AI token demand
Summary
金尚浩的核心判断是，市场把 AI token demand 错算成若干可枚举场景的总和，而企业真正上线前必须持续 post-train，因此算力需求被系统性低估。 Shorts cap demand with “$1T of IT spending × 30% replaced by coding agents = $300B,” but his hands-on conclusion is: “Attaching a knowledge base and directly using a model are not the same thing.” Companies need to label private data, evaluate it, train it into models, and retrain repeatedly as new business data arrives.
决定 AI 产出的不只是模型，而是程序、agent、agent 的任务拆分、模型与算力共同作用。 When rebuilding roughly 450K–500K lines of trading-grade code, he previously burned about $2,000–$3,000 of tokens a day and migrated Claude Code’s task-allocation approach to his own platform; mixing Opus, DeepSeek and Soul often worked better than betting on a single smartest model. “A fairly large task might call 70 or 100 agents,” so GPU, CPU, bandwidth and interconnect demand all scale with parallelism.
通用模型在企业私域里仍是“一个大鹦鹉”，真正值钱的是各公司独有且不断更新的数据闭环。 The common thread across online fiction, education, customer service, sales, pharmaceuticals and cross-border e-commerce is that success and failure records are not public or shared across peers; the final serving model can be small, but the teacher model, data labeling and continuous post-training still consume frontier models and substantial compute. Beyond the roughly 10% of data available publicly, most of the operating data companies have accumulated has never been properly used.
Muse 一类 To C agent 的胜负手可能不是准确率，而是响应率、互动感与流量入口。 Using 豆包 as an example, 金尚浩 argues that even with only 60%–70% answer accuracy, a product that always responds and interacts smoothly may still beat users searching on their own; “the model can be weak, as long as the response rate is high.” He is not convinced Meta can monetize quickly on its current path, but believes these products are more likely to erode Google Search first than to fully replace the pleasure of travel and shopping.
AI 造成的大规模失业尚未出现，不是替代逻辑失效，而是企业还没有把模型训练成可用的员工。 He estimates the current payroll impact in AI-sensitive industries at roughly 10K–15K jobs, partly offset by data-center construction; hiring has weakened in tech, Hong Kong investment banking and asset management, but general-purpose agents still cannot reliably take over business processes. “Current layoffs are nowhere near enough”; once companies finish organizing data, post-training models and matching them to workflows, layoffs could arrive in a concentrated wave.
Anthropic 与 OpenAI 的 ARR 没有在开源模型和云厂入场后明显掉下去，本身就是需求强度的证据。 The two companies consumed roughly 27% of compute over the past 2 years and once captured most of the monetization; now Google, Amazon and Microsoft can also monetize open models, yet the original restaurant is “still full.” Among the cloud providers, he favors Microsoft for enterprise certainty and Amazon for near-term reported performance, while remaining skeptical of Meta’s enterprise prospects: “Having compute does not guarantee you can build an enterprise business.”
半导体的核心预期差不是 backlog 有没有，而是市场给 backlog 打了多大的需求折扣。 金尚浩 sees post-training and multi-agent workloads as positive for GPUs, while CPUs and optical interconnects are equally indispensable; he is less bullish on DRAM in the near term because the whole industry is optimizing memory at elevated prices. At the portfolio level, he still likes the risk/reward in Microsoft and Nvidia, while AVGO could continue to face multiple compression despite sound results because of margins near 70%, powerful cloud customers and lower-margin adjacent markets.
短期风险资产环境并不舒服，但与高仓位、高估值时期的脆弱结构不同。 Rates, high-yield spreads, the Goldman Financial Conditions Index and market volatility gauges have all deteriorated, with another roughly 32B of pension-rebalance selling; but positioning and forward P/E are lower, while semiconductor demand contains a meaningful expectation gap, leading him to conclude that “there will be a negative impact, but it will not be catastrophic.” The 10-year Treasury is already materially oversold in the short run; over the longer term it remains pulled by higher for longer and GDP growth, while the real force suppressing buyers is diesel, crack spreads and disruption across the entire energy supply chain.
Deep dive
1. Actual token burn rewrote his view of demand
金尚浩 does not extrapolate demand from demo products. He starts with the burn rate from rebuilding a high-frequency trading system: roughly 450K–500K lines of trading-grade code in 3 months, with “zero tolerance for error”; the project previously consumed about $2,000–$3,000 of tokens a day, plus 2 Max plans.
He distilled the original program inside Claude Code, focusing on “how to split the tasks,” then had Sonnet or other models execute the complex work. His own code platform is not another foundation model; it is a migration of effective agent orchestration to other platforms, including Codex.
The experience also changed his earlier view. In March and April, he saw Anthropic and OpenAI models “get stupid” when compute was constrained and briefly thought coding-model ARR would fall. He now believes the work required before a company can truly use a model is far greater than expected, and represents recurring spend.
His framework has 4 components: his own program, the agents, how agents split tasks, and the model itself; compute is tied mainly to the model layer. The market tends to focus on models and infrastructure, while potentially missing the agents, task allocation and post-training that determine whether a system is actually usable.
2. Agent orchestration often matters more than the intelligence of one model
In his coding workflow, Opus remains the primary model, DeepSeek handles some execution, and Soul sometimes handles planning; Astro, in his words, is better at “3D.” The key is not picking one winner, but having agents call different models at the right steps.
Claude Code looks powerful partly because the platform has already handled convergence, task division and dispatch for the user. 金尚浩’s view is that many people mistake the agent layer’s capabilities for the model’s own capabilities: a strong agent does not mean the underlying model can independently complete the entire engineering project.
His idea is to distill Claude Code’s task-splitting pattern onto other platforms: if a better agent can be organized and deployed on platforms such as Codex, the user experience could be very strong. The cost of KV cache remains an issue, however. The more models and context involved, the more important data scheduling becomes.
3. A general-purpose model without post-training is still a “giant parrot”
In tens of thousands of lines of zero-tolerance trading code, models make “simply unbelievable” low-level mistakes. No single model can take a large task from start to finish; the target must be continuously narrowed, outputs checked and the work broken into smaller modules.
Coding looks mature because model vendors such as Anthropic and OpenAI have already done extensive scenario-specific post-training for users and optimized the workflow into the models. Common tasks such as creating PPTs work similarly; they do not prove that every enterprise use case is ready out of the box.
金尚浩’s words are that once removed from these optimizations, a model is “often just a giant parrot.” It can repeat, imitate and generate within a limited structure, but it does not automatically understand a company’s success criteria or error tolerance.
The enterprise problem is therefore not “buying tokens completes deployment.” It is defining the output, labeling historical data, building evaluations and training the patterns back into the model. Connecting a knowledge base supplies material; it does not replace behavior shaping.
4. The online-fiction experiment made the need for post-training obvious
金尚浩 tried having Soul, Astro and Fable[?] write online fiction directly, and also fed them passages, structures and context from popular works. The results were either “very literary” or tightly converged along a preset architecture, unable to diverge where the story required it.
His workable approach is to train a relatively small Qwen 27G model. The final execution model does not need to be the strongest; it needs to read the large volume of articles on platforms such as 起点 and convert paragraph-level performance, based on reading results for each page, into training labels.
Labeling can be done by humans or AI; he uses the latter. A teacher model first reads dozens or roughly 300 articles, then evaluates several thousand texts generated divergently by the execution model and identifies which outputs have vectors closest to the teacher model.
Even a final 27G model does not mean a light training burden. He estimates that 2 512 GB Macs might need 1–2 days; with an MoE model, “you definitely have to use Nvidia cards.” A small model does not imply a small compute chain.
5. “Self-iteration” is really an ever-faster data loop
金尚浩’s explanation of “ASI” is not fantastical: when a new article or chapter is published, the system reads it faster, labels it faster and then automatically post-trains the AI. The iteration speed comes from how quickly feedback enters the model, not from the model spontaneously creating a better version of itself.
The loop requires new data, a teacher model, generated samples and evaluation criteria to exist at the same time. The teacher reads roughly 300 articles, the execution model writes several thousand pieces, and similarity scores and labels filter the results into the next round of training data.
The final user-facing model may be cheap and lightweight, but it still needs a stronger model resembling Fable[?] for interpretation and evaluation. Demand for advanced models therefore does not disappear when companies adopt small open models; it moves from front-end inference to back-end teaching, labeling and validation.
6. The real enterprise moat is data that cannot be obtained publicly
金尚浩 estimates that publicly available online data accounts for only about 10%; more importantly, “99%” of companies’ historical operating data is theoretically unusable in its current form. AI’s value is to lower the bandwidth cost of communication between people and databases, turning dormant records into rules a model can execute.
Sales is the clearest example. Even if 2 companies sell the same product, their wins, losses, scripts and customer feedback differ. Every new close or failure changes the data distribution, so the model needs continued training rather than a one-time deployment.
Education and tutoring cannot simply be absorbed by a general-purpose model either. Institutions such as New Oriental hold students’ score-improvement curves, exams, question-setting methods and teaching feedback. The future may have fewer hands-on teachers but more teams organizing exams, analyzing curves and post-training models.
The same applies to customer service, pharmaceuticals, cross-border e-commerce and ad placement. Transaction flows, returns, financial data and customer behavior are interconnected and often require on-premises deployment; “attaching a knowledge base” does not automatically turn the material into a qualified employee.
7. Capping token demand with coding misses continuous training
The common market calculation is that coding represents about $1T of IT spending, 30% of the work is replaced by coding agents, and Anthropic and OpenAI therefore have a roughly $300B serviceable market; AI for Science is then added separately based on R&D spending.
金尚浩 believes this underestimates demand because it counts only vertical use cases that have already been productized. Coding is “the crown jewel,” valuable enough that model vendors will optimize it themselves; no one will train a company’s proprietary data and business standards for it.
Enterprise AI is therefore a continuous process: employees require ongoing training, and models require ongoing post-training. Company data updates, competitor differences and changing business standards all generate new labeling and training demand rather than a one-time purchase.
The cloud feedback he has heard is that corporate AI is not the same thing as corporate coding; post-training orders are already stacked in backlog. Open models let enterprises buy cards and train themselves, while on-premises deployment may continue to benefit Dell and the smaller suppliers around Dell’s enterprise hardware.
8. Multi-agent workloads bring CPUs and data movement back to the foreground
金尚浩’s rough intuition for the scaling law is about the 1/10th power of X, but he explicitly says he is “not very sure” and needs Silicon Valley practitioners to validate it. Even if models keep getting smarter, complex engineering will not be completed serially by one super-intelligent person; it will require layered decomposition.
In the past, one agent might serially handle several tasks. With Claude Code, one large task may call roughly 70–100 agents, many of them running in parallel. The more complex the task, the more agents it requires.
His analogy is a cloud bus: one bus can carry many requests, but a user starting multiple agents at once is like “buying 20 tickets right at the start.” Scheduling requests, sending data and moving context all require CPU, not just GPU.
Optical interconnect capacity may expand from 2024, followed by storage, while CPUs scale with actual agent demand. Every agent carries CPU, compute and bandwidth requirements, so parallelization amplifies the entire hardware stack in sync.
9. Compute is a loop, not an isolated GPU
金尚浩 repeatedly emphasizes that interconnects, CPUs, GPUs and storage form one demand loop: “You cannot say there is only a GPU and no CPU.” Betting on one link while assuming the others will not grow is, in his view, the most one-sided position.
Post-training and multi-agent workloads are positive for GPUs. Even if GPUs are no longer as extremely scarce as before, the market may still be underpricing their next leg. Compute demand will be released further once enterprises complete post-training and move into production use.
He is relatively less optimistic on DRAM because prices are already high and the entire industry has started optimizing storage; as tasks are distributed, storage use should decline relatively. He acknowledges that long-term storage demand will still rise, but the near-term risk/reward is less obvious than when the market was most supply-constrained.
The semiconductor discount comes from the same question: cloud providers have large backlogs, but investors are discounting future orders against coding revenue visible today. 金尚浩 believes the discount does not fully reflect enterprise post-training, private-domain data and continuous training.
10. Muse may post strong operating data without necessarily making money the current way
金尚浩 downloaded and tested Muse, and believes its traffic and operating metrics could remain strong, but is uncertain that profits will keep pace. Meta has a large pool of under-monetized compute, while Microsoft and others have already directed more compute toward enterprise business; that helps explain why the product appeared first at Meta.
He classifies Muse as “non-serious work.” Consumer users do not necessarily require complex code to be completed; they care more about the product always responding and the interaction feeling natural. Model quality is not the first metric; response rate is the operating core.
A typical 豆包 interaction is to give the wrong floor in a mall and then say: “Sorry, you caught me—I made a mistake.” The response does not improve factual accuracy, but it provides emotional value and avoids ending the conversation with “I don’t know.”
金尚浩 estimates that products of this type could work even with 60%–70% accuracy, because ordinary users also make mistakes when searching for themselves. “The model can be weak, as long as the response rate is high”—enough to pull large numbers of users into a chat and search entry point.
11. Consumer agents are more likely to take search first than run people’s lives
金尚浩 does not buy the simple narrative that an “AI butler” will fully replace travel and shopping. Planning a trip and browsing products are themselves part of the pleasure of consumption, and users may not want to hand over the entire process to an agent.
The more realistic opening is comparing products, performing the final price check and shortening the payment path. Muse can connect email, apps and entry points, but high-frequency daily use may still center on chat and Q&A.
He therefore believes Muse is more likely to challenge Google Search from the side. Meta historically did social, Google did search and Microsoft did enterprise services, with relatively clear boundaries; once agents connect the entry points, competition becomes “warring-states chaos.”
Apple could also enter through its devices and app distribution. Token consumption per consumer user may not be large, but every company will invest and penetration will rise; “traffic is money” may eventually prove true, but current monetization should not be overestimated.
12. Employment impact is temporarily hidden by the lag in enterprise preparation
金尚浩 estimates that AI-sensitive industries currently account for an impact of roughly 10K–15K jobs on payroll, partly offset by employment generated by data-center construction. Aggregate data therefore has not yet shown severe unemployment.
Hiring contraction has appeared first. Technology workers clearly feel it, while Hong Kong investment banks, asset managers and research teams are also “not hiring much.” But hiring less is different from mass replacement, which requires AI to reliably complete actual business work.
Companies cannot yet replace people simply by “using agents casually,” because general-purpose models remain parrots outside coding. Replacement capacity will emerge only after private data is organized, post-training is completed and the model is connected to the business workflow.
His timing call is explicit: current layoffs are nowhere near enough. Once many companies finish the preparation, layoffs could arrive in a concentrated wave. That is both a social risk and a lagging indicator of whether post-training has actually been deployed.
13. Anthropic’s ARR resilience matters more than one-month growth
金尚浩 says Anthropic and OpenAI together used roughly 27% of compute over the past 2 years because they are model companies and monetization and profits were once concentrated there. Most compute was not in their hands; they had a lot of compute, but not all of it.
Open models let Google, Amazon and Microsoft begin monetizing as well, but Anthropic’s ARR has not fallen materially. His restaurant analogy: “3 or 5 restaurants opened next door,” yet the original remains full. That is positive for industry-wide demand.
Open models serve cloud buyers’ interests: with substitutes available, the four hyperscalers do not have to remain fully dependent on a single closed model. 金尚浩 also does not believe political forces can easily suppress the trend, given the scale of the cloud providers’ interests.
On an Anthropic IPO, he separates company quality from the transaction path. A large financing could pull liquidity out of the market, while the CEO’s repeated discussion of a “10% probability of extinction” tied to the product creates valuation risk. But because his own P/E entry price is low, he is not especially worried about losing principal.
14. Microsoft wins on enterprise stickiness; Amazon wins on near-term reporting purity
Post-training expansion should benefit both Microsoft and Amazon. If forced to choose, 金尚浩 sees Microsoft as more certain and Amazon as purer, with better-looking near-term reported numbers.
Microsoft’s advantage is not a single reproducible program, but a complete SaaS service system and customer relationships. “SaaS is not something you can hand-build” casually; enterprise deployment, support, feedback and renewals together create stickiness.
Enterprises are more likely to buy cloud-provider tokens with measurable ROI. After a customer-service deployment, whether return rates fall and costs decline can be tracked through existing SaaS and commercial systems; buying model tokens directly requires the enterprise to verify the investment and payoff separately.
He does not have the same confidence in Meta’s enterprise business: “Having compute does not guarantee you can build an enterprise business.” Otherwise Oracle, neoclouds and others could also win naturally. Meta lacks the enterprise-services DNA, not another larger model.
15. Near-term market setup is weak, but not as dangerous as when positioning was high
金尚浩 sees rates rising, single-stock vol failing to rise in tandem, the Goldman Financial Conditions Index deteriorating, and high-yield spreads and equity vol moving higher. That combination is unfriendly to risk assets for the week.
There is also roughly 32B of pension-rebalance selling. By his account, about 10B is weekly rebalancing to be completed before Friday, while about 22B is monthly rebalancing to be completed by Tuesday or Wednesday; the original remarks did not specify the currency.
He believes the fundamental setup differs from earlier high-risk periods: overall positioning is not elevated, many investors are avoiding an Anthropic IPO, forward P/E is not high, and the low valuation is mainly because semiconductor orders are being discounted.
The conclusion is not an immediate precise buy point, but rather: “There will be volatility, but it is okay.” The amount the selling can push prices down is unknowable; the tradeable medium-term variable remains the gap between market expectations and actual compute demand.
16. Momentum is a price-volume strategy, not an explanation for post-training
Asked about ETFs such as SPMO, MTUM and VFMO, 金尚浩 first points out that SPY itself is a form of long-term momentum: it retains roughly 500 better companies and continuously removes those deteriorating; the Dow does not fully follow this mechanism.
Whether momentum deserves a place depends on stock selection, single-name caps and the rebalancing method; the final test is Sharpe, not the label. Momentum and SPMO have Sharpe ratios around 1; his own strategy of 20–30 stocks is around 1.2, slightly above the roughly 1.1 reported for SPMO.
Reducing the number of stocks can raise vol to roughly 30%–40%. If Sharpe remains near 1, expected returns could reach roughly 40%–50%, but the cost is a larger drawdown. The high returns of some products are therefore mainly the result of a larger risk budget, not free alpha.
The earlier momentum crash was severe because the strategy held few stocks and they were all crowded into semiconductors. Common risk controls include stop-losses and vol control; the latter mainly reduces position size and drawdown, and may not change Sharpe much.
17. Tesla runs on trust; Microsoft and Nvidia run on risk/reward
On Tesla, 金尚浩 does not offer a valuation answer but a judgment about the person: “It depends on whether you believe Elon Musk.” Everything he says sounds promising, but the valuation is too high for conventional frameworks to constrain, and holders must also calculate the alternative cost of missing other assets.
His answer on SpaceX is shorter: the company and its cloud-deployment capabilities are “very strong,” but a strong company does not mean it should be bought at any price; “it can be bought after it falls.” The distinction from Tesla still comes down to entry price and trust in Musk’s execution.
He views Microsoft and Nvidia as attractive holdings within a roughly 30-vol framework. Investors willing to tolerate 100 vol will naturally look for higher-beta names such as AEHR and FORM; conservative investors who sell their large-cap core holdings may not truly be able to bear the risk of the replacements.
He believes AI profits will ultimately shift from “selling shovels” toward cloud providers and downstream companies using AI. Cloud free cash flow is currently compressed by capex, like a “nanny taking the hits”; once cash flow is released, market-cap levels should not be constrained by today’s absolute scale.
18. AVGO’s results are fine; the issue is that customers will eventually pressure margins
AVGO could earn gross margins near 70% because the chips involved were not previously a core Google production input, so Google cared less about how much the supplier made. As TPU becomes more closely tied to revenue, every dollar saved flows into Google’s earnings.
If the relevant AI chips or adjacent markets reach roughly 2T with gross margins of only about 40%, the market will infer that AVGO cannot maintain 70% forever. This does not mean margins collapse immediately; it means the multiple is discounted in advance to a lower long-term profit margin.
金尚浩 still believes AVGO’s results, interconnect capabilities and recent supply are fine. Precisely because the fundamentals have not broken, it could rebound when short interest in the AI momentum trade becomes crowded.
The longer-term risk is that the customers themselves are powerful. The four hyperscalers may not design chips as well as AVGO today, but no one can say they never will; that possibility alone is enough for the market to keep applying a valuation discount to the supplier.
19. Component opportunities depend on demand timing and supply bottlenecks
金尚浩 describes Wolfspeed as “a piece of garbage from the bottom that has become not quite as bad,” while warning that capacity at China Tianyu and others is growing quickly. Demand for 800V and SiC could be large, but the timing has shifted from the original expectations to Q4 this year or early Q1 next year, with the exact quarter still uncertain in his words.
Nokia has had roughly 1.5 years of backlog in optical interconnects and scale-across infrastructure; once demand arrives, that could extend toward 2 years. Ciena is the main comparison. Even if large data centers face obstacles and are forced to shrink in scale, inter-region and inter-data-center connectivity remains part of the demand chain.
In his view, the fundamental correlation between Coherent and Lumentum is “equal to 1”; both are severely supply-constrained. Differences in valuation at a given moment do not need to be explained as structural divergence. If Google launches the “Pro 4” model referenced in the discussion and continues increasing investment, the relevant optical supply chain should benefit.
For Micron, he believes gross margin may be near a relative high, but that does not mean volumes have peaked; storage has been thoroughly modeled by the market and is priced on supply and demand. If forced to choose within storage, he might prefer NetApp, whose cycle could be somewhat longer, but he does not forecast sentiment moves around earnings.
20. Optical-electronics mass production will move the bottleneck from assembly to testing
The need for active alignment will not disappear, and components such as 天孚通信 are unlikely to be simply displaced. But mass-production routes such as the TSMC approach and Intel OCI require PIC and EIC wafers to be joined through passive alignment and bonding.
Pluggable products could previously rely to a greater extent on manual “hand assembly.” Wafer-level mass production cannot tolerate misalignment or latent defects. As assembly efficiency improves, the quality bottleneck naturally moves to testing.
FormFactor covers wafer-level testing, using many windows across a wafer rather than testing only from the edge; AEHR covers early failures and burn-in. 金尚浩 stresses that the 2 are not alternatives but are required at different stages.
The investment case for testing is therefore not that one company captures all the value. The more standardized the production route and the larger the scale, the more important each layer of verification becomes. “Testing at different stages—those tests themselves will be highly important.”
21. The 10-year Treasury has a long-term upward pull but is already sharply oversold in the short run
金尚浩 does not believe the 10-year Treasury must mechanically equal GDP growth, but thinks the market may be using roughly 6.5% GDP growth as a reference ceiling for yields. As long as the Fed does not cap rates, the 10-year still has a long-term upward pull.
The 30-year is more of a pure supply-and-demand market. Insurance companies and pension funds are temporarily unwilling to buy, and auction demand is weak. Even with 5.1%–5.2% offering clear value to long-duration capital, buyers can still wait for cheaper levels.
The current pressure is not just crude oil, but the energy supply-chain disruption reflected in diesel prices discussed at around 100 and a very high crack spread. Europe lacks refining capacity and faces tighter inflation constraints; Japan is also repricing. The market is therefore moving toward a higher-for-longer stance among global central banks.
War and energy shocks are making both central banks and long-duration buyers hesitate, with Europe and China bearing the larger real cost. The 10-year yield has risen too quickly in the short run and bonds are clearly oversold, but its volatility will force hedge funds to adjust positions and transmit pressure into equities. Over a 2–3 year horizon, he still believes roughly 1.5% efficiency growth will support a higher rate center.