Authored by Aaron Filbeck, CAIA, CFA, CFP®, CIPM, FDP, Managing Director, Content & Community Strategy
The title of the book-inspired Netflix show, 3 Body Problem, refers to a physics problem. An advanced alien race, the Trisolarans, live on a planet caught in the gravitational pull of three suns simultaneously. The problem, as the title suggests, is that two bodies in gravitational interaction produce stable, predictable orbital patterns. However, once a third body is added to the equation, those patterns become chaotic and unpredictable. Small changes in initial conditions compound into wildly different outcomes over time. The Trisolarans experience this as alternating “eras” of stable conditions and catastrophic instability. Unfortunately, they never know how long either era will last and are constantly rebuilding their civilization as a result.
The Trisolarans never know whether they’re entering a period of stability or heading toward catastrophe until it’s already becoming obvious. Watching AI develop sometimes feels similar. The direction of transformation is disruptive and massive, but the terminal state is unknown. To foreshadow my thinking a bit, three forces are interacting simultaneously in ways that make the system unpredictable, and it is their interaction that creates chaos.
But first, let’s talk about the investment stack itself. Not all AI investment opportunities are created equally, so we should define a few things first. Figure 1 maps the four layers of the AI investment stack, from the raw compute infrastructure at the base to the application layer at the top. Each layer has a fundamentally different investable thesis, as well as risk and return characteristics.
Reading the stack from bottom to top is also roughly the order of maturity: infrastructure is the most established, and applications are the most nascent. The verdict labels on the right tell the investment story at a glance. A note on scope: compute infrastructure and GPUs — effectively NVIDIA, which holds roughly 92% of the discrete GPU market — sit underneath these four layers as the foundational layer, but with a different investability logic that warrants separate treatment.
Figure 1: The Four Layers of the AI Investment Stack
Source: CAIA Association
Looking at the stack, the instinct for most investors is to ask which layer wins. That is the wrong first question. The better question is how much of the spending flowing through each layer translates into something you can own or lend against at a reasonable risk/return. This differs across each layer.
Layer 1 — Models: Enterprise LLM API spending reached $12.5 billion in 2025 and is growing fast, but the two primary direct beneficiaries — OpenAI and Anthropic — are private companies valued at $852 billion and $380 billion respectively. A meaningful portion of their reported revenue also flows through cloud resellers on a gross basis, making the true capturable economics harder to isolate than the headline figures suggest. Most institutional allocators cannot access this layer directly at any reasonable risk/return…. yet.
Layer 2 — Platform Companies: The hyperscalers are spending $725 billion in AI capex in 2026, but that spending is mostly an outflow and compresses their free cash flow rather than generating a targeted AI return. The revenue from AI features flows into broader cloud and services businesses where it is diffused across a much larger P&L. You can own the equity, but you are not making a targeted AI bet.
Layer 3 — Adjacency Plays: Power, cooling, data center real estate, and networking are technology-agnostic — revenue is tied to the physical buildout regardless of which platform wins. This layer has the cleanest capturable story, particularly in private markets.
Layer 4 — Applications: Enterprise AI application spending reached $19 billion in 2025, and this is where the most durable long-term value may sit, but only for the subset of companies that have built something the platforms cannot replicate. Generic wrappers and feature-thin applications are not investable over a longer period.
With that map in place, let’s walk through what is actually happening at each layer, and what’s investable.
Layer 1 – Investing in the Models
The model layer of the AI stack had one of its most disruptive moments in January 2025 when DeepSeek released R1, a competitive model to OpenAI, built for approximately $500 million and using a fraction of the chips assumed necessary, with an open-source release to add salt in the wound. Fast forward 18 months, and GPT-4 equivalent performance, which cost $20 per million tokens in late 2022, now costs approximately $0.40. Budget-tier models have fallen to $0.075 per million input tokens. While this cost curve is great for the consumer, it has made competition even more intense as different foundational models compete for user attention, effectively commoditizing them.
Indirectly illustrating this point, Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, spent a weekend in late 2025 building a tool that routes any question simultaneously to GPT, Gemini, Claude, and Grok, lets the models critique each other, and has a designated “chairman” model synthesize the final answer. What he found was that none of the answers consistently came from one model, as the application does not know or care which company provides the intelligence. As these models leapfrog each other, we benefit…but the companies don’t.
OpenAI and Anthropic, which reached $24 billion and $30 billion in annualized run-rate revenue respectively by April 2026, will eventually need to grow revenue to justify their valuations. Revenue is growing at incredible growth rates. But profitability is years away, and the gross-to-net gap matters when assessing true economic capture. When a company becomes large enough that OpenAI or Anthropic notices, it also becomes vulnerable enough to be copied. Of course, this is how competition works, but the barriers to entry are very low.
That said, the profitability timeline has compressed dramatically. Anthropic reported $4.8 billion in revenue in Q1 2026 and projects $10.9 billion in Q2 2026 — with its first operating profit of $559 million expected in that quarter, years ahead of earlier guidance. The company itself cautioned that full-year profitability is not assured given planned infrastructure spending commitments. One quarter of operating profit does not resolve the long-run question of whether the model layer can generate durable margin at scale — but the trajectory has materially changed.
The economic moat of this layer is pretty thin. This is different from advancements in ability, which is still happening, but advances no longer translate into durable competitive advantage. The frontier is accessible enough that it generates no pricing power for any single provider. Just like a good company doesn’t make for a good stock, transformative technology and good investment opportunity are not the same thing.
Layer 2 – Investing in the Platform Companies
Meta, Microsoft, Google, and Amazon are collectively planning over $725 billion in AI infrastructure investment in 2026. Their motivations are broader than direct financial return. In fact, they are likely not building because they expect the model layer to be profitable in isolation. Like Apple taking on music streaming and services to compete with Spotify, AI creates a flywheel for their existing businesses. But there is also non-financial logic behind this. These are companies that cannot afford to be strategically outmaneuvered by a competitor, regardless of near-term ROI. The investment is partly a Cold War approach towards defensive positioning, partly internal capability building, and partly ensuring they see everything that gets built on top of their platforms.
This creates a specific problem for AI infrastructure equity investors. If model efficiency continues to improve, the assumption that the buildout will be fully utilized becomes less certain. As of early 2026, capital expenditures are expected to reach anywhere from 45-57% of revenue across the major hyperscalers, compressing free cash flow to keep up.
In other words, the equity story is increasingly reflected in prices and perhaps priced asymmetrically to the downside. As an illustration of this, Joachim Klement, wrote in a recent FT article that the ROI math is fundamentally broken: to justify current capex levels, hyperscalers would need to sustain a pace of revenue growth that is, in his words, “mathematically improbable, if not impossible.” Deutsche Bank has made a similar observation, noting that AI capital expenditures have reached levels where they are single-handedly preventing the U.S. from tipping into recession — a dependency that creates its own systemic risk.
The debt story is slightly different. Corporate bond investors in hyperscaler-backed paper are not underwriting the AI thesis in any speculative sense. As is typical for credit investors, they just want to be repaid. As John Medina, a JPMorgan infrastructure finance executive, put it, most of the investment is supported by companies with very profitable existing lines of business that are not going away.
From a supply standpoint, a lot of debt financing has happened in private markets or through non-traditional vehicles. Despite some of the recent scares around AI, Morgan Stanley estimates private credit markets could supply over half of the $1.5 trillion needed for the data center buildout through 2028.
Like corporate bonds, the counterparty and collateral backing these loans matter quite a lot, and investors should ask whether the debt is pure play, a diversified play, or something else entirely.
To understand what that means in practice, consider a recent deal between Meta and Blue Owl to fund the Hyperion Data center. Meta needed to finance a massive data center expansion but did not want to load up its own balance sheet. So instead of borrowing directly, it created a Special Purpose Vehicle (SPV) that sits between Meta and the lenders. The SPV took out the $27 billion loan, used that money to build the data centers, and then agreed to lease those facilities back to Meta. Meta makes regular lease payments to the SPV, and the SPV uses those payments to repay the debt. The lenders’ counterparty is technically the SPV, but their real credit exposure is Meta.
Layer 3 – The Adjacency Plays
There is one layer of the AI investment stack that does not require you to predict a winner, or perhaps even care. You just need the buildout to continue. That’s not a bet without risk, but given the competitive dynamics among the hyperscalers, where pulling back on spending means ceding ground to a rival, it’s a reasonable one to make.
Power, cooling, data center real estate, and networking are some examples of adjacency plays. They are technology-agnostic by design, and their revenue is tied to the physical construction of AI infrastructure, no matter which company sits on top of it.
In 2025, 113 data center transactions closed globally, representing more than $69 billion in total deal value, approximately $8 billion above the prior annual record. That same year included the $40 billion acquisition of Aligned Data Centers by a consortium including BlackRock and MGX, which was the largest digital infrastructure transaction of its kind. As an aside, 84% of that deal value was private equity funded. There are numerous other examples of massive private capital commitment in the adjacency plays as well, including Blackstone, KKR, and Brookfield.
Data center investing has effectively converged with energy infrastructure investing. Power availability is now the top opportunity for many GPs. Electric and gas utilities are forecasting a record capital expenditure increase of 22% year over year to $212 billion in 2025, a sharp rise from the longer term 7.6% CAGR over the prior decade.
The Caution
Goldman Sachs notes that equity gains have already been concentrated in AI infrastructure companies, with their infrastructure basket returning approximately 44% year to date as of their most recent report. Their forward view is that the next phase of the AI trade shifts toward platform stocks and productivity beneficiaries, implying that the adjacency trade has partially run in public markets. The entry point for public infrastructure equity may be less attractive than it was 18 months ago. Additionally, private credit has seen its fair share of AI-focused loan origination, which may mean we’re in for some credit pain if defaults continue to rise in SaaS and adjacent sectors.
The more interesting opportunity today may be in private credit and direct infrastructure investment, where the risk/return profile relative to public market equivalents remains more attractive. However, the risk over the long term is stranded asset risk. The more computing power evolves, and companies go in and out of business, the more likely we’re to see the value proposition of physical assets change or be abandoned.
So, the opportunity is great for investors with infrastructure underwriting capabilities who understand how to evaluate power contracts, counterparty quality, and construction timelines.
Layer 4 – The Application
The hyperscalers can see every profitable application built on their platforms. When a company proves itself in a positive way, be it user retention, monetization, or switching costs, the platform provider faces a straightforward decision: acquire it or replicate it.
The question for investors is not which application layer companies have built something the hyperscalers can observe but cannot replicate. That requires a specific kind of structural advantage embedded in data, relationships, and domain expertise that cannot be reconstructed from outside the organization.
Unfortunately, the internet has been scraped, and the frontier models have seen essentially all of it. But a hospital’s clinical notes, a law firm’s case files, and a financial institution’s transaction records are not part of that model. The moat is the cost of replication, not the secrecy of the method.
Additionally, applications that actually do the work, sometimes referred to as agentic AI, represent another competitive advantage. Traditional SaaS software hit a ceiling at tasks that were multi-modal, language-heavy, or highly specialized because they were too complex for rule-based software but too variable for rigid workflows. Vertical AI is the first technology that can execute those tasks, not just organize them. Healthcare, legal, and housing companies have reached $100 million in annual recurring revenue within a few years, faster than any prior software category.
Enterprise AI application spending reached $19 billion in 2025, more than half of all enterprise generative AI spending, growing at 5x year-over-year. The companies leading that growth are those that have become the “system of record” through accumulated outcome data that general models cannot replicate.
The application layer is not a panacea, however. There are some real challenges that will lead to dispersion of outcomes. Research by Kai Wu at Sparkline Capital, published in May 2026, provides quantitative support for this framework. Studying over 30 years of disruption cycles, from e-commerce to cloud to AI, Wu finds that traditional value metrics (price-to-earnings, price-to-book) have consistently failed when applied to stocks facing technological disruption, actually working in reverse by systematically buying value traps and selling disruptive winners.
The factor that did work was “intangible value,” which incorporates IP, brand equity, human capital, and network effects alongside traditional assets. For your portfolio, this means the selection question at the application layer is not which software stocks look cheap on traditional metrics, many do, after a 30%+ selloff over the past year, but which ones possess the complementary intangible moats that allow them to survive and profit from disruption rather than become its casualties.
- The cost problem. Many AI application companies are paying more in model API fees; the per-query charges to use someone else’s model, than customers are paying in subscriptions. You charge $50 a month; it costs $70 a month to serve that customer. Volume makes it worse, not better.
- The platform problem. Just like the hyperscalers, any model provider can see every API call. They know exactly which applications are gaining traction. They face a straightforward decision: build the same feature themselves or acquire at a price that reflects your vulnerability rather than your potential.
- The user problem. Most application layer companies charge by the seat per user per month. But if AI agents start doing the work that humans used to do, you have fewer human users. Fewer users mean fewer seats means less revenue, even if the product is working exactly as intended. IDC predicts that by 2028, pure seat-based pricing will be obsolete, with 70% of software vendors refactoring their pricing strategies.
The 3 Body Problem and Externalities
What made the Trisolarans’ world so unpredictable wasn’t any one sun. It was the fact that three powerful forces were acting on the system simultaneously. If we get more meta about this analogy, three gravitational and interdependent forces are colliding simultaneously in AI, and it is their interaction, not any one of them in isolation, that creates the chaotic system.
- Power and physical constraints. The primary bottleneck in AI has shifted from chips to electricity. Global data center consumption is projected to exceed 1,000 terawatt-hours by 2026, nearly double the 460 TWh consumed in 2022. New projects requesting more than 100 megawatts are facing connection delays of five years or more. Fed Chair Jerome Powell noted in early 2026 that data center construction is pushing inflation higher, which means power constraints are becoming a monetary policy issue.
- Regulation. The global regulatory picture is far more fragmented than the US/EU binary most commentary focuses on. The US and EU are moving in opposite directions: the Trump administration’s July 2025 AI Action Plan frames AI as a space race and calls for reviewing all prior FTC and DOJ enforcement actions, while the EU AI Act’s high-risk system requirements become applicable in August 2026. But the more instructive contrast may be outside the West. China amended its Cybersecurity Law effective January 1, 2026 to bring AI explicitly into national legislation for the first time, while simultaneously announcing comprehensive AI legislation as a 2026 National People’s Congress priority, moving from soft-law guidance toward hard-law obligations at an accelerating pace. The UAE took the opposite posture: in January 2026 it became the first country in the world to formally adopt a National AI System as an advisory member of Cabinet, embedding AI directly into federal executive decision-making as a strategic asset rather than a risk to be managed. Singapore, meanwhile, unveiled the world’s first governance framework specifically for agentic AI at the World Economic Forum in January 2026, establishing global standards that others are already adopting. For application layer companies building across jurisdictions, the regulatory patchwork is an operational reality.
- Labor transition. The productivity gains from AI seem to be occurring, though that depends on who you ask. According to EY, 96% of organizations investing in AI report gains, and 57% describe them as significant. But if margin expansion primarily accrues to capital rather than reaching consumers or workers, the political winds of the next regulatory cycle could change.
All three forces interact in ways that make the system unpredictable. Which comes first, and how does one influence the other? And what are the knock-on effects?
Power constraints are already affecting regulatory conversation. Regulation affects the pace of deployment, which affects the pace of labor displacement, which affects the social license for continued buildout. Labor transition shapes how aggressively governments intervene, which reshapes the regulatory landscape that application layer companies are trying to build moats within. The forces pull on each other and are often non-linear.
A fourth force that’s emerging is index concentration in public markets.
Over 78% of U.S. market capitalization is now in companies facing AI disruption. As AI winners concentrate into cap-weighted indices, broad-based exposure to equity markets makes AI hard to avoid on the upside or downside. If major indices become dominated by a handful of AI winners and losers, the political and regulatory response could reshape the landscape in ways that are difficult to model today.
To this end, the most actionable near-term opportunity may not be investing in AI companies at all but rather identifying companies across sectors that are using AI to structurally widen their margins or competitive position. Looking for companies where AI adoption is already showing up in margin expansion and competitive moat widening is more broadly accessible than any of the four layers described above, and does not require navigating private markets or timing a venture cycle. Similarly, applications and adjacency plays where moats are wider but harder to identify may be a good investment thesis.
3-Body Problem ends its first season with humanity less certain than when it started. A lot of plans were made, but humanity realizes this is far more complicated than anticipated.
That sounds about right for AI investors in 2026.
Resources:
Kai Wu, Sparkline Capital, “AI Disruption: Moats and Value Traps,” May 2026. https://www.sparklinecapital.com/post/ai-disruption
Jon Peddie Research, “GPU Market Share Q3 2025,” cited in TechSpot, “Nvidia dominates discrete GPU market with 92% share despite shifting focus to AI,” December 2, 2025. https://www.techspot.com/news/110464-nvidia-dominates-discrete-gpu-market-92-share-despite.html
Menlo Ventures, “2025: The State of Generative AI in the Enterprise,” December 9, 2025. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
Philipp Dubach, “AI Model Commoditization: $400B in Capex, 97% Price Drop,” philippdubach.com, updated March 15, 2026. https://philippdubach.com/ai-model-commoditization
VentureBeat, “A weekend vibe code hack by Andrej Karpathy quietly sketches the missing layer of enterprise AI orchestration,” December 22, 2025. https://venturebeat.com/ai/a-weekend-vibe-code-hack-by-andrej-karpathy
Sacra Anthropic revenue tracker, April 2026; SaaStr analysis, April 2026. https://sacra.com/research/anthropic-revenue
WSJ confidential financials reported April 2026; PM Insights Anthropic revenue tracker.
CNBC, “Anthropic set to hit $10.9 billion in revenue during second quarter,” May 20, 2026. https://www.cnbc.com/2026/05/20/anthropic-revenue-explosive-growth-ipo-…
Startup Fortune, “Anthropic’s path to profit is coming into view,” May 2026. https://startupfortune.com/anthropics-path-to-profit-is-coming-into-vie…
Google, Amazon, Microsoft, Meta Q1 2026 earnings guidance compiled by Financial Times / Tom’s Hardware, May 2026.
Invezz / TradingView, “Looking ahead to 2026: why hyperscalers can’t slow spending without losing the AI war,” December 26, 2025. https://www.tradingview.com/news/invezz:751717ae0094b:0-looking-ahead-t…
Joachim Klement, “The impossible maths of the AI boom,” Financial Times, May 20, 2026. Summary at: https://klementoninvesting.substack.com/p/the-impossible-maths-of-the-a…
Deutsche Bank Research, cited in “Is the AI boom a house of cards?” MEXC / BitcoinEthereumNews, May 2026. https://www.mexc.com/news/112616
John Medina, SVP, JPMorgan Global Project and Infrastructure Finance, quoted in Bloomberg / Insurance Journal, “The $3 Trillion AI Data Center Build-Out Becomes All-Consuming for Debt Markets,” February 2, 2026.
Morgan Stanley, “Energy Markets Race to Solve the AI Power Bottleneck,” February 27, 2026. https://www.morganstanley.com/ideas/ai-energy-demand
TradingKey / Petar Petrov, “Private Credit Funding for AI Data Centers,” January 2, 2026. https://tradingkey.com/blogs/insights/private-credit-funding-for-ai-dat…
S&P Global Market Intelligence, cited in Data Center Dynamics, “S&P Global: Data center M&A topped $69bn in 2025, neoclouds in unenviable position,” March 11, 2026. https://www.datacenterdynamics.com/en/news/sp-global-data-center-ma-top…
John Dinsdale, Chief Analyst, Synergy Research Group, quoted in Data Center Knowledge, “Data Center M&A Outlook Robust in 2026 Despite Power, AI Risks,” February 27, 2026. https://www.datacenterknowledge.com/investing/data-center-m-a-outlook-r…
Deloitte, “AI Infrastructure Survey 2025”; Summit / NYU Stern AFIRE Data Center Report, November 2025. Utility capex figure from EEI / Deloitte industry analysis.
Goldman Sachs, “Why AI Companies May Invest More Than $500 Billion in 2026,” December 2025. https://www.goldmansachs.com/insights/articles/why-ai-companies-may-inv…
CAIA Association. “Private Credit Redemptions, Defaults, and Wrappers, Oh My!” April 20, 2026.
Menlo Ventures, “Software Finally Gets to Work: The Opportunity in Vertical AI,” April 2026. https://menlovc.com/2026-the-state-of-generative-ai-in-the-enterprise
Bessemer Venture Partners, “The State of AI 2025,” August 2025. https://www.bvp.com/atlas/state-of-ai-2025
Menlo Ventures, “2025: The State of Generative AI in the Enterprise,” December 9, 2025. https://menlovc.com/2025-the-state-of-generative-ai-in-the-enterprise
IDC, “FutureScape: Worldwide Agentic AI 2026 Predictions,” October 23, 2025. Cited in IDC press release via HPCwire: https://www.hpcwire.com/off-the-wire/idc-futurescape-2026-predictions-reveal-the-rise-of-agentic-ai-and-a-turning-point-in-enterprise-transformation/
Brookings Institution, “Global energy demands within the AI regulatory landscape,” April 2, 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-…; IEA projections 2026.
Federal Reserve, FOMC Press Conference Transcript, March 18, 2026. https://www.federalreserve.gov/mediacenter/files/FOMCpresconf20260318.pdf
Greenberg Traurig, “AI Antitrust Landscape 2025: Federal Policy, Algorithm Cases, and Regulatory Scrutiny,” September 5, 2025. https://www.gtlaw.com/en/insights/2025/9/ai-antitrust-landscape-2025
China Cybersecurity Law amended October 2025, effective January 1, 2026. White & Case, “AI Watch: Global Regulatory Tracker – China,” September 2025. https://www.whitecase.com/insight-our-thinking/ai-watch-global-regulato….
China’s National People’s Congress announced AI as a 2026 legislative research priority: FreshFromChina, March 9, 2026. https://freshfromchina.com/chinas-2026-ai-legislative-roadmap
UAE National AI System adopted as advisory member of Cabinet, January 2026. Bird & Bird AI Regulatory Horizon Tracker – UAE. https://www.twobirds.com/en/capabilities/artificial-intelligence/ai-leg….
UAE Regulatory Intelligence Office established April 2025: CMS Law, “Middle East TMT – Key Legal Developments in 2025 and 2026,” January 2026.
Singapore IMDA, Model AI Governance Framework for Agentic AI, unveiled at World Economic Forum, January 22, 2026. https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeche….
EY, “AI-driven productivity is fueling reinvestment over workforce reductions,” EY US AI Pulse Survey, December 9, 2025. https://www.ey.com/en_us/newsroom/2025/12/ey-survey-finds-ai-investment…
Photo Credit | iStockphoto: ByM
