This morning I saw that fiat24 has stopped cryptocurrency deposits, and many people's uCards have been suspended. Honestly, the uCard is very convenient for daily life. OKX Pay is working on its own uCard, waiting for the opportunity to get one when it launches

#财报观察员：微软Meta亚马逊今夜交卷 Over the past two years, the core change in the AI wave can actually be seen from capital expenditures. Cloud providers are investing unprecedented funds to build AI infrastructure, while GPU manufacturers are on the other end reaping orders. Google's capital expenditure grew from $6.8 billion in Q3 2021 to $44.9 billion in Q2 2026, an increase of more than six times in five years; Microsoft's latest quarterly capital expenditure also reached about $30.9 billion, up from $16.7 billion a year ago. Behind these numbers is the continuous investment by cloud giants in servers, data centers, and energy facilities for AI computing power. On the other hand, Nvidia's capital expenditure has remained low for a long time, about $1.8 billion in the latest quarter, because it does not build data centers but shifts manufacturing and infrastructure pressure to the supply chain through a fabless model. Thus, the same AI revolution is creating two completely different cash flow outcomes: companies selling GPUs are quickly collecting money, while companies buying GPUs are continuously increasing their investments. From the perspective of free cash flow, this difference is even more obvious. Nvidia's operating cash flow has exceeded $50 billion, but capital expenditure is only about $1.8 billion, with free cash flow close to $48.6 billion; Microsoft’s business cash flow is still growing, but the expanding AI investments are compressing cash flow space; Google has become the market focus, with capital expenditure reaching $44.9 billion in Q2 2026, while operating cash flow is only $39.1 billion, marking the first time capital expenditure exceeds operating cash flow, turning free cash flow negative. Tesla also saw a sudden rise in capital expenditure due to increased investments in AI, robotics, and factories, with quarterly free cash flow turning negative again. This is why the market reacted noticeably after Google and Tesla released their earnings last week. The issue is not that investors suddenly think AI has no value; Google's earnings report itself was very impressive: revenue of $119.8 billion, up 24% year-over-year, Google Cloud revenue of $24.8 billion, up 82%, and operating profit reaching $40.8 billion. But the market started focusing on another sheet—the cash flow. Google invested $44.9 billion in a single quarter to build AI infrastructure, while the company generated only $39.1 billion in operating cash flow, and it raised its full-year capital expenditure forecast, stating it will continue to increase investments in the coming year. The real market concern is when the massive capital investments in the AI infrastructure war will translate into corresponding revenue returns if the competition continues to escalate. In the past two years, the market traded on the idea that "AI will change the world"; but now, investors are starting to calculate "how much it will cost to change the world." The biggest winners in AI may still emerge, but in the coming months, the market’s focus will not only be on AI revenue growth but on who is still burning cash and who has already started collecting money.

Feeling down, many people have cut losses in US stocks, feels like there will be a rebound Bottom-fished $SNDK $MU in the afternoon

$CXMT's market capitalization has exceeded 500 billion USD, making it the most valuable listed company in China. AI-driven DRAM shortages are pushing operating profit margins above 70%, but the real test is whether it can narrow the technological gap with $MU, $SKHY, and Samsung without access to $ASML EUV machines.

Samsung Expands Hwaseong Factory Scale, Plans to Increase Production Capacity by 15% by Year-End Samsung Electronics is consolidating its DRAM production facilities in the Hwaseong campus while expanding capacity to increase the supply of commodity DRAM. This move is seen as an effort to maximize commodity DRAM output—DDR4 prices surged over 80% in the first half of this year alone—and to meet the demands of major tech clients like Apple, which is considering sourcing DRAM from Chinese suppliers amid storage shortages. According to industry insiders on the 27th, Samsung Electronics' Memory Manufacturing Technology Center formed a special task force this month and began establishing a commodity DRAM back-end factory at Hwaseong Campus Plant 1 (H1). The company plans to relocate equipment currently scattered between Campus Plant 2 (H2) and the Tianan back-end factory to H1 and add production lines in the freed-up space. The back-end factory is responsible for connecting transistors and other memory components manufactured in the main factory, representing the final stage of front-end wafer processing. Historically, Samsung has expanded Hwaseong production facilities around the main factory, while some back-end operations were dispersed in Tianan. The physical separation between the main and back-end factories inevitably extended the DRAM production cycle, hindering efforts to increase commodity DRAM supply. Jun Young-hyun, head of Samsung Electronics' Device Solutions Division, recently reviewed how to utilize the cleanroom space left idle at Hwaseong H1 after dismantling old memory production lines and ultimately selected it as the new back-end factory hub. By renovating existing cleanrooms instead of building new factories, Samsung can reduce capital expenditure and construction time, enabling faster DRAM supply expansion. This reorganization is not just a simple capacity increase. Its key feature is integrating upstream and downstream commodity DRAM production processes at a single location, significantly improving efficiency. Previously, components manufactured at Hwaseong's main factory had to be transported to the Tianan back-end factory for completion, causing considerable production inefficiency. In fact, Samsung is introducing the advanced production model used at the Pyeongtaek campus to Hwaseong, where the main and back-end factories have been co-located since construction. "By the end of this year, Samsung Electronics' commodity DRAM production capacity is expected to be about 15% higher than at the beginning of the year," an industry official said. "Integrating the main and back-end factories will also enhance logistics efficiency through collective wafer transfers to downstream processing lines."

$NVDA, together with $PLTR, $SPCX, $CRWD, $PANW, $NET, $SKHY, $IBM, and dozens of other tech leaders, has launched the Open Secure AI Alliance.

#长鑫科技上市，全球存储竞争添变量 The total market value's circulating ratio on the first day is only 6.63%, essentially still a pricing game of high FDV and low circulating supply. The regulatory warnings circulating in the market seem more like controlling the trading rhythm on the first day rather than denying the company's valuation. The lesson from PetroChina back then was too deep: the listing was the peak, followed by long-term lock-up, and the negative impact remains to this day. Regulators obviously hope to first create some profit effect, then use policy tools to regulate institutional behavior, avoiding the first-day pricing hitting the top in one step. The optimism on the institutional side is not low; expectations above 3 trillion are common. Indeed, some institutions have received risk warnings from regulators or brokers, being asked not to concentrate buying on the first day of listing. But if the opening valuation is only about 2.5 trillion, a sufficiently low price may still make some funds willing to take the risk and enter. The problem is that the idea of "big money entering late" has spread too widely. Once retail investors generally believe that institutions will take over later, fewer people will be willing to sell on the first day. The low circulating supply combined with reluctance to sell may push the price up prematurely. When the valuation jumps to the target level in one step, institutions originally prepared to enter later will have no buying space. Currently, the market's favored concentrated valuation range is about 3 trillion to 4 trillion, with few views below 3 trillion. Sentiment has been fully ignited, and a valuation surge above 6 trillion is not impossible.

According to RootData's statistics, 99 crypto projects will "die" in 2026. These projects will cease operations, go bankrupt, or have their websites unavailable for a long time in 2026. This includes centralized exchanges BitMart, BitMEX, the wallet Family, and the DeFi project Zapper. I only use the OKX wallet

The cycle will definitely end, but it may not be next year Whenever someone mentions the undervaluation of $SKHYNIX, $MU, $SNOW, and Samsung, the market always responds with a rebuttal: cyclical stocks have the highest profits when their P/E ratios are often the lowest. This statement is not wrong, but it avoids the real issue. The market has long known that the current profit margins will not be maintained permanently, which is why it is only willing to assign such low valuations. The current disagreement is not about whether the storage cycle will end, but about how long this round of supply tightness can continue. If the industry peaks this year and then quickly enters a phase of oversupply, price declines, and margin contraction, then the low valuations we see now are indeed a trap. But if supply tightness continues for another two to three years, the cash these companies accumulate before the cycle reverses could far exceed what current valuations reflect. The signals released by suppliers currently are closer to the latter scenario. SK Hynix believes that some storage shortages may continue until the end of this decade. Samsung's judgment is relatively cautious but also believes that obvious supply tightness will last at least until 2027. These companies obviously want the market to believe the industry outlook remains strong, so management's statements cannot be taken at face value. However, they do have complete information that outsiders cannot access. Customer commitments, equipment orders, wafer planning, and packaging capabilities will ultimately be reflected in their capacity arrangements. Repeated orders are also a common market concern. Lower actual utilization rates sometimes mean customers are exaggerating demand, or it could mean customers are competing for limited capacity and must lock in supply beyond short-term needs in advance. It's hard to distinguish between the two just by looking at orders; contract structures may be more informative. In this cycle, some customers are willing to sign multi-year agreements, accept price floors and ceilings, provide prepayments, and even participate in funding new capacity. Short-term demand usually does not commit capital years in advance, and customers are even less likely to bear expansion risks for suppliers lightly. The manufacturing characteristics of HBM also make supply difficult to increase rapidly like in traditional storage cycles. HBM consumes wafer capacity far more than standard DRAM, and yield rates and advanced packaging impose new constraints. As the industry moves from HBM3E to HBM4 and HBM4E, the manufacturing difficulty of products continues to rise. Some of the new capacity added by suppliers is likely absorbed by the higher manufacturing intensity per product. The same logic applies to $TSM and $ASML. More advanced AI chips require leading-edge processes, EUV equipment, and advanced packaging to work together. TSMC cannot replicate a mature fab in a few quarters, and ASML's equipment delivery and customer expansion also require years of coordination. Supply can increase, but the speed may be far slower than the market imagines. On the demand side, $NVDA, $AMD, and $AVGO face another kind of skepticism. The market worries that $META, $GOOGL, $AMZN, and $MSFT are over-purchasing accelerators and custom chips. When data center construction slows, inventory, pricing, and margins will all be under pressure. This risk is real, but AI computing demand is also changing. Training still consumes a lot of resources, inference scale is expanding, and Agent and custom chip projects bring new loads. Even if the procurement growth of a certain chip category slows, new demand may continue to extend the entire construction cycle. In the past two years, bottlenecks have been shifting among different parts of the supply chain. Initially it was GPUs, then pressure shifted to HBM and advanced packaging. Later, optical connections, power, cooling systems, and data center capacity became constraints. The continuous migration of bottlenecks at least indicates the entire industry is still expanding on multiple physical levels and no clear endpoint has appeared. The cyclicality of semiconductors certainly has not disappeared. Supply will eventually catch up with demand, prices will ease, and margins will fall. What really needs to be compared is whether the timeline implied by current market pricing aligns with contract terms, expansion progress, and supplier feedback. If AI demand slows down prematurely, then today's low valuations are a warning. But if physical bottlenecks still exist and new capacity is delayed, the market may be underestimating not these companies' profitability but the duration this cycle can last. The cycle will definitely end, but not necessarily collapse starting next year as current valuations imply. #韩国存储双雄获AI双巨头大单