Aston在下不求

$NVDAB $GOOGLB
Marvell’s earnings and guidance beat expectations, yet its stock drops more than 7%? Google’s $100 billion order won’t ramp until 2029

Marvell Technology delivered better-than-expected earnings and guidance, and raised its revenue outlook for fiscal 2027 and 2028 for two consecutive quarters. Although the company did not meaningfully increase its long-term targets, Google’s $100 billion order will not bear fruit until 2029, leaving some investors disappointed who had expected faster growth. The stock fell as much as 7% during Thursday’s trading session.

Marvell’s earnings and guidance both beat! Raises long-term outlook for two straight quarters
The U.S. data-center chipmaker Marvell Technology reported its fiscal 2027 second-quarter results: revenue of $2.739 billion, a record high, up 37% year over year. According to the company, data centers are its main profit engine. Revenue from that segment grew 46% year over year and accounted for 79% of total revenue, higher than 74.4% in fiscal 2026.

Why couldn’t Google’s $100 billion order lift the stock?
Reuters reported that CEO Matt Murphy said that the company’s custom revenue target for fiscal 2028 already includes part of the revenue from that deal. A substantial ramp will not occur until fiscal 2029. He also said that revenue from custom chips will more than double next year. The prior target of “more than $10 billion” for fiscal 2029 still has room to be revised upward, but he declined to provide a new target.

Driven by AI this year, Marvell (MRVL) stock has already risen nearly 189%, and it was added to the S&P 500 in June, drawing passive funds and additional buying.

Bob O’Donnell, chief analyst at TECHnalysis Research, said expectations for customized AI chips have been pushed very high in the market, and he believes expectations are generally running ahead of reality.

AI shifts from training to inference—custom chips are in demand
The motivation for big tech companies to design their own chips is that, compared with processors from Nvidia that are expensive and face tight supply, their own solutions are cheaper to build and run. This trend has made Marvell’s custom silicon intellectual property (IP) and ASIC business one of the winners amid the data-center buildout wave.

Another tailwind is a change in workload structure. As AI application focus moves from model training toward inference, custom chips often deliver better performance-per-task and energy efficiency than general-purpose processors, driving demand upward as well.

This article—“Marvell’s earnings beat expectations, yet its stock drops more than 7%? Google’s $100 billion order won’t ramp until 2029”—first appeared on .

Silicon Valley private schools don’t just host fundraising galas—they even build their own venture capital funds! Early investment in Snap returned over 2,000x

In the United States, school fundraising commonly includes parent donations, charity sales, auctions, charity runs, and various campus events. But in parts of Silicon Valley, some top private schools are taking it one step further: they’re bringing parents’ and alumni’s venture capital expertise directly into the school itself. According to a report by Fortune, Crystal Springs Uplands School, Saint Francis High School, and Menlo School have all established venture-capital funds in different forms. The most astonishing case involves Saint Francis: it invested just $15,000 in Snap, before the company was public, and ultimately generated returns of more than 2,000x for the school (about $34 million). What these schools truly have may not be only more donations, but rather Silicon Valley’s hardest-to-get resources—networks, expertise, and investment opportunities.

Crystal Springs established a growth fund; venture-capital parents bring investment opportunities into the school. Crystal Springs Uplands School’s middle school is in Belmont, while its high school is in Hillsborough—well-known private schools on the San Francisco Peninsula. The school currently has about 569 students, and for the 2026–2027 academic year, annual tuition is $68,600.

In the eyes of many friends around me, this school has long been considered one of the top private schools on the Silicon Valley Peninsula. But what really caught my attention wasn’t its acceptance rate—it was how the school has recently leveraged its parent community.

Crystal Springs set up its own Crystal Growth Fund, or Crystal growth fund. This money isn’t invested using tuition paid by students, nor is it redirected from the school’s regular operating funds. Instead, parents and alumni donate separately, and parents and alumni familiar with investment markets help identify opportunities to invest in private companies.

People involved include executives and investors from well-known Silicon Valley venture firms such as Lightspeed Venture Partners, Notable Capital, and Sequoia Capital. Some parents working at venture capital firms can even allow the school to participate using a small portion of the investment allocation they originally secured.

As of the 2023 fiscal year, Crystal Springs’ publicly available information indicates it had not invested in private markets; but by June 2025, among the school’s investment portfolio of about $61.1 million, roughly $1.75 million had already been invested in private equity.

$1.75 million may not seem huge relative to the overall fund. But in my view, what truly matters isn’t the number—it’s that the school has already obtained a “ticket” to the private investment market. And this ticket is often not something you can buy just by having money.

Saint Francis invested in Snap—returns up to 2,000x.

One of the most classic examples of this model appeared more than a decade earlier in nearby Mountain View. Saint Francis High School established its own growth fund in the 1990s. It was initially funded with about $250,000 from two parents working in venture capital, and investment opportunities were then identified with help from parents and alumni familiar with technology startups.

In 2012, one of Lightspeed Venture Partners’ founding partners, Barry Eggers, noticed that his child and classmates were using a then-new app in large quantities: Snapchat. Lightspeed later became one of Snapchat’s early investors, and Eggers also enabled Saint Francis’s growth fund to invest $15,000.

Five years later, Snap went public. In its official materials, Saint Francis currently states that the $15,000 investment ultimately generated returns of more than 2,000x for the school (estimated at about $34 million, assuming the entire amount was sold at the IPO price).

That money was later used to increase student tuition assistance, build a new innovation center, and provide bonuses for faculty and staff. According to Fortune, Saint Francis currently invests in about 10 companies per year, with each investment typically ranging from $25,000 to $50,000. Eggers estimates that cumulative investment returns created since the fund was established are approaching $50 million.

Snap is, of course, an extremely rare success story—and you can’t assume that investing in any new startup will yield 2,000x returns.

But what this really changed is that the school began to realize: what parents can donate to the school isn’t only money.

Menlo also set up a venture-capital school fund; Nueva can’t be directly equated.

Menlo School, located in Atherton, also created a similar setup. Based on the latest publicly available information, Menlo’s venture-capital school fund holds about 36 venture funds and early-stage company investments, totaling under $1 million. By comparison, as of June 2025, Menlo’s overall endowment is about $118.4 million, so the venture-capital allocation still represents only a small portion.

However, among Menlo’s investment oversight team are people from investment firms such as Bessemer Venture Partners, Scale Venture Partners, and Sobrato Capital.

This is actually more notable than how much money was invested.

One phenomenon I’ve seen living in Silicon Valley is that many private school parents here are, in themselves, quite unusual. One parent might work at an artificial intelligence company; another might be a senior executive at a tech firm. Some have founded companies, while others spend their days evaluating which startups are worth investing in.

So even if it’s the same kind of parent meeting, in other places they might talk about…

$AAPLB
Idle Mac to make money? Darkbloom decentralized AI inference tutorial + earnings

That Mac of yours sitting around gathering dust might be missing out on monthly income of over a hundred dollars. Built by Eigen Labs—the team behind EigenLayer—Darkbloom turns idle Apple Silicon Macs into compute nodes for a “decentralized AI inference network.” When others pay to call AI models, your Mac helps do the computations, and you get paid. The project has recently been upgraded to support paid providers on OpenRouter as a model gateway. According to the official X post, there are currently 250 Macs online, with a cumulative 4.5 billion tokens serviced. For node operators, each Mac can earn between $120 and $200 per month. This article explains what it is, how to join, how much you can realistically make, and what risks you should watch for.

What is Darkbloom? Turn an idle Mac into an AI inference cloud
Darkbloom (open-source project codename d-inference) is a “private inference network” built specifically for Apple Silicon. The idea is straightforward: since 2020, the world has shipped over 100 million Apple Silicon devices. On average, they’re idle for about 18 hours per day—and these M-series chips are already well suited for running AI. Darkbloom pools this idle compute power and provides services externally through an OpenAI-compatible API. Developers don’t need to change their code to connect; Mac owners earn revenue by lending compute. For users, its selling point is “cheap”—official pricing is about half that of mainstream API providers.

How it works: three roles in the network
The whole network consists of three roles:
- Developers (consumer side) send inference requests via an API compatible with OpenAI/Anthropic.
- Coordinators run by Eigen Labs handle identity verification, routing, billing, and hardware attestation—but they can’t see the contents of the requests.
- Providers are the people lending their Macs. They run the models on the Apple Silicon GPU using the MLX framework, then send back the results.

Privacy is the core of the design
Requests are end-to-end encrypted before transmission (using NaCl Box, X25519 + XSalsa20-Poly1305). Even if a provider’s machine receives the compute work, it can’t read your prompts or responses. Inference also runs in-process, with debuggers and memory reading blocked, and hardware attestation is performed via Apple’s Secure Enclave. The official promise is: “The coordinator can route, and the provider can compute—but neither can obtain the usable request content.”

Which Macs can join? Hardware requirements at a glance
Be sure to note a gap: the technical “minimum requirement” and whether you can actually run the currently supported models are two different things.

| Project | Technical minimum (open docs) | Practical recommendation (runs current models) |
|---|---|---|
| Chip | Apple Silicon (M1 and above) | M1 Pro/Max/Ultra or newer |
| Unified memory | 8GB | 48GB or above (based on existing model needs) |
| OS | macOS 14 (Sonoma) or later | Latest stable version |
| Disk | 50GB available | 100GB or above |

In other words, an 8GB entry-level Mac can install the software, but it can’t run the models currently online. The official documentation also states that the main models at this stage require unified memory of 48GB or more—this is why high-memory machines like the Mac Studio are better suited as providers. Models are loaded only when they can fit after subtracting roughly 2GB of headroom from available memory.

How to become a provider? Installation and setup steps
On the provider side, there’s a Swift command-line tool. Installation and startup are very simple:
1. Install: run the official installation command in Terminal. It verifies file hashes and code signatures, and provisions Secure Enclave helper components.
2. Start: darkbloom start runs as a background service (launchd) and opens an interactive model selection. Adding --foreground runs it in the Terminal’s foreground.
3. Login and bind your account: darkbloom login uses a device code flow to link your account, which is required to claim earnings.
4. Check status: darkbloom status shows your configuration, hardware, and trust decisions, while darkbloom doctor performs local diagnostics.

The config file is at ~/.config/darkbloom/provider.toml, where you can adjust an availability schedule.

There are also two advanced modes:
- Add the header X-Darkbloom-Route: self to use only your own machine (free, no paid relay node).
- Use darkbloom start --local to run an OpenAI-compatible endpoint directly on your local machine, skipping the coordinator—lowest latency, and it can be used offline.

Can you really make $120 to $200 per month?
The range given in the official X post is “$120 to $200 per Mac per month,” and it recommends running the Gemma 4 26B model. To understand this number, you need to know three variables:

Revenue side: In the public alpha stage, the platform takes a 0% cut, and 100% of inference revenue goes to the hardware owner.
Revenue depends on your “duty cycle”—the percentage of time you’re willing to let your machine accept jobs. The default is 5%, but you can adjust it up to 100%. The more jobs you accept, the higher your potential income, but you’ll also use more electricity and stress the hardware. The actual amount you earn also depends on...

Stablecoin Oversight Compared Across Five Jurisdictions: MiCA or GENIUS—Who’s Tougher? BIS Breaks It Down

As countries increasingly bring stablecoins under regulation, the same term—“stablecoin issuer”—can mean very different things in different jurisdictions. In August 2026, the Bank for International Settlements (BIS) Financial Stability Institute (FSI) released FSI Briefs No. 33, which offers a cross-jurisdictional comparison of rules for stablecoin issuers in the EU (MiCA), Hong Kong, Singapore, the UK, and the United States (GENIUS Act). The report highlights a key loophole: most of these restrictions apply only to the “issuing entity,” not to the entire group. Market stagnation, a duopoly: about $320 billion, with two players taking 90% The report first describes the current stablecoin market. The total market value of stablecoins has shifted from rapid growth to stagnation; since October 2025 it has hovered around $300 billion to $320 billion, still accounting for only about 7% of the overall crypto market. The market is also highly concentrated: two issuers account for roughly 90% of stablecoin market value—Tether’s USDT and Circle’s USDC. The report focuses on “money-like” payment stablecoins, meaning those pegged one-to-one with fiat currency and designed to be redeemable at face value. Who can issue? Different thresholds for banks vs. non-banks, locals vs. foreigners Whether a stablecoin can be issued depends on the local regulatory framework. The report notes that banks and non-banks follow different paths. In the EU, banks only need to notify the competent authority, while Hong Kong and the UK and the US require separate stablecoin-specific licenses; moreover, the UK and the US require banks to issue via subsidiaries. For non-bank issuers, in the EU and Singapore, holding payment licenses is sufficient; in Hong Kong and the UK/US, stablecoin-specific authorization is required across the board. Cross-border issuance rules are even more crucial. Most jurisdictions do not allow a foreign-registered issuer to issue directly in their territory. Hong Kong is the only exception, allowing a foreign “authorized institution” (i.e., a bank) to issue through its Hong Kong branch after obtaining a license from the Hong Kong Monetary Authority. The US sets a transition period: after July 2028, foreign stablecoins can only be issued by licensed entities that are part of “designated countries,” and they must register with and be supervised by the US Office of the Comptroller of the Currency (OCC). For “systemic” stablecoins of sufficient size, the EU designates them as “important” via the EBA; the UK designates them as “systemic” via the Treasury and then subjects them to supervision by the Bank of England; the US uses a threshold of more than $10 billion in circulation, requiring transfer to federal supervision within 360 days (with possible exemptions). Core activities: custody and redemption fees differ; all five prohibit paying interest On core activities—issuance, custody of reserves, and redemption—overall directions are similar, but details diverge. The ability to “self-custody” reserves is the most obvious difference: the EU and the US allow it (subject to segregation requirements), Singapore directly prohibits it, and the UK sets a 20% cap on intra-group custody. Redemption fees also differ: Hong Kong, the UK, and the US allow charging reasonable fees, while the EU prohibits redemption fees in all cases except under stress events, to ensure redemption at face value. Redemption timelines vary as well: Singapore is most permissive at up to five business days; the US is proposed as two business days; Hong Kong is the next business day; and the EU requires redemption to be possible “at any time.” One point is identical across all five jurisdictions: they all prohibit issuers from paying interest to holders. This reflects a shared policy choice—to separate payment stablecoins from “interest-bearing investment products,” preventing stablecoins from being used as de facto investment vehicles. The UK’s FCA final crypto rules take the same stance. Non-core activities: lending, pledging, and custody—three regulatory models The real gap comes down to whether issuers can do “business beyond their core,” such as lending, pledging collateral, proprietary trading, or providing custody of crypto assets for third parties. The report summarizes two models. One is a “restrictive” model: it tightly boxes the issuer’s activities, achieved in Singapore through explicit bans and in the US via the GENIUS Act’s “positive list” approach (if it’s not listed, it can’t be done). The other is an “obliging/constraint” model: it does not ban diversification, but each new line of business requires additional authorization or must meet the corresponding industry regulatory requirements. Hong Kong, the UK, and the EU follow this approach. For bank issuers, since they are already covered by existing prudential supervision, restrictions on these activities are largely replaced by the bank framework—effectively an exemption. BIS’s key warning: restrictions apply only to the issuing entity; non-banks may use group structures to circumvent them The report’s sharpest point is that all these activity restrictions apply only to the “issuing entity” itself—not to the entire group. For banks, this gap has been filled through consolidated supervision and mechanisms such as intra-group exposure limits. But for non-bank issuers, there is no equivalent group-level framework. As a result, restrictions can be bypassed through corporate structures: the prohibited activities are moved to a sister company within the same group. The report therefore suggests that stablecoin frameworks, or related prudential requirements, may need to extend oversight to the group level for non-bank issuers—especially for larger groups. The report’s final reminder: stablecoins can circulate freely across borders, and the market is highly concentrated among a few issuers, which amplifies both regulatory arbitrage and the difficulty of cross-border supervision. This is why BIS emphasizes cooperation among national regulators and alignment with international standard-setting bodies—the key to making stablecoin oversight truly effective. For regulators in various jurisdictions that are currently drafting and implementing stablecoin rules, this comparison provides a clear side-by-side reference table—and points to the loophole that should be addressed next. This article Stablecoin Oversight Compared Across Five Jurisdictions: MiCA or GENIUS—Who’s Tougher? BIS Breaks It Down First appeared on .

$TRX $NVDAB $MU
Should we hoard memory or gold? Data reveals that 1 kilogram of DRAM is worth 620 grams of gold, and it has increased 12-fold over three years

AI demand is squeezing HBM capacity. South Korea’s DRAM export average price has risen more than 4 times year-on-year, and surged 12.5 times from the 2023 low. Global investment banks and major memory manufacturers said: “High prices not only have not eased the shortage; the supply gap next year may be even more severe than this year.” (Is all the money going into memory? A research note: In 2027, memory will account for nearly 70% of cloud giants’ capital expenditures.) How much gold is 1 kilogram of DRAM worth? It is currently worth 620 grams of gold.

Gold has long been one of the standards for measuring wealth. According to data from the Korea Trade-Investment Promotion Agency, from August 1 to 20 this year, the average export price of South Korea’s DRAM (excluding modules) reached 127.4 million won per kilogram (about US$92,183). Over the same period, export value grew 504.8% year-on-year to 13.56 trillion won (about US$9.807 billion).

Converted, 1 kilogram of South Korea–made DRAM is worth about 620 grams of gold. That is equivalent to 1.53 times the value of platinum of the same weight, or 41.7 times the value of silver.

After analyzing South Korea’s DRAM export statistics, research institute Sitri Research said the current average export price is about 12.5 times the historical low in January 2023. During the same period, gold’s price increase has reached 2.4 times, but comparatively it is still far behind.

However, DRAM is different from gold. It cannot be priced simply by weight, because performance and specifications vary across product generations. The mainstream spot product tracked by the market at the beginning of 2023 was DDR4 8Gb, while the benchmark currently used by TrendForce is DDR5 16Gb. As process shrink allows more data to be packaged on chips of the same area, it naturally raises the per-kilogram conversion unit price, which has no direct relation to market supply and demand.

Industry insiders openly admit: “Even after accounting for product-generation upgrades, the rate of increase in recent memory prices remains steep.” Data from the Bank of Korea also shows that in July the year-on-year increase in DRAM export prices reached 270.3%. Micron’s earnings report also revealed that Micron’s average selling price (ASP) for DRAM in the third quarter of fiscal 2026 grew by about 60% compared with the previous quarter.

Why can’t high prices fix supply?
HBM is “consuming” capacity

Goldman Sachs’ latest forecast indicates that the DRAM supply shortage rate will widen from 5.0% this year to 5.9% next year. Compared with the previous forecast, the estimated value for 2027 has been raised by 3.4 percentage points, and it expects that DRAM supply shortages will continue through 2028.

Memory market prices have hit record highs repeatedly, yet the shortage is becoming even more severe due to HBM’s displacement effect across the overall DRAM ecosystem. HBM is manufactured by vertically stacking multiple DRAM chips. Compared with producing the same amount of commodity DRAM, this requires far more wafer input.

According to TrendForce’s estimate, by the end of next year, HBM will account for 30% of the DRAM wafer投入 (wafer investment) by Samsung, SK hynix, and Micron. But in terms of actual conversion into bit supply, it will account for only 13%. This means that once large-scale capacity flows into HBM, the supply space available for commodity DRAM used in PCs, smartphones, and general servers will be significantly compressed.

Goldman Sachs estimates that the global HBM market size next year will expand by an additional 108% compared with this year, and this trend will not ease in the near term.

The expansion plans of the three major players can’t keep up with demand; 2027 may become the most shortage-ridden year in history

Memory manufacturers are not failing to expand capacity, but the problem lies in the time lag between starting new plants and when the capacity actually ships. TrendForce points out that given the time needed for new factories to ramp up and stabilize production processes—combined with continued growth in demand for HBM and server DRAM—next year’s supply is highly likely to still fall short of demand.

All major companies have expressed consistent views. In June this year, Micron said that tight DRAM and NAND Flash supply-demand conditions driven by AI demand and structural supply constraints will continue beyond 2027. Last month, SK hynix president Kwak Noh-jung went further, saying that 2027 is likely to be the most severe shortage year in the history of the memory industry.

The market has already begun to respond. Some of Nvidia (Nvidia) and cloud service providers (CSPs) are evaluating adjustments to the HBM configuration or installation base of next-generation AI accelerators to reduce reliance on constrained capacity.

(“Nvidia plans to lower the Rubin Ultra HBM specifications: mass production of HBM4e is not confirmed yet; DRAM shortage is the main reason.”)

When will the shortage end? Three key variables determine the outcome

The duration of this memory shortage will ultimately depend on three yet-uncertain variables:
- how much and for how long HBM capacity expansion will squeeze other DRAM supplies;
- when the new capacity planned by each company will actually be converted into shipments;
- whether rising high prices can, at some critical point, suppress end-user demand and trigger self-correction on the demand side.

Current market consensus suggests that the negative effects of the first two variables will not see any clear easing before 2027. The third variable is uncertain, especially since the investment heat in AI infrastructure has shown no signs of cooling.

This article “Should we hoard memory or gold? Data reveals that 1 kilogram of DRAM is worth 620 grams of gold, and it has increased 12-fold over three years” first appeared on .

$BTC $ETH
Machi Big Brother angrily calls out the media for fake news? In five days, he earned nearly 300 million TWD—here you go

Machi Big Brother (Huang Licheng) today on X posted two messages and went off on Taiwan media. One cited ETtoday’s “Uncle makes good money, goes lazy X…” and the other captured a TV news segment and wrote, “Taiwan news says my portfolio up 84x. Fake news.” Is it true or not? Let’s break down the on-chain data and show everyone.

Taiwan news headline says that I have earned several dicks. pic.twitter.com/gwbcTr0mbL — Machi Big Brother (@machibigbrother) August 26, 2026

According to the on-chain data from hyperbot, before this market rally started, the account for this “sustainable contract” that Big Brother used had almost been wiped out—only the original 1.4% principal remained. Over the next five trading days after 8/17, that remaining 1.4% was rolled out to nearly 300 million TWD, a 29x increase!

Below I’ll restore the actual numbers point by point—how much this Big Brother cashed out, and where the leverage and risks in his current position are. (For easy calculation, the following figures are automatically converted to TWD, using an exchange rate of 1 USD / 31.8477 TWD.)

Crypto currency is volatile; unrealized gains/losses can make media headlines. “Taiwan news says my portfolio up 84x. Fake news.” pic.twitter.com/X7XKVqmGdq — Machi Big Brother (@machibigbrother) August 26, 2026

Based on the media’s algorithm, Big Brother’s account net value peaked at 390 million TWD at noon on August 24, while the mid-August low on August 14 was only 11.45 million. Dividing the two yields roughly 86x, which is close to the 84x mentioned in the news.

However, to assess an account’s real performance, looking directly at cumulative deposits and withdrawals is more intuitive and practical. According to on-chain records, this address has cumulatively received 1.132 billion TWD and cumulatively withdrawn 292 million TWD, for a net investment of 841 million.

These funds moved in and out over more than three months, during which the account’s net value fluctuated wildly. Just in August, it went from 2.22 million on the 1st all the way to 390 million on the 24th—an 175x difference.

So it can only be said: let Big Brother bring everyone to experience the beauty of the crypto market. Big Brother has a big principal and big leverage—and on top of that, crypto is extremely volatile. No matter which angle you look at it from, it’s astonishing, and that’s where the media headlines come from.

Losing 99% of principal, leaving 11.45 million: rolling out 295 million in five days

If you put Big Brother’s net investment of over 800 million next to it, on the day BTC started rising on August 17, only 1.4% of the principal remained in the account. In other words, before this rally began, roughly 99% of the funds he put into this contract account had evaporated—leaving only his last breath.

Based on the trade records, after August 17 there was a dead-cat bounce. So far he has realized and cashed out 291 million TWD. Just across the five trading days from August 18 to 22, he cashed out 295 million. Although on August 23 he then lost some money back again.

Source of the image: https://hyperbot.network/trader/0x020ca66c30bec2c4fe3861a94e4db4a498a35872

Worth noting is the source of the profits. One position in Ethereum (ETH) alone contributed 215 million TWD. Platform token HYPE contributed 83.3 million, while his Bitcoin (BTC) position actually lost 30.01 million. Even now, his BTC long position is still underwater. What he really made in this run was from the upward rebound in Ethereum and HYPE—not from Bitcoin itself.

He gives back over 70 million in two days: Big Brother can afford it; regular people can’t

At the moment, all four positions are long. Their nominal values total 3.97 billion TWD. Compared with the account net value, overall leverage is about 12.5x. His Bitcoin position is opened at 40x leverage, and the withdrawable balance is zero—meaning all the margin has already been locked into the positions, with no buffer left.

Using the same playbook, it took him nine days to turn it into nearly 30x. And before that, the same playbook turned over 800 million in principal down to a little above 1%. Big Brother’s premise is that he can afford the losses—and he has already lost once. If ordinary investors copy the same leveraged approach with the same leverage, a single misjudgment is enough to get them out of the game.

In short, just the fact that he still dares to keep running high leverage even with only 1% left shows that Big Brother’s psychological resilience is nothing short of extraordinary. That kind of霸氣 (“save/hold a XX”) isn’t a normal person’s game. Hopefully Big Brother can one day become the Pirate King!

This article, “Machi Big Brother angrily calls out the media for fake news? In five days, he earned nearly 300 million TWD—here you go,” was first published on .

Pelosi’s earnings report disclosure lays out positions in AI-related stocks; heavy stake in Bloom Energy triggers a nearly 6.5% pre-market jump

In her latest congressional financial disclosure, former U.S. House Speaker Nancy Pelosi revealed that her family built positions in AI clean-energy company Bloom Energy during July, while also significantly increasing its stake in chipmaker Intel. With this investment information and strong revenue results coming in simultaneously, Bloom Energy’s pre-market share price surged by nearly 6.5%, with the stock currently trading at $217.45 per share.

Pelosi family completes multiple Bloom Energy equity allocations in July

According to financial disclosure documents made public by the U.S. House on August 24, Pelosi’s husband, Paul Pelosi, executed two common-stock transactions in Bloom Energy (NYSE: BE) in July. The first trade occurred on July 24, buying 10,000 shares at approximately $184.89 per share; the reported value ranges between $1 million and $5 million. He then added 5,000 shares on July 28 at approximately $166.84 per share; the reported amount ranges between $500,000 and $1 million.

The disclosure information also includes the purchase of 200 call options with a strike price of $100 and an expiration date of June 17, 2027. (Pelosi’s exclusive analysis: the strategy behind buying Intel call spreads and Uber calls)

Bloom Energy fuel-cell technology meets AI data centers’ high energy demand

Bloom Energy’s main business is producing fuel-cell systems that can provide independent on-site power for high-energy facilities. As global AI computing capacity expands and data center power demand surges, traditional grid supply faces bottlenecks, making on-site power generation equipment a key component in building AI infrastructure.

Notably, Intel has long been a cooperation customer of Bloom Energy’s data-center power systems. With the Pelosi family simultaneously positioning in two companies, it indicates that its investment portfolio is highly focused on AI supply chains and supporting power infrastructure.

Bloom Energy’s second-quarter revenue first surpassed $1 billion, up about 166% year over year. The company’s management team also raised its full-year revenue guidance to between $3.9 billion and $4.2 billion, indicating that the company’s penetration in the AI power market is accelerating quickly.

As the broader U.S. market strengthens, Bloom Energy shares rise

News that the Pelosi family established large new positions became a catalyst for Bloom Energy’s share price increase. Traders picked up the information ahead of the open, driving a surge in buy orders. The reference price for the stock on the previous trading day was $204.02; after the congressional disclosure, the pre-market trading jumped by about 6.4%. The stock is currently hovering around the $217 level.

The overall U.S. stock market environment also appears optimistic: the S&P 500 rose 0.4%, the Dow Jones Industrial Average rose 0.4%, and the Nasdaq index rose 0.7%.

This article: “Pelosi’s earnings report disclosure lays out positions in AI-related stocks; heavy stake in Bloom Energy triggers a nearly 6.5% pre-market jump” was first published on .

Encrypted Custody Firm Copper Seeks Sale, Asking Price Only 200 Million—Far Below the Hype Price

Consolidation in the crypto custody industry is being driven forward at shrinking valuations. According to a report from CoinDesk on August 25, the London-based crypto custody firm Copper is said to be looking to sell. While it has attracted buyer interest, its current asking price is far below the $500 million asking price from a few months ago.

Having previously been valued at $2.5 billion, Copper now has offers of only about $200 million from two to three potential buyers. Copper was once valued as high as $2.5 billion, and a few months ago it was shopped around by Cantor Fitzgerald at a price of roughly $500 million. As per the report, the company currently has two to three buyers, and all bids are in the range of about $200 million—significantly discounted from the original asking price.

Copper has raised venture funding of over $300 million during periods of high valuation. Now it faces challenges related to preferred shares, with its valuation being marked down from a peak of $2.5 billion—reflecting pressures in the crypto industry’s market conditions in recent years.

Core Asset ClearLoop: Over 1,000 Counterparties, Monthly Trading Volume Exceeding $50 Billion

Copper shut down its corporate custody business in 2023 and instead focused on ClearLoop—a settlement system for institutions that enables network participants to complete “delivery versus payment (DvP)” settlement directly in a custody environment without moving assets on-chain, thereby eliminating settlement risk. The platform currently has more than 1,000 active counterparties and monthly trading volume exceeding $50 billion. It is also the core asset that buyers are truly interested in for this potential transaction.

This article, “Encrypted Custody Firm Copper Seeks Sale, Asking Price Only 200 Million—Far Below the Hype Price,” first appeared on .

$AAPLB
Apple M5 Ultra chip debuts, AI compute up to 4.5x the previous generation

Apple is pushing its “on-device AI compute” to new heights. According to an Aug. 25 announcement from Apple’s official newsroom, Apple has introduced a new M5 Ultra chip, which it claims is the most powerful chip in its lineup to date. The focus is a major leap in AI computation and graphics performance, and it is included exclusively in a new Mac Studio unveiled the same day. First sighting of a quad-die architecture: two M5 Max chips are stitched together with UltraFusion to form the largest M5 Ultra structure. This is the first time the M series uses a “quad-die” design. Using Apple’s own UltraFusion packaging technology, Apple connects two M5 Max chips—each with two dies—into a single processor. This delivers inter-die bandwidth of more than 4.4TB/s, with connection density more than 6x higher than the previous generation, enabling the four dies to coordinate and operate like a single chip. CPU up to 36 cores, made up of 12 super cores and 24 performance cores. Compared with the M3 Ultra, single-thread performance is up to 1.25x, and multi-thread performance is up to 1.3x. A 80-core GPU with a Neural Accelerator built into each core; AI peak compute reaches 4.5x that of the M3 Ultra. Graphics and AI are the main battleground of this generation. The M5 Ultra features up to an 80-core GPU, and every core includes a Neural Accelerator to accelerate AI workloads. Apple claims its AI peak GPU compute is up to 4.5x that of the M3 Ultra, and more than 6x higher than the M1 Ultra. Graphics performance is up to 40% faster than the M3 Ultra. Paired with a 32-core Neural Engine, up to 512GB of unified memory, and memory bandwidth up to 1.2TB/s (50% higher than the M3 Ultra), this combination of high-capacity unified memory and neural acceleration is designed specifically to run large on-device AI models and generative workloads. Mac Studio starts at $5,499, ships starting Sept. 22. Alongside it, Apple also unveiled 2nm M6 and M5 Ultra chips. The M5 Ultra is currently available exclusively in the new Mac Studio. The Mac Studio with M5 Ultra starts at $5,499, while the M5 Max version starts at $2,499. Both open for preorders on Aug. 25 in 30 countries and regions including the United States, with shipments beginning on Sept. 22. On the same day, Apple also announced the M6 chip, its first 2nm process chip, targeting performance and energy efficiency for mobile devices. When on-device compute can handle multiple times the AI power of the previous generation—along with up to 512GB of memory—this trend of moving AI inference from the cloud to the desktop gains another major piece. This article “Apple M5 Ultra chip debuts, AI compute up to 4.5x the previous generation” first appeared on .

$TRX $METAB $MU
HBM: Past and Present—Looking at HBM Across Generations to Understand Bandwidth, Stacking, and Thermal Bottlenecks; Why Packaging Is the Critical Battlefield?

One of the world’s three memory makers, Micron (MU), delivered a talk at Hot Chips 2026, openly stating that AI compute grows by 3x every two years—outpacing the expansion speed of HBM bandwidth, which increases by less than 2x per year. As a result, the “Memory Wall” continues to worsen. During the training of Meta Llama 3, as much as 17% of unexpected interruptions exposed HBM defects, further intensifying the memory reliability crisis. Micron worries that without disruptive breakthroughs across three layers—process technology, packaging, and IO design—HBM’s current roadmap may face physical limits even sooner.

Key takeaways: AI compute continues to grow (3x every two years), relentlessly outstripping HBM bandwidth growth (under 2x every two years), and the memory wall keeps deteriorating. In Meta Llama 3 training, about 17% of unexpected interruptions are blamed on HBM failures. In typical AI accelerator packaging, the HBM silicon area is 8x that of the GPU. HBM4 bandwidth reaches 2,800 GB/s—15 to 17 times higher than DDR5. However, the cost of silicon area per unit capacity is about 3x that of DDR5. As density increases, efficiency worsens, and stacking higher isn’t a cure-all.

Micron is betting on disruptive packaging technologies such as Fusion Bonding, Glass Substrate, and CPO to break through. The current problem is hard for any single vendor to solve; it requires end-to-end collaboration across the entire industry chain: HBM suppliers, wafer foundries/OSATs, SoC designers, software, and academia. (Why can the memory bull market still rise for another five years? Which metrics explain the differences among HBM, DRAM, and NAND?)

Scaling Law perspective: Memory’s core role in LLM progress
Micron starts by citing OpenAI’s research paper, “Scaling Laws for Neural Language Models.” It explains the core issue with three power-law curves: LLM test loss declines smoothly as compute, dataset size, and the number of model parameters increase, but all three must expand in sync to achieve optimal performance.

Frontier figure on compute efficiency
In other words, memory is not a supporting character—it is one of the key variables driving continuous progress in LLMs. Without synchronized memory scaling, increased compute is ineffective.

The memory wall is worsening: bandwidth can’t keep up with compute growth
Based on real product data from 2019 to 2027, GPU compute keeps climbing along A100, H100, MI300X, and B200, with growth staying around 3x every two years. In contrast, HBM bandwidth evolves from HBM2E to HBM4, but its growth is only under 2x every two years.

Memory wall: the widening gap in GPU compute growth trajectories between 2017 and 2027
With the compounding effect, every two years the compute demand gap relative to memory expands by another 1.5x, and over the long term the divide becomes substantial.

Why HBM can’t be replaced: Roofline Model’s engineering logic
Micron uses the Roofline Model to illustrate HBM’s core value. In the massive computations of AI training and inference, many workloads land in the “Memory Bound” region, where the processor can’t deliver full performance while waiting for data to arrive.

Roofline Model: By raising the bandwidth ceiling, HBM lifts the roofline, enabling memory-intensive AI workloads to reach higher performance
HBM’s role is to move the overall bandwidth ceiling upward, so more AI workloads enter the compute-saturation region before they ever reach the memory limit—fully unlocking the GPU’s theoretical peak performance.

HBM six-generation spec evolution at a glance: shifting from “how much you store” to “how fast you transfer”
Since HBM debuted in 2014, the core specs of six generations have leapt significantly.

HBM six-generation specifications comparison table: covering channel count, data rate, nominal bandwidth, DRAM density, stack height, and Cube capacity, etc.
- Data rate rises from 1 Gbps to 11 Gbps: 11x growth
- Nominal bandwidth jumps from 128 GB/s to 2,800 GB/s: 22x growth
- HBM4 is the first to double the number of Data I/O channels from 1,024 (the first five generations) to 2,048—one of the main reasons for the major bandwidth leap.

Worth noting: the single-die DRAM density of HBM4 and HBM3E is both 24 Gb. HBM4’s bandwidth jump comes from doubled IO and increased data rate—not from higher die density. This indicates that the competitive focus across HBM generations has shifted from “how much you store” to “how fast you transfer.”

Silicon area is 8x that of the GPU: HBM’s real role as the main body of AI accelerators
Micron also points out that in a typical AI accelerator packaging equipped with 8 HBM4 chips, the memory silicon area is 8x that of the GPU. In other words, HBM is the largest physical component by area in an AI accelerator package; the GPU die is actually the minority.

Silicon area comparison: the left side shows an SiP (System-In-Package) packaging where 8 HBM4 chips surround 2 GPU die; the right side compares the silicon area of the GPU and HBM4 (12-high). The latter exceeds the former by more than 8x.

HBM vs. DDR5: 15x bandwidth lead, but the silicon-area cost is 3x
In a typical GPU configuration, total HBM system bandwidth reaches 5.3 TB/s, while DDR5 is about 300 GB/s—a gap of roughly 16x. For single-die bandwidth, HBM3E’s 256 GB/s compared with DDR5’s 8 GB/s expands the gap further to 32x.
AI high-order access...