Summary

In H1 2026, the AI trade spread from model stocks into a wholesale repricing of legacy IT hardware. 庄明浩 tracked roughly 17 CPU, memory, storage, hard-drive and networking vendors, whose shares gained an average 192% in the first half; SK Hynix was up more than 700% at one point. 14 of the 16 best-performing S&P 500 stocks were IT hardware companies. “The old friends are back,” but a 13% one-day drop in SpaceX and an 8% to 9% slide in Micron also show that the repricing has entered a high-sensitivity zone.
Capex is not peaking; it is beating prior expectations for a 3rd straight year. The market initially forecast 19% growth in 2024, versus an actual 54%; it expected roughly 20% growth in 2025, versus an actual 73%; 2026 now looks capable of reaching 80% or even 100%. Google, Meta, Amazon, Microsoft, Oracle and Apple are set to spend roughly $800B in 2026, with 2027 potentially clearing $1T; Agents are broadening the beneficiary pool to CPUs, memory, storage and inference chips.
Storage is the strongest—and most obviously cyclical—link in this infrastructure rally. Storage’s share of data-center costs has risen from roughly 2% before the AI boom to 18% in 2026, second only to compute chips at 42%; Micron’s gross margin rose from -33% in 2023 Q1 to 85% in May 2026. The “this time is different” demand curve may prove real, but an industry that has historically cycled roughly every 4 years and tends to overshoot into excess supply after capacity expansions cannot be valued on linear growth alone.
The bulls’ real weak point is not the demand story but cash flow and debt. 庄明浩 says that, as of Q2 2026, industrywide revenue “just barely covers depreciation” and has yet to cover fixed costs and operating expenses, let alone produce profit. Several tech giants’ free cash flow has fallen rapidly since Q3 2025 and could turn negative in Q3 2026 on current forecasts. The megacaps can still issue debt smoothly, but highly levered neoclouds such as CoreWeave and Nebius have already fallen 30% to 50% from their highs. “The equity is usually fine, but the debt keeps breaking.”
The model frontier is moving from “answering better” to working autonomously for extended periods, but valuations are outrunning profits. The latest frontier models, including Claude 5 and GPT-5.6, can now execute continuously for 10-plus hours without human involvement; OpenAI and Anthropic are taking different routes toward the same endpoint: AI training AI. At the same time, Anthropic and OpenAI have reached valuations of more than $900B and $800B, respectively; OpenAI generated roughly $25B of revenue and lost about $14B in 2025, implying a P/S multiple of roughly 40x on that year’s numbers.
The US and China have formed a bipolar market of “expensive closed-source frontier models” and “efficient open-source diffusion,” but the technology has not truly decoupled. Leading US companies are spending $700B to $800B annually, potentially approaching $1T; Alibaba, Baidu, Tencent and ByteDance together are spending just over $100B in China, a gap of roughly 1 order of magnitude. Yet HARVEY took the top spot in a legal benchmark by post-training the open-source GLM-5.1, while Cursor was found to have post-trained on Kimi K2.5. “As long as it is open source,” the technology is difficult to divide into 2 absolute camps.
Agent, not Chatbot, is the product consensus for 2026, with Coding the first entry point to be swallowed. Codex’s active users rose from 5M in May to 6M, 7M, 8M and 9M from July 12 through July 15; 庄明浩 judged that the figure should have reached 10M later that day. Anthropic engineers reportedly saw average code submissions rise 8x after Claude Code launched. “Chatbot is completely behind us,” but there are still no standard answers on what a general-purpose Agent does, who pays, how it is priced, or whether it is primarily To B or To C.
The biggest unknown is no longer how many tokens the industry can burn, but which outcomes are worth paying for and which responsibilities must remain human. Top software companies among America’s leading enterprises are already spending $7,000 to $7,500 a month on tokens, roughly half the $15,000 monthly salary of an engineer, but tokens may be like page views in the early internet or data traffic in mobile internet: an intermediate metric rather than the end business KPI. Governance conflicts, employment backlash and the redefinition of human roles will all intensify. “Once intelligence is no longer scarce, where exactly do humans fit?”

Deep dive

1. “The Old Friends Are Back”: AI Is Repricing the Entire IT Supply Chain

张震岳 and 蔡依林 winning the 2026 Golden Melody Awards’ Male and Female Singer of the Year awards gave 庄明浩 a sense of temporal dislocation—“What year is it?” The same feeling has surfaced in the capital markets: Intel, AMD, Samsung, Micron, SanDisk, Western Digital, Seagate, Nvidia and Cisco, names from the PC era, have returned to the center of the AI trade.
His sample of roughly 17 companies gained an average 192% in H1 2026, while SK Hynix was up more than 700% at one point. 14 of the 16 best-performing S&P 500 stocks were in IT hardware; the other 2 were pharmaceutical companies. In his shorthand, the technology market now has only “AI and everything else,” with drugs the largest part of everything else.
The repricing has not been one-way. In the week after the episode was prepared, SpaceX fell roughly 13% in a single day and Micron dropped 8% to 9%. 庄明浩 sees the volatility as evidence of 2 narratives continuing to collide, not of the trend having been conclusively validated.

2. The Renaissance Was About Collapsing Replication Costs, Not Inventing Knowledge from Nothing

庄明浩’s core analogy is that the printing press did not invent knowledge; it collapsed the cost of reproducing it. Today, falling inference costs are turning intelligence into infrastructure. Once scarcity changes, “everything gets repriced.”
The second parallel is the patronage system. Wealthy patrons supported artists then; today, megacap capital is concentrating around AI, while top researchers depend on the protection of compute and money. The 3rd parallel returns to humanism: the old question was “What is a human?” The new one is “What does a human still need to do?”
The Medicis map to capital and debt, the printing press to open source and low-cost inference, the Northern and Southern Renaissances to the US and Chinese approaches, perspective to world models, workshops to multi-Agent division of labor, and the “Bonfire of the Vanities” to regulation and social backlash. That mapping is the episode’s organizing framework.

3. Trillion-Dollar Capex Is Still Accelerating, and Agents Are Broadening the Infrastructure Beneficiary Set

A Goldman Sachs lookback has made 庄明浩 wary: at the end of 2023, the market forecast 19% capex growth for 2024, versus an actual 54%; it then expected roughly 20% growth in 2025, versus an actual 73%; even after raising its 2026 forecast, reality could still come in at 80% or even 100%.
Google, Meta, Amazon, Microsoft, Oracle and Apple are expected to spend roughly $800B in 2026, and 2027 will “definitely clear $1T.” If the market is still forecasting only 20% to 30% growth at that point, could actual growth again reach 50%, 60% or 70%? 庄明浩 calls the possibility “a little terrifying.”
The demand mix has also changed. The pretraining era mainly bought GPUs; the Agent era also needs infrastructure to allocate tasks, retain long contexts and execute repeatedly. CPUs, memory, storage, TPUs and other custom inference chips are now on the procurement list. The “token seller” opportunity therefore extends across foundries, chips, storage, networking, cloud, energy and land.

4. Storage Has Risen to 18% of Costs, but It Still Carries the Original Sin of a Cyclical Industry

Before the AI boom, storage accounted for roughly 2% of data-center costs; by 2026, that share could reach 18%. In the same cost breakdown, compute chips account for 42%, energy 15%, buildings 15% and cooling 10%, making storage the 2nd-largest line item after chips.
Each new generation of Nvidia chips delivers more compute but also requires storage on a geometric growth curve. Both unit volumes and prices are rising, expanding storage’s absolute dollar value and cost share at the same time. Supply is also highly concentrated, allowing revenue and gross margin to accrue rapidly to a small group of vendors.
Micron is the extreme case. Its gross margin reached 85% in May 2026 after falling as low as -33% in the previous trough. Micron’s CEO has also complained that Apple can buy memory for $8 and sell the resulting incremental product value for more than $200.
庄明浩 has not forgotten the counterargument: storage has historically followed a roughly 4-year cycle, with rising demand triggering new factories just as demand may be weakening when the capacity comes online. SK Hynix’s more-than-700% rally, its US listing and the record fundraising amount it achieved as an overseas company listing in the US already embed a substantial “this time is different” premium.

5. Intel and Cisco Have Cleared Their Dot-Com Bubble Highs; the Other Side of the Ledger Is the Need for External Demand

Intel and Cisco were icons of the 2000 internet bubble and failed to break those highs for roughly 26 years. By late 2025 and into 2026, both had finally completed “A Long Way Back.” It is both a return of value to old-line hardware and a market vote for a new infrastructure cycle.
SpaceX’s xAI may no longer be pursuing large-scale model training. When SpaceX listed, it signed a compute contract with Anthropic, effectively allowing Elon Musk to provide his own compute capacity for Anthropic’s use. Meta, which has no traditional cloud business, has also begun pitching a neocloud and making some of its compute available to partners.
庄明浩 brings the question back to the scale of the spending: large amounts of hardware and data-center capacity have already been built, but AI revenue has not yet generated enough profit to absorb the investment. External usage and cloudification offer one way to create that demand, while also exposing whether revenue growth can keep pace with capex.

6. AI Revenue Currently Covers Only Depreciation; Cash Flow and Debt Will Decide When the Bubble Is Tested

庄明浩’s key judgment is blunt: “Current revenue just covers depreciation.” A normal business model should cover depreciation first, then fixed costs and operating expenses, before producing gross margin and profit. As of Q2 2026, aggregate AI revenue remains stuck at the 1st layer.
The free cash flow of the tech giants has fallen sharply since Q3 2025 and could turn negative in Q3 2026 on current forecasts. Google, Meta and Nvidia can still issue debt smoothly; even long-dated bonds can attract several times the offering size. That financing privilege does not extend to every participant.
Neoclouds such as CoreWeave and Nebius carry extremely high leverage, and some companies’ revenue may not even be sufficient to cover interest expense. Their stocks have fallen 30% to 50% from their highs. 庄明浩 repeatedly stresses: “The equity is usually fine, but the debt keeps breaking.”
The bull case is that models will continue to improve, unit inference costs will fall, token consumption will rise by multiples, and technological revolutions typically require investment before monetization. The bear case is that enterprise adoption is moving far more slowly than supply expansion, with circular financing and cross-guarantees eventually becoming a problem. The verdict remains open: “A Medici golden age, or the eve of a tulip bubble?”

7. Models Can Now Execute Autonomously for 10-Plus Hours; 2 Doctrines Converge on the Same Self-Improvement Loop

Claude 5, GPT-5.6 and the other latest models cited by 庄明浩 can already execute tasks continuously for 10-plus hours without human involvement. A multi-model world is also becoming normal: coding, reasoning, multimodality and long-context tasks each have their own temporary leaders, and no company can maintain an absolute lead indefinitely.
OpenAI’s L1-to-L5 roadmap defines L3 as Agent and L4 as researcher; Anthropic’s “When AI Builds Itself” describes humans gradually exiting the training loop. The companies represent “2 doctrines, 1 consensus”: put AI inside the training process and create AutoResearch, RSI and a continuous iteration Loop.
The capability is also accelerating release cycles. Within months, the market moved through Opus 4.6, 4.7, 4.8 and 5, as well as GPT-5.5 and GPT-5.6. Meta and xAI have not exited the race, while Mistral, Thinking Machines and a large number of new labs founded by researchers leaving leading laboratories continue to attract tens of billions of dollars in capital.

8. Anthropic Has Temporarily Pulled Ahead of OpenAI, While Near-Trillion-Dollar Valuations Arrive Before Profits

庄明浩 says Anthropic overtook OpenAI in B2B subscriptions in April or May 2026. Around the same time, its new financing round valued it at more than $900B, above OpenAI’s valuation of more than $800B. Neither company is public, yet both are already approaching the trillion-dollar mark.
The earnings comparison is stark. OpenAI generated roughly $25B of revenue and lost about $14B in 2025, implying a P/S multiple of roughly 40x on that year’s numbers. The figures show how far the valuation has moved ahead of current revenue and losses.
Apple took 42 years, Google 21 years and SpaceX 24 years to reach the trillion-dollar range; OpenAI has taken roughly 10 years and Anthropic roughly 5. SpaceX fell below its IPO price in fewer than 10 trading days after listing. Whether OpenAI and Anthropic go public as planned will also depend on market sentiment.

9. The US and China Have Formed 2 Distinct Paths, but Open Source and Post-Training Keep Them Interpenetrating

庄明浩 uses the Northern and Southern Renaissance as an analogy: Italy relied on patrons, expensive commissions and masterpieces by individual geniuses, while the North relied on prints, printing and urban buyers. Silicon Valley maps to extremely expensive compute and closed-source flagship models; China leans toward open source, efficiency, low prices and mass diffusion.
The capital gap is tangible. Leading US companies are investing $700B to $800B a year, potentially approaching $1T; Alibaba, Baidu, Tencent and ByteDance together are spending just over $100B in China, a gap of roughly 1 order of magnitude. Will chip restrictions and company scale widen the model gap further? His answer is: “I don’t know. Watch and see.”
China’s field includes Zhipu, Qwen, MiniMax, DeepSeek and Kimi, with the overall market leaning more open source. The secondary market is equally unforgiving: Zhipu once fell 30% in a day, its market cap pulled sharply back from around HK$1T, and MiniMax at one point fell below its IPO price. “These swings kill people.” 庄明浩 also cautions that it is hard to know whether OpenRouter can cover even 1% of global traffic, yet the market still treats it as an important third-party gauge.
Integration cases matter more than slogans. HARVEY post-trained the open-source GLM-5.1 and beat Opus 4.8, GPT-5.5, Sonnet 4.6, GPT-5 and Kimi K2.6 on a legal benchmark; Cursor’s proprietary model was found to have been post-trained on Kimi K2.5. Open source makes absolute isolation difficult to sustain.

10. World Models Are Like Perspective: The Contest Is Over the Next Representation of Reality

Perspective did not invent pigment; it rewrote painting, architecture and maps through a new system of representation. World models likewise seek to move AI from “predicting the next word” to “predicting the next state of the world,” combining text, images, video, geometry, physics and semantics from the start of training.
李飞飞’s taxonomy divides the field into 3 categories: renderers that generate content for human vision, simulators that reproduce physical rules, and planners that select actions for robots and autonomous vehicles. Video models, 3D generation, embodied intelligence and autonomous driving therefore converge on a single narrative.
The technology has not converged. 杨立昆 argues for predicting states in latent space rather than generating pixels; 哈萨比斯 and DeepMind lean more toward multimodal understanding; 李飞飞 places greater emphasis on 3D spatial representation. World Labs, autonomous-driving companies and video-model developers are all competing to define the category, and the absence of consensus may itself be the norm for now.

11. Agent Workshops Are Replacing Chatbots, While Claude Code and Codex Rewrite Software Production

Renaissance paintings often came out of workshops: the master handled composition while apprentices executed specialized tasks. For Agents, the equivalent is planning, tool calling, execution, error correction, evaluation and looping, moving from single Agents toward multi-Agent orchestration. “Pure Chatbot has run its course”; coordinated division of labor is the new production capacity.
After OpenAI updated Codex and GPT-5.0, Codex became the core ChatGPT experience. The original ChatGPT was renamed “ChatGPT Classic,” and ChatGPT Work was added. 庄明浩’s judgment is unequivocal: “Chatbot is a chatbot. That chapter is completely over.”
Codex’s active users were roughly 5M in May, then rose to 6M, 7M, 8M and 9M from July 12 through July 15; 庄明浩 judged that the number should have reached 10M later that day. “Adding 1M a day” shows the interaction model shifting from prompt to answer toward task delegation, tool calls, repeated evaluation and human supervision from the sidelines.
Claude Code could become a historical milestone on par with the arrival of ChatGPT. Anthropic’s internal technical staff and software engineers reportedly saw average code submissions rise 8x. Code is not just for programming; it is becoming the harness through which AI executes tasks and trains itself. “Code is swallowing everything.”

12. Agents Do Not Need Traditional UIs; Model Companies Are Expanding into Every High-Value Software Category

As software shifts from being used by people to being called by Agents, the UI matters less while interfaces and protocols such as CLI, MCP and A2A matter more. The open-source ecosystem is also migrating from LLM tooling toward Agent execution tooling, with Coding, personal assistants, Runtime and Agent Infra projects growing rapidly within 6 months.
The model boundary is expanding at the same time. After Opus 4.7, Anthropic rolled out Claude Finance, Claude Legal, Claude for Science, Claude Design and Claude Enterprise, and entered areas including security. Finance, legal, healthcare, design and enterprise services already have traditional or AI-native leaders; model companies are now entering directly to compete for those markets.
The market framework cited by 庄明浩 puts subscriptions plus advertising at roughly $300B, Coding at another $2T and non-Coding white-collar work at another $2T. Those figures cover only the digital world; the physical world represents another potential market. That is why Agents and world models are heating up at the same time.

13. Chinese Agents First Converged on a Common Shell, Then Moved Toward Consolidation Inside the Big Tech Platforms

China has also reached a provisional consensus that the “Chatbot phase is over.” 庄明浩 lists ByteDance’s Trae and Coze, Alibaba’s Coder, MoonRun and Wukong, Tencent’s WorkBuddy, QClaw and Mavis, Baidu’s Z Code and AutoGLM, as well as Kimi Code and Kimi Work. New paid capabilities in Doubao, Yuanbao and Qwen are also shifting toward Agents.
There is still no standard answer on use cases, customers, monetization, or whether the products are To B or To C. “The only thing with a standard answer is what these products should look like at this exact moment.” He credits Codex as the industry’s “teacher” for first turning the appearance, shell and interaction model of a general-purpose Work Agent into a template.
An Analysys report groups WorkBuddy, Trae, QClaw, Coder, AutoGLM, Coze and Wukong as AI-native workplace Agents. 庄明浩 sees these products as general-purpose Work Agents launched by different vendors after OpenClaw appeared.
New entrants also mean old products are being retired. Doubao and Qwen shut down their earlier Agent businesses, which were closer to prompt wrappers; regulation may be part of the reason. 庄明浩’s view is that those products and today’s task-executing Agents are not even the same generation.
DingTalk has already integrated MoonRun, Wukong and Coder. Whether ByteDance’s multiple similar applications, Tencent’s WorkBuddy and QClaw, and Mavis will also be merged is now the next question. Their functions, interfaces and target users overlap heavily, so organizational consolidation will proceed alongside product competition.

14. The Governance Backlash Is Only Beginning; Efficiency and Employment Are the Hardest Tension to Reconcile

The “Bonfire of the Vanities” came from a moral backlash within Florence, not from an external enemy. The person who organized the burning of paintings, mirrors and luxury goods was ultimately burned at the stake as well. 庄明浩’s warning is that anti-AI sentiment will grow, but an excessive backlash is dangerous too.
He identifies 3 tensions: innovation versus safety, which requires distinguishing the destruction of forgery capabilities from the destruction of image generation itself; sovereignty versus interoperability, because laws differ by country while models and data naturally cross borders; and efficiency versus employment, which is becoming most immediate as AI expands from code into finance, law, healthcare and enterprise services.
According to 庄明浩, Claude 5 and GPT-5.6 went live only after government review, a process that could become standard. As tech companies continue to cut jobs, the social backlash “is only just beginning” and will persist for a long time.

15. Tokens Are Not the Ultimate KPI; the Fourth Pillar Is Still the Redefinition of the Human

Token maxing was the theme of Q1 2026. Top software companies among America’s leading enterprises were already spending $7,000 to $7,500 a month on tokens, while the average software engineer earned roughly $15,000 a month. By Q2, the industry had begun questioning whether burning more tokens necessarily creates more value, and whether the logic would reverse once token costs exceeded labor costs.
History shows that intermediate metrics are often discarded. Electricity is ultimately measured in kilowatt-hours, not brightness; the internet moved from PV to sessions and clicks; mobile internet moved from data traffic to DAU and MAU. “No one knows” what the corresponding commercial metric for the intelligence age will be. Outcome-based pricing is only a direction to be tested.
After technology, capital and knowledge, the “human profile” has not appeared automatically. AI is making thought cheaper, while AlphaFold-style knowledge unlocks will continue to spread and megacaps will keep playing the role of modern Medicis. But the 4th pillar can only be built deliberately by specific people.
庄明浩 leaves the ultimate question with humanism: “Are humans still the measure of all things?” By 2050, people may spend years refining a craft, assume responsibility in the physical world, maintain long-term obligations to a small circle of people, and inherit and transmit traditions. This is only a forecast, but it points toward the question of which tasks must remain human.
If someone traveled 500 years into the future, they might not remember a particular model. They might remember that humanity first seriously asked: “Once intelligence is no longer scarce, where exactly do humans fit?”

16. The 4 H2 Themes Remain Unresolved, While Real-World Penetration Suggests the Cycle Is Only Beginning

The 1st theme is ROI: when will the massive capex curve inflect, and what should replace tokens as the measure of outcomes? The 2nd is open source versus closed source: will the US-China gap narrow or widen, and is Kimi K3’s first lead in certain niche benchmarks sustainable?
The 3rd is world models and Physical AI: will autonomous driving, embodied intelligence, games or another use case become the first verifiable entry point? The 4th is governance and legislation. One-size-fits-all controls are only a temporary measure; could more violent conflict arrive at the same time as a bubble burst?
The final chart uses a grid of the entire human population: green represents people using free AI, yellow represents paid users, and red represents paid users already using Agents. All categories grew visibly from March through June, but even added together they still represent a small share of humanity. “It really seems that everything is only just beginning.”

Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271