Citrini is restructuring its model portfolio, retaining most core themes while adding reversals, pair trades, and investment directions overlooked by the market.

In 2023, the core question for AI investing was: how much did the market underestimate the upcoming spending? Now, Citrini believes the question has become: how much does revenue need to grow to support the investment already committed, and who is willing to continue financing that investment?

This is also the starting point for restructuring its model portfolio Citrindex in the tenth issue of "State of the Themes." AI is still creating real demand, but between technological progress, corporate revenue growth, and stock returns, there are multiple tests of valuation, financing, and execution capability. The approach of relying on identifying structural trends and buying a basket of beneficiary companies over the past few years is facing a higher threshold.

From underestimating investment scale to overdrawing growth expectations

When Citrini launched its thematic research in 2023, the market had several favorable conditions: stock valuations were low, investment sentiment was depressed, and a technological transformation had just begun to be incorporated into earnings expectations.

After ChatGPT launched in late 2022, investors began to believe that generative AI's commercial applications could extend beyond writing poetry and emails. Nvidia's strong results subsequently provided clearer evidence for computing power investment. Even though the business models for end applications are not yet fully clear, chip procurement and infrastructure construction have become visible demand.

At that time, high-quality tech companies that had experienced significant drawdowns had relatively reasonable valuations, while the market was still underestimating future investment scale. Tech earnings recovery, falling inflation, and expectations of a monetary policy pivot provided further support for thematic investing. Since then, the flow of fiscal spending, the regulatory environment, and market risk appetite have pushed capital to spread into defense, energy, robotics, and healthcare, among other areas.

According to Citrini, since its inception on May 31, 2023, Citrindex's cumulative total return has exceeded the S&P 500's cumulative total return over the same period by more than 2.5 times. But over the past 12 months, thematic investing has become noticeably more difficult.

As potential growth is continuously priced into stocks, the market's search for new targets has gradually shifted from companies with proven competitiveness and reasonable valuations to companies that rely on 2030 total addressable market size to support valuations and are still unprofitable today. Total addressable market (TAM) describes the scale an industry could potentially reach; it does not directly indicate how much profit any single company will ultimately capture.

Undervalued opportunities still exist, and Citrini points to the summer inflection point in the life sciences sector as an example. But in some popular themes, the pressure to find new explanations for gains has begun to outpace the speed at which fundamentals can provide new evidence.

Changes in AI revenue are especially representative. Initially, billions of dollars in incremental revenue meant upside that the market had not fully understood; after the industry committed to investing trillions of dollars, strong revenue growth gradually became a basic condition for justifying existing valuations and spending. On this basis, Citrini judges that the risk-reward structure has changed, and downside risk deserves greater attention.

Who will fill the gap between capex and revenue?

By Citrini's estimate, hyperscalers' annualized capital expenditure is about to reach $1 trillion, a measure that still does not cover spending across the entire AI ecosystem. At the same time, the AI ecosystem's annualized revenue scale is expected to reach at least $200 billion, already far exceeding most people's expectations in 2023.

The statistical scope and economic meaning of these two figures are not the same: assets formed by capital expenditure can serve for many years, and annualized revenue is not equivalent to cash flow available for investment. But in Citrini's analysis, they reveal the same constraint: sustaining the current pace of investment expansion requires revenue and financing capacity to keep up.

There are two main paths to narrow the gap: revenue rising by another order of magnitude, or capital expenditure growth slowing. Citrini believes both could happen simultaneously, but the trillion-dollar-scale revenue required to continue doubling investment is difficult to achieve at the pace the market previously expected.

Therefore, its core judgment is that the growth rate of capital expenditure may have peaked. This judgment concerns growth rates and expectations; it does not mean total capital expenditure will immediately decline, nor does it mean AI demand has stopped expanding.

Citrini points out that the most intensive phase of massive investment commitments has passed, and quarterly growth in remaining performance obligations (RPO) has also slowed markedly. RPO refers to contracted amounts yet to be recognized as revenue, an indicator for observing future business, but not current-period revenue or cash receipts.

In the past, announcing $500 billion in investment commitments was enough to boost the market. Now, investors' sensitivity to commitments is declining. The market often invokes the Jevons Paradox, arguing that efficiency gains and lower unit costs will drive total usage higher, but increased total demand does not automatically solve every company's financing and return problems. Changes in model companies' commitments to additional compute capacity also deserve to be observed alongside usage data.

The gap between revenue and investment is not a new problem. Citrini reviewed Sequoia's earlier $200 billion and $600 billion questions: as the industry expanded, those figures now seem small. But this skepticism has never truly interrupted the rally, because as long as the market still has confidence, momentum, and capital, construction can continue.

Now, the capital condition is becoming more important. Citrini believes that some companies' ability to support new projects with internal cash flow can no longer match their investment plans, increasing the importance of external bond and equity financing. Each financing round requires new buyers to absorb it, and equity financing dilutes existing shareholders' interests. As financing scale continues to expand, the returns investors demand may also rise.

Once willingness to finance declines, projects become harder to fund on original terms; rising financing costs in turn lower projects' expected returns. Strong long-term industry demand alone is no longer sufficient to fully answer this question.

The bottleneck shifts from a price-hike theme to project risk

In the early stage of AI buildout, shortages of power, equipment, and supply chains were mainly understood by the market as suppliers' pricing power and order opportunities. Now, the same shortages are beginning to expose another side: they slow construction, delay cash flow, and amplify mismatches between debt maturities and project progress.

Citrini summarizes these constraints as power, interconnection, and policy. Grid interconnection, construction labor, equipment delivery, and various permits can all delay data center commissioning. Even if demand is real, obstacles at the construction execution level are enough to slow the pace of spending.

If gas turbine orders cannot be delivered for a long time and input costs continue to rise, investors will have to compare the continuously delayed completion dates with the already determined debt maturity dates. Whether construction-period loans can be refinanced with long-term financing and whether projects can be completed in accordance with financing agreements will both become specific credit risks.

These bottlenecks may also provide companies with reasons to adjust investment plans: by delaying projects or revising delivery expectations, they can gradually scale back overly aggressive expansion arrangements.

Policy further affects costs and timing. Citrini believes that gubernatorial elections in key data center states such as Texas, Ohio, and Georgia reflect local disputes over construction expansion. It expects that new administrations may not fully halt data centers, but local community compensation, clean energy requirements, consumer electricity bill protections, and permitting approvals could all raise project costs.

The materials also cite a Texas case, stating that an Anthropic-backed data center project stalled due to construction suspension measures, making it difficult for banks to transfer out the $15 billion bridge loan they had underwritten as originally planned. For banks, exposure that cannot be distributed or replaced will directly reduce their willingness to continue providing similar financing.

On this basis, Citrini is more cautious about financing chains lacking credit support from large cloud providers, including emerging GPU cloud service providers (Neoclouds), high-yield private credit, and business development companies (BDCs) that provide financing to enterprises. Its concern is that tighter financing conditions will first affect participants that rely more heavily on external capital and have weaker buffers.

Usage growth must also withstand the test of price competition

Beyond funding constraints, the revenue quality and pricing power of model companies are equally important.

Citrini's assessment of Anthropic is not pessimistic. The materials state that its financial condition as presented in its prospectus is better than some market expectations: the company has retained substantial cash, let partners bear a considerable portion of capital-intensive construction risk, and achieved positive cash flow relatively early.

But model services still face price competition. When product capabilities are similar, customer migration costs are low, and suppliers can even be switched instantly through automation, price changes can quickly alter revenue and market share.

The materials cite OpenAI's reduction of model token prices on July 30 as an example, arguing that the price cut helped it compete for share while dealing a blow to Anthropic's previous rapid growth. A token is the basic unit of measurement for model processing and text generation, and also a commonly used billing unit for model services. Here it refers to model invocation pricing and has nothing to do with crypto tokens.

This mechanism has a direct impact on capital expenditure returns: increased call volume does not guarantee revenue growth at the same proportion. As long as unit prices fall fast enough, the benefits of usage growth may be partially offset, and profit margins will also come under pressure.

Citrini believes that token prices had temporarily stabilized at the time, Anthropic's Opus 5.5 was also performing well, and both major model companies may possess considerable long-term value. However, to continue supporting high-intensity capital expenditure growth on this basis, greater revenue realization capability is still needed.

When trillions of dollars in investment and asset values are built on the assumption that a handful of companies will grow revenue tenfold in the coming years, the standard of scrutiny should naturally be raised.

At the same time, AI's impact on traditional business models is also beginning to become more concrete. Citrini believes that some of the disruptive effects in software-as-a-service (SaaS) and search have already gradually shown up in operating data. The new revenue created by AI needs to be examined alongside the revenue and profits that incumbent enterprises may lose.

Therefore, the most noteworthy variables to watch next are whether model companies' annualized revenue growth rate next year will be higher or lower than this year, whether data center deliveries in 2027 to 2028 will accelerate or be delayed, and whether financial markets' willingness to provide new financing will strengthen or weaken.

AI trade cools, economy may still expand

Peaking expectations for AI infrastructure trades does not mean that the stock market and the economy will immediately enter a broad downturn.

Citrini prefers to understand the current stage as a transition period: AI usage, related spending, and application penetration continue to rise, but the enterprises that captured most of the investment returns in the past may not still be able to deliver returns in the same way. Performance across different segments will diverge more clearly.

Its observations on the real economy still include signs of broad expansion. The material notes that overall job vacancies have rebounded, cyclical manufacturing employment has improved, overall readings from services and manufacturing surveys are at multi-year highs, and price pressures are also rising.

AI investment is one of the driving factors. In addition, tax cuts, easing tariff pressure, and production reshoring may also provide support. Citrini acknowledges that a genuine collapse in AI investment would deal a major blow to the economy, but there is still a large gap between slower growth and a full-scale collapse.

Labor supply especially needs to be observed separately. The material cites estimates from Federal Reserve researchers saying that the monthly new jobs needed to maintain labor market balance fell from 155,000 during the immigration surge in 2023 to 2024 to 85,000 in 2025; as net immigration turns negative, it may further decline to 5,000 to 10,000 by the end of 2026.

Based on the data presented in the material, nonfarm payrolls have averaged approximately 51,000 per month over the past three months and about 68,000 per month year-to-date. If labor supply growth has already slowed markedly, this pace of job gains may still be sufficient to absorb new entrants into the labor force. Weaker employment growth than in past cycles does not necessarily mean wage pressures have simultaneously disappeared.

Existing data center projects will take years to complete. Storage and power supply can be gradually improved by expanding capacity, but construction workers such as skilled electricians cannot be trained quickly. As long as there is no full-scale collapse, labor bottlenecks may persist and continue to support wages and construction costs.

The market can preemptively压低 stock valuations due to lowered storage price expectations for 2029, yet a real halt in data center construction sites may still be far off. Financial assets trade future changes in advance, while physical investment unfolds on a slower timeline.

This also creates a more complex valuation environment. Citrini believes the economy is exhibiting characteristics of a late-cycle expansion; among the 14 asset pairs tracked by its macro model, 9 carry stagflationary tendencies—that is, the market is beginning to worry about relatively weakening growth coexisting with inflationary pressures.

Among them, cyclical sectors such as consumer discretionary, industrials, and materials have fallen 2.4%, 2.9%, and 6.9% over the past one, three, and six months, respectively, relative to defensive sectors such as consumer staples, utilities, and healthcare. Although growth and inflation data still point to an expansion accompanied by inflation, asset price performance and the 10-year real yield have already pushed the model's judgment toward a more stagflationary direction. The real yield is the yield after subtracting inflation expectations, and its rise typically raises the discount threshold for forward earnings.

But the rotation into defensive sectors has lasted for several months, and valuations of some consumer staples stocks have already risen markedly. At this point, continuing to chase defensive assets also requires considering price; meanwhile, AI and momentum stocks may also see periodic rebounds after adjustments.

For portfolios, the difficulty is that high-beta assets need stronger growth and financing conditions to sustain further gains, while low-growth defensive assets may already be very expensive. Citrini therefore turns to more granular stock selection, reversal, and pair trades. Pair trades typically express the performance divergence between two assets by simultaneously holding the relatively favored asset and shorting the relatively weaker one.

Its plan retains most core themes while changing the manner of positioning and adding new directions overlooked by the market. This adjustment targets cyclical changes that have accumulated over several years, as well as the judgment that subjective observations and quantitative signals are gradually converging. High-beta stocks are more sensitive to market swings, amplifying gains during upswings and amplifying losses when risk appetite declines. In the next phase, judging the trend will remain important, but entry price, funding sources, and the path to realization will also more directly determine returns.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia