Over the past two years, AI-related investments have dominated global equity markets. Assets linked to the AI infrastructure chain—including Nvidia, semiconductor equipment, HBM, advanced packaging, data centers, power equipment, transformers, cooling systems, and gas turbines—have been repeatedly revalued by the market. Rather than fading, this trend has intensified, forcing investors to confront a more challenging question: having richly rewarded the first-phase winners of the AI supply chain, can the rally continue?
Two reports—one from Goldman Sachs and another from SemiAnalysis—offer sharply contrasting views at this inflection point.
Goldman Sachs analyst James Covello presents a cooler outlook: he argues that the first phase of AI infrastructure has already been fully priced in, with chipmakers and "picks-and-shovels" suppliers capturing the bulk of early, certain profits. However, enterprise return on investment (ROI) remains unproven at scale, and cash flow pressures are mounting for cloud providers. Under this logic, the better relative trade is not to chase semiconductors further, but to overweight hyperscale cloud providers and underweight semiconductors.
SemiAnalysis offers a nearly opposite view: if Agentic AI truly turns tokens into means of production, model labs begin improving their margins, and cutting-edge models retain pricing power, then AI infrastructure is not "overpriced"—it has yet to be fully repriced according to the new value of tokens. Companies like Nvidia, TSMC, memory suppliers, Neocloud, and model labs still have room to capture incremental value.
This is not a debate about whether AI has a future. AI capital expenditure continues to rise, and AI infrastructure stocks have not cooled. The real question is: having captured the first wave of profits, has the chip layer already been fully valued by the market? If Agentic AI amplifies token value further, will the next round of profits continue to accrue to hardware players, or will they be redistributed to model labs, cloud providers, and enterprise software layers?
Goldman Sachs is focused on an industry chain that has not yet closed the loop. The report’s most striking argument is not a denial of AI user growth or technological progress. Covello acknowledges two facts: consumer adoption of AI has been faster than expected, and cloud providers—despite stock pressure—have not cut AI capital expenditures as anticipated but have instead raised them. AI has not cooled, and capex has not receded.
But Goldman looks further downstream. Many consumers still use AI at the free tier. User growth demonstrates product appeal but does not directly pay for GPUs, data centers, electricity, networks, or model inference. Enterprises are key to closing the AI economic loop: their willingness to pay continuously, and their ability to cut costs, raise revenue, or boost productivity via AI, will determine whether the entire chain can sustain today’s capital spending levels.
Goldman’s answer is cautious. The report notes that while enterprises are investing heavily in generative AI, many organizations have yet to see verifiable returns. Meanwhile, global IT spending continues to rise—AI has not reduced overall tech budgets. For investors, this raises a practical issue: companies are buying, testing, and discussing AI, but it has not yet broadly reached the profit statement.
This contrasts sharply with the profits already realized in the AI infrastructure chain. Chip companies are earning money; memory, power, and data center-related firms have been repeatedly revalued. Cloud providers, however, bear much of the capital expenditure—data center construction, GPU procurement, power access, networking, and server racks all hit their books first. Goldman notes that hyperscale cloud providers have already consumed some of their operating cash flow surplus and have begun funding data center builds via debt, with data center debt issuance expected to double to $182 billion in 2025.
This is the imbalance Goldman sees. In a normal semiconductor cycle, chip suppliers’ high profits usually indicate that their customers are also expanding. Profitable customers buy more chips, sustaining the cycle. This AI cycle is more awkward: chip-layer profits are clear, but returns for customers and application layers are not.
Thus, Goldman’s view is not that "AI is useless," but that "the current profit distribution is difficult to extrapolate linearly." Semiconductor firms have captured the most certain profits of the first phase. The question is whether downstream customers will generate enough profit to continue supporting such high upstream capex and profit concentration.
Goldman’s trade recommendation essentially bets on mean reversion: relatively overweight hyperscale cloud providers, underweight semiconductors. This rests on two potential paths. First, enterprise AI ROI materializes. If companies prove AI drives revenue and efficiency gains, the market will reassess cloud providers’ capex—spending once seen as a drag on free cash flow could be reappraised as future revenue and platform control. Semiconductors would also benefit, but having already been richly rewarded, their relative upside may be limited. Second, if enterprise ROI remains elusive, cloud providers may cut capex under cash flow and investor pressure, and the market would reward better cash flow discipline. Semiconductor suppliers would then face order downgrades.
Goldman believes both paths support cloud providers over semiconductors. The trade would fail only under a third scenario: enterprise ROI stays unclear, yet cloud providers continue spending indiscriminately, and semiconductors keep capturing the lion’s share of profits—precisely the dynamic familiar over the past two years. Thus, Goldman’s critique targets market pricing, not AI technology: AI infrastructure benefits have been fully traded, as have cloud providers’ drawbacks. Next, the market will watch for reversals in both directions.
SemiAnalysis approaches from a completely different angle. It does not deny that from 2023 to 2025, AI value flowed mainly to infrastructure—Nvidia, power, data centers, and memory were the clear first-phase winners. Model companies and inference service providers struggled early on; many AI products resembled better search boxes with unattractive margins.
But SemiAnalysis argues that things began changing after late 2025, driven by Agentic AI. Previously, tokens were更像 "Q&A costs"—users asked, models answered, saving time but offering limited value. Now tokens are entering complex workflows: coding, financial modeling, dashboard generation, earnings analysis, data cleaning, and chart creation. The firm uses itself as an example: its analysts now use agents daily for research and modeling tasks that once took junior analysts hours or went unaddressed. SemiAnalysis discloses annualized token spending on Anthropic Claude reached $10.95 million at one point—about 30% of employee compensation.
These figures may not represent all enterprises, but they signal change among marginal users. For average consumers, AI subscriptions may cost just tens of dollars monthly. For high-intensity knowledge workers, tokens are becoming means of production. A few dollars in tokens can yield not just text, but models, charts, code, data cleaning, and financial analysis—even work that would otherwise not get done. Users’ perception of AI costs shifts: they ask not "how much per million tokens?" but "how much labor did these tokens replace, and how much output did they add?"
This is where SemiAnalysis diverges from Goldman. Goldman sees unclear ROI for the average enterprise; SemiAnalysis sees power users already consuming tokens heavily and willing to pay for superior models.
SemiAnalysis’ second key argument is that model labs’ unit economics are improving—contrary to past concerns. Model companies were once seen as squeezed between chip suppliers and cloud providers: revenue grew quickly, but training and inference costs grew faster. More users meant higher costs; stronger models meant heavier capex. The model appeared high-growth, low-margin, and cash-burning.
Agentic AI changes this. On the pricing side, cutting-edge models handle higher-value tasks, and users pay premiums for stronger performance. On the cost side, hardware迭代, inference optimization, caching, and software engineering lower per-token costs. On the product side, model firms can tier pricing via premium SKUs, faster response, and better reasoning.
SemiAnalysis cites the example of running DeepSeek on B300: different software optimizations can raise throughput from around 1,000 or 8,000 to roughly 14,000 tokens/second/GPU. Combined with hardware upgrades, an optimized GB300 NVL72 configuration offers ~17x higher FP8 throughput than H100; switching to FP4 (unsupported natively on Hopper), the gap reaches 32x, with total cost of ownership per GPU up only ~70%. This means model labs can simultaneously increase the economic value created by tokens and lower their production cost.
SemiAnalysis notes that Anthropic’s ARR rose from $9 billion to over $44 billion, with inference infrastructure gross margins improving from 38% to over 70%. Even as model list prices fall, rising adoption of premium models, better cache hit rates, and hardware efficiency gains could further expand margins.
If this holds, the second phase of the AI supply chain will not simply see "chips keep winning" or "cloud providers rebound." Model labs could transform from cash-burning entities into new value-capture layers.
The core disagreement lies in which sample better represents the future: the average enterprise or the marginal user. Goldman focuses on average firms—with complex legacy systems, compliance needs, and approval processes. Many deploy AI chatbots, internal assistants, or pilot projects to showcase AI strategy to boards. Spending is real, but business processes may not change—and without process change, ROI struggles to reach financial statements.
This is why Goldman emphasizes data structure and orchestration layers. Without integrated inventory, customer profiles, and recommendation systems, an AI customer service agent might recommend out-of-stock items. Without model routing layers, even simple queries go to expensive frontier models, driving costs out of control. AI implementation stalls not because models are weak, but because enterprises are not ready to embed models into operational systems.
SemiAnalysis focuses on marginal users—tasks like research, coding, modeling, charting, and financial analysis are naturally agent-friendly. They are text-heavy, digital, structured, easily evaluated, and amenable to workflow integration. Such organizations will see ROI sooner and ramp token consumption faster.
The question for capital markets is whether this leading-edge adoption will diffuse widely. If SemiAnalysis’ sample is just a few super-users, Goldman’s framework may prevail: AI capex will face growing cash flow constraints, semiconductor chains will digest high expectations, and cloud providers may gain relatively due to spending discipline and compressed valuations.
If SemiAnalysis’ observations represent leading indicators ahead of broad diffusion, markets cannot dismiss the AI chain based on today’s low average ROI. Once Agentic AI enters more white-collar workflows, token demand, model revenue, cloud sales, and hardware needs could rise together.
This judgment is more important than being "bullish or bearish on AI." Markets trade not on static averages, but on whether marginal changes become mainstream.
The biggest capital market disagreement between Goldman and SemiAnalysis centers on Nvidia and the semiconductor chain. Goldman’s view is straightforward: semiconductors have captured the largest, most certain first-phase profits. After the "picks-and-shovels" narrative is priced in, risk-reward deteriorates. Any softening in cloud provider capex would expose semiconductor firms to valuation and order pressure.
SemiAnalysis contends that Nvidia and TSMC control the AI era’s scarcest resources but have not fully repriced according to value. The article notes that memory prices have risen ~6x over the past year, and Neocloud’s one-year H100 lease rates are up ~40% from October 2025 lows. Yet Nvidia and TSMC have not repriced as rapidly as downstream token value.
SemiAnalysis calls Nvidia the AI ecosystem’s "central bank"—an apt analogy. Nvidia controls compute liquidity. It can raise prices, but draining the system excessively would spur clients toward custom ASICs, TPUs, or Trainium chips and attract regulatory scrutiny. TSMC behaves similarly: advanced nodes are extremely scarce, but the company values long-term client relationships and ecosystem stability, avoiding cashing in all scarcity during an upcycle.
Restraint does not mean no room for increases. SemiAnalysis points to Rubin VR NVL72 as evidence of Nvidia’s lingering pricing power. Its model suggests that for Neocloud to achieve a 15.6% IRR on VR NVL72 similar to GB300 projects, lease rates would need to reach ~$4.92/hour/GPU. If priced parity per PFLOP with GB300, the theoretical ceiling is ~$12.25/hour/GPU. Even using a conservative $0.55/PFLOP, the rate would be ~$9.63/hour/GPU—nearly double the cost-based threshold. The implication is clear: if downstream token value keeps rising, Nvidia’s new systems have pricing power, Neocloud can remain profitable, and end-users may accept the costs.
Thus, the disagreement sharpens: Goldman believes semiconductor dominance is unsustainable because downstream profits are insufficient. SemiAnalysis argues the downstream profit pool is expanding, so the hardware layer is not overearning but undercharging relative to value.
The decisive variable is whether AI’s new profit pool grows large enough to support model labs, cloud providers, Neocloud, Nvidia, TSMC, memory, and power chains simultaneously. If the pie is too small, Goldman wins. If it keeps expanding, SemiAnalysis wins.
Cloud providers occupy the most delicate position. They are both the largest capex bearers and the platform most likely to monetize AI demand. They face pressure from Nvidia, memory, and power suppliers, yet possess enterprise clients, cloud services, model APIs, custom chips, and software ecosystems.
Goldman’s overweight on cloud providers reflects that market prices already incorporate much of the negative—capex suppressing free cash flow, AI ROI doubts, and valuation pressure. If either enterprise AI monetization materializes or capex contracts, cloud providers have a recovery path.
SemiAnalysis views cloud providers from the demand side: if token demand keeps growing, model labs and enterprise clients will need more compute, which is constrained by advanced processes, memory, power, and rack-level systems. Buyers’ main worry shifts from cost to availability.
Thus, cloud providers are neither pure victims nor automatic winners. They must demonstrate with financials that AI capex translates into revenue, profits, and customer stickiness. Metrics like cloud revenue reacceleration, clearer AI revenue disclosure, inference utilization improvements, custom chip progress reducing Nvidia dependence, enterprise pilots shifting to long-term deployment, and free cash flow stabilization will grow in importance.
If these metrics improve, Goldman’s relative overweight thesis strengthens. If they lag, cloud providers remain a capex-squeezed layer between Nvidia and enterprise customers.
The software layer will determine whether ROI diffuses from leading samples to the average enterprise. Goldman’s emphasis on "data structure" and "orchestration layers" may be closest to enterprise reality. AI will not remain limited to employees querying chat interfaces. Financially impactful AI must enter customer service, sales, finance, procurement, R&D, risk management, supply chain, and IT operations—each with its own data, permissions, compliance, approvals, legacy systems, and accountability boundaries. No model, however powerful, can bypass these.
This is where enterprise software regains importance. Low-risk, high-frequency tasks can use lighter or open-source models; high-risk, high-value tasks need frontier models. An intermediate layer must classify tasks, access data, enforce permissions, select models, monitor costs, and return results.
Traditional SaaS firms bring industry expertise, customer relationships, data access, and workflow experience—but also technical debt and slower iteration. AI-native companies offer product speed, model integration, and cost efficiency—but lack enterprise entry points and domain context. Frontier model companies provide top-tier intelligence—but lack control over business processes.
The software layer will not simply be absorbed by AI. Software firms without data or process control may be abstracted away by models. Those mastering data structures, workflows, and model routing could turn AI into a larger market—shifting from selling seats to selling productivity.
The diffusion of ROI from power users like SemiAnalysis to average enterprises will largely depend on this layer.
Looking ahead, capital markets will monitor six key variables: 1. Whether token value continues rising—if Agentic AI spreads from coding, research, and analysis to more white-collar workflows, model labs and inference chains will be revalued. 2. Whether model lab margins keep improving—revenue growth alone won’t suffice; markets will watch inference costs, cache efficiency, SKU upgrades, and frontier model pricing power. 3. Whether cloud providers can convert capex into revenue—AI spending will no longer automatically be seen as positive; only capex that drives cloud revenue, inference margins, and enterprise contracts will be rewarded. 4. Whether Nvidia can continue raising prices via system-level bottlenecks—GPUs are just the first layer; Rubin, SOCAMM, networking, rack-level systems, software stacks, and supply chain leverage will determine its ability to extract value. 5. Whether TSMC and memory suppliers can reprice scarcity—if advanced nodes, HBM, DRAM, SOCAMM, and advanced packaging remain supply constraints, value will stay upstream. 6. Whether enterprise software captures AI implementation entry points—firms without process access may be compressed; those with entry points, data, and orchestration capabilities could gain value.
After AI infrastructure’s market dominance, the real debate is just beginning. The trade has not failed—it has risen so sharply that it forced the Goldman-SemiAnalysis disagreement. Goldman warns that chip chain benefits are fully priced; if enterprise ROI delays, cloud provider cash flow could curb capex, correcting semiconductor profit dominance. SemiAnalysis cautions against judging 2026 Agentic AI by 2024 experiences—tokens are becoming production factors, model labs are improving margins, compute supply remains tight, and Nvidia/TSMC may not have fully value-priced.
Together, these views signal a shift in AI trading focus. The past two years rewarded scarcity; next, markets will judge which players can sustainably retain AI-created economic value on their income statements.
If SemiAnalysis identifies a marginal inflection, the AI chain’s profit pie will keep growing, justifying continued value capture by model labs, cloud providers, Nvidia, TSMC, memory, and power chains. If Goldman’s average-enterprise reality prevails, capex will hit cash flow limits, semiconductor chains will digest excessive expectations, and cloud providers may benefit from valuation compression and potential spending discipline.
The current state likely lies between these extremes. Power users are aggressively buying tokens; average enterprises are still calculating ROI. Capital markets will first trade the marginal change driven by power users, then wait for average firms to validate with financials. The faster validation comes, the closer SemiAnalysis’ world becomes; the slower, the greater Goldman’s trade advantage.
AI infrastructure still rules the market, but the question has shifted from "who sells the picks and shovels?" to a new ledger: who has earned enough, who can still raise prices, and who will become the next true rent-collector.
Comments