a little yesterday and rose over 6% today. The market is now either extremely optimistic or extremely pessimistic; one piece of news seriously affects the direction, and a couple of days later it comes back.
Models split in two: top price for creativity, cheapest for repetition
Now blind tests cannot tell the models apart, so people start choosing on price. But the enterprise side actually splits into two scenarios: creativity-driven scenarios, such as writing code or doing long-horizon analysis to produce reports, where no matter how cheap it is you will not use it — you will definitely use the top tier, because the sunk cost is too high and redoing it after a bad job is not worth it, wasting both time and money. The other is work like content analysis, where the output directly determines enterprise revenue, so you must choose the model with more reasonable ROI. It is like hiring people: for cleaning, even at a monthly salary of 10,000 the best still feels unnecessary, so among the cheapest you find one that is passable; for an engineer, you are willing to pay a higher salary to buy creativity. Two ends — stand in the middle and you are backed by neither.
The official website is read by AI; more words is an advantage
A report says 70% of internet traffic is already AI. If the official website is read by AI, it must be written more complexly and give complete information to gain an advantage from the GEO angle. SEO used to be about stuffing keywords, but AI does not look at those; if it cannot summarize, it skips over you, so you must write it clearly with a relatively complex and complete narrative logic — do not fear too many words, fear not covering it. Technically, the old dynamic pages no longer work; AI scraping on the device side can only get an empty page, so you must use SSR backend rendering and put all the content into HTML returned in one go. There is another approach: the large model sends header information about acceptable content types; if it is TXT plus markdown, return pure markdown with nothing in it that could cause hallucination.
In their own words · checked verbatim
This is not a question of technology selection; it is even a question of how to answer to the capital markets.
这不是一个技术选型的问题，这甚至是一个如何向资本市场交代的问题。
So right now in this market, first, futures are all fake, 100% all fake, it is simply impossible for them to be real. Because he just turns around and sells spot.
所以现在这个市场第一期货全是假的，百分之百全是假的，就是不可能是真的。因为他转手就拿现货卖了。
A lot of businesses do not make money because they show too much time, just like a restaurant: if you can be full 24 hours a day, your margin is also guaranteed very high, very very high.
很多的生意不赚钱，是因为他显的时间太多了，就跟饭馆一样，你要二十四小时丢人都能坐满车，你的率也保证的很高，很高很高。
Actually it is just two ends, two ends, you cannot stand in the middle; if you stand in the middle you are backed by neither, and you have to be cheaper than the badass.
其实就两头就是两头，你就不能站中间，你站中间就两边不靠，要比牛逼要便宜。
In the future, if your content or your website's interface is not agent-friendly, you will be eliminated from this game's rules.
将来如果你这个内容或者你这个网站的接口对agent不友好，你就会被从这个游戏规则里干掉。
All product logic, your engineering logic, will be flattened by AI.
所有的产品逻辑，你的工程逻辑都会因为ai就抹平了。
Figures
| B300 eight-card server spot price in China | over 14 million RMB | 6:01 |
| Same-configuration server price in the US | just over 500,000 US dollars (about six to seven million RMB) | 6:01 |
| OpenAI compute rental amount | 45 billion US dollars | 4:59 |
| Alibaba fundraising amount | 80 billion HKD | 10:10 |
| Quote for 512 machines | over 7 billion | 10:10 |
| Model maker gross margin | 60-70% | 19:21 |
| Share of internet traffic that is AI | 70% | 39:14 |
| Nvidia single-day gain | over 6% | 24:44 |
Glossary
- CUDA
- Nvidia's software middle layer connecting programming and the GPU, the core of its moat.
- KYC
- Identity verification under export controls, requiring proof of a non-Chinese buyer and an IDC location.
- Letter of credit
- A bank-guaranteed paper version of Alipay, used for bulk trade, requiring a third-party warehouse and warehouse receipt.
- SSR
- Backend renders content into HTML and returns it in one go, making it easier for AI on the device side to scrape.
- GEO
- Content optimization for AI generative engines, replacing traditional SEO keyword strategies.
- MCP
- A scheme that wraps data services into a protocol for AI to call.
How to listen
Founders and engineers watching the compute supply chain, the GPU black market, enterprise AI deployment and product design logic.
The model evaluation and blind-test section in the middle (27:00-31:00) can be fast-forwarded.