Key Takeaways:
- Cheaper AI is not necessarily bad for the AI investment thesis. The key question is whether increased usage can offset falling token prices and sustain overall AI spending.
- Pricing pressure is building across the AI ecosystem. Open models, enterprise cost discipline, and efficiency gains are all pushing token prices lower, which could reshape where value is captured.
- Portfolio construction matters more as the AI trade matures. If demand keeps pace, infrastructure spending could remain durable. If it doesn’t, companies financing the buildout may be more vulnerable than the businesses using AI.
The AI investment cycle has been driven by an assumption of ever-higher spending on models, data centers, electricity, chips, and compute. However, falling token prices at frontier models could challenge that assumption. Lower token prices may accelerate overall adoption and expand the range of economically viable AI use cases. But more AI adoption may not automatically translate into higher profits everywhere.
The question for investors is whether usage grows fast enough to offset falling prices. The answer will help determine how value flows through the AI ecosystem, which parts of the value chain benefit, and which may come under pressure.
The AI Value Chain: How $1 Flows
To understand where falling token prices matter most, it helps to follow the money through the AI value chain.
Exhibit 2 traces how $1.00 of foundation-model API spending flows through model companies, cloud providers, data centers, power, chips, memory, and fabrication.
Watching the flow of dollars through the value chain, it becomes clear that not every layer participates in AI economics in the same way. The foundation-model layer, despite sitting closest to the end user, runs at an operating loss—funded by external capital rather than revenue [Exhibit 4]. Cloud and neocloud providers capture thin margins, with standalone neoclouds earning almost nothing once financing costs are counted [Exhibit 4].
The real profit pools sit further upstream, in the silicon: chip designers like Nvidia, AMD, and Broadcom, memory makers like Micron and SK Hynix, and the foundries that fabricate the chips all earn margins well above the rest of the chain [Exhibit 4]. In other words, as AI spend flows through the ecosystem, value concentrates not at the layers closest to the user but in the hardware vendors near the bottom of the stack.
Why Token Prices Are Falling
Three mechanisms appear to be contributing to the decline in token prices, and each could affect the AI value chain differently; a price war looks very different from a decrease in prices because of compute efficiency.
From Open Weight Models
The most visible driver is competition: open-weight models2, increasingly led by Chinese labs, are narrowing the capability gap with the closed frontier labs3 while undercutting them sharply on cost. Open model performance now lags closed model performance by a matter of months, and DeepSeek’s training cost was reported at roughly 30x cheaper than OpenAI’s o1 [Exhibit 5].4 As models become more capable and better integrated into workflows, leading AI users increasingly route work to cheaper but sufficiently capable models.
Beyond cost, open-weight models offer control, customization, and data ownership, making them attractive for organizations with specialized workflows where those attributes can matter as much as raw capability. The recently announced Palantir-Nvidia partnership, built around deploying open models in security-sensitive government environments, illustrates the tradeoff.5
Ramp estimates open model penetration remains low, at just over 5% of enterprise token usage. But the heaviest token consumers are already the most active open-model users, suggesting adoption should broaden as enterprise AI usage accelerates [Exhibit 6].6
From Enterprise Customers
The second driver of token pricing pressure sits on the buyer’s side of the table: enterprise customers appear to be pushing back on surging AI bills. As companies’ AI usage moves from pilot projects into production workloads, buyers may increasingly evaluate models based on cost per task rather than benchmark leadership. For many companies, a model that is 95% as capable at half the cost would be economically preferable to the frontier alternative. Silicon Data’s Large Language Model (LLM) Token Expenditure Index, a usage-weighted measure of what users are willing to pay for LLM access, has declined significantly in the last few months, suggesting that customers are becoming more price-sensitive [Exhibit 8].
From Efficiency Itself
In an AI context, efficiency can mean two different things. The first is that customers may require fewer tokens to complete a task. For example, a company might cut token use by 30% per task, but if it starts using AI for three times as many tasks, total compute demand still goes up.
The second form of efficiency comes from providers requiring less computing power to generate each token. Lately, inference costs have been falling quickly as technology improves and hardware gets better [Exhibit 9]. For example, as AI moves from frontier research into routine production workloads, companies are using more specialized chips, called application-specific integrated circuits (ASICs), that can run those tasks at lower cost than general-purpose GPUs.
Both forms of efficiency can reduce the cost of delivering AI services and put downward pressure on token prices. Whether lower costs ultimately reduce infrastructure spending depends on the price sensitivity of users. If cheaper AI unlocks substantially more usage, total compute demand can continue growing despite falling prices. Otherwise, efficiency gains could reduce spending throughout the stack.
Scenario A — If Demand Keeps Pace (The Bull Case)
In the bull case, token prices keep falling, but usage grows fast enough to keep aggregate AI spending at or above expectations.
Nvidia provides a useful example. Custom chips may reduce Nvidia’s share of the AI accelerator market, but the overall market has expanded fast enough for Nvidia to continue growing significantly. For investors, the more important variable may be market growth rather than market share. In this scenario, the clearest beneficiaries would likely be companies that either control multiple layers of the AI stack or benefit from rising aggregate compute demand:
- Hyperscalers could remain among the best positioned because they monetize AI across cloud services, proprietary models, custom silicon, enterprise relationships, and distribution.
- Memory and custom silicon providers could benefit from sustained compute demand. Memory content is climbing toward 25-30% of AI server rack cost, while custom ASIC programs could benefit from both growing compute and the continued push toward efficiency.7
- Application-layer businesses could also benefit. While AI may pressure some software moats, cheaper AI could also reduce input costs and improve the economics of differentiated AI-enabled products and services.
Scenario B — If Demand Doesn’t Keep Pace (The Bear Case)
The bear case is not that AI fails. It is that AI demand emerges on a different timetable than the infrastructure being built to support it. If token prices fall and usage does not rise enough to offset them, aggregate AI spending could disappoint, and the winners and losers across the ecosystem could look very different. In that scenario, some of the most exposed businesses may be those most dependent on continued infrastructure investment. Recent volatility bears this out: since early June, markets have priced a growing risk premium into AI-linked market segments even as less AI-exposed areas made new highs [Exhibit 10].
Standalone neoclouds may be particularly vulnerable. Unlike hyperscalers, they often rely on concentrated customer relationships and have fewer alternative revenue streams to absorb a slowdown in compute demand. Their financing structures can add another layer of risk, particularly when growth expectations are embedded in long-duration contracts or asset-backed financing.
History offers a useful caution. During the dot-com boom, fiber networks were built ahead of demand, creating financial distress for some capital providers even though the internet ultimately transformed the economy. Railroads produced a similar lesson a century earlier. In both cases, the technology proved transformative, but many early infrastructure investors struggled to capture the long-term upside.
AI is not there today. Demand remains strong and capacity remains constrained. But the comparison highlights a key risk: the companies financing the buildout may not be the same companies that ultimately capture the value.
Capital Markets Are Becoming Central to the AI Buildout
If the bear case is fundamentally a story about infrastructure spending disappointing expectations, the next question is who ultimately bears that risk. Increasingly, the answer is not just hyperscalers. It is the broader capital markets financing the AI buildout.
Capital markets are becoming increasingly important to the AI ecosystem. As AI investment expands beyond what hyperscalers can comfortably fund through internally generated cash flow, a growing share of data centers, power infrastructure, and computing capacity is being financed by public and private investors [Exhibit 11]. At the same time, many of the industry’s most highly valued AI companies remain private, meaning a significant portion of AI’s equity value has yet to face continuous public-market scrutiny.
The New AI Financing Model
The AI buildout is increasingly being financed by a broader ecosystem of investors, including REITs, infrastructure funds, insurers, pension funds, private-credit vehicles, and structured-finance arrangements.
Transactions such as Meta’s Hyperion project with Blue Owl illustrate the shift. Rather than funding every project directly from their balance sheets, hyperscalers are increasingly partnering with outside capital providers to secure capacity while transferring more of the financing burden to investors.
AI Data Center Financing Landscape
AI infrastructure is no longer funded through a single model. Exposure increasingly spans hyperscaler balance sheets, public infrastructure vehicles, and private-capital structures, each with different risk and return characteristics. The table below summarizes the major financing models currently used across the ecosystem.
Distributed Risk Is Not Necessarily Reduced Risk
As exposure spreads beyond hyperscaler balance sheets, risk becomes more widely distributed across REIT shareholders, infrastructure funds, insurers, pension funds, and private-credit investors. While broader participation can improve resilience by reducing reliance on any single balance sheet, it also creates additional channels through which stress can be transmitted. Distributed risk is not necessarily reduced risk, particularly if investors accumulate concentrated exposure to the same assets, counterparties, or assumptions about future AI demand.
From Technology Risk to Credit Risk
A slowdown in AI adoption would likely first affect equity valuations, but heavily financed infrastructure projects could create broader credit pressures if expected cash flows fail to materialize. In that scenario, losses would extend beyond technology companies and their shareholders to the investors financing the buildout. The key question is not whether AI spending ultimately proves excessive, but who ultimately owns the downside if it does. As AI infrastructure financing increasingly resembles traditional infrastructure and project-finance markets, technology risk has the potential to migrate into the broader financial system.
The Unexpected Winners of Cheaper AI
A weaker AI infrastructure scenario would not hurt everyone. If the greatest risks sit with the structures financing the buildout, the largest beneficiaries may be the businesses that consume AI rather than produce it. For AI-native and application-layer companies, tokens are a direct input cost, so cheaper models feed straight into gross margins; more broadly, greater AI abundance could lift productivity across private equity and venture-backed portfolio companies. The losers may be those who financed the buildout assuming persistently scarce, highly profitable compute.
Owning the Upside, Managing the Whipsaw
The portfolio question is not whether AI adoption continues. It is how to stay invested in a theme where adoption and monetization may diverge, where different layers of the value chain benefit under different scenarios, and where the ultimate winners may not be the same companies financing the buildout.
This is where structured investments can fit naturally. They are designed for environments where investors want continued exposure to a compelling long-term theme but face uncertainty around timing, valuations, and downside risk. Rather than requiring a precise view on whether the bull or bear case plays out, structured strategies can help investors participate in continued AI adoption while providing some protection if market expectations prove too optimistic.
Buffered and defined-outcome strategies can help mitigate drawdowns if AI-linked equities re-rate, while preserving participation if adoption and usage continue to expand. More conservative structures may offer a stronger downside floor in exchange for limiting upside participation. The objective is not to predict the winning scenario, but to remain invested while reducing the cost of being wrong.
Falling token prices need not undermine AI adoption. But they could reshape where value accrues across the AI ecosystem. For investors, that makes portfolio construction as important as forecasting. The next phase of AI investing may depend less on predicting the best model and more on understanding where the economics of cheaper intelligence ultimately accrue.
1. A neocloud is a specialized, AI-first cloud provider dedicated to renting out high-performance GPU compute for training and running artificial intelligence models.
2. Open-weight models are AI models whose core components are publicly released, allowing anyone to download it. This lets users run the model on their own computers, study how it works, and even modify it for their own specific needs.
3. AI organizations that train and build state-of-the-art models from scratch. Examples include OpenAI and Anthropic.
4. Epoch AI, What Went Into Training DeepSeek-R1?, January 31, 2025.
5. Nvidia, Open Models, Closed Environments: Palantir Brings Secure AI to US Agencies With NVIDIA Nemotron, June 29, 2026.
6. Ramp, with data as of June 1, 2026.
7. Morgan Stanley, Selloff of US memory stocks creates a compelling entry point, July 2026.
Appendix: AI Value Chain Flowchart
Methodology: This exhibit is an illustrative model of how $1 of AI spend may flow through the AI value chain, from foundation model providers through cloud infrastructure, compute, physical infrastructure, and upstream suppliers. The model combines publicly available company filings, earnings releases, management commentary, industry research, trade publications, media reporting, and iCapital analysis to estimate operating profit retained at each layer and dollars routed downstream. Where direct company disclosures were unavailable, figures were derived using third-party research and industry benchmarks. A limited number of allocation assumptions, including server cost shares, facility-branch allocations, merchant versus custom silicon exposure, and hyperscaler versus neocloud compute mix, are reasoned estimates based on the available source material rather than directly reported figures. The exhibit is intended to illustrate the relative distribution of economic value across the AI ecosystem and should not be interpreted as the economics of any single company, a forecast of future profitability, or a prediction of future market-share, valuation, or investment outcomes. Estimates are based on publicly available information available as of July 2026 and may differ materially from actual company results.
Flow Chart Sources:
1. Reuters and Bloomberg, reporting Greg Brockman sworn testimony in Musk v. OpenAI, May 5, 2026.
2. Ed Zitron (Where's Your Ed At), leaked audited financial statements verified by the Financial Times, June 16, 2026.
3. The Information, as reported via Ed Zitron (wheresyoured.at), OpenAI Had a Negative 122% Operating Margin in Q1 2026 and ChatGPT Growth Has Stalled.
4. The Wall Street Journal, Mind-Blowing Growth Is About to Propel Anthropic Into Its First Profitable Quarter, May 20, 2026.
5. Amazon, Q1 2026 earnings release, April 29, 2026; Alphabet, Q2 2026 earnings release, July 22, 2026. [ir.aboutamazon.com], [abc.xyz]
6. SemiAnalysis (Dylan Patel et al.), TPUv7: Google Takes a Swing at the 900lb Gorilla, November 2025.
7. Synergy Research Group (Jeremy Duke), Neocloud Market Forecast to Approach $400B by 2031, Driven by Surging AI Infrastructure Demand, April 2, 2026.
8. Gartner, as reported via CIO Dive (Makenzie Holland), Neocloud Providers Will Gain Greater Slice of AI Cloud Market by 2030, June 23, 2026.
9. The Information, as reported via CNBC, Oracle Stock Slips on Report Company Is Seeing Thin Cloud Margins From Nvidia Chips, October 7, 2025.
10. CoreWeave, Q1 2026 earnings release and Form 8-K, May 7, 2026. [marketbeat.com]
11. Nvidia, Q1 FY2027 earnings release, May 20, 2026; AMD, Q1 2026 earnings release, May 5, 2026. [investor.nvidia.com], [ir.amd.com]
12. Broadcom FY2025 Form 10-K.
13. Tom's Hardware, Amazon's Trainium and Microsoft's Maia custom silicon are developed in partnership with Marvell Technology; Broadcom's confirmed custom-silicon partners are Google (TPU) and Meta (MTIA), May 21, 2026.
14. Moor Insights & Strategy, as reported via tech-insider.org, Google TPU 8t/8i, Broadcom, MediaTek, Nvidia (2026), April 23, 2026.
15. TSMC, Q2 2026 earnings release, July 16, 2026; TSMC FY2025 annual report. [marketbeat.com]
16. SK Hynix, Q1 2026 results, April 22, 2026; Micron, Q3 FY2026 results, June 24, 2026. [news.skhynix.com], [investors.micron.com]
17. Super Micro Computer, Q4 FY2025 Form 8-K, August 5, 2025; Super Micro Computer, Q1 FY2026 Form 8-K, November 4, 2025; Dell Technologies ISG, Q1 FY2027 earnings release, May 28, 2026; Foxconn (Hon Hai), Q1 2026 earnings release, May 14, 2026. [ir.supermicro.com], [last10k.com], [investors....logies.com], [foxconn.com]
18. Vertiv Q2 2026 results; Eaton Q2 2026 results; Arista Networks Q2 2026 results; Schneider Electric FY2025 annual results.
19. Q1 2026 earnings releases and Form 8-K filings: NextEra Energy; Vistra, May 7, 2026; Constellation Energy, May 11, 2026. [investor.v...racorp.com], [constellat...energy.com]
20. Equinix, Q1 2026 earnings release; Digital Realty, Q1 2026 commentary.
21. Quanta Services FY2025/Q1 2026 filings and Morningstar coverage, MasTec, Quanta: Market Is Overestimating Data-Center Grid Tailwinds.
22. Amelia Michael and Ben Cottier, Epoch AI, Servers Account for 60% of the Total Cost of Ownership of a One-Gigawatt AI Data Center, May 2026.
23. SemiAnalysis (Dylan Patel and Daniel Nishball), H100 vs GB200 NVL72 Training Benchmarks: Power, TCO, and Reliability Analysis, Software Improvement Over Time, August 20, 2025.
24. Silicon Analysts, NVIDIA B200 Cost Breakdown: What Blackwell Really Costs to Manufacture, March 2, 2026.
25. Anthropic newsroom announcements, 2025-2026.
26. CNBC, OpenAI Partners With Broadcom on Custom AI Chips Alongside Nvidia, AMD, October 13, 2025.
INDEX DEFINITIONS
Bloomberg Magnificent 7 Total Return Index: The Bloomberg Magnificent 7 Total Return Index is an equal-dollar weighted equity benchmark consisting of a fixed basket of 7 widely-traded companies classified in the United States and representing the Communications, Consumer Discretionary and Technology sectors as defined by Bloomberg Industry Classification System (BICS).
Hyperscaler Index: The Hyperscaler Index is defined as Meta Platforms, Oracle Corporation, Microsoft Corporation, Amazon.com Inc., and Alphabet Inc. Weightings as of July 29, 2026.
KOSPI 200: The KOSPI 200 is a stock market index that tracks the 200 largest and most liquid companies listed on the Korea Exchange (KRX), representing about 70% of the total South Korean stock market value.
NASDAQ 100 Index: An index comprised of equity securities issued by 100 of the largest non-financial companies listed on the Nasdaq stock exchange. It is a modified capitalization-weighted index.
Philadelphia Semiconductor Index: The Philadelphia Stock Exchange Semiconductor Index is a modified market capitalization-weighted index composed of companies primarily involved in the design, distribution, manufacture, and sale of semiconductors.
S&P 500 Index: The S&P 500 is widely regarded as the best single gauge of large-cap U.S. equities. The index includes 500 of the top companies in leading industries of the U.S. economy and covers approximately 80% of available market capitalization.
S&P 500 Equal-Weighted Index: The equal-weight version of the widely used S&P 500. The index includes the same constituents as the capitalization weighted S&P 500, but each company in the S&P 500 EWI is allocated a fixed weight - or 0.2% of the index total at each quarterly rebalance.
Russell 1000 Growth Index: The Russell 1000 Growth Index measures the performance of US large cap growth stocks. The index includes US large cap stocks with relatively higher price-to-book ratios, higher 2-year I/B/E/S forecast growth and higher historical 5-year sales growth. The index is reconstituted fully in June to ensure accurate representation of the US large cap growth style, with updates for parent index membership changes in December and quarterly IPO inclusions in March and September. Since March 24, 2025, the index applies quarterly capping if constituent weights exceed target RIC thresholds. IPOs with investable market cap above the Russell Top 500 breakpoint are eligible for fast entry; breakpoints are set semi-annually and market-adjusted quarterly.
Russell 1000 Value Index: The Russell 1000 Value Index measures the performance of US large cap value stocks. The index includes companies with relatively lower price to-book ratios, lower 2-year I/B/E/S forecast growth and lower historical 5 year sales growth. The index is reconstituted fully in June to ensure accurate representation of the US large cap value style, with updates for parent index membership changes in December and quarterly IPO inclusions in March and September. Since March 24, 2025, the index applies quarterly capping if constituent weights exceed target RIC thresholds. IPOs with investable market cap above the Russell Top 500 breakpoint are eligible for fast entry; breakpoints are set semi-annually and market-adjusted quarterly.
Russell 2000 Index: The Russell 2000 Index measures the performance of the small cap segment of the US equity market. The index includes approximately 2,000 of the smallest companies based on a combination of their market cap and current index membership. The Russell 2000 Index is a subset of the Russell 3000 Index, which was designed to represent approximately 98% of the investable US equity market. Semi-annual reconstitution and quarterly IPO inclusions ensure newly eligible companies are represented and that larger stocks do not distort the performance and characteristics of the true US small cap opportunity set.
IMPORTANT INFORMATION
The material herein has been provided to you for informational purposes only by Institutional Capital Network, Inc. ("iCapital Network") or one of its affiliates (iCapital Network together with its affiliates, "iCapital"). This material is the property of iCapital and may not be shared without the written permission of iCapital. No part of this material may be reproduced in any form, or referred to in any other publication, without express written permission of iCapital.
This material is provided for informational purposes only and is not intended as, and may not be relied on in any manner as, legal, tax or investment advice, a recommendation, or as an offer or solicitation to buy or sell any security, financial product or instrument, or otherwise to participate in any particular trading strategy. This material does not intend to address the financial objectives, situation, or specific needs of any individual investor. You should consult your personal accounting, tax and legal advisors to understand the implications of any investment specific to your personal financial situation.
ALTERNATIVE INVESTMENTS ARE CONSIDERED COMPLEX PRODUCTS AND MAY NOT BE SUITABLE FOR ALL INVESTORS. Prospective investors should be aware that an investment in an alternative investment is speculative and involves a high degree of risk. Alternative Investments often engage in leveraging and other speculative investment practices that may increase the risk of investment loss; can be highly illiquid; may not be required to provide periodic pricing or valuation information to investors; may involve complex tax structures and delays in distributing important tax information; are not subject to the same regulatory requirements as mutual funds; and often charge high fees. There is no guarantee that an alternative investment will implement its investment strategy and/or achieve its objectives, generate profits, or avoid loss. An investment should only be considered by sophisticated investors who can afford to lose all or a substantial amount of their investment
STRUCTURED INVESTMENTS ARE CONSIDERED COMPLEX PRODUCTS AND MAY NOT BE SUITABLE FOR ALL INVESTORS. Please note that there is no public secondary market for structured investments. Although the issuer may from time to time make a market in certain structured investments, the issuer does not have any obligation to do so and market making may be discontinued at any time. Accordingly, an investor must be prepared to hold such investments until maturity. Any or all payments are subject to the creditworthiness of the issuer. Before investing in any product, an investor should review the prospectus or other offering documents, which contain important information, including the product's investment objectives or goals, its strategies for achieving those goals, the principal risks of investing in the product, the product's fees and expenses, and its past performance.
iCapital Markets LLC operates a platform that makes available financial products to financial professionals. In operating this platform, iCapital Markets LLC generally earns revenue based on the volume of transactions that take place in these products and would benefit by an increase in sales for these products.
The information contained herein is an opinion only, as of the date indicated, and should not be relied upon as the only important information available. Any prediction, projection or forecast on the economy, stock market, bond market or the economic trends of the markets is not necessarily indicative of the future or likely performance. The information contained herein is subject to change, incomplete, and may include information and/or data obtained from third party sources that iCapital believes, but does not guarantee, to be accurate. iCapital considers this third-party data reliable, but does not represent that it is accurate, complete and/or up to date. iCapital makes no representation as to the accuracy or completeness of this material and accepts no liability for losses arising from the use of the material presented. No representation or warranty is made by iCapital as to the reasonableness or completeness of such forward-looking statements or to any other financial information contained herein.
Securities products and services are offered by iCapital Markets LLC, an SEC-registered broker-dealer, member FINRA and SIPC, and an affiliate of iCapital, Inc. and Institutional Capital Network, Inc. These registrations and memberships in no way imply that the SEC, FINRA, or SIPC have endorsed any of the entities, products, or services discussed herein. Annuities and insurance services are provided by iCapital Annuities and Insurance Services LLC, an affiliate of iCapital, Inc. "iCapital" and "iCapital Network" are registered trademarks of Institutional Capital Network, Inc. Additional information is available upon request.
© 2026 Institutional Capital Network, Inc. All Rights Reserved.
