Pang Dong Lai, a popular Chinese retailer, is preparing to close one of its older stores when its leases expire. Online, the story has been framed in a much simpler way: greedy landlords raised the rent, and the company refused to bow down or negotiate. It just walked away.

     A similar story has circulated about Tencent leaving the Kexing Science Park in Shenzhen. In that version, landlords get greedy, companies stand firm, and everyone watching gets the satisfaction of seeing someone finally say, “Enough.”

       But once you set the emotion aside, the situation looks more complicated.

1. This Is Not Just a Story About Higher Rent

       The older Pang Dong Lai store sits in a building divided among dozens of individual owners. The company had to negotiate separate leases with separate landlords instead of dealing with one property owner or management company.

       That setup creates obvious problems. An owner with a storefront near the main entrance may prefer a short lease and demand more at renewal. An owner on an upper floor or in a less desirable corner may accept very different terms. Even if the retailer wants one unified agreement, not every owner has a reason to go along.

       The most widely shared claim is that annual rent jumped from 8 million yuan to 18 million or even 24 million yuan. According to the original discussion, a rent dispute of that scale did happen at another Pang Dong Lai location years earlier. It may not have been the direct cause of this store’s closure.

       Combining numbers, events, and locations makes for a stronger viral story. It also makes the facts harder to sort out.

2. The Old Store Has Problems Rent Alone Cannot Fix

        This store has been operating for more than 20 years. It covers about 23,000 square meters, and it faces a more basic problem: it is no longer built for the traffic it attracts.

       The facilities are old. The layout is tight. There is no underground parking. Customers and cars now overwhelm a space designed for a different era.

       A successful store can still be held back by its building. If getting there means sitting in traffic and the inside layout cannot be meaningfully improved, strong sales do not solve everything.

       Major renovation would also be difficult. With dozens of owners involved, even basic changes require negotiation after negotiation. Adding parking, redesigning traffic flow, or rebuilding the property would be far harder than simply updating the décor.

       So closing the old store may not be a dramatic act of defiance. It may be a business decision: instead of endlessly patching a building that cannot be fully modernized, put resources into a space designed for the future.

3. The Bigger Shift Is From Renting Space to Building It

       Pang Dong Lai is building a new project called Dream City. It is planned to cover about 580,000 square meters, with an investment of 6.5 billion yuan. The company says it will use its own funds and plans to open in 2029.

       Tencent is making a similar move. Kexing Science Park had long struggled with congestion and pressure on its surrounding infrastructure. Tencent’s new Penguin Island campus began construction in 2021, with an investment of more than 30 billion yuan. By the end of June this year, about 28,000 employees had moved in. When the second phase is complete in 2028, the campus is expected to hold 80,000 to 100,000 people.

       Both companies are making the same basic choice: moving from rented space to space they can own and plan themselves.

       That is expensive. But it gives a company more control over parking, transportation, office layout, future expansion, and supporting services. It also reduces the time and uncertainty involved in negotiating with many landlords every time a lease comes up.

4. A Company’s Value Does Not Stop at Its Front Door

       Large stores and corporate campuses bring customers, employees, restaurants, housing demand, and other services. As a result, they can also increase nearby property values.

       That is one reason companies care so much about land and self-built campuses. They are not only trying to lower rent. They want to retain more of the long-term value created by their brand, foot traffic, and business activity.

       Warehouse clubs such as Sam’s Club offer a familiar example. Some locations are built in areas that are not yet fully developed. Once the store and its parking infrastructure bring in regular traffic, nearby land and housing can become more valuable.

       In the best case, this is not a zero-sum game. Developers, retailers, local businesses, and customers can all benefit from a new commercial center.

5. Why the “Walk Away” Version Spreads So Easily

       Because it gives people an emotional release.

        Many people have jobs, families, bills, and responsibilities that make it hard to simply walk away from an unfair situation. They may feel frustrated, but they cannot always afford to quit, move, or start over.

        That is why a story about a company saying, “I’m done,” is so appealing. The landlord becomes the obstacle. The business owner becomes the person who finally pushes back. It is simple, satisfying, and made for social media.

       The problem is that the more satisfying the story becomes, the more likely it is to leave out the difficult parts: Why did the old building stop working? Why is ownership so fragmented? Why were those leases signed in the first place? What risks did each side accept when they did?

6. Turning Landlords Into Villains Does Not Solve the Real Problem

       If some owners ask for higher rent and the tenant refuses, that is a normal commercial conflict. The company can leave. The owners can risk vacancies. Whether either side made the right choice will be decided by the market.

       The more useful questions are different. Why can a building that once worked become so hard to upgrade? Why does fragmented ownership make modernization nearly impossible? Why does a company that creates enormous traffic still have to spend so much energy on short-term lease negotiations?

      The Pang Dong Lai story matters for more than rent or corporate bravado. It shows what happens when an old space can no longer support a growing business. Sometimes leaving is not about winning an argument. It is simply the next practical step.

       The news was met with cheers across the Chinese internet. Then a new meme appeared and poured cold water on the celebration: “Li Shufu blasts the 5-million export figure.” Online summaries boiled his supposed comments down to three sharp accusations: either manufacturers are colluding with domestic used-car dealers to push excess inventory overseas through foreign ports; or they’re gaming the 13% export tax rebate with fake shipments; or they’re shipping batteries disguised as complete cars, then stripping them out for overseas energy projects.

       Three cuts, each aimed at the heart of the record. Let’s walk through what actually sits behind those 5 million cars.

1. First, did Li Shufu even say it?

       The circulating version claims Li made these remarks during a keynote at a Chongqing forum in June. Public records show he talked mainly about Geely’s restructuring and succession planning — not these three points. Whether he said something similar in a closed-door session without cameras is impossible to verify. It’s equally possible that media mixed earlier comments from Li with last year’s “auto-industry Evergrande” warning from Great Wall’s Wei Jianjun.

       So the attribution remains uncertain. What is certain is the number itself: 5.096 million cars really did leave China. And the official role those cars played is spelled out clearly in a mid-year review by the China Automobile Dealers Association — “exports supporting domestic demand.”

       In plain language: the home market could no longer carry the weight, so exports had to hold it up from below.

2. Why couldn’t the domestic market hold?

       In 2024 and 2025 China ran large-scale trade-in subsidy programs. Real money was spent to pull forward car-replacement demand that many families would otherwise have delayed.

       But cars are durable goods. A household doesn’t replace its vehicle every year, and the pool of households that can afford to is finite. Once that demand was pulled forward and exhausted, the domestic market cooled sharply. New-energy vehicle sales fell 20–30% year-on-year. As the clear industry leader, BYD felt the drop first and hardest — when you sell 200,000–400,000 cars a month, even a modest percentage decline hits the absolute numbers hard.

       So the cars had to go somewhere else.

       Of the 5.09 million exported in the first half, roughly 54% were still gasoline vehicles, many of them from joint-venture or foreign brands. SAIC’s MG badge is the clearest example — large volumes of China-built MGs are shipped back to Europe. That is also why the EU slapped SAIC with the highest countervailing duty. The remaining 46% were new-energy vehicles, but fluctuating oil prices, clogged ports, and incomplete local after-sales networks all make it hard to claim pure overseas demand is the full story.

       The real pressure remains at home: trade-in subsidies burned through replacement demand while the production lines kept running.

3. What the three cuts actually hit

• Cut one: zero-kilometer used cars and the tax rebate

       A “zero-kilometer used car” is a brand-new vehicle that a manufacturer pushes onto dealers, 4S shops, or affiliated financiers purely to inflate sales figures. The car is registered, then sits in a lot without ever being driven. Some even collect national, provincial, and local subsidies along the way.

       These cars do get exported, but usually not to tightly regulated markets like Europe. They tend to head for Southeast Asia, the Middle East, South America, or Russia. That helps explain why some overseas buyers complain that Chinese cars lack proper after-sales support — a portion may have come through channels with no real overseas service network.

       Until recently these zero-kilometer cars could still claim the 13% export VAT rebate. Authorities have now tightened the rules: a vehicle must be registered for 180 days before a normal export application, or the original manufacturer must issue a formal after-sales confirmation. The goal is not only to close a loophole but to clear space for legitimate manufacturers to sell new cars through proper channels.

       Rebate fraud has also occurred. A common tactic is to declare a car actually worth 150,000 yuan at 250,000 or even 350,000 yuan, because the rebate is calculated on the declared value. The current priority, however, is simply to slow the zero-kilometer pipeline so regular exports can move first.

• Cut two: payment terms under pressure

       Suppliers who delivered parts to BYD used to receive an internal instrument called “Di-Chain” rather than cash. Payment could be delayed two, three, or even four months. Anyone who needed money sooner had to discount the instrument, losing part of its face value as financing cost.

       After the government required large firms to settle with suppliers within 60 days, BYD switched to commercial acceptance bills. “Acceptance” does not mean immediate cash; the bills still carry their own maturity periods. By the end of 2025 BYD’s commercial acceptance bills had surged 727%. Many people assume a bill is backed by a bank. In reality a commercial acceptance bill is ultimately backed only by the issuing company’s own credit. The payment delay never disappeared — it just changed its name.

       As the biggest seller, every industry-wide pressure is magnified on BYD.

• Cut three: the economic logic of shipping batteries inside car shells

       China tracks a figure called the “installation rate” — the share of power batteries that actually end up in vehicles. In 2021 it was still 70%. By 2025 it had fallen to 44%. In May 2026 it stood at just 38%. More than 60% of the batteries produced never go into cars.

       At the same time the export tax rebate on complete vehicles remains 13%, while the rebate on batteries has already dropped to 6% and will fall to zero on 1 January next year. On a car priced around 150,000 yuan the battery can account for two-thirds of the cost. The arithmetic creates an obvious incentive: install the battery in a car shell, claim the higher vehicle rebate, ship it overseas, then remove the battery for other uses.

       No solid public evidence has yet confirmed that any company is doing this at scale. The economic motive, however, is clear. BYD already operates plants in Hungary, Brazil, and Thailand and therefore has the physical capability. The point is only that the logic exists, not that it has been proven.

4. The world is becoming a reservoir for Chinese overcapacity

       On the surface BYD looks formidable. Three days before this writing it took delivery of its seventh car carrier, the Zhengzhou. These ships are named after cities where BYD has factories and sail the world with giant red BYD logos. The Hungarian plant is already running; Brazil and Thailand are expanding. This is the glossy face of a globalizing champion.

       Underneath sit weak domestic replacement demand, the cleanup of zero-kilometer cars, tighter payment rules, and a 727% jump in commercial acceptance bills. Every industry ailment is amplified on the market leader.

       What the government is doing now is putting those problems on the table one by one. BYD looks less like a company about to collapse and more like a strong patient wheeled into the operating room. The surgery will hurt and there will be blood, but the underlying constitution is robust enough that recovery remains likely.

       The simplest conclusion is this: domestic demand has stalled while production lines keep turning out cars and batteries. The surplus has to go somewhere, so the rest of the world is being asked to serve as China’s reservoir. Cars are pushed outward; batteries travel inside car shells. Policy is simultaneously closing the gray channels and clearing space for the legitimate manufacturers. All of it amounts to drilling extra floodgates in an overflowing capacity pool.

       Whether those gates will eventually drain the pool, only time will tell.

       Most people think the way capital takes over a foreign company is simple: just buy enough shares. After the 1997 Asian Financial Crisis, Samsung was a clear example — yet even then, American capital never fully took control. Thirty years later, the playbook has been upgraded. It is no longer just about buying stocks.

       Let’s look at the full path through Samsung and SK Hynix. These two Korean companies control the most critical part of the AI industry — HBM (think of it as the “heart” of AI chips). Without it, even the most expensive AI chips cannot run. Together they hold nearly 80% of the global market. In theory, an asset this important should stay firmly in Korean hands. Capital, however, is taking a quieter and far more thorough route.

Step 1: Wash the shares out of local hands first

       In 2024, when HBM was in severe shortage and Samsung and Hynix were making record profits, their stock prices suddenly plunged 40%. Wall Street quietly bought during that period. Almost no one noticed.

Then in 2025, global media suddenly began shouting in unison: “AI memory is the golden decade,” “Missing Hynix is like missing Nvidia in its early days.” Ordinary Korean office workers, retirees, and newly married couples poured in. Many borrowed heavily from banks and brokers — in simple terms, they had 1 million of their own money but borrowed enough to buy as if they had 2 million or more.

       When prices fell, those who borrowed could not hold on. In just half a year, Korea’s major banks had already lent out 85% of their entire annual lending capacity. Retail investors could no longer borrow to cover losses. They were forced to sell their shares at rock-bottom prices to repay the debt. 300,000 accounts were wiped out. 1.2 million people were left with debts.

       The real goal of this step was never simply to crash the stock price. It was to force the shares that had been scattered among ordinary Koreans back onto the market through panic and forced liquidation. American capital slowly picked them up at the bottom.

       Note that the 1997 crisis already proved that simply holding shares is not enough to truly control a company. That is why U.S. holdings were deliberately spread across hundreds of funds, each owning less than 5%. Real voting power still sat with the Korean founding families, who held more than 20% in concentrated blocks. On the surface, the companies remained Korean-controlled.

       But capital had completed the first step: it had stripped the chips out of local hands.

Step 2: Move the “pricing power” to its own home ground

       Owning the shares is still not enough. The more important question is: who gets to decide how much the company is actually worth?

       In July 2026, SK Hynix listed in the United States. By then its share price had already fallen 43%, and the Korean won had depreciated 30% against the dollar. Americans could buy the stock at roughly a 60% discount.

       Previously the stock could only be traded in Korea, so its price was set by Korean money. Korean retail investors buying $65 billion in a whole year was already their limit. The U.S. stock market, by contrast, turns over $200 billion in a single day. Once global capital starts trading the same stock in the United States, the price is decided where the trading volume is deepest.

       A simple analogy: Imagine you used to sell fruit in your own neighborhood wet market. You and the local residents set the price. Now the same fruit is sold in the city’s largest supermarket. The price is set by the supermarket and the entire city’s buyers. Even if you are still the farmer, you have to follow the supermarket’s lead.

       That is the transfer of pricing power. From now on, when Hynix wants to raise new capital, issue new shares, or borrow money, it must watch the reaction of the U.S. market. Capital does not need to buy the entire company. It only needs the place where the price is discovered to be its own home ground.

Step 3: Physically move the factories and production capacity

       Once the shares and the pricing power are in hand, the final step is to move the actual production lines onto its own soil.

       The United States put a clear choice on the table using subsidies and tariffs: Want to keep selling chips to the world? Either build factories in the U.S., take the subsidies and tax breaks, or face a 100% tariff. A 100% tariff means your chips instantly become twice as expensive and no one will buy them. Samsung went to Texas. Hynix went to Indiana.

       Once the factories are on the ground, the subsidies become new leverage. Intel has already shown how this works: it received billions in government subsidies to build plants. When construction was only halfway done, the government suddenly said, “Give us a portion of the company’s shares so we become a shareholder.” Refuse? The subsidies would be cancelled and the half-finished factories would become ruins. In the end, Intel signed.

       When the real factories, equipment, and workers are all on American soil, even the strongest technology becomes increasingly constrained by American rules. At that point, the company’s nominal nationality no longer matters much.

The real method capital uses

       Capital does not take foreign core assets through a single forced acquisition overnight. It works through three progressive layers:

1. Use market swings and leverage to wash the shares out of local retail investors’ hands;
2. Through a U.S. listing and a deeper market, move pricing power to its own home ground;
3. Use subsidies and tariffs to physically relocate factories and production capacity onto its own soil.

       Koreans still hold the surface-level voting rights. Americans already hold the real control over price and future production capacity. This is capital operating at a higher dimension — it does not rush to change the company name. It simply peels away control, layer by layer.

       The next company or country that holds critical technology and keeps capital awake at night — will it walk the same path? Worth watching closely.

1. How Manus Reached a $2 Billion Valuation So Fast

       Manus started as Beijing Butterfly Effect Technology. Its founder, Xiao Hong (born in 1993 and a Huazhong University of Science and Technology graduate), had already built earlier tools like Monica and Yiban Assistant. In March 2025 the product launched publicly as the “world’s first general AI Agent” — basically an AI that can automatically handle complete workflows, such as writing code or managing repetitive tasks with minimal human input.

       Money came in quickly. ZhenFund joined the earliest seed round. In November 2024 Sequoia China led the A round with Tencent following. In April 2025 Silicon Valley’s Benchmark led the B round with $75 million, pushing the valuation to $500 million. That fast climb set the stage for what happened next.

2. Meta Wanted It — Then Regulators Blocked the Deal

       In December 2025 Bloomberg and CNBC reported that Meta planned to spend over $2 billion to acquire Manus and fold it into its super-intelligence lab. Mark Zuckerberg reportedly got personally involved and the deal was reportedly wrapped up in just ten days. It was called one of Meta’s biggest AI moves at the time.

       Chinese regulators saw it differently. In January 2026 the Ministry of Commerce began reviewing the deal. By late April the National Development and Reform Commission’s foreign investment security review office blocked it outright. The clear reason: Manus was founded by a Chinese team, serves many Chinese users, and holds local data and usage scenarios. Regulators did not want that technology and data moving directly to a foreign company. In May Meta agreed to cancel the transaction, started refunding money, cut off data sharing, and removed Manus employees from its internal systems.

3. Old Investors Now Want to Buy It Back at the Same Price

       By mid-June reports surfaced that early Chinese investors were planning to repurchase the company from Meta at the original $2 billion price. In early July the Financial Times and Bloomberg reported that Tencent is now leading the effort, bringing in ZhenFund, Sequoia China and other original backers to form a new consortium. They would buy Meta’s shares at the unchanged $2 billion valuation and prepare the company for a Hong Kong listing after reorganization.

       The valuation stayed frozen the entire time — from Meta’s attempted purchase through the rumored domestic buyback. That single unchanged number is the most telling clue in the whole story.

4. The Real Driver: Keeping the Investment “Alive” on Fund Books

       Venture capital funds don’t treat every investment the same way regular people might expect. They know most startups fail, but for the fund itself, a company going out of business and the investment being completely written off on paper are two different things. As long as the company isn’t formally liquidated, the investment can still sit on the books at a high value.

       Here’s how it works in plain language: A fund’s managers (called GPs) raise money from outside investors (called LPs — think pension funds, university endowments, or wealthy individuals). When a startup exits at a high price, the managers receive a performance bonus (called “carry”). They can also use the big exit story to raise their next fund from new LPs. Later, when the fund reaches its exit phase, it needs to return money to the original LPs — and a high-value exit makes those numbers look much better.

       When regulators killed the Meta deal, the funds suddenly faced a problem: bonuses already paid out, fundraising stories already told, and money already returned to LPs would all look bad if the investment dropped to zero. The cleanest solution was to keep the valuation exactly where it was by having the original Chinese investors buy the company back among themselves. That way the $2 billion number stays on the books and everyone’s accounting stays tidy — at least for now.

5. Why Is Tencent Leading the Buyback? It’s Probably Not Mainly for Internal Use

       Tencent has plenty of cash on hand (it recently sold a large chunk of Kuaishou shares for roughly $1.6 billion and continues investing in AI, including a stake in Kling). But buying an external AI team isn’t always straightforward inside a big company.

       Tencent had just finished an internal competition between its WorkBuddy and QClaw AI tools — QClaw lost and WorkBuddy became the main internal product. Adding another external team like Manus could easily create conflicts over resources, direction, and credit. A previous example shows the pattern: Tencent invested in the Windsurf team (project name Antigravity) with plans to integrate it, but it clashed with the internal CodeBuddy team and has remained separately operated ever since. Big companies often prefer to keep external teams at arm’s length rather than force integration that disrupts existing groups.

6. Does Manus Still Have Strong Technology?

       Before tools like Claude Code, Codex, and xAI’s Grok Build appeared, Manus had a genuine edge — it could automatically connect AI agents to real-world workflows. That originality helped it stand out early.

       Its biggest limitation, however, is the lack of its own foundational AI model. Companies that control their own models (such as Zhipu’s Zcode, MiniMax’s code model, or Tencent’s CodeBuddy) have more leverage and can build their own agents on top. Startups without their own model often find their valuation already high while later investors become reluctant to pay more, and much of the long-term value ends up captured by the model owners instead. Similar agent-focused companies — Perplexity, Cursor, Windsurf, and Character.AI — have each followed different paths, with varying degrees of success.

7. Who Might Actually Be Taking the Risk?

       Right now this looks less like a traditional business acquisition and more like a coordinated effort to keep a nearly written-off investment looking valuable on paper at the original $2 billion level. Tencent, Meta, and the regulators are all playing roles in that larger story.

        If the company eventually lists on the Hong Kong stock exchange, everyday retail investors would be the ones deciding whether the technology can actually support that valuation going forward. As long as the project stays “not dead,” the high number can remain on fund balance sheets. Once public shareholders are involved, the judgment shifts to whether the product can keep leading in a fast-moving field.

        Capital narratives sometimes move faster than the underlying technology. Paying attention to the incentives behind the numbers helps us read these headlines with clearer eyes.

       Have you ever wondered what happens when two companies team up to build electric cars, the cars keep selling, yet one partner reports a massive loss while the other keeps making money hand over fist? That’s exactly the situation with Seres (the company behind the AITO / Wenjie brand) and Huawei right now. Seres just dropped its half-year earnings warning: an expected loss of 1.5–1.8 billion RMB. Last year at the same time it made 2.94 billion RMB profit. Meanwhile, Huawei’s smart vehicle solutions business is still growing fast and collecting revenue from multiple streams with almost no inventory risk.

       Let’s break this down like we’re chatting over coffee — what really caused the loss, how Huawei structured the deal so it almost always wins, and what it means when a carmaker hands over too much of its “soul.”

1. Seres’ Loss Breakdown: It’s Not Simply “Sales Collapsed”

       First, clear up the biggest misconception. Seres’ cars are still selling. First-half volume was actually up a bit. The pain comes from three places: raw material costs, asset write-downs, and the removal of a big government subsidy cushion.

        In Q1 Seres still showed a 754 million RMB profit — but 600 million of that came from government subsidies. These subsidies go straight into the profit line with no matching cost. Strip them out and Q1 operating profit was only around 100 million RMB, already down 74% year-on-year and barely above break-even.

       For the full first half the company now expects a 1.5–1.8 billion RMB loss (or 2.2–2.5 billion RMB after removing all non-recurring subsidy income). The official reasons given are rising raw material prices and asset impairment.

Raw materials matter, but not equally. Lithium carbonate (the key battery material) really did push costs up — roughly several thousand RMB per car. Storage chips and industrial metals (copper, aluminum) added far less — maybe a few hundred to a thousand RMB per vehicle. A normal car doesn’t carry nearly as much memory as an AI server, so the chip story is mostly a distraction.

       The bigger chunk is asset impairment. Companies don’t expense the full cost of machines and molds in year one; they depreciate them over several years. When technology moves fast or old tooling won’t generate future profits, management can choose to write the remaining book value down immediately and book the loss now. Many listed companies prefer to dump all the bad news into one already-ugly quarter rather than spread it out. That’s likely where a large part of Seres’ extra loss came from.

2. The Two Hidden “Cash-Flow Reservoirs” That Got Drained at the Same Time

       Even before raw materials and write-downs, Seres (and many other carmakers) lost two important sources of breathing room.

       One was supplier payment terms. In the past, car companies could stretch payments to suppliers for three to six months — essentially using other people’s money as free working capital. New rules now require payment within 60 days. For a company that buys a lot of parts, that sudden loss of float hurts cash flow badly.

       The second was “zero-kilometer used cars.” These are brand-new vehicles that have already been registered but haven’t reached end customers yet. Dealers used to hold them in inventory or quietly export some through parallel channels. That channel has now been tightened. Previously these cars made up roughly 12% of volume; now the inventory and cash pressure lands squarely on the carmaker.

       Seres also has a harder time exporting than peers like BYD or Geely. Because so many core components come from Huawei, which faces U.S. sanctions, many overseas markets are reluctant to accept the vehicles. Other “界” brands have alternative export routes through their parent groups; Seres has fewer options.

       Faced with all this pressure, Seres chose to recognize as much loss as possible in this single reporting period. If sales pick up in the traditional “Golden September–Silver October” season, it can book profits later. That’s one legitimate accounting approach — but it makes the headline number look brutal.

3. Huawei’s Three “Pockets”: A Business Model Designed to Collect, Not to Risk

       Huawei, on the other side of the street, structured the relationship so it has three reliable ways to take in money while bearing almost none of the inventory or market risk.

• Pocket 1 – Selling core components.

• Pocket 2 – Taking a cut of every car sold.

• Pocket 3 – Occasional big-ticket deals.

       Huawei’s smart vehicle solutions revenue reached 45.018 billion RMB in 2025, up 72.1%. That figure doesn’t even include most of the store commissions or the brand/equity transactions. Crucially, Huawei never has to hold finished car inventory or worry about unsold stock.

4. Handing Over the “Soul” – What Does It Really Cost?

       Seres spent years learning from Huawei and reached heights it probably couldn’t have reached alone. But the core intelligence (driving algorithms, cockpit software, key chips) now sits with the partner. Once that gap in capability and talent density opens, it becomes very hard to close. Internal teams struggle to get the same resources or priority when the “partner” is seen as the expert.

       It’s a bit like the Microsoft–OpenAI relationship: Microsoft has enormous resources, yet it still finds itself largely iterating around someone else’s leading model because “we can just use OpenAI’s.” The same dynamic can appear when a carmaker treats its tech partner as the permanent “soul” of the product.

       Seres is not a pure victim. Before the Huawei partnership it was a small van maker (formerly Sokon) that had already received government support but was struggling to move upmarket. The collaboration genuinely lifted it into the premium EV conversation. Now it must live with the consequences of that strategic choice.

       SAIC’s former chairman once famously said he didn’t want Huawei to become the “soul” while SAIC became just the “body.” Five years later SAIC also partnered with Huawei on the Shangjie brand. The point wasn’t that Chen Hong was wrong — it was that giving away too much core control carries long-term costs.

5. Apple’s Asset-Light Model vs Huawei’s — Two Very Different Flavors

       People often compare Huawei to Apple because both are relatively asset-light. The differences are telling.

       Apple designs its own products, owns the brand, pays suppliers up front, holds inventory, and maintains a clear, non-competing product roadmap. Huawei goes further: it sells complete solutions to multiple competing car brands, takes a percentage of revenue regardless of the carmaker’s profit, carries almost no finished-vehicle inventory, and lets the various “界” brands sort out their own positioning.

       When the market is hot, everyone grows together. When price wars and inventory pressure hit, almost all the pain lands on the carmakers. It’s closer to a hub-and-spoke model where the center collects steadily while the spokes absorb the shocks.

6. A Final Thought

       The Seres–Huawei story isn’t a simple morality tale. Huawei brought real technology and market access that helped Seres grow fast. At the same time, the partnership was structured so one side bears heavy fixed costs, inventory risk, and market cyclicality while the other side enjoys multiple protected revenue streams.

       For any company considering a deep “borrow the ship to sail” partnership, the lesson is clear: keep enough independent capability and bargaining power that you can still steer when the weather turns. Cooperation is powerful — but handing over the entire soul rarely ends well for the one doing the handing.

       On July 10, 2026, Apple filed a lawsuit in a California federal court accusing OpenAI of systematically stealing its trade secrets. The defendants include OpenAI itself, the hardware company io Products (founded by former Apple design chief Jony Ive and later acquired by OpenAI), OpenAI’s Chief Hardware Officer Tang Tan, and former Apple engineer LiuChang.

       Apple is asking for five things: a jury trial, an immediate injunction to stop the alleged infringement, destruction of all confidential materials, redesign of any unreleased hardware products to remove Apple technology, and damages. The most painful demand is the forced redesign — in past cases, Apple has used similar rulings to force deep audits of a company’s entire technical roadmap, exposing secrets and dragging out timelines for years.

       The case quickly grabbed attention across Silicon Valley. Elon Musk jumped in with a jab at his longtime rival Sam Altman: “You stole my open-source charity OpenAI, and now you’re stealing from Apple too?” Altman fired back that Musk’s public company would face SEC trouble if his big promises on compute centers didn’t materialize. Musk replied he would begin “test launches” next year and might even invite Altman — if his parole officer allowed it.

1. The central figure: Tang Tan

       At the heart of the case is Tang Tan. A Malaysian-Chinese engineer who studied at Imperial College London and MIT, he joined Apple around 1999 and spent 25 years there. He rose to vice president of iPhone product design, with his name on over 270 Apple patents. He played a key role in the design of the iPod, multiple iPhone generations, and the Apple Watch.

      In 2024 he left Apple to co-found io with Jony Ive. OpenAI acquired io for roughly $6.4 billion in 2025, and Tang became OpenAI’s Chief Hardware Officer. Apple’s lawsuit treats him as the key link — not just another departing employee, but someone in a senior position who allegedly helped turn knowledge transfer into a repeatable process.

2. The evidence trail

       Apple’s complaint lays out a timeline. In January 2026, LiuChang left Apple after eight years and joined OpenAI. He kept a company-issued laptop that could still access Apple’s internal network. He messaged a colleague still at Apple: “Haha, I can still get into the company storage — so funny.” The finally reply came back: “I’m ready.”

       The two later switched to the Line app to communicate, but Apple’s systems still captured the activity. Liu allegedly downloaded dozens of confidential files covering technical specs, hardware designs, manufacturing details, and testing processes.

       Apple says this alone could be treated as individual misconduct. The bigger problem, according to the lawsuit, is what it claims Tang Tan did:

• During job interviews with Apple employees, he allegedly used internal project codenames that only insiders would know and asked candidates to bring physical parts (batteries, system-in-package chips, motherboards) to the interview for him to examine.
• He reportedly coached people planning to leave Apple on how to “leave gracefully” — advising them not to announce their departure immediately so they could gather more information while they still had access, before devices and permissions were revoked.
• Even before leaving Apple, Tang allegedly emailed sensitive supplier information to his personal account.

      Taken together, Apple argues these actions show a pattern originating from the top of OpenAI’s hardware team rather than isolated employee behavior.

       There is also a separate but damaging claim about the supply chain. Former Apple engineers reportedly approached Apple’s long-time manufacturing partners (the so-called “fruit chain” companies in China) and suggested they could continue working together — sometimes claiming they had “Apple authorization” to use proprietary processes or materials. Some suppliers reportedly agreed to informal arrangements. For Apple, this strikes at the heart of its quality advantage, which has always depended on tight control over specialized manufacturing know-how.

3. Why Apple and OpenAI fell out

       Apple and OpenAI used to cooperate. In 2024 Apple integrated ChatGPT into Siri. Apple expected the usual revenue-sharing model it has with partners: it provides the huge base of iPhone users and the App Store platform, and takes a cut (30% in the first year, dropping to 15% in the second).

       OpenAI saw it differently. It complained that conversion to paid ChatGPT subscriptions ($20/month) was too low and that serving millions of Apple users was expensive because every AI response consumes significant computing power. Unlike Google Search — where adding more users costs almost nothing extra — OpenAI’s marginal cost rises with every additional query.

       OpenAI reportedly pushed back hard and even threatened legal action. Apple responded by switching Siri’s AI partner to Google’s Gemini in the latest iOS versions. The partnership was over.

4. What Apple is really afraid of

       Apple’s premium pricing has long rested on three pillars: superior build quality, distinctive design, and a tightly controlled ecosystem. The lawsuit suggests OpenAI’s talent raid and alleged information gathering threaten all three.

       Quality comes from Apple setting strict standards and working with its supply chain partners to execute them. With reportedly hundreds of Apple hardware engineers now at OpenAI (and supply chain relationships shifting), that knowledge edge is leaking.

       Design appeal is not entirely separate — it depends on the overall experience of quality and software. Without a clear quality or ecosystem advantage, aesthetic differences alone are harder to maintain.

       The biggest long-term threat is the ecosystem. Traditional iOS apps run in “sandboxes” — isolated environments where each app can only access its own data. This protects privacy and security, and has been a core reason Apple’s ecosystem feels more controlled than Android’s.

       AI Agent phones work differently. Instead of opening multiple apps, users speak to a single AI that can pull together data from many sources to complete complex tasks (for example, planning a trip that involves navigation, restaurant booking, food delivery, and hotel reservations). If this model succeeds, the value of Apple’s traditional app sandbox system could shrink dramatically.

       OpenAI is reportedly accelerating its own AI phone plans, aiming for a 2027 launch with annual sales targets in the tens of millions. Other players are exploring similar hardware. For Apple, this is not just another competitor — it is a potential shift in how people interact with their devices.

5. How the case might play out

       Some individuals will likely face consequences, possibly including criminal liability for the most serious leaks. OpenAI is unlikely to accept a court-ordered full audit of its code and hardware designs.

       The more probable path is a negotiated settlement: cutting ties with certain people while offering Apple concessions such as licensing, equity, revenue sharing, or cash. Similar past disputes in Silicon Valley have often ended this way rather than in total victory for either side.

       Whatever the legal outcome, the larger story is the arrival of AI-driven hardware. This lawsuit is Apple’s attempt to slow down a fast-moving competitor during a critical transition. The technology itself — AI agents that can actually get things done across services — is not going to wait for court rulings.

       The real question going forward is not just who wins this case, but what kind of devices we will all be using in a few years.

       On the evening of July 6, a former Huawei “genius youth” posted online about his second-round interview at DeepSeek. The interviewer stared at his left monitor and said, “You’re copying code, right? If you can’t prove you’re not, the interview is over.” He felt publicly accused of cheating, walked out on the spot, and shared the story. It shot to the top of trending topics the next day. That same afternoon, a well-known investor jumped in and called him “the most untrustworthy founder I’ve ever met.” The former Huawei talent fired back overnight, saying the investor had only wired half the promised money and then stopped.

       Two shiny labels — “Huawei genius youth” and “idealistic, national-level AI model company DeepSeek” — crashed into each other and went viral. But the real story isn’t about who won the argument. It’s a mirror that shows what often happens when AI companies get big funding and try to scale fast: systems break, the wrong people get interviewed for the wrong roles, and everyone ends up frustrated. The core issue isn’t that one interviewer was rude. It’s that DeepSeek’s hiring process hasn’t kept up with its growth, and we keep treating both “genius” individuals and big-name companies as if they’re perfect.

1. What actually happened in that interview

       Li Bojie, born in 1992, is now chief scientist at a startup called Pine AI (he specifically clarified he is not a co-founder). He wanted to do research work at DeepSeek. Instead, the first two rounds were coding tests. When the interview finally happened, the interviewer was late and seemed uninterested in his actual research. The only question that kept coming up was “What engineering challenges have you solved?”

       The most explosive moment was the copying accusation. Because it was a remote interview, he had two monitors on his desk — a completely normal setup for programmers. The interviewer accused him of looking at the left screen to copy code and threatened to end the interview if he couldn’t prove otherwise. Li explained it was just dual monitors. The interviewer didn’t believe him and turned it into a cheating allegation.

       A lot of people’s first reaction was sympathy: even a Huawei genius youth gets treated this way, so how hard must it be for everyone else? But if you look closer, this isn’t mainly about one interviewer’s attitude. It’s about deeper problems in how DeepSeek is building its team right now.

2. Why live coding tests feel outdated

       These days, a lot of code is written with AI assistance. Asking someone to write large chunks of code from scratch on a blank screen, with no tools, has become less useful for judging real-world ability. Turning that directly into a “you’re cheating” accusation makes it worse.

       This kind of mismatch has happened before. In older programming interviews, candidates were sometimes asked to hand-write a keyword that the development environment automatically generates. People who actually wrote code every day never typed it manually, so the test ended up measuring the wrong thing. In the AI era, clinging to old “prove you can code blind” methods often misses the point and can unfairly penalize strong candidates.

3. When interviewers skip the resume, bigger problems follow

       In normal hiring, you read someone’s background first, then talk about what they’ve done and whether it fits the role. In this case, the interviewer clearly hadn’t looked at Li Bojie’s resume. His history as one of the first eight Huawei genius youths, his top-conference papers, his training at a top university — none of it seemed to matter. He was just another person going through the same standardized process.

       Other candidates have reported the same pattern at DeepSeek: multiple HR people adding them on WeChat at once, interviewers arriving late and acting impatient, and questions that have nothing to do with the candidate’s actual experience. When hiring turns into “run everyone through the same script,” you waste time and often end up with poor matches.

4. The “Huawei genius youth” label cuts both ways

       The program started in 2019 when Huawei’s founder personally pushed for it. It offered very high salaries — up to around 200 million RMB a year for the top tier — to attract exceptional young talent. Li Bojie was in that first group of eight. He later left Huawei reportedly turning down a 3 million RMB annual package.

       But a flashy label only proves one thing: strong research ability. It doesn’t automatically mean someone is also great at engineering execution, leading teams, or running a company as CEO. Some former genius youths have stayed and contributed quietly inside big organizations. Others left to chase bigger dreams. The public tends to notice the dramatic exits more than the steady work happening behind the scenes. Treating the label as proof of overall perfection sets everyone up for disappointment.

5. Research and engineering are very different jobs

       Li Bojie wanted a research role. The interview kept pushing toward engineering tasks. This is a common mismatch.

       Think of it this way: research is like working in a lab — exploring new methods, running experiments, publishing papers, and focusing on “can we make this work at all?” Engineering is like running a restaurant that serves customers every day — the food has to be consistent across different kitchens, handle rush hours, fix problems quickly, and keep customers happy long-term. The skills don’t overlap as much as people assume. Code written for a research paper can “run” in a demo, but it often falls apart when you try to make it stable for real users. Putting pure research talent into an engineering-heavy interview without understanding this difference leads to exactly the frustration we saw here.

6. The investment side showed another expectation mismatch

       On the same day, a different story broke. An investor (co-founder of Huobi) publicly accused Li Bojie of taking funding for a project called Metagent and then going silent — not sharing updates or financials. Li replied that the investor had promised $1.5 million but only sent $500,000 and then stopped.

       Both sides seem to have had unrealistic expectations. Investors who back young research stars know the risk is high: great at papers doesn’t automatically mean great at being a CEO or handling investor communications. Smart investors usually release money in stages — give some upfront, then more after clear milestones and transparency. This is called milestone-based funding. It protects the investor while giving the founder a chance to prove progress.

       Li apparently felt the full amount was promised and got frustrated when it didn’t arrive. The investor felt the founder wasn’t communicating properly and chose to limit further risk. Neither side had aligned expectations or built enough trust. The result was public finger-pointing instead of a productive relationship.

7. Why he actually wanted to join DeepSeek

       Li later posted a video clarifying his situation. His current company is running fine — he wasn’t job-hunting because of failure. He wanted to join DeepSeek because almost everyone is now building “harness agents” (AI systems that can use tools, write code, and handle complex tasks on top of large models). Many of the most useful capabilities, however, are locked inside the model companies: how they build training data, how they evaluate performance, and the long system prompts that guide behavior.

       He wanted to go inside and see how things actually work. This isn’t unusual. Even high-profile people in Silicon Valley have left startups to join base model companies just to understand the foundations better. Without an easy external collaboration path at DeepSeek, he tried the normal job route — and ran straight into the hiring issues described above.

8. What happens when a company rushes hiring after funding

       DeepSeek completed its first external funding round on June 16. After that, it clearly wanted to quickly build its own harness agent capabilities. The normal way to grow a new team is to first design the structure, identify the right experienced people, and let them help build the rest of the team layer by layer. Instead, it appears they went straight to headhunters for stacks of resumes and ran everyone through a fast, uniform interview process.

       This “get it done quickly” approach often backfires. You can end up with mismatched roles, interviewers who don’t understand the work, and a team that looks busy but struggles to deliver real results. A famous earlier example is what happened at JD.com during one of its fast-growth periods. They hired a lot of “good enough for now” people to keep the website running. It worked short-term, but later, when they needed to level up and bring in stronger talent, many key positions were already occupied by people who couldn’t handle the next stage. The company had to spend years fixing the gaps.

       DeepSeek risks repeating a version of this story if it keeps prioritizing speed over thoughtful team building.

9. A few final thoughts

This whole episode really comes down to three simple reminders:

• Don’t treat job titles or company names as proof of overall perfection.
• They only show strength in one area. Don’t imagine companies like Huawei or DeepSeek as single, flawless personalities. Every organization has strong parts and weaker parts.
• Nothing in the real world is perfect. Even the most impressive systems have rough edges and rough periods.

       DeepSeek has the money now. Whether it builds a strong team or ends up carrying hidden problems forward depends on how it handles hiring and role matching in the next phase. The world is basically one big makeshift team — some groups know they’re still figuring things out and keep improving, while others act like they’ve already built a perfect temple. The first kind usually travels farther.

      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.

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.

DeepSeek just closed a 51 billion RMB (~$7.4 billion) round — the largest single financing ever for a Chinese AI company, and its very first round. Naturally, it drew a lot of attention.

Compared with the U.S., the gap is still wide. Anthropic has raised about $116.3 billion cumulatively, OpenAI around $150 billion. DeepSeek’s round is roughly 1/15.7 of Anthropic’s and 1/20 of OpenAI’s. In China, however, this is the biggest single round on record.

Chinese AI companies enjoy lower labor and power costs, so they simply don’t need to burn as much cash. That explains part of the difference.

1. Investor Lineup: Liang Wenfeng Leads with 20 Billion RMB

The money came from ten parties:

• Liang Wenfeng: 20 billion RMB (largest single check)
• Tencent: 10 billion
• CATL-affiliated fund: 5 billion
• NetEase and JD: 3 billion each
• Legend Capital and IDG: 3 billion each
• Zhengxin Valley and Shixiang Tech: 1.5 billion each
• National AI Industry Investment Fund: 980 million

Total ≈ 50.98 billion RMB. As lead investor, Liang secured pricing rights and the ability to set the main terms. Others who wanted in had to accept the framework he put forward.

2. Key Deal Terms: How Limited Partnership Locks in Control

Instead of direct equity into DeepSeek, the round used a limited partnership structure with GP (general partner) and LP (limited partners).

Liang Wenfeng is the GP; everyone else (Tencent, CATL system, JD, etc.) are LPs. LPs can review financials but have no voting rights or say on strategy, suppliers, or major decisions. For example, CATL couldn’t insist DeepSeek use its batteries, and Tencent couldn’t force DeepSeek to rent its cloud — they’d first have to convince the GP.

This concentrates decision-making power with Liang.

A second term is a 5-year lock-up: LPs cannot exit early. Standard exit routes include IPO, acquisition, or negotiated extension/rollover/impairment at the end. Funds legally cannot guarantee “no losses,” so everyone understands the risks.

Third, LP qualification review. DeepSeek wants to hold its own licenses (ICP, ISP, etc.) as a pure domestic entity. Any foreign capital, even 1%, could complicate things. Screening LPs upfront avoids future headaches and smooths the path for a potential A-share listing.

Fourth, a no-poaching clause: investors cannot hire away DeepSeek employees. Tencent and JD are both building their own large models, yet they agreed not to recruit from DeepSeek after signing. The absence of Alibaba, Xiaomi, and ByteDance from the list may be related.

3. The National AI Fund’s Special Role

The National AI Industry Investment Fund stands out. It contributed only 980 million RMB but invested directly into DeepSeek (not as an LP), giving it board seats and veto rights over new financing, IPO/bankruptcy/liquidation, charter amendments, major asset deals, and more.

This is not the well-known big semiconductor fund. It’s a new AI-focused fund established in January 2025 (≈60 billion RMB scale). Its involvement provides both regulatory support and a clearer route to A-share listing — which is why other LPs were willing to accept stricter terms.

4. Where the Money Goes: Compute, Team, and Commercialization

The capital is earmarked for three interconnected priorities:

• Inner Mongolia compute center — Located in Ulanqab (one of the “East Data West Computing” hubs) where electricity costs less than 0.3 RMB/kWh. DeepSeek is already hiring senior data-center operations engineers at ~30k RMB/month + 14 months.

• Harness Agent team — To be built starting May 2026, focusing on inference systems.

• Commercialization — The model is open-source under MIT license, so anyone can deploy it for free. DeepSeek’s edge will come from running the lowest-cost, highest-efficiency inference itself. The plan is to let others deploy through DeepSeek’s infrastructure (a bit like a 4S dealership model: the brand sets standards and pricing, partners handle sales and share revenue). Tencent Cloud has reportedly leaned toward this approach.

5. Why All RMB? A Direct Path to A-Share Listing

The entire round was done in RMB with domestic entities. The goal is twofold: stay fully domestic to secure licenses easily, and prepare for an A-share IPO.

Many other Chinese AI firms raised in USD for higher valuations but face Hong Kong or overseas exit routes. DeepSeek’s RMB + national fund route prioritizes A-share certainty. A-shares offer deeper liquidity and scarcity value (still partially approval-based), making exits more attractive for investors.

6. “Circumstances Trump Individuals”: Who Actually Sets the Rules

Some commentary frames this as “Liang Wenfeng outsmarted capital.” A more accurate view is that circumstances favored the founder. When a company has a credible A-share path, national-fund backing, and a lead investor willing to write the largest check, investors are more willing to accept GP-led structures, lock-ups, and no-poaching terms.

7. Bottom Line: Funding Is Done — Rhythm Matters Most

DeepSeek has always been low-key and execution-focused. With this capital in hand, if it continues at its own steady pace on technology and commercialization — rather than rushing into heavy PR and frequent launches — the round should prove genuinely helpful.

The national fund brings resources and listing support, along with a degree of oversight. Overall, the structure gives the founding team meaningful control while giving capital a clear path to returns. DeepSeek’s next chapters are still worth watching.