In late June 2026, Xiaohongshu quietly filed for a Hong Kong listing. Backed by Goldman Sachs and CICC, the move would mark one of the largest Chinese internet company IPOs in Hong Kong in nearly a decade — with a potential valuation of $30–50 billion. Note the currency: US dollars. Earlier deals by MiniMax and Zhipu were priced in Hong Kong dollars and were significantly smaller in scale.

       But shortly after the filing, a whistleblower letter from a former employee landed at both the Hong Kong Stock Exchange’s listing division and the Securities and Futures Commission. Regulators have confirmed receipt of the email.

       The sender, Chen Hao, previously headed Xiaohongshu’s South China commercial direct-sales team. He holds a court judgment that found the company unlawfully terminated his employment — plus evidence that could directly challenge the integrity of its VIE structure.

1. Xiaohongshu’s Long Road to IPO

       The company had been preparing for years. Back in 2018, a co-founder told Bloomberg it might list within two to three years. In 2021 it hired a former Citi executive as CFO and was reportedly preparing a US roadshow to raise $500–1,000 million.

       Then Didi’s troubled US listing triggered tighter scrutiny on Chinese companies. Xiaohongshu paused its US plans, along with several other high-profile names including Soul, Huolala, Keep, and Himalaya.

       For the next few years, rumors of a Hong Kong listing kept surfacing — and Xiaohongshu repeatedly denied them (the market started calling it the “three denials”). Only in 2025 did it rent office space in Hong Kong, build out a major commercial team, and clean up its financials. The secret filing finally came at the end of June 2026.

2. Why Rush to File Before June 30?

       A listing application needs audited financial statements, and those statements are generally valid for only six months. Xiaohongshu was working with its fully audited 2025 annual results. It had to submit before the end of June — otherwise the audit would expire and the company would need to start over.

       For a company of this size, redoing an audit is expensive and time-consuming. That’s why many firms sprint to file by June 30: it’s not always about market sentiment, but about a hard document deadline.

3. A “Bleeding” Listing?

        In the secondary market, Xiaohongshu shares had been bid up to roughly $50 billion in valuation. Yet the IPO range is being discussed at $30–50 billion.

       The gap exists because today’s Hong Kong market favors AI, robotics, and chip-related stories. Xiaohongshu’s core business — social commerce, livestreaming, and advertising — belongs to an earlier internet era. At this moment, it may have to accept a lower valuation to get listed while the window is still open. Market watchers sometimes call this a “bleeding IPO.”

       It’s not unique to Xiaohongshu. Many companies outside the hottest tech themes face the same pressure right now.

4. ChenHao’s Whistleblower Letter: Three Sharp Questions

       ChenHao joined in June 2022. According to him, he quickly lifted team performance, exceeded targets by 20% in 2023, won a company innovation award, and signed a 30,000-share option agreement with founder and CEO Mao Wenchao (vesting over four years: 50/25/25).

       In December 2023 the company fired him for “incompetence.” Because he hadn’t completed two years, the first 50% of options were cancelled under standard policy. He received statutory severance, but believed the dismissal lacked proper grounds and took the case to arbitration.

       Guangzhou’s Tianhe District Court ruled in his favor in both the first and second instance, finding the termination unlawful. Xiaohongshu was ordered to pay roughly 800,000 RMB (including the option portion). For a company heading toward a multi-billion-dollar valuation, the amount itself is small. The real problem is the unresolved loose end it left behind.

     ChenHao’s letter raises three main issues:

i. VIE structure clarity (the most sensitive)

       ChenHao’s employment contract was with “Shu Yi Shu Er Culture Media,” a domestic Chinese company. To defend itself in the labor case, Xiaohongshu submitted documents stating that this entity had “no direct relationship and no contractual control” with its offshore structure.

That statement touches the heart of the VIE model.

       VIE structures (first popularized by companies like Sohu) let Chinese internet firms list overseas while complying with rules that restrict foreign investment in sensitive sectors. An offshore holding company (usually in the Cayman Islands) lists on the stock exchange. It controls a domestic licensed entity through a series of agreements so that profits can flow upward to foreign shareholders.

       The critical link is that the domestic company’s revenue must be fully captured by those agreements. If one domestic entity is shown to sit outside the control chain, regulators may question whether the structure is watertight — and whether shareholders can actually receive the economic benefits they expect.

       ChenHao’s evidence points straight at this potential weak spot.

ii. Disclosure of the labor dispute

       Since a court has already ruled the termination unlawful, did Xiaohongshu fully disclose the case and its potential impact in the listing documents? Incomplete disclosure would be a straightforward regulatory issue.

iii. ESG and labor compliance

       Public companies are expected to meet labor standards. Systemic employment problems can hurt ESG assessments and investor confidence.               Once shares are sold to the public, any later penalty or scandal over past labor practices would be borne by ordinary shareholders. Regulators therefore examine this area closely.

5. Can It Still Get Through? Time Is the Real Test

       These issues are unlikely to cause an outright rejection. The exchange and regulator will probably send written questions and ask for more information or clarification.

That process carries two costs:

• Money:
  Lawyers, accountants, and sponsors will need to review everything again — potentially several million RMB in extra fees.
• Time:
  The back-and-forth could take two to three months. If Xiaohongshu misses its original window, it may need a fresh audit and could lose the current positive sentiment around tech-related listings in Hong Kong.

       Right now the market is excited about AI-related names like MiniMax and Zhipu. Xiaohongshu hoped to ride that wave. If momentum fades — or if external shocks hit — the valuation or even the ability to complete the deal could suffer.

6. A Lesson Before Listing: Don’t Leave Loose Ends

       Xiaohongshu’s situation sends a clear message to any company planning to go public: once you step into the public markets, historical problems must be cleaned up.

       If you need to let people go, do it properly with full documentation. If you don’t want to do that, settle the matter cleanly. Leaving a “landmine” behind means it can explode at the most vulnerable moment — right when regulators and investors are scrutinizing everything.

       Listing isn’t the finish line. It’s the start of public accountability. Any unfinished business tends to surface under that spotlight. Handling the small things properly is what actually keeps the process on track.

      On June 10, a hacker group called WorldLeaks posted roughly 204,000 files — about 630GB of data — on the dark web. Within hours, the files were freely available for anyone to download.

      The data came from Tata Electronics, one of Apple’s key suppliers in India. The level of detail goes far beyond typical product leaks.

1. What Exactly Was Leaked? More Than Just Looks

The files include:

• Detailed motherboard schematics for the iPhone 18 Pro, created in Siemens engineering software with confidential watermarks — showing layer structures, chip placement, and routing.
• Information on how the foldable iPhone works.
• Technical manuals for the A20 Pro chip (2nm process and packaging).
• Details on Apple’s in-house baseband chip (reportedly codenamed Ganymede).
• Bill of Materials (BOM) lists for hundreds of components, clearly naming suppliers and alternative vendors.
• Factory drop-test photos and videos of the iPhone 18 Pro.
• Copies of some employee passports.

      Because this happened before the iPhone 18 launch, much of the usual September event suspense has already disappeared.

2. Why This Leak Matters More

      Most past Apple leaks focused on design, features, or release timing. While those reduced some surprise, they were often seen as part of the marketing buildup and didn’t seriously harm Apple’s business.

      This leak is different. It reveals how the product is made, who supplies the parts, and at what cost — the foundation of Apple’s supply chain advantage.

      For years, Apple has relied on information asymmetry: each supplier only knows its own piece and pricing, not what others are doing. This gave Apple strong negotiating power to push prices down or switch partners easily.

      The BOM lists and supplier comparison tables have now torn that veil. Suppliers can benchmark each other’s pricing, and competitors get a much clearer view of Apple’s product roadmap and cost structure.

3. Apple’s Secrecy Culture: Jobs Era vs. Cook Era

      Under Steve Jobs, Apple was famous for extreme secrecy. Buildings were divided like apartment blocks — employees’ badges only worked in their own section. Before a launch, staff involved in training were often isolated until after the event.

      Under Tim Cook, product-level leaks became much more common. Renders, mold shots, and analyst reports turned into almost routine parts of new product cycles. Many observers believe Apple sometimes tolerated or even welcomed certain leaks because they generated buzz and acted as free marketing.

      The key difference: earlier leaks mostly revealed “what the product is.” This one exposes “how it’s actually made” — the deeper commercial logic.

4. The Other Side of Supply Chain Power

      Shortly after Apple announced price increases, Micron’s Chief Commercial Officer gave an interview to The Wall Street Journal. He noted that aggressive pricing pressure in the past had not been constructive and had led to canceled capacity investments in 2023.

      As a major customer, Apple has long held strong leverage in price negotiations. This pushed some suppliers to cut back on investment during low-margin periods. When AI demand surged, capacity couldn’t keep up, prices rose, and Apple eventually had to raise its own prices.

      It shows that information asymmetry and tough bargaining can cut both ways. The Tata leak further weakens Apple’s position in this dynamic.

5. Why Tata’s Factory?

      Tata Electronics was only founded in 2020. In just five years, its workforce grew from a few hundred to 75,000 people. Rapid scaling often leaves gaps in security systems and processes.In particular, its iPhone production line combines the factories of Pegatron and Wistron.

      By contrast, Apple’s supply chain in mainland China has had over a decade to mature, with better network separation, access controls, and supplier portal management.

      Historically, Wistron’s Indian factory suffered major losses from labor issues and violence before selling at a steep discount. Pegatron also gradually sold down its stake. Tata inherited and combined elements of these operations, resulting in a less hardened security setup.

6. Will Apple Move Production Back to China?

      Unlikely in any major way. India has become Apple’s second-largest iPhone production base, accounting for roughly a quarter of global output in 2025. Apple views India as a key long-term growth market — not primarily because of lower manufacturing costs, but because of its large population and future consumer potential.

      Unless Indian demand stays disappointing for an extended period, Apple is unlikely to make big shifts in its global production footprint over a single security incident.

7. What’s Really at Stake?

      The leak caused real damage to both Tata and Apple. But the deeper impact is on the information advantage Apple spent decades building in its supply chain.

      As Tim Cook prepares to step down as CEO on September 1 and hand over to hardware chief John Ternus, this incident leaves his successor with a supply chain whose protective information barriers have been partially breached.

      That may be the part Apple needs to take most seriously.

      In these two years, the world was obsessed with one dramatic question: Will AI hurt humanity? Will it replace us? Will it rule us? Elon Musk, Sam Altman, Dario Amodei — everyone was shouting.

      But while the debate raged, China quietly built an advantage in a different contest — the race over algorithms. And it did so in a very Chinese way: not by inventing the best chips or the strongest models, but by creating the world’s most intense competitive arena.

1. China’s internet ecosystem

      People call it the “gu jar.” In an old legend from southwest China, you throw various poisonous insects — centipedes, scorpions, and the like — into a jar. They fight and bite until only the strongest, most toxic one remains: the gu king.

      China’s internet ecosystem works a bit like that jar.

      Inside go hundreds of millions of atomized individuals and countless small merchants. A group of highly compliant platforms runs the jar. Fierce algorithms battle inside. And sitting on the lid is the “jar keeper” — the government. It doesn’t lift the lid, but it controls the temperature and sets the rules that decide who ultimately wins.

      After brutal elimination, the survivors that crawled out include global powerhouses like TikTok, Temu (Pinduoduo’s overseas version), and Shein — plus rising stars like Meituan and Didi that are now expanding worldwide.

2. How does the government manage this jar?

       In February 2025, China’s State Administration for Market Regulation issued the Anti-Monopoly Compliance Guidelines for Internet Platforms. It was followed by waves of meetings with companies, clearly defining eight types of new monopoly risks.

      In plain terms: the insects in the jar were told — these eight ways of “biting” are not allowed.

The eight categories roughly cover:

• Algorithm collusion between platforms:
  Platforms share data and coordinate against users. Search for something on Taobao, and suddenly Douyin recommendations become eerily targeted.
• Helping merchants form price-fixing agreements:
  Platforms assist sellers in setting unified prices or price ranges.
• Unfair high pricing by platforms:
  Using dominant position to raise commission rates that merchants can’t escape (often discussed in the context of Meituan).
• Selling below cost:
  Aggressive subsidy wars, such as Alibaba’s zero-yuan purchase campaigns against Meituan.
• Blocking and shielding:
  Preventing mentions of competitors (e.g., saying “Dou something Yin” instead of Douyin on WeChat).
• Exclusive dealing (“choose one or the other”):
  Merchants can only operate on one platform.
• “Lowest price across the web” (most-favored-nation treatment):
   Influencers sign agreements requiring the absolute lowest price.
• Differential treatment by platforms:
  Big-data price discrimination — new users get better deals than returning ones.

      These rules aim to stop algorithms from being weaponized for monopoly power.

3. Consumers can fight back too — the “add to cart and wait” trick

      Interestingly, consumers sometimes turn the tables. The simplest move: see something you like, add it to your cart, but don’t check out immediately.

      Why does it work? Once the item is in your cart, the algorithm shifts into full persuasion mode — flooding you with coupons, reminders, bundle suggestions, and loyalty points. The goal is simple: get you to complete the purchase.

      Algorithms operate in stages. First they figure out who you are and what you like. Then they bombard you with content. Once you engage, they double down. The moment you add to cart, the mission changes to “close the deal.”

      Platforms would rather give bigger discounts and more coupons than risk an uncompleted transaction. Their core metric is completed orders, not maximizing profit per order.

4. Why can only China grow these kinds of “monsters”?

Three key conditions make this system especially effective in China:

i. Increasing social atomization
      China now has around 320 million flexibly employed people — nearly half of its 700+ million labor force. Many lack stable employers and work in highly individualized, gig-style arrangements. This gives algorithm platforms a massive pool of atomized users and workers. The West hasn’t reached this stage yet, largely due to unions, industry associations, and legal protections. But in the AI era, societies worldwide are likely heading in this direction — exactly the environment these platforms are built to thrive in.

ii. Few traditional brakes or constraints
      Western platforms face unions, lawsuits, and public scrutiny that limit how aggressively they can compete. Chinese platforms operate with fewer external restraints because the system favors atomization for social stability reasons.

iii. Highly concentrated government power
      Western governments that want to regulate algorithms often hit limits from checks and balances. China’s government can set rules, summon companies, and adjust enforcement granularity more directly. This gives the “jar keeper” strong control over the arena.

      Inside this jar, algorithms don’t just target consumers — they also manage delivery riders and small merchants (tiny restaurants, shops, etc.). The government uses frequent guidelines and meetings, mainly targeting big platforms (because they’re large enough and responsive), to keep the whole system stable. High-profile actions, like pulling Didi back from the stock market, serve as strong signals.

5. How powerful are these “monsters” once they go global?

       Once they leave the jar, these companies show remarkable strength. TikTok dominates global short video. Temu disrupts Western e-commerce with ultra-low prices. Shein has taken fast fashion worldwide.

      Meituan and Didi are expanding into Hong Kong, Southeast Asia, Brazil, and the Middle East.

      Local traditional businesses struggle to compete. These companies were forged in China’s most brutal market conditions: they understand human psychology deeply, squeeze efficiency relentlessly, and know exactly how to dance on the edge of rules.

      They often expand using a “scout” approach — small overseas-registered teams or entities test the waters while parent companies provide resources and subsidies. Once they scale, they become hard to dislodge. TikTok, for example, set up a U.S. data-management company, but the core algorithm and operations remain under ByteDance control.

6. The cost — and the open question

      This gu-jar system is undeniably powerful. It has allowed China to lead in adapting algorithms to an atomized society and has produced a batch of globally competitive platforms.

      Yet the jar is fed by hundreds of millions of ordinary people who have been repeatedly “bitten”: delivery workers facing tighter time windows and lower pay, small merchants operating on razor-thin margins, and consumers navigating information asymmetry and algorithmic nudges.

      When a society is largely connected and managed through algorithms, and the main restraining force rests with the government alone, life inside the jar can feel precarious.

      The AI era is only beginning. Atomization will deepen, algorithms will grow stronger. The surviving “monsters” will only become more potent.

      The real question is this: Besides the jar keeper’s hand on the lid, can the people inside also hold onto a rope of their own — some form of meaningful counterbalance?

      There is no clear answer yet. But for anyone who has been shaped by these algorithms — and who has shaped others through them — it’s a question worth keeping in mind.

        There’s another twist: Apple has been talking to the US government, asking for an exemption so it can buy chips from Changxin Memory, which is currently on a Defense Department list. The US government is also a big Apple customer (iPads and MacBooks). That conversation is still ongoing.

        Apple has spent years being urged to “hurry up with AI,” yet it seems content to keep selling computers and phones. Reality shows that even if you don’t jump in, the ripple effects of AI on resources can still reach you.

        This reminds me of several big industrial shifts in history. Every major wave tends to create three groups: the earliest movers, the ones who join midway, and those who refuse to move at all. Their outcomes differ, but society often ends up with real, lasting changes after the “bubble” bursts. Here are four stories that show the pattern.

1. Case 1: British Railway Mania (1840s)

        Once steam locomotives appeared, Britain went all in. In 1846 alone, Parliament passed 272 railway acts authorizing nearly 9,500 miles of new track. Even priests, widows, and lawyers who never touched stocks put their life savings into railway shares. Prices doubled in just two years.

        Railways needed massive amounts of iron and coal. Britain’s iron output jumped from over 2 million tons in 1840 to nearly 3 million tons by 1852 — more than the rest of the world combined. Iron and coal prices soared. Industries completely unrelated to railways — shipbuilding, construction, farm tools — suddenly had to pay more for materials. Sound familiar?

        Among the earliest investors, some who backed lines with real passenger demand made money. But the ones who built the biggest empires on borrowed money (like “Railway King” George Hudson) collapsed when the bubble burst. He misused around £750,000 (roughly £74 million today) and ended up in prison.

        People who jumped in at the peak got crushed. Ordinary middle-class investors who bought after prices had already doubled watched shares fall 66% by 1850. Many families lost everything. Ironically, only about two-thirds of the authorized lines were ever built.

        Those who refused to participate and stuck with canals? Railways were faster and cheaper, so canal traffic dried up. They cut prices desperately but couldn’t compete. Many were acquired or abandoned. The canal companies were the “Apple” of their day — they just wanted to protect their existing business and got swept aside.

        Society, however, came out ahead. The tracks stayed on the ground. Transport costs collapsed, Britain became a single unified market, and industrialization and urbanization accelerated sharply. Shareholders lost; society gained real infrastructure.

2. Case 2: US Factory Electrification (1880s–1920s)

        When electricity arrived, factory owners expected huge efficiency gains by replacing steam engines with electric motors. What actually got expensive was copper for wiring.

        For the first 30 years, productivity barely moved. Why? Factories simply swapped the steam engine for an electric motor but kept the old layout — one giant overhead shaft with belts running everything. Nothing about the workflow changed, so nothing improved.

        Around 1910, a new generation of factory owners finally got it: put a small electric motor under each machine, rearrange the floor around actual production flow, and let machines start and stop independently. Efficiency exploded. By the 1920s, electrification accounted for roughly half of US manufacturing productivity growth.

        Factories that refused to redesign and clung to steam power were gradually driven out by lower-cost competitors.

        On a societal level, electricity didn’t stop at factories. It transformed homes and cities — electric lights, refrigerators, elevators, phones, radio. Modern urban life was literally redrawn by electricity.

        This feels a lot like how many companies are using AI today: they bolt it onto old processes (writing emails, summarizing meetings) and then wonder why nothing changes. The real gains come only when people redesign the entire business around the new technology.

3. Case 3: US Auto Industry (1900–1920s)

        At one point there were 253 car companies in America — more chaotic than today’s AI startup scene. Tires needed rubber, so rubber prices were bid sky-high. Capital rushed to plant rubber in Southeast Asia; acreage exploded nearly 30-fold in just a few years.

        Most of those 253 early car companies died. By 1929 only 44 remained.

        The winners were later entrants: Ford (with the Model T and assembly line), General Motors, and Chrysler. The three eventually took about 80% of the US market.

        What happened to the carriage and wagon makers who refused to touch cars? America once had over 13,000 businesses tied to horse-drawn transport. By 1920 almost all were gone. Whip makers collapsed because cars don’t need whips. A few smart carriage makers survived by pivoting to make car seats, bearings, and brakes for the new auto factories. They adapted instead of fighting the wave.

        Society was transformed: suburbs, highways, gas stations, motels, supermarkets. The entire geography and daily life of America were reshaped.

4. Case 4: Internet & E-commerce Bubble (1995–2002)

        This one is closest to us. Add “.com” to any company name and the stock soared. Nasdaq rose 600% from 1995 to March 2000, then fell 78% by October 2002.

        Most first-wave internet companies disappeared. Webvan, an online grocery service, burned through nearly $400 million before going bankrupt in 2001.

        A few who survived the crash became giants. Amazon’s stock dropped more than 90% and nearly died, but Jeff Bezos used the crisis to strengthen fundamentals. It went from an online bookstore to the massive company it is today. The overbuilt fiber from the bubble era later became cheap broadband that enabled YouTube, cloud computing, and mobile internet.

        Traditional retailers? Borders, once a bookstore giant, liquidated in 2011. Toys “R” Us and others followed. Walmart was slow at first but eventually invested heavily in e-commerce and survived.

        Society changed permanently. Shopping, entertainment, social life, and work all moved online. Can you imagine your daily life without the internet now?

5. Companies That Saw the Future but Refused to Change

        History also has cases of companies that clearly saw what was coming yet clung to past success:

• Kodak invented digital camera technology but dragged its feet because film was so profitable — and eventually died.
• Nokia dominated feature phones and had early smartphone efforts, yet mocked the first iPhone. Everyone knows how that ended.
• Blockbuster had the chance to buy a small Netflix and turned it down. It went bankrupt in 2010.
• Sears was once America’s “catalog Amazon” — housewives ordered from printed catalogs. It should have been perfectly positioned for e-commerce, but decided its old model was too good to change. It disappeared.

These companies didn’t fail because they missed the future. They failed because past success became the heaviest anchor holding them back.

6. Back to Apple and Today’s AI Wave

        We’re back where we started. Apple has been cautious on generative AI, yet memory and storage prices are still rising because of AI demand. You don’t have to participate — the wave still affects the inputs you need.

        History shows the same pattern again and again: bubbles burst, but the infrastructure and societal shifts that remain are real. Railways got built, electricity spread, cars changed how we live, and the internet moved everything online.

        Many people today ask, “Is AI actually useful? Why hasn’t productivity jumped yet?” It sounds almost exactly like factory owners in 1900 who replaced their steam engine with an electric motor and saw no immediate gain.

        The reason is usually the same: most are still bolting AI onto old workflows instead of redesigning the whole business around it. The big payoff comes later, once enough people do what the successful factories did in the 1920s.

        Companies that swear they’ll never touch AI are making the same bet as the canal operators, steam-only factories, carriage makers, and catalog retailers who eventually vanished.

7. Practical Tips for Ordinary People

• Don’t ignore the wave. History shows standing on the shore is never free.

• Don’t mock the people getting swept up in it. Some are pioneers; others are doing the necessary trial-and-error that later benefits everyone.

• Stay open-minded and adapt quickly. The more successful your past model has been, the more carefully you should watch whether it’s becoming the very thing holding you back.

What do you think about Apple’s price increase — short-term cost blip or a sign of bigger supply-chain shifts in the AI era? Drop your thoughts in the comments.

    On June 23 in Hamburg, Germany, the global supercomputing "summit showdown" — the TOP500 list — was updated. China's supercomputer LingSheng quietly took the top spot, pushing the U.S. machine El Capitan, which had held the lead for over a year, down to second place.
    This marks the first time since 2017's Sunway TaihuLight that a Chinese supercomputer has reclaimed the world #1 position after more than eight years. Many people's first reaction might be "way ahead"? This does indded have something to do with Huawei, but it's not that simple.
    Some ask: Isn't this just AI computing power? Actually, it's quite different. Others wonder: Where are Google and Microsoft? How did a machine this huge even get scored without being shipped to Germany? Today, let's break it all down clearly and objectively — no hype, just the facts.

1. What Exactly Does TOP500 Test?

    The TOP500 list has been around since 1993 and is updated twice a year (June and November). It's basically the supercomputing world's report card. It only has one test: LINPACK — solving an extremely large system of linear equations and measuring how many floating-point operations can be done per second.
    Two key terms to remember:

• Measured performance (Rmax): 

• Double precision (FP64): 

• Computing speed: 2.198 exaflops of double-precision performance (currently the world's fastest)
• Processor: LX2 (ARMv9 architecture), 304 cores per chip, two chips per node
• Scale: Over 20,000 nodes, more than 40,000 CPUs, totaling 13.79 million cores, housed in 92 cabinets
• Power consumption: 42 megawatts at full load — about 370 million kWh per year, costing over 200 million RMB in electricity (at Chinese rates), plus dedicated substations and a full liquid cooling system

• The test itself includes built-in verification — if the error exceeds 16, the result is automatically invalidated.

• A single global scoring standard.

• Official verification rights with random on-site audits.

• Reputation — the community is small, and getting caught faking data would embarrass the entire institution and country.

• Tianjin (Tianhe-1 & Tianhe-3): Petroleum, aerospace, biomedicine, controlled nuclear fusion, high-end equipment, animation rendering
• Guangzhou (Tianhe-2): Biomedicine, automotive, shipbuilding, film/animation, finance, nuclear power, oceanography, public safety
• Wuxi (Sunway TaihuLight): Oceanography, oil & gas, climate, industrial design, animation rendering (the 2016 fully domestic champion)
• Jinan (Sunway Blue Light): Oceanography, modern agriculture, oil & gas, drug screening, finance
• Zhengzhou (Songshan): Digital economy, precision medicine, biological breeding, environment, AI-related computing
• Kunshan (Yangtze River Delta Major Science Facility): AI, biomedicine, materials, atmosphere, oceanography

Micron’s latest earnings report caught a lot of attention. People are asking: how did the company make this much money in just three months? Getting clear on where the revenue actually came from also helps us think more realistically about whether the AI sector is in a bubble right now.

Here’s the headline: In the March–May 2026 quarter, Micron posted $41.4 billion in revenue. That’s more than its entire 2018 full-year revenue of $30.3 billion. Net profit reached $28.24 billion (up nearly 14× year-over-year), and gross margin hit 84.6% — higher than NVIDIA and Meta. Wall Street analysts, including Goldman Sachs, described the results as “across-the-board beats” on revenue, margins, segment performance, and stock reaction.

So where did the money come from? Micron’s own breakdown points to three connected layers.

1. AI Data Centers: The Core Money Driver

AI-related cloud storage and data-center business accounted for roughly 61% of total revenue. In simple terms, Micron is selling the “data pantry” that AI training and inference need — High Bandwidth Memory (HBM).

HBM has moved from being a supporting component on memory modules to a strategic bottleneck in data centers. A single NVIDIA B300-class AI chip typically needs 8–12 HBM chips around it, each with 48 GB or 64 GB capacity and data throughput thousands of times higher than standard memory. That performance comes at a price.

HBM production shares the same basic process as conventional DRAM, but it consumes the highest-quality wafer capacity. Each HBM chip takes up several times more fab space than a regular DDR chip and has lower yields. By shifting advanced capacity heavily toward HBM, Micron (and the industry) created a supply gap in the broader DRAM market. That gap shows up in higher prices for LPDDR5 and DDR5 memory used in phones, PCs, and servers.

Result: Micron earns high margins on the AI-specific HBM side while also benefiting from rising prices on the standard memory side. That dual effect is a big reason a single quarter now exceeds what used to be a full-year record.

2. Long-term Supply Agreements Lock in Future Visibility

One under-appreciated detail in the report is Micron’s disclosure of long-term supply contracts for 2026–2028. The total value has already passed $120 billion, with more than half backed by non-refundable customer prepayments.

Last year at this time they had signed just one such deal; this year they have 16. Microsoft, Google, Meta and other large AI players are prepaying tens to hundreds of millions (sometimes billions) to secure wafer capacity years in advance and avoid future bottlenecks.

Because so much of this revenue is already contracted and paid for, Micron felt comfortable guiding next quarter’s revenue at $50 billion — well above consensus expectations. That contracted portion gives the numbers more durability than a typical beat-and-raise quarter.

3. Who Ultimately Pays? The Three-Layer Chain

Follow the chips:

Micron’s HBM → sold to NVIDIA and AMD → packaged into GPUs → turned into servers sold to the six major cloud hyperscalers (Microsoft, Amazon, Google, Meta, Oracle, Tesla) → deployed in new AI data centers worldwide.

The hyperscalers are not the final payers — they are essentially reselling the compute. The real money comes from three layers below them:

• Top layer (most profitable today): Large enterprises, banks, pharmaceutical companies, and retailers that spend millions to tens of millions per year renting cloud capacity to train their own private AI models.
• Middle layer: AI startups (OpenAI, Anthropic, etc.) that pay hyperscalers based on API token usage and pass the cost to their own customers.
• Bottom layer: Everyday consumers — ChatGPT Plus subscriptions, Microsoft Copilot enterprise licenses, Tencent’s Doubao, and similar services. These smaller payments add up like steady tributaries feeding the larger river of cloud revenue.

4. Is AI in a Bubble? Demand, Supply, and Historical Context

The biggest skepticism centers on whether end-user spending can ever cover the hyperscalers’ massive hardware bills. The straightforward answer is that today’s gap is being filled by public-market investors and low-cost debt. Hyperscalers are using elevated valuations and cheap financing to fund today’s capex, betting that enterprise, startup, and consumer spending will scale dramatically over the next three to five years.

If that bet fails, the problem shows up on the hyperscalers’ balance sheets, not on Micron’s income statement. If it succeeds, Micron (and similar suppliers) effectively collected rent in advance.

To judge the nature of any “bubble,” it helps to compare with past technology buildouts:

• Demand side: 

Micron’s HBM is being bought primarily by Microsoft, Google, Amazon and peers that are laying down broad infrastructure — not just one app, but the equivalent of a rail network for the AI era. Model sizes have reached hundreds of trillions of parameters, and HBM demand per training run grows exponentially. This isn’t optional spending; it’s competitive necessity. The first movers who can train the biggest models at scale are more likely to set the rules for the next era.

• Supply side: 

HBM requires advanced Through-Silicon Via (TSV) stacking — essentially building 8–12 memory layers vertically like a skyscraper. Yields improve slowly, and the process depends on scarce advanced packaging capacity. Competitors can’t simply copy capacity in a few quarters; it takes years. In a market where supply is physically constrained, high prices and high margins rest on real bottlenecks rather than pure speculation.

There is obviously froth in parts of AI right now, and corrections will happen. Micron’s 84.6% gross margin is unlikely to stay at that level forever. Some investors will buy at the emotional peak and lose money.

But this looks different from a classic speculative bubble (think Dutch tulips). It resembles earlier infrastructure buildouts that left behind lasting physical assets even after painful corrections:

• 19th-century railway manias produced bankruptcies and wasted capital, yet Britain ended up with a dense rail network that slashed transport costs and created a unified market.
• Late-1990s fiber-optic overbuild bankrupted telecom giants, but the “dark fiber” they laid made global internet bandwidth nearly free and enabled Google, Amazon, and Netflix.

Right now, companies like Micron are pushing HBM capacity far ahead of current demand at high prices and margins. That will almost certainly create oversupply and price pressure at some point. The willingness of customers to prepay suggests they are securing optionality for a future where missing capacity could mean losing the next round of AI leadership.

When the froth settles, what remains is expanded HBM production, advanced packaging facilities, and a larger pool of experienced engineers — the physical foundation AI needs to scale. In that sense, even if parts of today’s spending look excessive, the outcome is more like building the steel and rails than planting tulip bulbs.

Capital is willing to fund ahead of proven demand because the cost of being short capacity when the next leap in model scale arrives is seen as higher than the risk of temporary oversupply.

What’s your take — is this AI spending cycle mostly necessary infrastructure with some bubble characteristics, or closer to pure speculation? Curious to hear different views in the comments.

On June 22, Japanese company Sakana AI released Fugu (meaning “pufferfish”). It scored 73.7 on the SWE-bench Pro coding benchmark, beating Anthropic’s Claude Opus 4.8 (69.2) and OpenAI’s GPT-5.5 (58.6). Social media quickly filled with excitement that Japan had finally arrived as a serious AI contender.

Before we celebrate, let’s slow down and look at what actually happened. Here’s a clear breakdown.

1. Who is Sakana AI?

Sakana AI was founded in Tokyo in 2023. Unlike labs that train massive models from scratch, Sakana specializes in combining and orchestrating existing models.

The company has three founders with complementary strengths:

• CEO David Ha: Chinese heritage, born in Hong Kong. He studied at the University of Toronto and earned his PhD from the University of Tokyo. He previously worked on neural networks and evolutionary algorithms at Google Brain and served as research lead at Stability AI.
• CTO Léon Jones: British engineer and one of the eight authors of the landmark “Attention Is All You Need” Transformer paper. He spent 12 years at Google and drives the company’s technical direction.
• COO Ito: Law graduate from the University of Tokyo and NYU, with prior experience at Japan’s Ministry of Foreign Affairs. He brings government connections and experience scaling Japanese tech companies globally.

It’s a blend of strong technical talent and deep roots in the Japanese market.

2. Sakana’s three main products so far

• Their Japanese large language model (released March 24), paired with a free Sakana Chat interface.
• An autonomous agent (released June 15) that can run unsupervised for about 8 hours and generate 60–100 page strategic reports.
• Fugu (released June 22) — the multi-model orchestration system behind the recent benchmark results.

3. Fugu is not a single model — it’s a system

Fugu’s core is a “Loop Agent” workflow. Instead of building one all-powerful model, it routes tasks across multiple existing models in a repeating loop:

Task arrives → one model breaks it down → another executes the work → another reviews the output → if the result isn’t good enough, the loop repeats until it meets the standard.

Sakana claims Fugu partially surpasses their own previously announced Mythos and Fable models (these were announced by Sakana; there are no independent third-party public benchmarks available, and access is currently restricted — non-US citizens may not be able to use them). It also outperformed Claude Opus 4.8 and GPT-5.5 on SWE-bench Pro.

4. The important detail: Fugu doesn’t disclose which base models it uses

Fugu doesn’t reveal exactly which models it calls internally. It only confirms it is not using its own latest Fable 5 and Mythos 5 versions (for competitive reasons and due to regulatory constraints).

It is very likely combining several strong publicly available models — for example, using Claude Opus 4.8 for planning, GPT-5.5 or DeepSeek for execution, and looping back for verification. This “mix and improve” approach helps explain the strong benchmark numbers.

5. Does Sakana have any models of its own?

Yes, but they are not frontier base models.

Inside the Fugu system is a lightweight “conductor” model fine-tuned from Qwen 2.5 7B that mainly handles task routing and loop management.

Their earlier Japanese LLM was fine-tuned on DeepSeek V3.1 Terminus and also drew from the Llama 3.1 405B open-source base. Its main strength is localization: it significantly lowered refusal rates on Japan-related historical and political topics (from around 72% in the base model to nearly zero).

In short, Sakana excels at adapting existing models for the Japanese market rather than pushing the absolute frontier of base-model intelligence.

6. Why this doesn’t create a lasting moat

Model orchestration itself is not a high barrier to entry. On June 12, OpenRouter — the largest model aggregation platform — launched a very similar feature called Fusion. It can combine up to 8 models on the same query and fuse their answers with a judge model. They claim it can approach top-tier quality at roughly half the cost without needing the single strongest model.

OpenRouter reported that about 75% of the performance gain came from the fusion step, with only 25% coming from model diversity.

This shows that multi-model orchestration is an open engineering technique available to anyone — not a unique advantage for Sakana.

The real competitive moat comes from owning frontier base models + controlling token pricing and access. Those currently sit with OpenAI, Anthropic, Google, and similar labs. They decide when new models are released, run their own internal orchestration systems (already through many iterations), and set the prices everyone else pays.

Companies like Sakana can only use publicly released models and have little control over costs. The big labs can adjust pricing, limits, or even restrict API access at any time. This situation closely resembles the early internet advertising industry: many smaller companies built clever optimization tools, but the traffic and pricing power stayed with Google and Meta. Most were eventually squeezed or acquired.

7. What’s likely next for Sakana AI?

Sakana is unlikely to disappear, but it probably won’t become a global AI leader on its own.

A more realistic path is acquisition — potentially by Google Japan, OpenAI or Anthropic’s Japan teams, or another player focused on the Japanese market.

For Sakana, being acquired could actually be a positive outcome. Buyers would likely value three things:

• The elite technical team (including a co-author of the Transformer paper).

• Strong access to Japanese government and corporate relationships.

• Proven ability to localize models for Japanese users and sensitive topics.

Japan has a history of nurturing domestic champions. If Sakana ends up deeply embedded in the Japanese market rather than trying to outcompete the global leaders head-on, that could be a sustainable position.

8.Quick takeaways

• For founders: 

Pure orchestration / Loop Agent products are hard to scale long-term without owning a base model, proprietary data, or a unique local/enterprise position. The best realistic outcome is often acquisition by a larger player. Treat orchestration as a useful capability, not the entire business bet.

• For regular users: 

When possible, stick with models and services from the original providers. Third-party orchestration systems can only work with already-released models from others and usually don’t disclose exactly which models they’re combining. This makes consistency and reliability harder to verify.

• For investors and creators: 

Ask two questions first: Does the company control its own base model? Does it control its own pricing power? For orchestration-focused companies, valuation should reflect realistic acquisition potential rather than assuming they will become the next foundational model leader. Benchmarks can be assembled; real moats have to be owned.

9. Bottom line

Sakana AI’s Fugu release looks like a breakthrough for Japanese AI on the surface. When you look closer, it’s a smart orchestration company that leveraged existing resources to deliver strong results.

It underscores a key reality in today’s AI landscape: the companies that own frontier base models and pricing power hold the real leverage. Orchestration techniques are clever and practically useful, but on their own they don’t create durable competitive advantages.

For Sakana, the most probable future is becoming a meaningful player in Japan’s domestic AI ecosystem — likely through partnership or acquisition — rather than suddenly emerging as a peer to the global leaders.

What do you think about model orchestration approaches like this? Smart short-term tactic or long-term limitation? Share your thoughts below.