I started buying gold after Silicon Valley Bank collapsed on March 10, 2023. The FDIC later confirmed that SVB was closed by California regulators on March 10, 2023, and announced on March 13 that all depositors would be protected; the FDIC’s deposit insurance reform report that same year also placed this event within the context of re-examining the U.S. deposit insurance system. [G1][G2] Since then, gold has tripled.

Gold is about to return to its essence and challenge the authority of the dollar once again.

Today, June 12, 2026, SpaceX goes public. The IPO that tests the market’s temperature has arrived.

Stagflation

The first time I realized something was wrong was when a glaring detail appeared in the Q3 2025 earnings reports of major U.S. banks. Originally, global economic growth after the pandemic was like a capsized boat in a ditch—half dead.

Note: The World Bank stated that 2020–2024 would be the weakest half-decade of global growth in 30 years; the IMF also described the world economy at the time as “still limping along.” [WG1][WG2]

That day, several major U.S. banks released their third-quarter earnings. On the surface, the market heard a very nice story: the U.S. consumer remained “resilient,” bank performance was still strong, and card spending was still growing. Reuters later reported that Wall Street’s big banks highlighted consumer resilience in their Q3 earnings, with customer activity remaining solid, including increased credit and debit card spending, and credit trends still very good. [C1] That same day, bank stocks were also traded as good news: Reuters’ headline was “Wall Street ends mixed; banks rally on upbeat results,” with Wells Fargo rising on better-than-expected Q3 profit, Citigroup also rising, and the S&P 500 bank index strengthening. [C2]

But what really struck me as odd was that credit cards were not a peripheral item in the earnings reports; they were directly written into the performance structure of these banks.

JPMorgan’s Q3 2025 earnings report showed that its debit and credit card sales volume grew 9% year-over-year; Card Services & Auto net revenue reached $7.2 billion, up 12% year-over-year, driven in part by higher Card Services net interest income from increased credit card revolving balances. [C3] Wells Fargo’s Q3 materials also showed that card fees increased by $127 million, up 12% year-over-year, due to higher merchant processing card fees and increased consumer credit card activity; its Credit Card business grew 13% year-over-year, including higher loan balances and higher card fees. [C4] Citigroup’s Q3 report similarly noted that U.S. Personal Banking net income grew 64% year-over-year, and Branded Cards revenue reached $3.0 billion, up 8% year-over-year, driven by higher loan spreads, higher interest-earning balances, and higher interchange. [C5]

This is where I first saw “stagflationization” happening.

The “consumer resilience” in bank earnings might not be spending power in ordinary people’s lives, but rather revolving credit capacity.

Ordinary people continuing to swipe cards does not necessarily represent prosperity. More likely, cash flow has been squeezed tighter by the cost of living, and they are starting to use high-interest credit cards for debt rolling: using this card to sustain life, and using the next paycheck, another card, balance transfers, cash advances, or minimum payments to pay off the previous card, continuously rolling today’s inflation and high-interest rate pressures into the future.

The so-called resilient consumer might just be revolving debt that hasn’t blown up yet.

This is not an emotional judgment. Subsequent data began to confirm this direction. The Philadelphia Fed’s large bank credit card data later showed that in Q3 2025, large bank credit card balances and purchase volumes continued to rise year-over-year, with the most pronounced growth in average purchase volume among borrowers with credit scores below 660; at the same time, revolving balances hit new highs, and the average purchase APR on general-purpose credit cards remained near 24.5%, close to historical highs. [C6] The New York Fed subsequently also showed that by Q4 2025, U.S. household debt rose to $18.8 trillion, and credit card balances rose to a high of about $1.28 trillion; by Q1 2026, although credit card balances fell seasonally, they were still $70 billion higher than a year earlier, and the flow of credit card balances into serious delinquency was 7.10%, up from 7.04% a year earlier. [C7][C8]

This is not a simple story of “consumers are strong,” but more like “consumers haven’t broken yet, but they are already using high-interest debt to stay alive.”

If we pull the lens from bank earnings to the retail end, this “resilient consumer” story looks less like nationwide prosperity and more like a K-shaped consumer society: the upper arm is still spending, while the lower arm is already deforming. Walmart’s data is very telling. When Reuters reported in 2024 that Walmart raised its full-year forecast, it mentioned that Walmart’s market share gains spanned all income levels but were primarily driven by upper-income households. [K1] By Q3 2025, Walmart’s own earnings materials were even more blunt: Walmart U.S. comparable sales grew 4.5%, e-commerce grew 28%, and market share gains covered all income segments but were led by upper-income households; management also said on the earnings call that the U.S. business saw strength across all income levels, especially upper-income households, with growth primarily driven by middle- and upper-income households. [K2]

This is not prosperity in the traditional sense, but rather high-income consumers also starting to seek discounts. Channels that previously belonged only to low-income households as “money-saving channels” are beginning to absorb middle- and upper-income traffic. Costco is on the same track. Costco’s Q4 fiscal 2025 net sales grew 8.0% year-over-year to $84.4 billion, and full-year net sales grew 8.1% to $269.9 billion; Reuters’ report on that earnings release said that U.S. consumers, under pressure from inflation and tariffs, were flocking to membership-based warehouse retailers like Costco in search of affordable necessities, and Costco was also attracting price-sensitive consumers through its Kirkland private label, low-priced eggs and butter, and extended gas station hours. [K3] So, rather than saying the U.S. consumer is broadly strong, it’s more accurate to say that a portion of high-end consumption and traditional supermarket traffic is also migrating to “low-price but decent” channels like Walmart and Costco.

The low-income end is a different picture. By October 2025, the U.S. government shutdown threatened SNAP food stamp distribution, and Reuters reported that U.S. food retailers and food companies were preparing for a sales decline in November; if federal food aid were interrupted, it could cause a roughly $8 billion grocery revenue gap, with Walmart being the largest recipient of SNAP grocery spend, accounting for about 26.1%. The same report also cited research saying that if benefits were delayed, Walmart, Dollar General, and Dollar Tree’s Q4 sales could decline by less than 1% year-over-year, depending on the duration of the shutdown. [K4]

This is not pure extrapolation. Earlier, in 2014, the cut in U.S. food stamps had already left a sample for Walmart: Reuters reported at the time that severe weather and reduced U.S. food stamp benefits dragged down Walmart’s fiscal Q4 comparable sales, forcing the company to lower its quarterly and full-year earnings forecasts; Walmart’s official earnings release subsequently showed that Walmart U.S. Q4 comparable sales fell 0.4%, traffic fell 1.7%, and the company’s Q4 profit fell about 21% year-over-year. To be precise, that was not a comprehensive year-over-year decline in Walmart’s total revenue—Walmart U.S. Q4 net sales still grew 2.4%—what really got hit were the comparable sales, traffic, and profit metrics. [K5][K6]

This is the ugliest part of the K-shaped economy: banks see credit cards still being swiped and bad debts not yet exploding; the stock market trades on big bank profits and consumer resilience; but the retail end sees high-income groups sinking into discount channels, and low-income groups relying on food stamps to maintain basic consumption. On the surface, consumption hasn’t collapsed, but structurally it has already split. The so-called “resilient” is not nationwide strength, but the U.S. economy still managing to piece together a decent aggregate number through upper-tier consumption, discount migration, credit expansion, and government transfer payments.

And this is precisely the most insidious part of stagflationization. It doesn’t necessarily erupt in the form of a financial crisis, nor does it necessarily cause the job market to suddenly collapse. It’s more like a slow cash flow drain: inflation pushes up food, gasoline, rent, and daily consumption, high interest rates push up credit card and debt rolling costs, and wage growth cannot keep up with price increases. Macro data doesn’t look broken yet, but the household sector is already using future cash flow to pay today’s prices.

In Q1 2026, U.S. real GDP annualized growth was only 1.6%, but over the same period, the PCE price index rose at an annualized rate of 4.5%, and core PCE rose 4.4%; by May, CPI rose to 4.2% year-over-year, energy rose 23.5% year-over-year, and gasoline rose 40.5% year-over-year. [S1][S2] Employment data remained stable on the surface, with the U.S. unemployment rate holding at 4.3% in May and nonfarm payrolls increasing by 172,000; but average hourly earnings grew 3.4% year-over-year, already failing to outpace the 4.2% CPI, and the number of long-term unemployed reached 2.0 million, an increase of 524,000 from a year earlier. [S2][S3]

Forecasts and survey data are also converging in the same direction. The Philadelphia Fed’s Q2 2026 Survey of Professional Forecasters has already placed weaker growth, higher inflation, and higher recession probabilities in subsequent quarters into its forecast path. [S4] The University of Michigan’s preliminary May 2026 consumer survey showed that consumer confidence remained near 2022 lows, current conditions fell about 9%, and consumers continued to be hit by high prices, gasoline prices, and tariff pressures, with 1-year inflation expectations still at 4.5%; more importantly, groups with little or no stock holdings felt significantly heavier pressure from high prices. [S5] Reuters, based on the June Beige Book, reported that U.S. economic activity and inflation both rose in recent weeks, with retail visits declining, credit card usage increasing, demand for necessities stronger, and middle-income households stretching every dollar further; the same report also mentioned that 9 out of 12 Fed districts mentioned data center construction. [S6] Another Reuters economist survey also showed that, against the backdrop of persistent war-driven inflation, expectations for rate cuts within the year have further faded, and the Fed is likely to continue holding rates. [S7]

This is what I mean by stagflationary pressure: growth is not truly expanding at high speed, but the cost of living is pressing down on ordinary people again; employment has not collapsed across the board, but real purchasing power is being eroded; consumption has not stopped, but more and more consumption may be sustained by credit cards and debt rolling.

My original judgment was that capital did not want to directly trigger a financial crisis, but rather wanted to avoid a financial crisis erupting in an uncontrollable way. It was more like trying to manually engineer a controllable stagflation: letting inflation slowly steal ordinary people’s purchasing power, letting high interest rates suppress demand and valuations, and letting asset prices reprice over a long cycle. This way, the crisis would not arrive as a one-day crash, but would be broken down into years of costs—gasoline, food, rent, credit card interest, employment anxiety, and real wage declines—paid collectively by all ordinary people.

But what I didn’t expect was that in the end, it still went to a more direct step.

If it were just stagflation, then capital could at least pretend it was a macro cycle; but when AI companies, cloud providers, chipmakers, data center contractors, and giant platforms like SpaceX start centrally selling “the future” to the public market, things become blunt. They are not waiting for the bubble to naturally digest; instead, while the bubble still has narrative power, while retail investors still want to believe in the future, and while index funds and pensions still must allocate to growth assets, they are distributing the last round of risk as equally as possible.

This is what I call “harvesting the whole society.”

Not because AI has no future. On the contrary, AI is definitely the future. The problem is that the future itself cannot prove that all of today’s valuations are reasonable, nor can it guarantee that those who sell the future today will bear the same risk as those who catch the falling knife tomorrow.

Technological revolution and financial bubbles have never been mutually exclusive. Classic studies on financial instability, manias-panics-crashes, irrational exuberance, and IPO cycles repeatedly illustrate the same thing: when financing, narrative, leverage, and exit windows align, genuine technological progress can also be processed by capital markets into a bubble structure. [B1][B2][B3][B4] Railroads, electricity, the internet, and new energy have all proven: technologies that truly change the world often first create a bubble that destroys latecomers. The place where capital is most adept at substitution is packaging “directionally correct” as “reasonably priced,” packaging “the technology will exist” as “the current company will win,” and packaging “the industry will grow” as “people who buy in today will make money.”

AI is the same.

Even more cruelly, the companies that look most like winners today may not be the ones that truly remain after the bubble. Technology cycles never guarantee that first movers survive to the end, nor do they guarantee that the companies doing the best now will still be doing the best in the next paradigm. Today’s lead may just be a phased lead created by financing windows, GPU supply, brand credibility, and distribution channels; when cost structures, hardware forms, algorithm paradigms, or regulatory environments change, today’s winners can also become tomorrow’s obsolete assets.

OpenAI Infrastructure

OpenAI was the first place I felt something was off. As an AI giant, it should have been the closest to the future, but its actions looked less like a company that had found a profitable endgame and more like a capital node that had to bind everyone into the same growth curve.

The Stargate Project announced that over the next four years, $500 billion would be invested to build new AI infrastructure for OpenAI in the U.S.; participants included OpenAI, SoftBank, Oracle, MGX, with technology partners including Oracle and NVIDIA. [O1][O2] Subsequently, OpenAI reached an AI infrastructure agreement with CoreWeave worth up to approximately $11.9 billion and obtained equity in CoreWeave; Reuters also reported that the deal occurred on the eve of CoreWeave’s IPO. [O3][O4] Then came Oracle’s roughly $300 billion computing power procurement agreement, AMD’s 6GW GPU deployment, and Broadcom’s 10GW OpenAI-designed AI accelerators. [O5][O8][O9][O10]

These numbers are too large—so large that they are no longer ordinary business expansion but the financial structure itself.

If OpenAI already had a clear, stable, self-financing path to profitability, these agreements could be understood as capacity expansion. But reality is more complex. Reuters, citing a WSJ report, stated that as OpenAI sprinted toward its IPO, revenue and user targets fell short of expectations, raising external concerns about whether it could support its massive data center spending. [O11] HSBC analysts also estimated that OpenAI might need about $207 billion in new capital to meet its data center spending commitments through 2030 and might still be unprofitable by 2030; this is only an analyst estimate, not a company-confirmed fact, but it sufficiently explains why the market is starting to worry about its funding gap. [O12]

This is hardly ordinary commercialization. It’s more like binding cloud providers, chipmakers, data center financiers, U.S. policy narratives, and public market investors all to the same future story.

OpenAI might not be the most malicious actor. On the contrary, it might be the person at this table who first realized the bubble was inevitable. Sam Altman knew from the beginning that whether it was OpenAI or not, AI would be pushed into a bubble. Given that, the most rational approach was not to exit, but to turn oneself into the center of the bubble, binding the entire AI industry, the U.S. government, cloud providers, chipmakers, and capital markets.

It thought the U.S. government would bail out the market, and maybe eventually it will. But politics is not a stabilizer; politics itself can also become a transaction. Especially when Trump once again ties tariffs, industrial policy, capital market sentiment, and personal political narrative together, the market is no longer just a market—it becomes a K-line that can be drawn together by policy, public opinion, and liquidity.

This is the sharper point I wanted to make in my original text: when a capitalist-style president regains the ability to directly influence asset prices, many traditional “policy objectives” get reordered by asset prices. Tariffs can be negotiation tools, regulation can be a political tool, AI can be a national competition narrative, and the stock market itself can become a dashboard for demonstrating governance capability. Whether the government will bail out the market is no longer just an economic question, but a question of political price.

NVIDIA Circular Investments

Originally, every party at the table was gambling, betting that they wouldn’t be the loser. So everyone went all out, until everyone realized they were all exhausted. Profits had been compressed to the limit, while the AI economy was still circulating internally. Model companies needed computing power, cloud providers needed contracts, chipmakers needed orders, data centers needed financing, electricity and land needed policy, and capital markets needed growth stories. Each node was validating another node’s valuation, and each node needed another node to continue believing.

This is the most dangerous part of so-called circular trading.

It doesn’t have to be financial fraud, nor does it have to be a conspiracy, but it entangles real demand and capital demand together, allowing investments, revenues, equity, contracts, and market confidence to feed each other. NVIDIA invests in OpenAI, and OpenAI purchases NVIDIA systems; OpenAI signs massive computing power contracts with cloud providers, who use the contracts to finance, go public, and expand, in turn supporting OpenAI’s computing power narrative. NVIDIA’s announcement of up to $100 billion investment in OpenAI and its collaboration with OpenAI to deploy 10GW of NVIDIA systems is precisely the most representative node in this type of structure. [O6][O7][O13][O14]

This structure looks like a synergistic ecosystem during an upcycle, but during a downcycle, it becomes a chain of cross-collateralization.

NVIDIA is actually the most special. Unlike other companies that can cash out and exit through IPOs, secondary market share sales, or equity financing, as the hardest shovel-seller in the entire AI bubble, its cash-out method has always been selling hardware: selling GPUs, selling systems, selling networking, selling entire data center infrastructure. Its every move is analyzed by the market because it can no longer claim to be just a bystander; its customers, investment targets, revenue expectations, and the entire AI infrastructure bubble are tied together. It’s not that it can’t make money, but that it makes too much money, so much that it can only continue selling hardware, praying to become a second IBM—an infrastructure company that survives the bubble—rather than a hardware cyclical stock that gets revalued when the bubble bursts.

The Hardware Cost-Effectiveness of LLMs

There is also a technical judgment here that cannot be covered up by the capital narrative: what truly fails first for the LLM path may not be the technology itself, but hardware cost-effectiveness.

I believe scaling laws still exist. Larger models, larger data, longer training, more complex inference can indeed continue to bring capability improvements. But the problem is, if every capability improvement requires larger clusters, higher inference costs, more complex engineering scheduling, and higher electricity and cooling costs, then the bottleneck is no longer “can it be done,” but “is anyone willing to pay for this cost long-term.”

When a model must run across dozens or even more devices, it may indeed be stronger, but it also begins to move away from the economics of ordinary software. What is truly fatal is not that it isn’t intelligent, but that it is too expensive, too slow, too heavy, and too dependent on hardware cycles. API price wars will compress profits, inference costs will swallow cash flow, enterprise deployment will lengthen collection cycles, and open-source models will continuously lower the chargeable boundary. The technical path may not have failed, but the commercial cost-effectiveness will fail first.

So I say, the failure of LLMs is fundamentally not about technology, but about a cost-effectiveness mismatch on hardware. When better hardware appears, we may indeed have stronger models; but by that day, the algorithm paradigm itself may have already changed. History often goes like this: the first-generation technology path proves the direction, and the second-generation hardware and algorithms truly harvest the world. Those who build it first today may not survive to the end; those who do it best today may not continue to be the best in the next phase.

Companies like Oracle and CoreWeave are more like middle layers. Oracle can continue to hold up with its balance sheet, cloud contracts, and capital market narrative; neocloud companies like CoreWeave are more fragile. Their value comes from GPUs, data centers, contracts, and financing capabilities, but if the market begins to doubt the payment ability of model companies, or doubts that AI demand growth can cover capital expenditures, then these companies will shift from “core AI infrastructure assets” to “high-leverage cyclical assets.”

The biggest problem for this type of company is not a lack of assets, but that the assets are too heavy and the credibility too light. GPUs, server rooms, power contracts, and long-term computing power contracts all look like moats during an upcycle; but once the market is no longer willing to value GPUs and long-term computing power contracts as growth stocks, they will quickly turn from “future infrastructure” into “hardware contractors weighed down by debt, depreciation, and utilization rates.” Giants still have cash flow, ecosystems, and policy relationships to tell stories; the neocloud story is narrower: if model company demand gets discounted, they will be crushed into the road by the hardware cycle like a steamroller, becoming the sunk cost of the next round of infrastructure.

Executives Rushing to Cash Out

The most glaring thing about CoreWeave is not just its dependence on GPUs, data centers, massive debt, and a few large customer contracts, but the cash-out pace of insiders.

According to a Bloomberg report in June 2026, CoreWeave’s stock price had more than doubled since its IPO in March 2025; but over the same period, company executives had sold over $2.3 billion in personal holdings. Bloomberg, citing insider trading data from Washington Service, stated that the main sellers were precisely CoreWeave’s three co-founders: Michael Intrator, Brannin McBee, and Brian Venturo. [CW1] Subsequent market reports further noted that these transactions were executed through prearranged trading plans; among them, Brian Venturo sold over $1.1 billion after the lock-up period ended, and the three founders still held about 18% of the company’s shares after the sales, with Intrator remaining the largest shareholder at about 10.4%. [CW2]

This does not mean CoreWeave necessarily has problems, nor does it mean these transactions are illegal. Prearranged trading plans are a common share reduction arrangement used by executives of U.S. public companies; without further confirmation from individual SEC filings, I will not write this as a verified 10b5-1 fact. But when a company has just been pushed onto the public market by the AI infrastructure narrative, its stock price has skyrocketed, capital expenditures are huge, debt pressure is heavy, customer concentration is extremely high, and the founding team and executives have cumulatively cashed out over $2 billion, it is itself a very strong signal: the primary market and insiders have already turned the AI infrastructure story into cash, while public market investors are beginning to bear valuation, financing, and cyclical risks.

CoreWeave’s IPO documents also explain why this matters. Its S-1/A shows that the company issued 47,178,660 shares of Class A common stock in the IPO, and selling shareholders issued 1,821,340 shares; the company would not receive proceeds from the shares sold by selling shareholders. At the same time, after the IPO, the three co-founders Michael Intrator, Brian Venturo, and Brannin McBee still held approximately 37.0%, 23.2%, and 18.7% of voting power, respectively, totaling about 79.0%. [CW3] In other words, CoreWeave is not a company that has been fully publicized in the ordinary sense: control remains highly concentrated, but market risk has been publicized.

This is the core structure of the entire AI bubble: control is in the hands of insiders, and risk is in the hands of the public market.

IPO

Google / Alphabet also cannot stay out of it. It is, of course, a genuinely profitable company with search, advertising, cloud, and Gemini. But when Alphabet announced a massive equity financing to expand AI infrastructure and compute, its signaling significance was still very strong. [A1][A2] In the same phase, OpenAI also formally submitted a confidential draft S-1, putting the future listing option on the table. [I1] Even the most cash-flow-rich giants and the most core model companies are starting to more directly connect AI infrastructure, IPO paths, and capital markets, indicating that AI is no longer an internal R&D project but an infrastructure war that needs to be jointly borne by capital markets.

The truly dangerous signal is not that one company is financing, but that everyone is financing almost simultaneously. Every company is afraid of becoming the last one to stand up and ask for money, because the last financier is most easily identified by the market as “already unable to hold on.” So the rush itself becomes a run: Google first launches equity financing, OpenAI quickly puts out its confidential S-1, Anthropic also queues up to go public, and SpaceX packages the Musk platform into the public market. Every company says it is financing for the future, but from the market structure perspective, it looks more like everyone sees that the number of lifeboats is limited and suddenly starts running.

Anthropic is the same. On one hand, it represents the AI safety narrative; on the other, it submits a confidential draft S-1, preparing for an IPO. [H1] This does not prove Anthropic is hypocritical, nor does it prove it is deliberately harvesting. But it does show that discourses like safety, ethics, alignment, pause, and frontier risk are already coexisting with the capital exit cycle. When the Anthropic Institute discusses recursive self-improvement and proposes the need to establish verifiable mechanisms to slow down or temporarily pause frontier AI development, this may be sincere in terms of technical governance; but when it appears in the same period as an IPO, what the market sees is not just safety, but “the safety narrative is also going public.” [H2]

The problem is not that these safety discussions are necessarily fake, but that they can also become the emotional packaging most easily understood by ordinary investors before an IPO: the more you emphasize “this thing is too powerful and must be regulated,” the more you reinforce “this thing represents the future”; the more you talk about frontier risk, the more you tell the market “we stand at the frontier.” This is the most subtle part of the safety narrative: it can be sincere governance language, and it can simultaneously become fear marketing for the capital market. Whether you believe its motives or not doesn’t matter; the market will translate it into valuation language.

SpaceX takes this to another level.

It is no longer just an aerospace company, nor just a Starlink company. SpaceX’s S-1 divides its business into three segments: Space, Connectivity, and AI, and incorporates the AI compute, Grok, and X from the xAI acquisition into the AI segment: rockets and Starship are the physical capability to enter orbit, Starlink / Starshield are communications and government security networks, X is the data and distribution layer, Grok / xAI is the AI layer, and SpaceX becomes the capital and engineering parent body of the Musk company matrix. [X1]

The segment data in the S-1 is very blunt: in 2025, the Space segment had revenue of about $4.086 billion and an operating loss of $657 million; Connectivity / Starlink had revenue of about $11.387 billion, operating profit of $4.423 billion, and Segment Adjusted EBITDA of about $7.168 billion; the AI segment had revenue of about $3.201 billion, but an operating loss of $6.355 billion, and Segment Adjusted EBITDA of -$1.237 billion. [X1] Even more glaring is the capital expenditure: in 2025, AI segment capex was about $12.727 billion, and in Q1 2026, AI segment capex reached another $7.723 billion. [X1]

In other words, the SpaceX IPO is not simply selling rockets and Starlink to the market, but packaging Starlink’s cash flow, xAI/Grok’s losses, AI compute’s capital expenditure, X’s platform narrative, U.S. defense and communication networks, and Musk’s personal credibility all into a super-platform story that can be purchased by the public market.

So 2026 naturally becomes the harvesting year of the last dance.

This last dance most resembles a financing run. In a normal cycle, companies finance according to business maturity; at the end of a bubble, companies finance according to “who hasn’t had a chance to sell their story yet.” Everyone is afraid of becoming the clown scrambling for money on the eve of the deadline, because that means the market has already started asking: why do you need money now? Why has everyone else already sold the future, and you’re still standing at the door? So the rush itself exposes fear, and fear drives more rushing.

The FOMO sentiment in 2025 was clear to everyone. When OpenClaw exploded in popularity, nearly a thousand developers and AI enthusiasts lined up outside Tencent’s Shenzhen building for installation. Sina Tech reported that on March 6, 2026, nearly a thousand people lined up outside the Tencent building, and Tencent Cloud engineers completed free cloud installations of OpenClaw for users; because OpenClaw’s local deployment environment is complex, door-to-door installation services had already appeared on platforms like Xiaohongshu, with one-time installation fees ranging from 300 to 1,000 yuan in cities like Shenzhen, Guangzhou, Hangzhou, and Chongqing; platforms like Xianyu also appeared with tutorials, one-click installations, and remote assistance services, priced from tens of yuan to nearly 200 yuan. [OC1] 21st Century Business Herald also reported that nearly a thousand people lined up outside Tencent’s Shenzhen office area, and door-to-door installation pricing on domestic platforms like Xiaohongshu and Xianyu was mostly 300–800 yuan, with 500 yuan per session becoming the mainstream price, and remote installation also costing 50–100 yuan, with some online stores already selling 1,000+. [OC2]

This is not just a tool becoming popular, but the industrialization of FOMO. OpenClaw requires calling large model capabilities, can be deployed locally or in the cloud, and behind it requires model APIs, cloud computing power, servers, plugins, and configuration costs. Sina Tech mentioned that Tencent Lighthouse Cloud launched a one-click deployment template for OpenClaw at the end of January, and Baidu AI Cloud launched an ultra-simple version deployment plan for OpenClaw in February. [OC1] Cailian Press further reported that domestic cloud providers such as Tencent Cloud, Alibaba Cloud, China Mobile Cloud, China Telecom Cloud, JD Cloud, Volcano Engine, and Baidu AI Cloud all connected to OpenClaw; cloud providers, through one-click deployment, permission encapsulation, security hardening, etc., transformed the open-source project into a sellable SaaS service, essentially using computing power, storage, networking, and engineering capabilities to lower user barriers and seek commercial monetization loops. [OC3]

At the same time, cloud providers and model parties also simultaneously launched various Coding Plans. A Tencent Cloud developer article stated that agent workflows like OpenClaw consume tokens at high frequency, and a moderately complex task could trigger dozens of model calls; between the end of 2025 and March 2026, major domestic cloud providers and model companies intensively launched Coding Plan subscription packages that replaced per-token billing with a fixed monthly fee, including Alibaba Cloud Bailian, Volcano Ark, Tencent Cloud, Zhipu, Kimi, MiniMax, etc. [OC4] Sohu tech articles also summarized this phenomenon as the OpenClaw “national shrimp farming” era, with various cloud and model providers launching Coding Plans, exchanging a fixed monthly fee for model call quotas, attempting to solve the token cost anxiety of heavy OpenClaw users. [OC5]

But I believe this is precisely the most typical part of a bubble: at the time, they looked like entrances to a new era, like AI workflows that everyone must have; but once the craze recedes, many things will become negative assets. Subscriptions, quotas, computing power, plugins, workflows, training costs, team migration costs—all will become sunk costs after the bubble recedes. BBC Chinese later wrote in a report that a reversal from “raising lobsters” to “uninstalling lobsters” was emerging: some users felt that OpenClaw had high technical barriers, was expensive, and had little practical use, and began paying for uninstallation; experts also warned that OpenClaw is currently “not very practical,” and the costs and risks are disproportionate. [OC6] People’s Daily Online Shenzhen also warned that although OpenClaw represents a new trend in AI agents, the deployment threshold, long-term token costs, hardware costs, operation and maintenance costs, and security risks are not low, and ordinary users should not blindly follow the trend. [OC7]

So this is not as simple as “a tool got popular.” It’s more like a preview of the AI bubble at the ordinary person level: first create anxiety with a new entrance, then lower the barrier with free installation, then lock users into the ecosystem with cloud deployment, Coding Plans, plugin markets, and workflow migration. While the bubble is still there, these costs are explained as “learning the future”; after the bubble recedes, they will become bills.

The stock market is the same. As prices chased higher, everyone was tempted. Fund managers couldn’t miss AI, pensions couldn’t miss AI, indices couldn’t be without AI, and retail investors couldn’t watch others make money. So everyone knew it was expensive, but everyone was afraid of getting off too early. The strongest moment of a bubble is not when no one knows it’s a bubble, but when everyone thinks they can leave before the music stops.

This year, all model companies understood that they need to sell models, to directly sell this concept, to sell the credibility of Gemini, Claude, ChatGPT, Grok in their hands, in exchange for money, to survive the tough times ahead.

This sentence sounds harsh, but it might be the essence of capital markets. The models themselves may not be immediately profitable; API price wars will compress profits, open-source models will lower moats, enterprise deployment will lengthen sales cycles, and inference costs will continuously consume cash. But “models represent the future” itself can be securitized. As long as the market is still willing to pay for the future, model companies can sell the future to the market.

OpenAI sells the credibility of ChatGPT, Anthropic sells the credibility of Claude and safety, Google sells the credibility of Gemini, and SpaceX / xAI sells the composite credibility of the Musk platform, Grok, and X.

And what ordinary people buy is a future that others have already priced.

Index Funds

The SpaceX IPO is even more critical because of the index mechanism.

Reuters reported that SpaceX chose to list on Nasdaq, planning to use SPCX as its ticker, targeting about $75 billion in fundraising and a valuation of about $1.75 trillion, which, if completed, would be the largest IPO in history; the same report also mentioned that one of the backgrounds for SpaceX choosing Nasdaq was Nasdaq’s new fast entry rule, allowing newly listed large-cap companies to enter the Nasdaq-100 more quickly. [X2] Nasdaq’s official FAQ made the mechanism even clearer: starting May 1, 2026, eligible new Nasdaq-listed companies ranking in the top 40 can be included in the Nasdaq-100 15 trading days after listing, subject to a 3x float cap constraint; Nasdaq also adjusted the market capitalization ranking methodology so that both listed shares and unlisted shares can be used for eligibility/ranking market capitalization, but unlisted shares are used for eligibility and ranking, not for actual weight calculation. [X3]

This means SpaceX is not being “slowly chosen by the market after listing” in the ordinary sense. It embedded the buying mechanism of indexed funds from the listing design stage.

Active investors can at least

---