CalPERS AI Investments: 🔥 Magnificent Seven Pullback

A portfolio adjustment is not a blanket rejection

CalPERS Deputy Chief Investment Officer Anton Orlich told the Financial Times that public equity valuations looked expensive and that technology giants could underperform relative to other investments. The fund is also increasing exposure to private equity, aiming to spread investments across a wider set of companies and reduce reliance on a small group of listed leaders.

This is a portfolio decision made under long-term constraints, not a statement that AI lacks commercial value. Pension funds have to manage obligations over time, diversify risk, and keep enough liquidity. They may cut a concentrated holding if the valuation looks unattractive versus alternatives, even if they still expect the underlying businesses to matter.

It's also worth remembering the Magnificent Seven aren't one uniform bet. They differ in business models, revenue sources, capital needs, and where they sit in the AI supply chain. Cutting across the group doesn't tell you which company CalPERS thinks is most exposed, or whether the decision is driven by company-specific analysis, a broader allocation target, or both.

What the reported number cannot tell you

The 26% share-count figure is a useful data point, but it has limits. It doesn't prove that CalPERS sold 26% of the market value of its overall technology exposure, that it expects an imminent crash, or that it has fully exited AI-related investments. It also doesn't reveal the fund's full reasoning or the details of every position.

For large institutions, the distinction between direct holdings and total exposure is especially important. A fund can reduce the shares it owns directly while still maintaining exposure elsewhere. Without the full portfolio picture and investment mandates, the safest reading is straightforward: CalPERS reduced its reported direct share count in these seven companies and is increasing diversification via private equity.

Spending is visible; the payoff is harder to measure

Capex is easier to spot in company reporting. The payoff is harder. Revenue might come from cloud services, subscriptions, advertising, or AI features added to existing products. It can also be difficult to separate genuinely new AI-driven demand from activity that would have happened anyway.

That creates a timing problem for investors. A company can show rapid growth in infrastructure spending before it can prove that the investment is generating comparable incremental profits. That doesn't automatically mean the strategy is wrong. It does mean valuation depends more heavily on assumptions about future cash flows.

The Roundhill Magnificent Seven ETF was reported as up 7% year to date at the time of the report, with Apple up 34% and Nvidia up 12%, while Meta, Alphabet, and Tesla were down for the year. Those mixed results are a reminder that treating the group as one consistently rising trade can be misleading. And past performance still can't tell you whether current valuations are reasonable.

Expensive does not automatically mean overvalued

Saying a market is expensive isn't the same as predicting an imminent drop. High valuations can be supported by expected growth, strong competitive positions, and future profit margins. The real question is what has to go right - and what happens if growth, margins, or customer demand fall short.

A practical way to break it down is to test three areas:

• Business quality: Are there durable customers, a defensible position, and a credible path to profit?
• Investment burden: How much capital does it need to keep its advantage or meet projected demand?
• Return evidence: Are AI-related products bringing incremental revenue or improving unit economics, instead of just increasing usage and infrastructure costs?

These tests won't look the same across the seven companies. A software platform with a big existing customer base will have different economics than a company whose near-term prospects depend heavily on selling compute capacity. "AI stock" is too broad to replace that work.

Why financing matters so much

Building data centers and acquiring GPUs takes major capital. If companies borrow to fund expansion, lenders and investors need confidence that future cash flows can cover debt and ongoing operating costs. When financing gets more expensive - or harder to obtain - companies with limited revenue outside their expansion plans may struggle to keep scaling at the same pace.

The report also said Musk's SpaceX raised $40 billion in debt to buy GPUs for AI calculations. That gives a sense of the capital involved, but it isn't proof that every company can secure financing on similar terms. Access to capital depends on the borrower, the assets behind it, contracts, cash flow, and how likely lenders believe repayment will be.

Financing squeezes can turn into feedback loops. Companies slow investment, which reduces near-term demand for hardware and facilities. If investors then lower growth expectations, the cost of capital can rise further. Strong operators with contracted customers and manageable debt might keep investing, while weaker businesses - especially those concentrated in a narrow customer base - could be forced to sell assets, merge, or exit.

A stock decline is a warning, not a diagnosis

Sell-offs can come from many things: changes in financing expectations, broader market volatility, company-specific news, or simply profit-taking after a strong run. A drop in price doesn't, by itself, prove that AI demand has collapsed or that infrastructure will go unused.

The better questions are operational. How much revenue is contracted versus merely anticipated? How concentrated is the customer base? What portion of costs is fixed? How much debt is due soon? If demand for one AI workload slows, can the business shift to other workloads? The answers help separate a temporary market reaction from a structural weakness.

Use a four-part AI investment test

Before treating an AI-related company or project as attractive, check four areas that connect to each other:

1. Demand quality: Is demand coming from paying customers, signed commitments, and repeat usage - or is it mostly forecasts and pilots?
2. Economics: Does the use case reduce costs, increase revenue, or improve service enough to cover compute, integration, and ongoing operating expenses?
3. Funding resilience: Can the company or project handle slower adoption, higher borrowing costs, or delays in customer payments?
4. Flexibility: Can the infrastructure, contracts, and skills be reused if the original workload or supplier becomes less attractive?

This framework isn't a crystal ball. It's a way to surface assumptions early - before they become expensive. For example, a business evaluating a model deployment should estimate the full cost per completed task, including review and integration, rather than comparing model prices alone. A data-center operator should weigh customer concentration and financing maturities alongside expected capacity utilization.

Separate the investment horizon from the operating horizon

Public-market investors may focus on quarterly results, relative valuations, and portfolio risk. A business choosing an AI system needs to think in terms of multi-year operating commitments, migration costs, and dependence on a vendor. The same headlines won't answer both questions.

A company doesn't have to guess which technology stock will outperform to make a sound investment. It does need to understand whether a specific use case produces measurable value - and whether it can change course if the economics deteriorate. Pilot results should include a baseline, a defined success threshold, and a plan for deployment or shutdown. Without that, experimentation can quietly turn into an open-ended cost.