"AI shovel" dominates the market, who will be the next winner? Goldman Sachs and SemiAnalysis debate the "core difference"

In the past two years, AI trading has almost dominated the global stock market. Nvidia, semiconductor equipment, HBM, advanced packaging, data centers, power equipment, transformers, refrigeration, and gas turbines any asset that can be integrated into the AI infrastructure chain has been repeatedly re-valued by the market. This trading strategy has not failed but instead has risen to the point where investors are faced with a more difficult question: have the winners of the first stage of the AI industry chain been rewarded to the extreme by the market, and can they continue to rise in the next step? Two reports from Goldman Sachs and SemiAnalysis happen to stand at this crossroads. Goldman Sachs James Covellos judgment leans towards caution: the first phase of AI infrastructure has already been fully priced in, and the chain of chips and shovels have taken away too much certain profit, but the ROI on the enterprise side has not been universally achieved, and the cash flow pressure on cloud providers is increasing. According to this logic, the better relative trade next would not be to continue chasing after semiconductors, but to be bullish on the super large-scale cloud providers and low-spec semiconductors. The answer given by SemiAnalysis is almost the opposite: if Agentic AI really turns tokens into productive assets, the gross profit margin of model laboratories begins to improve, frontier models still have pricing power, then the AI infrastructure is not high enough, but has not yet been fully re-priced according to the new round of token value. Companies like Nvidia, TSMC, memory, Neocloud, and model laboratories still have reasons to continue siphoning off incremental value. This is not a debate about whether AI has a future. AI capital expenditures are still rising, and AI infrastructure stocks have not cooled down. The real problem has become: the chip layer has left the first round of profits on the books, and the market is now debating whether these profits have been fully priced in; if Agentic AI continue to increase token value, will the next round of incremental profits remain in the hardware layer, or will it begin to be reallocated to model laboratories, cloud providers, and enterprise software layers. Goldman Sachs is focusing on an industry chain that has not yet closed. The most striking aspect of the Goldman Sachs report is not the questioning of AI user growth, nor the denial of technological progress. Covello first admits two things: the speed at which consumers are adopting AI is faster than they originally expected; even though cloud providers are under stock price pressure, they are not cutting back on AI capital expenditures as much as they had anticipated, but instead are continuing to increase their investments. AI has not cooled down, and capital expenditures have not retreated. But Goldman Sachs is looking further ahead. Consumers are still using AI at a superficial level. User growth can prove product attractiveness, but cannot directly pay the bills for GPUs, data centers, electricity, networks, and model inference. The key to the closure of the AI economy is at the enterprise level: whether enterprises are willing to continue paying, whether they can save costs, increase revenue, and improve output from AI, will determine whether the entire chain can sustain today's capital expenditures in the long term. Goldman Sachs presents a cautious view. The report mentions that there has already been significant investment in generative AI, but many organizations have not yet yielded verifiable returns; at the same time, global IT spending is still on the rise, and AI has not yet reduced enterprise technology budgets on a total scale. For investors, this means a very real problem: enterprises are buying AI, testing AI, and discussing AI, but AI has not yet entered the profit column universally. This contrasts with the profit formation in the AI infrastructure chain. While chip companies are already making money, companies related to storage, power, and data centers have been revalued by the market. However, cloud providers are at the other end bearing the brunt of capital expenditures. Data center construction, GPU purchases, power access, network equipment, and server racks these expenses all initially fall on the books of cloud providers. The Goldman Sachs report states that super large-scale cloud providers have already consumed some of their operating cash flow surplus and have begun supporting data center construction through debt, with data center debt issuance expected to double to $182 billion by 2025. This is the imbalance in Goldman Sachs' eyes. In a normal semiconductor cycle, if chip companies are making big profits, it usually means that customers are expanding. Customers make money, continue to buy chips, and chip companies continue to prosper. But this round of AI is different: the profit in the chip chain is clear, but the return from customers and applications is not yet clear. As a result, Goldman Sachs' assessment is not that AI is useless, but that the current distribution method is difficult to linearly extrapolate in the long term. The semiconductor companies have already taken in the most certain profits from the first phase. The question now is whether downstream customers have enough profit to continue supporting the high capital expenditures and profit concentration in the upstream. Goldman Sachs' trading advice is actually betting on "mean reversion." Goldman Sachs' trading advice may seem counterintuitive: relatively bullish on super large-scale cloud providers and low-spec semiconductors. There are two paths behind this. The first path is that enterprise AI ROI begins to pay off. Companies prove that AI can bring revenue, efficiency, and cost advantages, the market will re-evaluate cloud providers' capital expenditures. The investments that were previously seen as dragging on free cash flow will now be seen as translating into future income and platform control. Cloud provider valuations will recover, and semiconductors will benefit as well, but since semiconductors have already been rewarded considerably by the market, their relative elasticity may not be as high. The second path is that enterprise ROI continues to be challenging. Cloud providers may reduce capital expenditures under pressure from cash flow and investor expectations, and the market will reward better cash flow discipline. The semiconductor chain will then face downward revisions in order expectations. Goldman Sachs believes that both of these paths support the idea that "cloud providers are relatively better than semiconductors". The scenario that would truly make this trade fail is the third path: enterprise ROI is still unclear, cloud providers continue to spend without regard to costs, and semiconductors continue to take up the majority of the industry chain's profits. This is precisely the state of the market over the past two years. For this reason, the most contentious point in the Goldman Sachs report is not AI technology, but market pricing. The benefits of AI infrastructure have already been traded to the fullest extent, while the downsides of cloud providers have been equally priced in. Next, the market will have to see whether these two directions will reverse. SemiAnalysis sees a sudden change in the value of tokens. SemiAnalysis approaches this issue from a completely different perspective. It does not deny that from 2023 to 2025, the value of AI will primarily flow to infrastructure. Companies like Nvidia, power, data centers, storage, have indeed been the big winners of the first stage. Model companies and inference service providers were not comfortable in the early days, and many AI products seemed to be nothing more than a glorified search box, with profit margins far from attractive. But SemiAnalysis believes that by the end of 2025, things have changed. The change comes from Agentic AI. Past tokens resembled "question and answer costs". Users ask a question, and the model provides an answer. It saves time but has limited value. Now, tokens are entering complex workflows: writing code, creating financial models, generating dashboards, analyzing financial reports, organizing data, creating charts. SemiAnalysis uses their own company as an example. Their analysts already use agents to handle research and modeling work every day, tasks that used to require junior analysts many hours or did not fit into the workflow at all. The article reveals that SemiAnalysis's annualized token expenditure on the Anthropic Claude platform once reached $10.95 million, with token spending accounting for about 30% of employee compensation. These figures may not represent all companies, but they represent a change in a certain type of marginal user. For the average consumer, an AI subscription may just be a tool that costs a few dollars a month. But for high-intensity knowledge workers, tokens are becoming productive assets. Tokens that cost a few dollars or tens of dollars are no longer just for a few lines of text, but for models, charts, code, data cleaning, financial report analysis, and even work that would never have been executed before. Users' perception of the cost of AI is also changing: they are no longer just asking "how much for a million tokens" but are asking "how many manual tasks have these tokens replaced, how much output have they added." This is the divergence between SemiAnalysis and Goldman Sachs. Goldman Sachs sees an average enterprise whose ROI is still unclear. SemiAnalysis sees that the strongest users are beginning to consume tokens extensively and are willing to pay for stronger models. The second key judgment from SemiAnalysis is that the economic efficiency of model laboratories is improving. This is contrary to the market's previous concerns. Before, model companies were seen as stuck between chips and cloud providers. Revenue was growing rapidly, but training and inference costs were growing even faster. The more users, the higher the costs. The stronger the model, the higher the capital expenditure. This model seemed to be high growth, low margin, and high burn rate. Agentic AI has changed this dynamic. In terms of pricing, frontier models are able to perform higher value tasks, and users are willing to pay a premium for stronger models. In terms of costs, continuous hardware iteration, inference optimization, caching mechanisms, and software engineering are constantly reducing the unit token cost. In terms of products, model companies can price based on different software optimizations and combinations, allowing them to offer tiered pricing. SemiAnalysis mentions a case where different software optimization combinations in the B300 running DeepSeek can increase the throughput of the same hardware from about 1,000 to 8,000 to approximately 14,000 tokens/second/GPU. With hardware upgrades, the optimized configuration of the GB300 NVL72 is about 17 times higher than the H100 in FP8; if switched to FP4, which Hopper natively does not support, the difference can be up to 32 times, while the total cost per GPU is only about 70% higher. This means that model laboratories can both increase the economic value of token creation and reduce the production costs of tokens. SemiAnalysis states that the Anthropic ARR has increased from $9 billion to over $44 billion, and the gross profit margin of the inference infrastructure has increased from 38% to over 70%. Even if model prices decrease, increased usage of high-end models, improved cache hit rates, hardware efficiency enhancements, etc., could continue to expand profit margins. If this judgment holds true, the second phase of the AI industry chain will not be just about "semiconductors continuing to win" or "cloud providers rebounding." Model laboratories may evolve from a cost center to a new value capture layer. The Real Divergence: Average Enterprises vs. Marginal Users On the surface, Goldman Sachs and SemiAnalysis are arguing about AI ROI, but in reality, they are debating which sample can better represent the future. Goldman Sachs is looking at average enterprises. These enterprises have complex data systems, historical IT burdens, permission management, compliance requirements, and approval processes. Many companies implement chatbots, internal assistants, and pilot projects to showcase their AI strategy to the market and the board of directors. The spending is real, but the business processes may not necessarily change. Without changes in processes, it is difficult to show ROI in financial reports. This is why Goldman Sachs emphasizes data structure and orchestration layers. For example, if a retail company does not integrate inventory, customer profiles, and recommendation systems, an AI customer service bot might recommend a product that is out of stock. If a company does not have a model routing layer, it may end up employing the most expensive frontier model for simple queries, leading to cost escalation. AI implementation barriers are no longer just about having powerful enough models, but about preparing enterprises to integrate models into their business systems properly. SemiAnalysis is looking at marginal users. Tasks like research, coding, modeling, chart creation, financial analysis these are tasks that are ideally suited for agents. They are highly textual, numerical, and structured, the results are easy to evaluate, and users are capable of embedding AI into their workflows. Organizations like this are likely to see ROI sooner and be more willing to increase their token consumption. The capital market will need to determine whether this leading sample will spread. If SemiAnalysis only sees outlier values from a few super users, the framework of Goldman Sachs will have the upper hand. AI capital expenditures will increasingly be constrained by cash flow, the semiconductor chain will need to digest high expectations, and cloud providers may get relatively better returns due to spending discipline and valuation compression. If SemiAnalysis sees leading indicators of a broader trend, the market cannot use today's low ROI of average enterprises to dismiss the AI chain. Once Agentic AI starts entering more white-collar workflows, the demand for tokens, model revenues, cloud revenues, and hardware requirements will increase together. This judgment is more important than simply being bullish or bearish on AI. Market trading is not about static averages but about whether marginal changes can become mainstream trends. Nvidia: Have they made enough profit, or are they still underpriced? The main capital market divergence between Goldman Sachs and SemiAnalysis ultimately falls on Nvidia and the semiconductor chain. Goldman Sachs' view is straightforward: semiconductors have already taken the largest and most certain profits from the first phase. The market has already factored in the logic of selling shovels, and the risk-return profile is starting to deteriorate. As long as there is slack in cloud provider capital expenditures, the semiconductor chain will face pressure on valuation and orders. SemiAnalysis believes that Nvidia and TSMC control the most scarce resources of the AI era but have not yet fully priced their value. The article mentions that memory prices have increased by about six times in the past year, and Neocloud's one-year H100 leasing contract has risen by about 40% from the low point in October 2025. At the same time, Nvidia and TSMC have not re-priced their products as quickly as downstream token values. SemiAnalysis refers to Nvidia as the "central bank" of the AI ecosystem. This metaphor is quite apt. Nvidia controls the flow of computing power. It has the ability to raise prices, but cannot drain the entire system. Raising prices too aggressively could stimulate customers to quickly switch to self-developed ASICs, TPUs, or Trainium, and could also bring regulatory pressure. TSMC is similar. Although advanced nodes are extremely scarce, the company has prioritized customer relationships and ecosystem stability, and will not aim to immediately profit from all its scarcities during an upturn. Restraint does not mean there is no room for growth. The Rubin VR NVL72 is a significant basis for SemiAnalysis' judgment that Nvidia still has pricing power. According to their model, Neocloud would need the VR NVL72 project to achieve a 15.6% IRR similar to the GB300 project, with rental fees of about $4.92 per hour per GPU; if calculated based on the per-PFLOP rental prices of GB300, the theoretical cap for the VR NVL72 would be about $12.25 per hour per GPU; even using a more conservative $0.55 per PFLOP, this would correspond to about $9.63 per hour per GPU, nearly twice the cost-based pricing threshold. The implication here is clear: as long as downstream token values continue to rise, Nvidia still has room to increase prices with their new systems, Neocloud may still be profitable, and end-users may still accept their offerings. The divergence between Goldman Sachs and SemiAnalysis becomes sharper as a result. Goldman Sachs believes that semiconductor profits alone are not sustainable because downstream customers have not generated enough profit yet. SemiAnalysis believes that the downstream profit pool is expanding, so the issue is not that the hardware layer is making too much profit, but that it has not yet fully charged based on its value. The variable that will determine the outcome is whether the new profit pool created by AI can sustain and nourish model laboratories, cloud providers, Neocloud, Nvidia, TSMC, storage, and power chains simultaneously. If the cake is not big enough, Goldman Sachs wins. If the cake continues to grow, SemiAnalysis wins. Cloud Providers in a Delicate Position Cloud providers are in the most awkward position in this debate. They are not only the largest spenders on capital expenditures but are also the platforms where AI demand is most likely to be monetized. They are squeezed by Nvidia, storage, power chains but also have enterprise customers, cloud services, model APIs, proprietary chips, and software ecosystems. They are not just victims, nor automatic winners. They must demonstrate through financial reports that AI capital expenditures can be translated into revenue, profits, and customer stickiness. The market will be looking closely at whether cloud business growth will accelerate again, if AI revenues will become clearer, if inference utilization rates can improve, if proprietary chip investments can reduce their reliance on Nvidia, if enterprise customers will transition from pilots to long-term deployments, whether free cash flow has stabilized, and all these indicators will be more important than ever. These indicators improving will strengthen Goldman Sachs' relative bullish view on cloud providers. If these indicators do not improve, cloud providers will continue to be seen as a layer stuck between Nvidia and enterprise clients with capital expenditure pressures. Software Layers Determine Whether ROI Can Transition from Sample to Average Goldman Sachs' emphasis on "data structure" and "orchestration layers" is perhaps the closest to enterprise reality in their report. Enterprise AI will not remain at the level of employees opening a chat box to ask questions forever. AI that truly has financial impact needs to enter customer service, sales, finance, procurement, R&D, risk management, supply chain, IT operations. Each process has data, permission, compliance, approval, historical system integration, and responsibility boundaries. Even with powerful models, these challenges must be addressed. This is where enterprise software layers become important again. Low-risk, high-frequency tasks can be handled by lightweight models or open-source models; high-risk, high-value tasks require cutting-edge models. In the middle is a layer of systems that determine task type, call data, control permissions, choose models, manage costs, and record results. The advantage of traditional SaaS companies is in industry experience, customer relationships, data access, and workflow expertise. The disadvantage lies in technical debt and iteration speed. The advantage of AI-native companies is in product speed, model invocation capabilities, and cost structure. The disadvantage is the lack of enterprise entry and industry context. The advantage of cutting-edge model companies is in having the most intelligent models. The disadvantage is the lack of control over enterprise processes. The software layer will not simply be absorbed by AI. Software companies without control over data and processes may be pushed out by models. Software companies that have control over data structure, workflow, and model routing have the opportunity to expand AI into a larger market, moving from selling seats to selling productivity. The ability for enterprise ROI to transition from the strong user sample seen by SemiAnalysis to average enterprises will largely depend on this layer. The Next Steps for the Capital Market For the past two years, the market has been asking: who is closest to computing power? This question is now too simplistic. In the next phase, the market will seek more detailed variables. First, will token values continue to rise? If Agentic AI expands from code, research, and analysis to