Goldman Sachs and SemiAnalysis Disagree on AI Infrastructure Valuation and Future Leaders

Over the past two years, AI trading has nearly dominated the global stock market.

NVIDIA, semiconductor equipment, HBM, advanced packaging, data centers, power equipment, transformers, cooling systems, and gas turbines—any assets that can be integrated into the AI infrastructure value chain have been repeatedly revalued by the market. This trade has not lost momentum; instead, it has risen to the point where investors must confront an even more difficult question: Have the winners of the first phase of the AI value chain already been fully rewarded by the market? Can they continue to rise further?

Two reports from Goldman Sachs and SemiAnalysis sit right at this crossroads.

James Covello of Goldman Sachs holds a more cautious view: the first phase of AI infrastructure has already been fully priced in; chips and the “selling shovels” chain have captured too much guaranteed profit, while enterprise ROI has yet to be widely proven, and cloud providers are facing rising cash flow pressures. Following this logic, the better relative trade going forward is not to continue chasing semiconductors, but rather to go long on hyperscale cloud providers and under-equipped semiconductors.

SemiAnalysis’s answer is almost the opposite: if agentic AI truly turns tokens into means of production, and model labs begin to see improved gross margins, while frontier models retain pricing power, then AI infrastructure has not yet been fully re-priced according to the new wave of token value—it has not been overpriced. NVIDIA, TSMC, memory providers, Neocloud, and model labs all still have valid reasons to capture additional 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 question has become: Has the chip layer already captured the first round of profits on its balance sheet, and is the market now debating whether these profits have been fully priced in? If agentic AI continues to amplify token value, will the next wave of incremental profits remain with the hardware layer, or will they begin to redistribute to model labs, cloud providers, and enterprise software layers?

The most striking part of Goldman Sachs' report is not questioning AI user growth, nor denying technological advancement.

Covello first acknowledged two things: consumers are adopting AI faster than originally expected; and cloud providers, despite pressure on their stock prices, have not cut AI capital expenditures as anticipated, but have instead continued to increase investment. AI is not cooling down, and capital spending is not receding.

But Goldman Sachs is looking further ahead.

Consumers are using AI, but many remain on free tiers. User growth can demonstrate product appeal, but it cannot directly cover the costs of GPUs, data centers, electricity, networking, and model inference. The enterprise sector is key to whether the AI economy can close the loop: whether businesses are willing to pay consistently, and whether they can reduce costs, increase revenue, and improve output through AI, determines if the entire chain can sustain today’s capital expenditures over the long term.

Goldman Sachs' answer was cautious.

The report notes that companies have already made significant investments in generative AI, yet many organizations have yet to achieve verifiable returns; at the same time, global IT spending continues to rise, and AI has not reduced overall corporate technology budgets. For investors, this presents a practical question: companies are buying, testing, and talking about AI, but AI has not yet broadly appeared on income statements.

This stands in stark contrast to the profitability of the AI infrastructure chain.

Chip companies are already profitable, while companies related to storage, power, and data centers are being repeatedly revalued by the market. Cloud providers, on the other hand, bear the brunt of capital expenditures—costs for data center construction, GPU procurement, power connectivity, networking equipment, and server racks all initially appear on cloud providers’ balance sheets. According to Goldman Sachs, hyperscale cloud providers have already consumed part of their operating cash flow surplus and are now financing data center development through debt, with data center debt issuance set to double to $182 billion by 2025.

This is the imbalance as seen by Goldman Sachs.

In a normal semiconductor cycle, chip companies making big profits typically indicates that their customers are also expanding. When customers earn profits, they continue buying chips, and chip companies continue to thrive. This AI cycle is more unusual: the profit margins along the chip supply chain are clear, but the returns at the customer and application layers are not yet equally clear.

Therefore, Goldman Sachs' assessment is not that "AI is useless," but rather that "the current revenue-sharing model is difficult to extrapolate linearly over the long term."

Semiconductor companies have already secured the most certain profits from the first phase. The question is whether downstream customers have sufficient margins to continue supporting the high capital expenditures and profit concentration at the upstream level.

Goldman Sachs' trading recommendation appears counterintuitive: overweight large-scale cloud providers, underweight semiconductors.

There are two paths behind this.

The first path: enterprise AI ROI begins to materialize. As companies demonstrate that AI generates revenue, efficiency, and cost advantages, the market will reassess cloud providers' capital expenditures. Investments previously seen as dragging down free cash flow will be reinterpreted as drivers of future revenue and platform control. This will lead to a valuation rebound for cloud providers, with semiconductors also benefiting; however, since semiconductors have already been heavily rewarded by the market, their relative upside may not be greater.

The second path: corporate ROI remains challenging. Cloud providers, under pressure on cash flow and from investors, are cutting capital expenditures, and the market will reward stronger cash flow discipline. The semiconductor supply chain faces downward revisions in order expectations.

Goldman Sachs believes that both paths support the view that cloud providers are relatively better than semiconductors. The scenario that would cause this deal to fail is the third path: enterprise ROI remains unclear, yet cloud providers continue to invest heavily without regard for cost, while semiconductors continue to capture the vast majority of profits across the supply chain.

This is precisely the state the market has been most familiar with over the past two years.

Precisely for this reason, Goldman Sachs' report is not targeting AI technology, but market pricing. The benefits of AI infrastructure have already been fully priced in, and the drawbacks of cloud providers have also been fully priced in. Next, the market will watch whether these two trends reverse.

SemiAnalysis approaches from a completely different angle.

It does not deny that from 2023 to 2025, AI value primarily flowed to infrastructure. NVIDIA, electricity, data centers, and storage were indeed the big winners in the first phase. Model companies and inference service providers did not fare well initially, as many AI products seemed merely like improved search boxes, with gross margins far from impressive.

But SemiAnalysis believes that things changed after the end of 2025.

The change comes from Agentic AI.

Past tokens were more like a "question-answer cost": users asked a question, and the model responded once. It saved time, but its value was limited. Today’s tokens are entering complex workflows: writing code, building financial models, generating dashboards, analyzing financial statements, organizing data, and creating charts.

SemiAnalysis uses its own company as an example. Its analysts have been using agents daily to handle research and modeling tasks that previously required junior analysts to spend many hours—or were simply not feasible within their workflow. The article reveals that SemiAnalysis’s annualized token spending on Anthropic Claude once reached $10.95 million, with token costs accounting for approximately 30% of employee salaries.

This set of numbers may not represent all businesses, but it reflects a shift among marginal users.

For average consumers, AI subscriptions may simply be tools costing a few dozen dollars per month. For intensive knowledge workers, tokens are beginning to become means of production.

Tokens costing just a few or tens of dollars bring more than just text—they bring models, charts, code, data cleaning, financial statement analysis, and even tasks that were previously never executed. Users’ perception of AI costs will shift accordingly: instead of asking only “How much per million tokens?”, they will ask, “How much human labor did these tokens replace, and how much additional output did they generate?”

This is where SemiAnalysis and Goldman Sachs diverged.

Goldman Sachs sees that the average company's ROI is still unclear. SemiAnalysis sees that the strongest users have already begun consuming tokens in large quantities and are willing to pay for more powerful models.

SemiAnalysis's second key insight is that the unit economics of model labs are improving.

This contrasts with past market concerns.

Previously, model companies were seen as caught between chip and cloud providers. While revenue grew rapidly, training and inference costs grew even faster. The more users, the higher the costs; the more powerful the model, the greater the capital expenditure. This model appeared to be high-growth, low-margin, and cash-intensive.

Agentic AI has changed this table.

• On the price side, frontier models can perform higher-value tasks, and users are willing to pay a premium for more powerful models.
• On the cost side, hardware upgrades, inference optimization, caching mechanisms, and software engineering continuously reduce the cost per token.
• On the product side, model companies can implement tiered pricing through higher-end SKUs, faster response times, and stronger reasoning capabilities.

SemiAnalysis noted that, in the case of running DeepSeek on the B300, different software optimization combinations can increase the same hardware’s throughput from approximately 1,000 to 8,000 tokens/second/GPU up to around 14,000 tokens/second/GPU. When combined with hardware upgrades, the optimally configured GB300 NVL72 achieves approximately 17 times higher throughput than the H100 under FP8; if switched to FP4—which Hopper does not natively support—the gap widens to 32 times, while the total cost per GPU increases by only about 70%.

This means that the model lab can simultaneously increase the economic value of tokens while reducing their production costs.

SemiAnalysis reports that Anthropic's ARR increased from $9 billion to over $44 billion, and its inference infrastructure gross margin rose from 38% to over 70%. Even as model pricing declines, higher adoption of premium models, improved cache hit rates, and enhanced hardware efficiency could further expand gross margins.

If this assessment holds true, the second stage of the AI industry chain will no longer be just about "chips continuing to win" or "cloud providers rebounding."

The model lab will transform from a cost center into a new value capture layer.

Goldman Sachs and SemiAnalysis are ostensibly debating AI ROI, but in reality, they are arguing over which sample better represents the future.

Goldman Sachs is looking at average companies.

These companies have complex data systems, legacy IT infrastructure, access controls, compliance requirements, and approval processes. Many organizations, in order to demonstrate an AI strategy to the market and board, start with chatbots, internal assistants, or pilot projects. Money is indeed spent, but business processes often remain unchanged. Without process transformation, ROI is unlikely to show up on financial statements.

This is why Goldman Sachs emphasizes data structures and the orchestration layer.

A retail business that hasn’t integrated its inventory, customer profiles, and recommendation system may have an AI customer service agent suggest an out-of-stock item. A company without a model routing layer might route simple queries to the most expensive state-of-the-art models, causing costs to spiral out of control. The bottleneck in AI adoption is no longer just about model strength—it’s that businesses haven’t yet prepared to integrate models into their operational systems.

SemiAnalysis focuses on marginal users.

Tasks such as research, coding, modeling, charting, and financial statement analysis are naturally suited for agents. They are highly textual, digital, and structured, with easily measurable outcomes, and users are capable of integrating AI into their workflows. Such organizations will see ROI earlier than typical enterprises and are more willing to increase token consumption.

The capital market must determine whether this leading sample will spread.

If SemiAnalysis is observing only outliers from a small group of super users, the Goldman Sachs framework will prevail. AI capital expenditures will become increasingly cash flow constrained, the semiconductor supply chain needs to digest high expectations, and cloud providers may gain relative returns due to spending discipline and valuation compression.

If SemiAnalysis is seeing leading indicators before the spread, the market cannot dismiss the AI chain based on today’s low ROI of average companies. Once agentic AI enters more white-collar workflows, demand for tokens, model revenue, cloud revenue, and hardware demand will all rise together.

This judgment is more important than “whether to be bullish or bearish on AI.” Markets don’t trade static averages—they trade whether marginal changes can become mainstream.

The biggest capital markets divergence between Goldman Sachs and SemiAnalysis ultimately centers on NVIDIA and the semiconductor supply chain.

Goldman Sachs' perspective is straightforward: semiconductors have already captured the largest and most certain profits in the first phase. Once the market has priced in the "selling shovels" logic, the risk-reward profile begins to deteriorate. Any easing of cloud providers' capital expenditures will subject the semiconductor supply chain to dual pressures on valuations and orders.

SemiAnalysis believes that NVIDIA and TSMC control the most scarce resource in the AI era, yet have not yet fully priced it according to its value.

The article notes that memory prices have risen by approximately six times over the past year, and Neocloud’s one-year H100 rental contract prices have increased by about 40% since the October 2025 low. Meanwhile, NVIDIA and TSMC have not repriced as rapidly as downstream token values.

SemiAnalysis refers to NVIDIA as the "central bank" of the AI ecosystem.

This analogy is apt. NVIDIA controls computational power liquidity. It has the ability to raise prices, but cannot drain the entire system. Pushing prices too high would incentivize customers to accelerate their shift toward in-house ASICs, TPUs, and Trainium, while also inviting regulatory pressure. TSMC is similar. Advanced nodes are extremely scarce, but it has long prioritized customer relationships and ecosystem stability, avoiding the temptation to fully monetize scarcity during periods of high demand.

Restriction does not mean there is no room.

Rubin VR NVL72 is a key basis for SemiAnalysis’s assessment that NVIDIA still maintains pricing power. According to its model, Neocloud would need to charge approximately $4.92 per hour per GPU to achieve an IRR of 15.6% for the VR NVL72 project, similar to that of the GB300 project; if priced equivalently to GB300 on a per PFLOP basis, the theoretical ceiling for VR NVL72 is approximately $12.25 per hour per GPU; even using the more conservative rate of $0.55 per PFLOP, this corresponds to about $9.63 per hour per GPU—nearly double the cost-based pricing threshold.

The meaning here is clear: as long as the value of the downstream token continues to rise, NVIDIA's new system still has room to increase prices, Neocloud may still profit, and end users may still accept it.

The divergence between Goldman Sachs and SemiAnalysis has become sharp.

Goldman Sachs believes that the semiconductor industry's excessive profits are unsustainable because downstream sectors lack sufficient profit margins.
SemiAnalysis believes that the downstream profit pool is growing, so the hardware layer isn't earning too much—it's simply not yet fully compensated for its value.

The only variable determining victory is whether the new profit pool created by AI can be large enough to sustain the model lab, cloud providers, Neocloud, NVIDIA, TSMC, storage, and the power chain simultaneously.

The pie isn't big enough—Goldman Sachs wins.

The cake keeps growing, and SemiAnalysis wins.

Cloud providers are the most awkward layer in this debate.

They are both the largest spenders on capital expenditures and the platforms most likely to monetize AI demand. They are under pressure from NVIDIA, storage, and power supply chains, yet they possess enterprise customers, cloud services, model APIs, in-house chips, and software ecosystems.

Goldman Sachs is bullish on cloud providers because many negative factors have already been priced into the market. Capital expenditures are pressuring free cash flow, investors are questioning the ROI of AI, and valuations are under pressure. Either of two developments could provide a path to recovery for cloud providers: the realization of enterprise AI revenue or a reduction in capital expenditures.

SemiAnalysis examines cloud providers from a demand-side perspective. As token demand continues to grow, model labs and enterprise customers will require more computing power. Computing power is constrained by advanced process technology, memory, electricity, and rack-level systems. Buyers are more concerned about not being able to secure it than about the price.

Therefore, cloud providers are neither mere victims nor automatic winners.

They must be proven through financial reports: AI capital expenditures can be converted into revenue, profits, and customer retention. Whether cloud business growth has re-accelerated, whether AI revenue disclosures have become clearer, whether inference utilization can improve, whether in-house chips can reduce dependence on NVIDIA, whether enterprise customers are moving from pilot projects to long-term deployments, and whether free cash flow has stabilized—these metrics will be more important than ever.

These improvements in the metrics will strengthen Goldman Sachs' relative bullish logic for cloud vendors.

These metrics have not improved for a long time, and cloud providers remain a capital expenditure pressure point caught between NVIDIA and enterprise customers.

The emphasis in Goldman Sachs' report on "data structures" and "orchestration layers" may be the closest to corporate reality.

Enterprise AI won't remain confined to employees opening chat windows to ask questions. Truly financially impactful AI must integrate into customer service, sales, finance, procurement, R&D, risk management, supply chain, and IT operations. Each process involves data, permissions, compliance, approvals, legacy systems, and defined areas of responsibility.

No matter how powerful the model is, it cannot bypass these things.

This is where the enterprise software layer regains its importance. Low-risk, high-frequency tasks can be delegated to lightweight or open-source models; only high-risk, high-value tasks require state-of-the-art models. A middle layer is needed to classify task types, access data, manage permissions, select models, monitor costs, and log results.

• The advantages of traditional SaaS companies are industry expertise, customer relationships, data access points, and workflow accumulation. The disadvantages are technical debt and slow iteration speed.
• The advantages of AI-native companies are product speed, model invocation capabilities, and cost structure. The disadvantages are the lack of enterprise entry points and industry context.
• The advantage of Frontier Model Company is its superior intelligence; its disadvantage is the lack of control over enterprise processes.

The software layer won't be simply consumed by AI. Software companies without control over data and process workflows may be abstracted away by models. However, software companies that control data structures, workflows, and model routing have the opportunity to turn AI into a larger market—shifting from selling seats to selling productivity.

Whether a company’s ROI can be generalized from strong user samples like SemiAnalysis to ordinary enterprises largely depends on this layer.

In AI trading, the question used to be: Who is closest to the computing power?

This question is too broad right now.

In the next phase, the market will ask about more granular variables.

First, will the token's value continue to rise? If Agentic AI expands from code, research, and analysis into more white-collar workflows, model labs and reasoning chains will continue to be reevaluated.

Second, whether the gross margin of the model lab continues to improve. Revenue growth alone is no longer sufficient; the market will focus on inference costs, caching efficiency, SKU upgrades, and pricing power for cutting-edge models.

Third, can cloud providers convert capital expenditures into revenue? AI-related capex is no longer automatically viewed as positive; only capex that contributes to cloud revenue, inference gross margins, and enterprise contracts will be rewarded by the market.

Fourth, can NVIDIA continue to raise prices due to system-level bottlenecks? GPU is only the first layer—whether NVIDIA can continue to extract margins depends on Rubin, SOCAMM, networking, rack-level systems, software stacks, and supply chain procurement capabilities.

Fifth, can TSMC and memory manufacturers reprice scarcity? Advanced nodes, HBM, DRAM, SoC RAM, and advanced packaging—if they continue to be supply bottlenecks, value will not easily move away from the upstream.

Sixth, can enterprise software secure an entry point for AI implementation? Software companies without process entry points will face compression, while those with entry points, data, and orchestration capabilities may become more valuable.

AI infrastructure trades have not expired.

It rose so sharply that it sparked a disagreement between Goldman Sachs and SemiAnalysis.

Goldman Sachs reminds the market that the benefits of the chip supply chain have already been fully priced in. If corporate ROI continues to be delayed, cloud providers' cash flow may turn against capital expenditures, correcting the current scenario where semiconductors are the sole beneficiaries.

SemiAnalysis reminds the market that the AI experience of 2024 should not be used to judge Agentic AI in 2026. Tokens are becoming means of production, model labs are beginning to improve gross margins, computing supply remains tight, and NVIDIA and TSMC may not yet be fully pricing according to value.

When these two judgments are considered together, the focus of AI trading has shifted.

Over the past two years, the market has rewarded scarce assets. Next, the market will focus on who can sustainably retain the economic value created by AI on their income statement.

If SemiAnalysis is seeing a marginal turning point, the AI value chain will continue to grow, giving model labs, cloud providers, NVIDIA, TSMC, storage, and power infrastructure further reason to share in the gains.

If Goldman Sachs is seeing a reality closer to that of an average company, capital expenditures will hit cash flow constraints first; the semiconductor supply chain needs to digest overly optimistic expectations, while cloud providers may benefit from better relative returns due to valuation compression and potential spending discipline.

The most likely current state is between the two.

The strongest users have already started buying tokens aggressively, while ordinary companies haven't even finished their accounting. Capital markets will first price in the marginal changes driven by the strongest users, then wait for average companies to validate those changes through financial reports. The faster the validation, the closer we get to SemiAnalysis’s world; the slower the validation, the higher the odds for Goldman Sachs’ trades.

The AI "shovel" still dominates the market, but the question has shifted from "who is selling shovels" to another ledger: who has already made enough, who can still raise prices, and who will become the next layer of true rent collectors.