Cruise橘子哥

NVIDIA’s earnings report left me stunned: revenue hits $96.2 billion, doubling year over year; net profit doubles too, and for the first time it even provides guidance a full year early—saying that in fiscal 2028 revenue will grow another 70%. Before the earnings, the stock had been falling for a week and the market was worried that growth might be peaking. Then, after the bell, it surged more than 4% straight away.

At the earnings call, Jensen Huang said that AI has reached an inflection point—compute power now equals revenue. It sounds like the same kind of talk from three years ago when ChatGPT first went viral. But there’s a catch for gross margin: HBM and DRAM prices have risen too aggressively, so next quarter gross margin is expected to drop to a low of 71%-72%. Capacity, power, and storage are all in tight supply. The company says shortages will last at least through the end of fiscal 2028.

The bottleneck is capacity, not demand. With growth still doubling at the same pace and guidance reaffirmed, after staring at it for a while, all I can say is this: this round of AI money—NVIDIA really knows how to make it.

Nvidia’s post-market earnings cash-out tonight (FY2027 Q2). Before the report, it took a cold shot: the head of JPMorgan downgraded it to a neutral rating. The reason? Nvidia is using its own balance sheet to pump cash into the entire AI ecosystem—coordinating financing partnerships totaling over $500 billion with six major financial giants. Guarantees, leasing, and revenue-sharing arrangements are all in play, risks that traditional metrics simply can’t capture. On Tuesday, the stock initially held steady, rising 2.19% to end a seven-day losing streak, with a market cap of $5.16 trillion.

Customers are also adding fuel. OpenAI said its in-house chip, Jalapeno, has outperformed Nvidia’s GB300 in performance tests—and then added a caveat that it won’t fully replace Nvidia, that it will still continue to place large-scale orders. The timing is right before the earnings release—this isn’t really a technical update; it’s plainly pressuring the negotiation. It left me stunned: the more aggressively the chips are sold, the more urgently customers are rushing to build their own.

The real test comes in the early hours of Thursday Beijing time: whether growth can hold up and keep the $5-trillion market valuation afloat—more important than that neutral rating from JPMorgan.

PDD’s Q2 revenue reached 112.4 billion yuan, up only 8% year over year—falling short of market expectations, and net profit even dropped 12%. In the same week, after Amazon released its earnings report, its stock price jumped 15% in a single day, and AWS growth was 37%. Two e-commerce giants: one seems like it’s put into reverse, the other like it’s pressing the accelerator.

PDD’s problem is overseas. Chen Lei himself said Temu has been tangled up by regulators and compliance requirements across different countries. As globalization hits new challenges, the company has shifted its main focus back to domestic grocery shopping; this year, revenue is aiming for around 400 billion yuan. Amazon, on the other hand, has been gorging on AI infrastructure dividends. It raised its full-year capital expenditures to $220 billion. Even Echo smart speakers had their prices increased due to higher storage chip costs—yet cost pressure continues to pass through.

One company is feasting on AI dividends, while the other is ramming into a compliance wall and finding it painful. If PDD can’t get its growth back on track, the valuation logic the market uses will need to be replaced with a new set of explanations.

Nvidia falls $208.48, down 2.91%; seven straight days of losses on the daily chart, setting the longest losing streak since 2022. The chip sector sinks across the board: the Philadelphia Semiconductor Index drops 2.7%, while SanDisk, Micron, and SK hynix all fall more than 5% together. With this setup, it looks like the AI narrative is about to be hauled in for inspection.

The inspection point is tomorrow night: after the close in U.S. Eastern time on Aug. 26, Nvidia will release its earnings report for fiscal 2027 Q2. As the bellwether for this round of AI infrastructure investment, the look of orders and guidance will directly determine whether the market still believes that compute demand hasn’t peaked. Wow—every bit of suspense built up over seven down days is being pinned on this earnings report. To be honest, I’m a little confused by it.

Outside the U.S. things aren’t much better: the Nasdaq also has seven straight down sessions, Tesla is down nearly 4%, and money is hiding in gold as the gold price breaks above $4,670. The script of a failed “AI faith” recap doesn’t look like it’s going to stop anytime soon.

Gold rises above $4,650 to a three-month high; New York COMEX gold futures intraday touched $4,700. Domestic gold futures surged 2.86%, breaking back above the 1,000-yuan mark. Four days ago, the gold price was still hovering around 4,500. This acceleration of $150 has been driven entirely by precautionary funds.

Meanwhile, Treasuries are facing growing doubts about the “risk-free” label. Last week, Bessent carried out the “U.S. Treasury-version reversal operation” to push down long-end yields, but it only worked for a day. Long-bond yields rebounded, and the 30-year rate remained stuck above 5.2%. Markets are starting to worry: administrative measures distort pricing, and in the end it is the credibility of the U.S. dollar that gets eroded. So funds “vote with their feet” — they sell Treasuries and buy gold. Even Asia has shifted from “crisis-era funds flowing back to the West” to a “local safe-haven pool.” The dollar’s old stronghold is loosening.

Gold and bonds are competing for the same “safe-haven” tag, and this time gold is fighting especially hard. Bonds are still bonds, but the definition of “safety” is being rewritten.

This earnings report from Pop Mart left me stunned. Revenue was RMB 17.173 billion, up 23.8%; net profit was RMB 5.038 billion, up 10.1%. The numbers aren’t too bad, but the market’s expectations were higher. They didn’t meet them—its stock price dropped more than 8% straight away, and it’s been cut in half from its peak. It’s really, truly miserable.

Wang Ning, on the other hand, is genuinely candid. He openly admits that last year’s surge had a luck component. LABUBU’s share is also declining. Then, he turned around and announced a share buyback plan worth RMB 2 to 5 billion. Goldman Sachs is still pouring cold water, saying demand is soft and inventory is running high. But the offline reality is totally flipped: the “Star” characters sold out in seconds, and second-hand reselling premiums are up 13 times. Even Duan Yongping says that store visits show business is doing exceptionally well. So who should you trust? I’m confused too.

The buyback is real money—at least the boss has some confidence. But growth rates and inventory are still two hurdles ahead. Whether there can be another breakout hit after LABUBU is the most urgent question for what comes next; the “Star” characters are just a sign, for now. Today it rebounded 4% to HK$155. Sentiment is recovering, but getting back in one go may be hard.

Alibaba completed the pricing of a new share placement of HK$80 billion on August 23, with the issue price set at HK$112.7 per share. A total of 710 million shares were issued, representing a discount of about 8% to the closing price of Hong Kong shares last Friday. This is the first time Alibaba has conducted a placement since its return to Hong Kong for listing in 2019. The subscription was limited to professional institutions located outside the United States. Settlement will take place on August 26.

Money is clearly going all-in on AI infrastructure. Sovereign wealth funds from the Middle East, Europe, and Asia all came to place subscriptions. My first reaction was: even with an 8% discount, they still get the allocation—cloud service providers are essentially using equity dilution to raise a whole batch of capital to buy AI firepower. The AI arms race is really willing to spend.

Coincidentally, Nvidia released its earnings report after market close on August 26. Servers have just been reported to be set to increase prices by more than 15%, with skyrocketing memory chip costs cited as the visible reason. The more expensive computing power becomes, the more big players stock up early—this money is not going to stop.

NVIDIA servers are going to get more expensive—in many cases by over 15%. The contract manufacturers that build the batch of servers for Microsoft, Google, and Oracle’s data centers have already notified their customers to prepare for price increases. The reason is simple: the cost of memory chips has skyrocketed, and NVIDIA itself hasn’t responded. Coincidentally, next Wednesday after the U.S. stock market closes (early Thursday Beijing time), it will release its Q2 earnings report, and the whole market is waiting.

Well, servers are already expensive. A 15% increase makes me wince. The price hike is basically handing a knife to the vertically integrated chips used by Amazon, Microsoft, Google, and Meta. But NVIDIA’s software ecosystem moat is too deep—new data centers still can’t get around its cards. On Friday, the stock fell 0.98%, and its market cap is still $5.2 trillion, keeping it firmly seated as the world’s top stock.

In short, NVIDIA turns around and passes the hit from rising memory prices downstream. How well the gross margin in next week’s earnings report can hold up depends entirely on what Huang Renxun has to say.

Interesting—this round of pricing for large language models between the US and China has gone in opposite directions. In an official announcement on August 21, OpenAI said that the API and credit pricing for GPT-5.6 Sol will be cut by more than 20% over the next three months. Domestically, it’s the reverse: DeepSeek leads the way, followed one after another by Zhipu, Kimi, and MiniMax. Morgan Stanley’s statistics show that in the second quarter, the average API input price of homegrown models rose to 4.9 yuan per million tokens; in the first quarter of 2025 it was still 3.3 yuan. The output price, moreover, has climbed to 21.9 yuan.
My first reaction was that I must be reading it wrong—but after double-checking, it’s correct. The key is who’s driving the increase: independent model vendors are leading the charge, while big companies with their own compute capacity stand pat or effectively adjust prices in another way. In plain terms, compute costs are essentially fixed; independent firms that are relying on lower pricing to gain market share can’t hold on, so they have to move their prices. OpenAI, meanwhile, is proactively cutting prices to grab volume—both sides are thinking differently.
This round of price hikes looks like a correction at first glance, but actually it’s independent vendors being unable to withstand it first.

SpaceX collected $134 yesterday, down more than 4%. It fell straight below its $135 offering price, and its market cap is now down to $1.77 trillion. The reason for the sell-off is simple: the lock-up period is still being lifted, with continuous additional shares being released. The whole arrangement won’t end until 2027, during which time about 88% of the total 13 billion shares will be released.

Put plainly, the unlocking was known in advance—everyone already knew the chips would come out slowly. If the selling pressure still drives the price below the issue price, it can only mean the current sentiment is weak. There are also hedges on the table: Trump has just signed a memorandum paving the way for 1,000 launches per year in 2030; Musk, meanwhile, says revenues by the late 2030s could rise 20x, and he also promises to buy 220,000 Nvidia super chips. The pie is being drawn pretty big, but the numbers haven’t been cashed in yet. Even the Starship recovery has been delayed by a few months.

My view: the unlock situation was already made clear on the market’s chessboard, and a drop below the issue price is the most tangible “thermometer” for sentiment. These stocks may swing around with the news in the short term—don’t take the bait and don’t panic-sell. Just wait for the supply to grind itself through. Not investment advice, though.

Well, the Dow got smashed by Walmart for more than 700 points overnight. Nvidia fell only 0.33%, closing at $216.85—by far the most resilient stock all day. After touching 225 on August 13, it kept sliding for three straight days, but the daily decline got smaller and smaller (-2.34% → -0.99% → -0.33%), and selling pressure appears to be running on empty.

Now everyone’s waiting for the end-of-month earnings report, and the market is focused on one thing: how tough the guidance is. Just looking at Walmart tells you that beating revenue expectations doesn’t matter—if growth and guidance aren’t up to par, it still gets hit. The news flow hasn’t stopped either: Jensen Huang’s daughter appeared at a robotics conference in Beijing; Physical AI is laying out its groundwork; and Google has also confirmed its next-generation liquid-cooling plan.

On levels, 215.7–216.2 looks like the short-term “bottom.” If it breaks, then 209–212 could be next. If it manages to reclaim 222–225 before the earnings report, it would suggest someone ran ahead and grabbed positions early—then the outcome on the ground could actually be easier to turn positive once the dust settles.

My take: before the earnings report, if it’s just low-volume sideways action, don’t go messing around. Betting on a one-way move with a heavy position is basically handing over money—wait for the earnings report to set the direction. Not investment advice, of course.

Walmart smashed the Dow. At the close of U.S. stocks on August 20, it was trading at $103.84, down 9.15%, posting its largest single-day drop since May 2022. Its market value fell to $826.4 billion, and the Dow was dragged down, plunging 703 points to close at 52,759.21 points.

Strangely, the earnings report itself wasn’t bad. In the second fiscal quarter, revenue came in at $187.9 billion and adjusted EPS at $0.81, both beating expectations. But the market wasn’t buying it, because U.S. same-store sales growth was only 2.6%—the slowest in six years. Pressure on pharmacy business pricing also held back results, and the company’s full-year profit guidance came in below expectations. Pretty revenue figures aren’t enough—growth rate is the real lifeline.

The one bright spot was China. Net sales were $7 billion, up 20.7% year over year, and Sam’s added 11 new stores over the past 12 months. Walmart can’t even prop up same-store growth in the U.S., making the chill in offline retail there clearly visible.

Moderna saw a dramatic single-day surge of 176.97% overnight, with its market cap increasing by about $44 billion in one day. The trigger point was its collaboration with Merck, known as the mRNA personalized melanoma cancer vaccine—after the first positive results from its Phase III clinical trial. For high-risk patients, the risk of post-surgery recurrence dropped significantly. With evidence backing it up, this became the second growth curve after the COVID-19 vaccine wave began to fade.

This move directly left the short sellers stunned: S3 Partners estimates that short sellers lost roughly $5 billion in a single day. The biotech/biopharmaceutical sector took off as well— the Nasdaq Biotechnology Index jumped more than 6%, recording the largest single-day gain in six years. On the same day, the U.S. Treasury Department announced during trading that it would expand its long-term bond repurchase program, bringing an end to the three-day streak of declines in the U.S. stock market; however, Nvidia, memory, and optical communications were still falling, indicating that investor enthusiasm was clearly shifting toward the medical/biopharma space.

The narrative is that mRNA technology is moving from infectious-disease vaccines into cancer treatment—and that story market participants bought into first.

AI 老二 surpassed老大. Anthropic’s Q2 revenue hit $11.6 billion, doubling quarter-over-quarter for the first time surpassing OpenAI’s $6.7 billion; its operating profit also turned positive for the first time. Annualized revenue climbed to $65 billion—more than 6 times higher than at the end of last year. The company then immediately said it plans to launch a US stock IPO within a few weeks, and it also increased its revolving credit facility from $2.5 billion to over $10 billion—making its intention to get a head start on listing clear.

The contrast is even more striking: OpenAI’s Q2 revenue was $6.7 billion, up 18% quarter-over-quarter, but its operating loss widened to $12.3 billion instead, prompting investors to publicly express disappointment. With ChatGPT’s name recognition the highest, its profitability efficiency is instead being matched and even overtaken; both companies are preparing for an IPO, and whoever goes first gets there first to make the money.

This also means the AI race has moved from competing on models and financing to competing on financial statements. The one that turns a profit first gets to shore up, while those that keep burning cash are becoming increasingly passive.

Well well, Nvidia cut the financial guarantee for OpenAI’s Ohio data center project from $250 billion down to less than $120 billion—more than a half—just like that. The revised guarantee mainly covers the first phase’s capacity of about 5 GW, aiming to ease investors’ concerns about their risk exposure.

While cutting the guarantee, Nvidia was also in talks to invest up to $3 billion in SB Energy, a unit of SoftBank (with $1.5 billion at signing and another $1.5 billion after SB Energy goes public). SB Energy is the developer of the Ohio project. Nvidia wants to use OpenAI’s compute purchases to drive demand for its own chips, but it doesn’t want to put its entire balance sheet on the line.

OpenAI, meanwhile, isn’t standing still either—there have been nearly five rounds of reorganization within the year, and an IPO, according to reports, has been pushed to next year. In this round of AI infrastructure financing, big players are starting to control risk collectively.

#termmax @TermMax After a long wait, it finally comes to light—TermMax is finally here. TGE on August 25. $TMX

With the official launch, it also signals that what it built over the past few years—fixed-rate infrastructure—has finally started to run in real use cases.

From 90M+ TVL, to 1.5M+ wallets, to continuous expansion onto new chains, the institutional market, bStocks, options, and yield strategies—the momentum over the past 100 days really has been impressive.

The interesting part about TermMax is this:
It’s not merely building another lending protocol—it’s trying to bring the financial logic of “fixed interest rates, fixed terms, and controllable risk” onto the blockchain for real.

If, going forward, institutional capital, RWA, and on-chain credit continue to grow, the room for imagination in this track may be even larger than plain DeFi lending.

On the early morning of August 13, DeepSeek V4 Pro’s official release launched and was updated to the API, with version number 0813. In the official evaluations, it outperformed Claude Opus 4.8 on Agent tests such as HLE and Terminal Bench, and in multiple areas it came close to Anthropic’s Fable 5—some leaderboards even saw it overtake directly. It also collided head-on with Grok 4.6, which Musk released the evening of August 12.

What really caused an uproar was the pricing. The official announcement said that, in the near term, the overall API pricing for the entire lineup will be raised. Under peak-and-off-peak pricing, the call price doubles during peak hours: for V4 Pro, the output price per million tokens increased from 6 yuan to 12 yuan, with peak time defined as 9:00–12:00 and 14:00–18:00 each day. After a two-year price war among domestic language models, DeepSeek is taking the lead in calling for price hikes.

After the increase, with output charged at 12 yuan per million tokens, it’s still far cheaper than Grok 4.6’s 6 USD. If the model is strong enough, it dares to charge. Chinese AI is shifting from competing on price to competing on intelligence, and Agent and programming scenarios have become the new battleground.

Well, it’s reported that NVIDIA is developing a new generation of the open-source model Nemotron 4, with a parameter scale of at least 1 trillion. The target is to go straight for the world’s leading open-source models. Chip giant joins the fray to ‘outdo’ in open source—yet the算盘 is still the old one: by opening up the ecosystem and expanding the AI application landscape, they can then keep driving demand for their own GPU computing power. On August 11, NVIDIA’s stock fell 0.02%; once the news came out, the share price hardly moved.

At the same time, Huang Renxun cooled down a $500 billion AI funding plan he posted on X: NVIDIA’s support would be no more than 25% of the opportunity for a single project. Based on residual value, it would only fill gaps and not replace independent underwriters. The market had previously worried that NVIDIA would take on too much risk exposure in this big plan. With the clarification, the relevant credit risk indicators dropped on August 11 in response.

The shovel-seller starts by burning the models themselves, but the ledgers are still counted in terms of computing power.

After watching Velvet’s action lately, it feels like it’s getting less and less like a simple DeFi trading tool.

In the past, everyone understood it as an on-chain trading terminal—swap, view tokens, do perps, manage positions.

Now it feels more like a SocialFAI trading entry point, bundling AI, social signals, smart money tracking, and trade execution together.

VelvetX is the most worth watching line recently.

Now, on-chain players are basically constantly switching between opportunities on X, Telegram, Dexscreener, wallet-tracking tools, and trading dapps.

What VelvetX wants to do is compress the steps of spotting opportunities, analyzing them, and placing trades into a single product.

When you see a token, you don’t just know it’s going up—you can also see who’s paying attention, who called the trade first, whether smart money is in, how the AI analyzes it, and then you can trade directly.

So its positioning actually goes beyond what SocialFi talks about and follower relationships—the core is solving which signals are truly valuable.

In crypto, information advantage is already part of trading. If it can help you judge signal quality and then connect you to the trading entry point, that direction has imagination.

Another highlight is Gems.

Trading, staking, recommendations, and daily active use all become Gems. Then use Gems to correspond to the $VELVET reward mentioned below— the more active you are, the more likely you are to get more allocations.

The Epoch 11 cycle is extended to three months, with distribution on August 10, which turns short-term “farming” into a long-term behavioral competition.

But the problem is also here: are these users here for the rewards, or because the product is genuinely useful and keeps them around?

Also, Velvet has brought in pre-IPO markets like SpaceX, OpenAI, and Anthropic.

This move doesn’t just feel like it wants to do token trading—it seems to want to bring more risk assets into on-chain trading scenarios.

In the future, you won’t only trade memes, DeFi, and AI tokens—you’ll be able to trade a broader range of assets in the same terminal.

Of course, pre-IPO or synthetic exposure doesn’t mean you’re buying company equity. It involves product structures, liquidity, settlement and liquidation risks, and regulatory risks—but in terms of narrative, it really does expand the boundary from DeFi tools to on-chain brokerage.

So right now, what I’m watching for is whether Velvet can run a closed loop:

Social discovery brings traffic, AI helps users judge, the trading product completes execution, Gems keeps users, and finally $VELVET captures the platform’s growth.

Most of the main narratives around generic perps still revolve around traders: opening positions, closing positions, boosting volume, and earning points.

Hertzflow is different. It splits users into several groups: traders, LPs, Vault users, and nodes. If the Permissionless Market really takes off later, it will add another category of people who organize liquidity around the new market.

Its positioning is a self-custodial RFQ perp. The price path is validated via an Oracle. Instead of placing orders and waiting for an order book counterparty, it interacts directly with LP pools at the price confirmed by the Oracle.

Execution also isn’t a single step that combines order placement and settlement. It’s split into two phases: the user first submits through the ExchangeRouter, and the funds go into the corresponding Vault;

then a keeper triggers execution, bringing the Oracle price to complete settlement.

By separating “request creation” from “price execution,” it reduces the space for being front-run.

Hyper Lev is more like it has redone the fee structure: the open/close position fee is 0, but funding, borrow, and price impact are still there.

It doesn’t charge profit share on losing trades; profitable trades are split with the protocol in layers based on ROI.

It also requires stop-loss to land between -80% and -30% PnL—covering only part of the market. It’s clearly designed for players with high confidence, short holding cycles, and high leverage.

On the LP side, it’s not just depositing money to earn subsidies; it’s about acting as the counterparty to traders.

When traders lose, the pool goes up;

when traders win, the pool has to pay.

Returns mainly come from trading fees, borrow interest, and trading activity. The risks are amplified by variables like traders’ unrealized profits, utilization rate, and OI bias.

In extreme cases, if traders’ unrealized profits get too large or the pool’s utilization is too high, LP withdrawals may even be temporarily restricted.

So compared with “pulling users with a high APY,” this is more grounded—but it also puts much more pressure on risk management.@HertzFlow

The biggest room for imagination is still the Permissionless Market

Its direction is: as long as there is a reliable Oracle, long-tail assets like niche tokens and RWA theoretically have a chance to be turned into perp markets.

Placed in the BNB Chain ecosystem—where both users and assets are abundant—the real value isn’t just adding a few more trading pairs. It’s about who can organize the new market’s liquidity and trading community first. Whoever does that is more likely to become the on-chain trading entry point for the market.

But the foundation is still LP risk and Oracle security. If the base isn’t stable, even a smooth narrative won’t be enough to get it moving.