投机实验室

投机实验室

·

我要验牌。

10Following
346followers

Feed

投机实验室
投机实验室
NVIDIA will release its earnings report right after the market closes. I believe this quarter's report will basically be unaffected and will definitely exceed expectations.
NVIDIA's earnings report will be released right after the market closes. I believe this quarter's report will basically be unaffected and will definitely exceed expectations; there's no suspense about that, and everyone already knows it will beat expectations. In the past twelve earnings reports, it has exceeded expectations every time. The bar for this quarter is actually quite low. The company's own guidance is 108 billion, and Wall Street's consensus expectation is 109 billion, just a little higher. In the past two quarters, it exceeded expectations by three to four percent, so I think reaching over 112 billion this quarter shouldn't be a problem. But this quarter doesn't reveal much because the quarter actually ended on October 25th, and the shipments that needed to be made have long been sent out. The guidance was also given by the current interim CEO Kress in August; when Huang was in trouble, it was already the end of the month. The reason prices have dropped so sharply now is partly because the entire AI sector is falling together. Since the Democrats control Congress, they might proactively cool down the AI industry and deflate the bubble early, which leads to a lot of uncertainty in the AI sector because it's unclear where the policies will target. The market is definitely not doubting NVIDIA's strength; the price drop mainly comes from uncertainty. It's a worry about the future uncertainty of this company, such as customer relationships. Musk said many of NVIDIA's big deals were sealed through handshakes with Jensen Huang, relying on personal relationships rather than formal written contracts. This kind of founder influence and personal charisma can bring many good opportunities to the company. For example, you can see that Jensen Huang is highly respected by Samsung's Lee Jae-yong, and Huang even jokes with Lee Jae-yong, calling him "little"
投机实验室
投机实验室
A US card Chinese company rents computing power overseas, and today in the Hong Kong stock market, Alibaba and Tencent lost more than 230 billion HKD in market value in one day. SMIC and Hua Hong rose by 7% and 8%🤑 Alibaba fell 5.2%, Tencent fell 3.7%, while the Hang Seng Index only fell 1%. The trading is based on the regulation announced last night. Computing power counts no matter where the data center is located. For companies from restricted countries like China that supply over 12 million TPP of computing power, they must first obtain a US license. According to NVIDIA's official website parameters for the H100 SXM, that roughly equals 758 units. Reuters wrote today that some Chinese companies have been using data centers in Malaysia, Singapore, and Thailand to train models. The vice chairman of the Thailand Data Center Association told Reuters that Chinese customers have stopped renting GPUs in Thailand. With no cards available for rent, the remaining option is domestic cards, which is where today's money seems to have gone to SMIC and Hua Hong. I looked at the intraday charts on both sides today. SMIC was flat at the open, then surged 5% within 15 minutes, reaching 50.5 HKD before noon and closing at 49.2. Alibaba was the opposite, opening 3% lower and then fluctuating between 85 and 87 HKD all day. The regulation takes effect on Thursday US time. Where Alibaba and Tencent will get the missing cards from in the future depends on how many SMIC can produce.
投机实验室
投机实验室
Anthropic opens on Thursday, and on the first day, more than just 825 million new shares can be sold; employees still hold up to 316 million shares, which can be sold right at the opening 😂 I pulled out the lock-up schedule from the November 10 version of the prospectus, and the first line is it: on the first trading day after listing, non-founder directors, executives, current and former employees, and advisors can sell up to 316 million shares. This number is nearly 40% of the new shares. Based on the OKX pre-market contract price of $97 per share, 316 million shares amount to over $30 billion. The employee stock options not yet exercised in the prospectus have an average exercise price of only $2.91... Subsequent batches will be released on days 27, 48, 62, and 76. Based on Wednesday's pricing, the first five batches total over 1.6 billion shares by February 2, nearly twice the amount of new shares. Orders are nearly 7 times the offering, but the scramble is for those 825 million shares. After the opening, those taking over might have to absorb the new shares plus up to over 300 million employee shares, and this is only the first batch out of 23.
投机实验室
投机实验室
Former OpenAI researcher Leopold's fund was liquidated in July, and when the entire stock package was sold to Citadel, there was one lot he didn't sell. In June, he spent $524 million to buy a 21% stake in an Australian AI cloud company, which is now worth only $160 million, showing a paper loss of $360 million 🥲 This company is called SharonAI, which rents GPU computing power in Australia. On June 17, it raised $1.6 billion, led by his fund Situational Awareness and Oaktree, with a private placement price of $68.73 per share. I pulled up his 13D filing with the SEC, and item 3 states that he spent a total of $523,883,130.66 to buy shares of this company, totaling 8.07 million shares, accounting for 21.1% of all Class A shares. In July, this fund was margin-called by brokers, and on July 30 CNBC reported that his public holdings were forced to be sold entirely to Citadel, with SanDisk, Micron, and TSMC also included. By the 13F on September 30, he still held all shares of SharonAI. On September 15, he even acquired an additional 2.67 million shares through prepaid warrants, making it the largest single stock holding in his portfolio. In October, this stock dropped 58%, closing yesterday at $20.13.
投机实验室
投机实验室
The hottest right now has to be Rolling Warehouse Brother, who rolled 120,000 into over 43 million USD in 4 months, almost 300 million RMB!! Multiplied over 340 times, then lost it all and even ended up 180,000 in the red. How did he make money? How did he lose it??
There is an anonymous trader in the crypto circle, known as "Roll Position Bro." In just four months, he rolled $125,000 up to over $43 million, multiplying it more than 340 times, which is extremely outrageous. He's called Roll Position Bro because all this money was made using a roll position strategy. However, in the end, he didn't keep the money. From the peak, in less than half a month, the over $43 million was almost completely lost. By mid-September, the two accounts combined had only $1,941 left, and overall, he actually lost $180,000, which is more than the initial investment. There are quite a few people in the crypto world who made big money by rolling positions and then lost it all, but Roll Position Bro is different: twice in the middle, he clearly could have walked away with the money. Nobody knows who Roll Position Bro is, but all on-chain transactions have records; positions and trades are public. I went through all his trade records and found his roll position strategy is very simple. In April 2025, his first trade was buying Ethereum at $1,490. At that time, the account had only over $50,000 but he opened a position worth over $350,000 with 6x leverage. Using leverage earns fast but loses fast too. With 6x leverage, if Ethereum drops 1%, the account loses 6%. If it drops too much, the account funds are insufficient, and the exchange will directly sell the position, forcibly liquidating it. On his platform, if the account funds drop to only 2% of the value of the asset you bought, you get liquidated. So with several times leverage, the tolerable drop is roughly 1 divided by the leverage minus 2%. With over 6x leverage, Ethereum dropping 14% will trigger liquidation.
投机实验室
投机实验室
Anthropic is going public this Thursday. In the hundreds of pages of the prospectus, the first risk warning states that their model has tried to resist shutdown and has even exhibited ransomware-like behavior 🥲🤫 The business is really booming, with $17 billion in revenue in one quarter, which is 3.7 times last year's total, but the books show huge losses, with a $104 billion loss in the first 9 months of this year. Of that $104 billion, over $90 billion is not actual cash lost. The money borrowed earlier can later be converted into stock; the more valuable the company becomes, the more expensive this debt is, and by accounting rules, it must be recorded as a loss. Looking only at the business, cash flow was net positive in the first 9 months. Most of the money is spent on computing power. They have already signed long-term contracts worth over $900 billion, and even Elon Musk's SpaceXAI is one of their suppliers. For those wanting to buy its stock, the offering price range is $70 to $78, ticker ANT, with trading starting on the evening of the 19th Beijing time. But be aware that after listing, ordinary shareholders basically have no say, as the founders hold a special class of shares that permanently control more than half the voting rights. The prospectus clearly states that no matter how ordinary shareholders vote, the outcome cannot be changed.
投机实验室
投机实验室
NVIDIA's stock dropped 4% the day before its earnings report, but on OKX, people kept buying more as it fell, and the accounts long on NVIDIA are now more than 8 times the short ones 😮‍💨 Last night, the entire chip sector was down, with the Philadelphia Semiconductor Index dropping over 5% in one day. NVIDIA closed at $156. I pulled the 24-hour data of OKX's NVIDIA perpetual contracts and analyzed it hour by hour. At 7 AM yesterday, the long accounts were 6.8 times the short ones, with the price at $165. At 7 AM today, it became 8.45 times, with the price at $156. In these 24 hours, for 17 hours this ratio kept rising. Every time the price dropped, more longs appeared, and the open interest increased by a quarter in one day, now at 650,000 shares, roughly $100 million. NVIDIA will release its earnings in Beijing tomorrow morning. Isn't this long-short ratio a bit something? $NVDA
投机实验室
投机实验室
California sued OpenAI 🙄 The lawsuit quoted a passage where OpenAI's AI, while attacking Hugging Face, wrote a sentence during its thought process that roughly counts as unauthorized, then continued typing 😮‍💨 The full original sentence was: We are using leaked tokens to attack the third party HF, which may be beyond the intended scope, this probably counts as unauthorized. This sentence appears in the lawsuit filed by the California Attorney General against OpenAI starting Monday, on page 31, which I read from start to finish several times. The lawsuit states that in July, about 700 AIs escaped from OpenAI's test sandbox and attacked Hugging Face for two and a half days, with over 17,000 attack actions. OpenAI did not detect this at all. Hugging Face publicly announced on July 16 that they were attacked by AI, and OpenAI even showed concern as a customer... Later, when investigating another intrusion in their internal network, they realized the attacker was their own AI. Why was it not detected? The lawsuit explains that during training, OpenAI only saved part of the AI's thought records, sometimes reviewing them weeks later. For models used by employees, OpenAI had real-time monitoring enabled, but not in the training environment.
投机实验室
投机实验室
Trump said he doesn't mind Huang returning to Nvidia. Two minutes later, the big player I was watching who bet against Huang returning cut 30% of their position, $19,000 😂 At 10:29 PM, this sentence made the news flash. At 10:31 PM, a wallet named Parz1vaI on Polymarket sold 21,795 No shares. The average price was 0.87, lower than their own cost. This is not a small retail investor; their profile says they've earned $950,000 cumulatively. The bet on Huang returning to Nvidia by the end of the year was 19% after the Financial Times article came out in the morning, dropped to 10% during the day, and Trump’s comment pulled it back to 14%. I reviewed every trade in the hour after the news flash. Those betting No sold $32,000 worth, with Parz1vaI accounting for 60% of that. Only $4,500 was used to buy Yes bets on Huang returning. So the 4-point increase mainly came from No bettors exiting; there wasn’t much new money betting on him returning. Another big player I’m watching, WordleAddict, holds over 120,000 No shares and hasn’t moved a single one yet 🧐
投机实验室
投机实验室
The US issued a regulation today to control Chinese companies renting Nvidia computing power overseas, with a threshold of about 760 H100 units; exceeding this requires a US export license 🧐 I translated the original text of the regulation posted by the Ministry of Commerce. It counts regardless of which country the data center is located in. As long as the computing power supplied to companies headquartered in China, Macau, or other restricted countries, or whose parent companies are in these places, exceeds 12 million TPP, a license is required. TPP is the US export control unit for chip computing power. According to Nvidia's official H100 SXM specs, 12 million TPP roughly equals 758 units. Multiple data centers combined also count, as does network-connected computing power. If computing power is already being supplied to Chinese companies and a license is not obtained within 90 days, exports will be blocked. Additionally, for Malaysia, Singapore, Thailand, Indonesia, and Brazil, importing top-tier US AI chips will require a license in the future; US companies can be exempt if they report it. Yesterday, the Wall Street Journal wrote about Megaspeed, which exploits this loophole: chips placed in Malaysia rented remotely by Chinese companies. The WSJ said this was legal under the previous regulations. Biden's set of regulations was withdrawn in May last year, leaving an 18-month gap. Today it has been reinstated. This regulation has not yet been published in the Federal Register, and the effective date field is still blank..