Where will the most value and profits accrue in the AI stack?

Thesis: the highest-quality profits remain upstream of applications—but the bottleneck is moving downward

As of 21 September 2026, the most defensible answer is not “AI applications” and not simply “ whoever owns the largest model.” The greatest near-term economic rents are accruing to the layers that are both capacity-constrained and difficult to substitute: accelerated-compute platforms, custom AI silicon and networking, leading-edge manufacturing, and high-density power and thermal infrastructure.

The key distinction is between industry spending and shareholder profit. AI spending can rise rapidly while customers, cloud operators and application vendors absorb much of the cost. The best businesses are those that sell a scarce input, have a differentiated ecosystem or technical qualification burden, and convert demand into operating cash flow before competition erodes pricing.

1. Compute platforms still hold the strongest current rent pool

NVIDIA remains the clearest example of a company monetizing scarcity rather than merely participating in AI growth. In Q2 FY2027, revenue was $96.2 billion, Data Center revenue was $89.0 billion, and GAAP and non-GAAP gross margins were both 75.0%. The company also reported $21.341 billion of free cash flow. [1]

Those numbers matter because they show that the current bottleneck is not just demand for compute; it is demand for a complete, deployable system with software, networking, memory and support. NVIDIA's economics reflect platform pricing power and integration, not commodity component pricing. Its Q3 FY2027 outlook called for $108.0 billion of revenue, plus or minus 2%, with a 74.0% gross-margin outlook, excluding Data Center compute revenue from China. [1]

The risk is that today's margin is mistaken for a permanent industry structure. Hyperscalers are economically motivated to design custom accelerators, use multiple suppliers and improve tokens-per-dollar. A Goldman Sachs analysis explicitly says it assumes NVIDIA accounts for 75% of total compute spend in each period; that is an analytical assumption, not an observed market-share fact. [2] The invalidation signal for the NVIDIA thesis is therefore not merely slower AI growth; it is sustained evidence that customers can obtain comparable performance at materially lower total cost without NVIDIA's software ecosystem.

2. Custom silicon and networking are the second major profit basin

Broadcom is positioned where hyperscaler scale creates a different opportunity from general-purpose GPUs: custom accelerators and the networking fabric required to connect large clusters. Broadcom reported Q3 FY2026 revenue of $29.6 billion, up 86% year over year. Semiconductor Solutions revenue was $20.839 billion, while AI semiconductor revenue reached $16.7 billion, up 221% year over year and 54% sequentially. Free cash flow was $13.7 billion, or 46% of revenue. [3]

Management's Q4 guidance was for approximately $34.8 billion of revenue and $21.7 billion of AI semiconductor revenue, with expected non-GAAP operating income of approximately 66% of projected revenue. [3] This is unusually strong evidence that value is spreading from one dominant accelerator vendor toward a broader architecture: custom silicon captures workloads that hyperscalers believe justify bespoke design, while networking becomes essential as cluster scale increases.

Arista is a credible but less bottlenecked networking exposure. It reported Q2 2026 revenue of $3.036 billion, up 37.7% year over year, with GAAP and non-GAAP operating margins of 45.4% and 49.9%. It introduced 1.6-terabit AI-fabric platforms and guided to approximately $3.3 billion of Q3 revenue with a 48–49% non-GAAP operating margin. [4] Arista's business quality is high, but its shareholder outcome depends more on maintaining share in a competitive networking market than on controlling the indispensable compute platform.

3. TSMC captures the manufacturing bottleneck with exceptional economics

TSMC is the broadest upstream beneficiary because leading AI designers still require its manufacturing scale and process capability. In Q2 2026, TSMC reported US$40.20 billion of revenue, up 36.0% year over year, a 67.7% gross margin, a 60.3% operating margin and a 55.6% net profit margin. Net income attributable to shareholders was NT$706.56 billion, up 77.4% year over year. [5]

TSMC's position is structurally different from NVIDIA's. It does not capture the full value of the platform, but it is a manufacturing gatekeeper for many competing designs. That makes it a diversified way to own AI-chip growth, though not a risk-free one: customers retain significant purchasing power, the company must spend heavily to remain at the leading edge, and geopolitical risk is material. Its Q3 2026 guidance was US$39.0–40.2 billion of revenue, with a 65.5–67.5% gross-margin range and a 56.5–58.5% operating-margin range. [5]

4. The next physical bottleneck is electricity, power delivery and cooling

The most important shift in the investment map is that AI capacity is increasingly constrained by the physical infrastructure around the chip. The IEA projects global data-center electricity consumption to rise from 485 TWh in 2025 to 950 TWh in 2030. It says AI-focused data-center electricity use is growing faster than total data-center demand and could triple over that period. [6] The IEA also identifies grid capacity as a critical bottleneck in many regions, slowing new generation, storage and demand deployment. [7]

This creates a rent opportunity for suppliers of power conversion, uninterruptible power, distribution and thermal-management systems—but only where technical qualification and execution prevent rapid commoditization. Vertiv's Q2 2026 results are useful evidence: sales rose 24% year over year to $3.274 billion, organic sales rose 18%, adjusted operating margin reached 22.6%, operating cash flow was $1.1 billion and adjusted free cash flow was $925 million. [8] Management said AI and general-compute deployments are becoming more infrastructure-intensive and raised full-year 2026 guidance to $13.8–14.2 billion of sales, 30–32% organic growth and $2.4–2.6 billion of adjusted free cash flow. [8]

Vertiv is therefore a more direct physical-bottleneck exposure than a generic electrical-equipment proxy. The counterargument is that high growth can attract capacity and competitors. Investors should watch whether margins remain near the current level as supply expands, rather than extrapolating management guidance indefinitely.

5. Cloud and applications can capture the largest absolute dollars—but they bear the largest costs

Cloud platforms are where AI revenue becomes visible to end customers. Microsoft reported FY26 Q4 revenue of $90.0 billion, Microsoft Cloud revenue of $59.3 billion, Azure and other cloud-services growth of 43%, and more than 30 million paid Microsoft 365 Copilot seats. Full-year additions to property and equipment were $115.948 billion. [9] This is genuine monetization evidence, but it is not proof that application-layer returns exceed infrastructure returns: cloud operators must fund servers, networking, power, depreciation, model access and customer acquisition.

Alphabet illustrates the cost side even more clearly. It reported $44.9 billion of Q2 2026 capex, with approximately 60% of technical-infrastructure investment directed to servers and 40% to data centers and networking. It raised full-year 2026 capex guidance to $195–205 billion, while Google Cloud revenue rose 82% to $24.8 billion. Q2 free cash flow was negative $5.9 billion because of capex, and management warned that higher infrastructure investment would pressure the income statement through depreciation and energy costs. [10]

The economic test for cloud and application profits is therefore revenue per unit of compute minus the full cost of serving that usage. Paid seats, cloud revenue and backlog are leading indicators; durable shareholder returns require inference utilization, pricing discipline and capital intensity to improve together. Gartner's market estimates show AI application software rising from $83.679 billion in 2024 to $269.703 billion in 2026, while AI-optimized servers rise from $140.107 billion to $329.528 billion and AI-processing semiconductors from $138.813 billion to $267.934 billion. These are market-spending figures, not profit forecasts. [11]

Investment conclusion

The most value is likely to accrue in a barbell:

1. Platform bottlenecks: NVIDIA remains the highest-quality direct exposure to accelerated-compute economics, but its valuation and long-term returns depend on retaining system-level differentiation as custom silicon improves.

2. Architecture and manufacturing beneficiaries: Broadcom and TSMC offer exposure to the widening AI architecture and to the manufacturing constraint beneath competing designs.

3. Physical infrastructure: Vertiv offers direct exposure to the power and cooling constraint that increasingly determines whether a data-center project can be energized and operated.

4. Selective network exposure: Arista is a strong operator with material AI-fabric exposure, but its economics are less uniquely scarce than NVIDIA's or TSMC's.

5. Cloud/application platforms: Microsoft and Alphabet may capture the greatest absolute revenue pools, but they also finance the buildout. Their returns depend on monetization outrunning depreciation, energy and compute costs.

The observable milestones are the next reported quarters: NVIDIA's ability to hold roughly mid-70% gross margins while custom silicon expands; Broadcom's conversion of AI-semiconductor growth into free cash flow; TSMC's ability to sustain margins while adding leading-edge capacity; Vertiv's conversion of exceptional growth into cash without margin dilution; and cloud platforms' ability to grow AI revenue without permanently worsening free cash flow.

This analysis distinguishes business exposure from valuation. No stock is described as cheap or expensive because a dated price and valuation comparison was not required to establish where the economic rents currently sit.