CITIC Securities Says China’s AI Trade Is Rotating Toward Energy and Chemicals

CITIC Securities has shifted the center of its China equity strategy after AI delivered much of its expected impact during the second quarter. The brokerage now expects energy and chemical companies to receive greater investor attention as rapid sector rotation tests expensive technology holdings.

The argument is not that demand for AI computing has peaked. CITIC says recent AI advances have strengthened the existing case for continued growth in computing demand. However, that trend alone has not resolved questions about long-term commercialization or justified unlimited valuation expansion.

That distinction creates the central conflict. China’s AI supply chain still offers growth, but investors increasingly want earnings evidence rather than another broad technology narrative. Energy and chemical companies offer a contrasting source of returns through lower valuations, improving profits, and tighter supply conditions.

The strategy therefore resembles a barbell. AI occupies the growth-oriented side, while energy and chemicals provide earnings support and valuation protection. It is a response to a market where momentum has become less reliable and durable leadership remains scarce.

CITIC’s AI and Energy Barbell Reflects a Faster Market

The allocation change is a response to unusually rapid sector rotation, not a rejection of China’s AI investment cycle.

A market bulletin published on August 30 summarized CITIC Securities’ latest strategy research. The brokerage said rapid structural rotation has become normal in China’s A-share market since 2022.

According to the report, periods of extremely fast rotation usually persist for only one or two months. Their common feature is narrow earnings revision breadth, meaning analysts are raising forecasts for too few industries to support a lasting market leader.

That backdrop changes how investors interpret a strong AI quarter. A concentrated earnings contribution can validate specific hardware suppliers without supporting every company carrying an AI label. It can also attract short-term capital that exits quickly when another sector produces stronger revisions.

CITIC describes AI as the aggressive side of its barbell. It appeals to investors seeking high growth, new computing architectures, and exposure to rising infrastructure demand. Energy and chemicals form the steadier side because their investment cases depend more directly on profits, supply discipline, and valuation.

A barbell portfolio does not require both sides to perform simultaneously. It combines assets driven by different expectations, reducing dependence on one continuous market trend. In this case, AI depends heavily on future growth, while energy and chemicals rely more on current operating evidence.

The distinction matters because investors often treat “AI demand” as one trade. In reality, the supply chain includes chips, optical networking, servers, power systems, cooling equipment, software, and data-center operators. Their revenue timing, margins, and capital requirements differ sharply.

Some suppliers recognize revenue when hardware ships. Data-center operators must first finance construction and secure power before workloads generate recurring income. Application developers face a separate test involving customer retention, inference costs, and willingness to pay.

CITIC’s thesis accepts the infrastructure demand signal but questions how far it can carry valuations without broader commercialization. The brokerage says recent AI progress reinforces computing growth, yet does not change the distant revenue narrative by itself.

The second side of the barbell is moving on firmer near-term data. China’s National Bureau of Statistics reported that profits at large industrial companies grew 18.7 percent during the first half of 2026.

The agency’s more recent industrial profit data showed an even sharper sector split through July. Computer and communications equipment profits rose 110 percent from a year earlier.

Profits from raw chemical materials and chemical products increased 56.6 percent. Coal mining profits rose 50.4 percent, while petroleum and natural gas extraction gained 15.4 percent. Those figures support both ends of CITIC’s allocation framework.

The data also show why the strategy is more nuanced than a simple move out of technology. Electronics profits remained exceptionally strong. The change concerns where additional valuation upside can emerge after investors have already recognized that strength.

CITIC is effectively separating earnings confirmation from multiple expansion. AI companies can continue posting growth while their shares become less responsive to familiar demand news. Lower-valued industrial businesses can rise as improving earnings remove reasons for discounted valuations.

That is the first reversal inside the strategy. Better AI fundamentals do not automatically make AI the strongest marginal trade. Once expectations become demanding, a smaller positive change elsewhere can produce a larger market response.

Narrow Earnings Breadth Is Keeping Investors in Motion

Rapid rotation occurs when the market has several credible trades but lacks one earnings trend broad enough to dominate them.

CITIC links the current rotation to limited earnings upgrade breadth. When revisions rise across many companies in one industry, investors can hold a theme through temporary volatility. When upgrades remain concentrated, capital moves between smaller groups of winners.

Trade friction adds another constraint. Export growth has been one of the most important earnings signals for many A-share companies. However, tariff risks, foreign policy uncertainty, and shifting supply-chain rules can reduce the valuation investors assign to overseas revenue.

Exchange-rate effects further complicate comparisons. Currency movements can raise or reduce reported earnings without reflecting a corresponding change in unit demand. They can also affect import costs, overseas assets, and the translated value of foreign sales.

These forces make market breadth harder to sustain. A company can report acceptable operating results while receiving a lower valuation because investors question the durability of its external markets. That encourages shorter holding periods and faster movement between sectors.

Momentum strategies suffer in this setting. Momentum depends on recent winners continuing to outperform long enough for the trend to compensate for occasional reversals. Rapid rotation repeatedly interrupts that process.

CITIC’s fuller strategy discussion says momentum and prosperity-chasing approaches are especially vulnerable when leadership changes quickly. These strategies work better when the market has a clear, persistent direction.

Low-valuation strategies generally fare better during rotation, but CITIC identifies an important exception. It says only the PB-ROE approach shows a statistically significant advantage as rotation speed increases.

PB means price-to-book value, which compares a company’s market value with its accounting equity. ROE means return on equity, a measure of the profit generated from shareholder capital.

A PB-ROE framework searches for companies whose valuations appear low relative to their profitability. It can also identify businesses where investors have priced in weak conditions despite evidence of improving returns.

According to a detailed strategy summary, CITIC’s traditional residual-based PB-ROE portfolio historically beat the CSI All Share Index by an average of 158 basis points monthly during fast rotation.

The brokerage says adding a distress-reversal factor raised that monthly historical excess return to 179 basis points. One basis point equals one-hundredth of a percentage point.

Those figures describe a historical model, not a guaranteed future return. They depend on CITIC’s definitions, portfolio construction, transaction assumptions, and selected sample periods. Investors cannot assume that a published factor result will survive real trading costs.

Still, the mechanism fits the present market. When investors abandon crowded leaders, they often seek companies with measurable profits and undemanding valuations. That process produces valuation repair rather than a lasting momentum wave.

CITIC places industrial metals, coal, and logistics among industries fitting its traditional PB-ROE approach. It associates chemicals, batteries, and animal breeding with the distress-reversal version, where supply rationalization improves industry economics.

This creates pressure across the AI trade. Chip and server suppliers must continue producing earnings upgrades strong enough to offset elevated expectations. Application companies must show that rising usage can become profitable revenue.

Fund managers face a different pressure. They cannot simply replace AI with a defensive sector and ignore computing growth. They need exposure to potential AI upside without relying on the same crowded holdings to lead every month.

Energy and chemical companies also face scrutiny. Low valuations alone cannot sustain a rerating. Investors need evidence that higher profits reflect durable supply conditions rather than temporary commodity prices or inventory movements.

The barbell structure addresses these competing demands. It preserves participation in AI infrastructure while adding companies whose valuations can recover through current earnings. It also accepts that neither side represents a permanent market leader.

Stronger Compute Demand Has Not Solved AI Commercialization

CITIC’s most consequential judgment is that additional computing demand confirms the AI buildout but does not settle who will earn attractive returns from it.

Computing demand is the clearest part of the AI investment case. Training larger models consumes substantial processing capacity, while serving those models requires repeated inference for every user request.

Agentic systems can multiply that load. A conventional chatbot might generate one response, while an agent can plan, search, call tools, inspect results, and revise its work. Each step can require additional model computation.

This pattern supports semiconductor designers, foundries, memory suppliers, optical networking vendors, server manufacturers, cooling providers, and electricity infrastructure. It also explains why China’s electronics profits can rise sharply during an uneven broader recovery.

However, revenue does not move evenly through that chain. Hardware vendors can benefit before software customers prove they will pay enough to cover inference expenses. Infrastructure can therefore boom while application economics remain unsettled.

CITIC’s report draws that line explicitly. Recent AI developments reinforce the established direction of compute demand, but they are insufficient to rewrite the long-term commercialization story.

For AI applications, commercialization requires more than growing usage. Providers must retain users, control serving costs, establish dependable workflows, and convert experiments into recurring contracts.

Enterprise buyers introduce another delay. They must test security, accuracy, integration, and governance before moving sensitive work into production. A successful pilot can still fail to become a large deployment.

Developers face similar friction. A stronger model can reduce the labor needed for one task while increasing the number of tasks attempted. The resulting compute bill can rise faster than the economic value captured by the software provider.

That uncertainty affects infrastructure valuations. A supplier may receive strong orders today because laboratories and cloud providers are racing to build capacity. Its distant valuation still depends on whether end customers generate enough income to sustain that spending.

CITIC identifies recursive self-improvement, or RSI, as one development that might expand the long-term valuation ceiling. RSI describes systems that materially accelerate the research and engineering needed to create stronger successor systems.

This idea has moved beyond purely speculative discussion, but it remains unproven at full scale. Anthropic has described early evidence that AI is accelerating AI research while emphasizing unresolved bottlenecks and serious control risks.

In its RSI research agenda, Anthropic argues that compute availability could eventually set the pace of development. It also notes that experiments, hardware, energy, fabrication, and human verification remain important constraints.

If AI systems substantially improve the process of developing new models, compute demand could rise for two reasons. More capable models would attract additional uses, while automated research would consume resources to generate further improvements.

The commercial implications would still be uneven. Model laboratories and infrastructure suppliers might capture significant value, but application providers could face faster product cycles and greater dependence on external platforms.

CITIC also mentions anti-distillation as a potential change. Knowledge distillation trains a smaller student model using outputs or behavior from a stronger teacher model. It can reduce costs, but it can also weaken the original developer’s commercial moat.

Anti-distillation techniques aim to discourage unauthorized extraction or make copied capabilities easier to detect. Their value is economic as well as technical because frontier developers spend heavily on training, data, and research.

Academic work has explored watermarking, output modification, and other defenses. One model protection study showed how invisible signals in generated text could help identify model misuse and resist removal.

A practical anti-distillation system could improve the ability of leading laboratories to charge for differentiated capabilities. It could also reduce the risk that competitors reproduce expensive performance through large-scale querying.

Yet defensive techniques involve tradeoffs. Changes to model outputs can affect quality or user experience. Detection systems can generate false positives, while determined attackers can adapt their extraction methods.

The legal position also differs across jurisdictions and contracts. Distillation is a standard machine-learning method, so disputes often focus on authorization, access methods, and terms of service rather than the technique alone.

This is why CITIC treats RSI and anti-distillation as possible valuation catalysts rather than established outcomes. Either development might strengthen the revenue moat around frontier models. Neither has yet resolved the entire commercialization problem.

The immediate AI case therefore remains infrastructure-led. Demand for chips and supporting systems is visible sooner than sustainable profits across the application layer. That favors selective exposure over an indiscriminate AI allocation.

Energy and Chemicals Offer Earnings Support, Not a Safe Haven

The energy and chemical side of CITIC’s barbell works only if improving profits survive weaker prices, policy shifts, and another inventory reversal.

The attraction begins with valuation. Many energy and chemical producers trade on mature-industry expectations, making them more responsive to evidence that profitability has stabilized.

Their earnings can also benefit from supply rationalization. When high-cost plants close or companies postpone new capacity, the remaining producers gain better utilization and pricing discipline.

CITIC’s distress-reversal framework looks for this combination. The market first discounts a sector because returns are weak. Supply exits, demand stabilizes, and surviving companies produce better earnings than their valuations imply.

China’s latest industrial data offer evidence for that recovery. Chemical manufacturing profits rose 56.6 percent during the first seven months of 2026, according to the national statistics agency.

Coal profits increased 50.4 percent, and petroleum extraction profits rose 15.4 percent. Petroleum, coal, and other fuel-processing businesses moved from an aggregate loss to a profit.

These are meaningful improvements, but they do not prove a durable cycle. Year-over-year growth can appear dramatic when the comparison period was weak. Aggregate industry data can also conceal large differences between companies.

Chemical profitability depends on feedstock costs, product spreads, plant utilization, and downstream demand. A producer can report higher revenue while earning less if raw-material costs rise faster than selling prices.

Energy companies face commodity-price volatility and policy intervention. Coal demand can remain strong while individual miners encounter production limits, transportation constraints, or higher safety spending.

Oil and gas producers depend on global prices, currency movements, and capital discipline. Refiners can experience different economics because lower crude prices reduce input costs but may also create inventory losses.

The sector therefore should not be treated as a passive defensive allocation. It carries operational, environmental, geopolitical, and regulatory risks that differ from those surrounding AI companies.

Its advantage in CITIC’s framework is relative. Energy and chemical valuations may require fewer optimistic assumptions than high-growth technology stocks. Improving profits can then trigger valuation repair even without a permanent commodity boom.

The timing also matters. CITIC says AI’s contribution was concentrated in the second quarter, while energy and chemicals could gradually gain strength afterward. This language describes a rotation sequence, not a definitive peak in AI earnings.

Investors can misread that sequence in two ways. The first is assuming that every AI holding has completed its earnings cycle. Hardware orders and data-center investment can remain strong beyond one quarter.

The second is assuming that every low-valued chemical producer will rerate. Companies with excess capacity, weak balance sheets, or poor cost positions can stay cheap despite better sector data.

A disciplined barbell therefore requires selection on both sides. AI holdings need demonstrable orders, margins, or customer adoption. Energy and chemical holdings need cash generation, improving returns, and credible supply conditions.

This approach differs from momentum chasing. It does not buy a company merely because its share price has recently risen. It asks whether profitability supports the valuation after the market’s leadership changes.

It also differs from a simple value strategy. A low price-to-book ratio can signal hidden opportunity, but it can also reflect permanently weak returns. Pairing PB with ROE seeks to separate recovering businesses from persistent value traps.

The strongest version of CITIC’s thesis would show simultaneous confirmation. AI infrastructure earnings would remain resilient, while chemicals and energy would continue posting wider profit improvements.

The weakest version would show deterioration on both sides. AI spending could slow before applications produce adequate revenue, while industrial profits could fade as capacity returns or commodity spreads narrow.

That possibility is the central risk of the barbell. Diversifying between two narratives does not eliminate losses if both depend on the same weakening macroeconomic environment.

Three Signals Will Test CITIC’s Rotation Thesis

The next stage will be decided by earnings breadth, evidence of an AI revenue moat, and the durability of industrial profit recovery.

The first signal is the breadth of earnings revisions after the second-quarter reporting season. Investors should watch whether upgrades spread beyond a small group of computing suppliers.

A broader set of revisions across software, cloud services, networking, power equipment, and AI applications would strengthen the technology side of CITIC’s thesis. It would show that computing investment is becoming a wider earnings cycle.

Narrow revisions would support the rotation argument instead. If only a few hardware companies keep receiving upgrades, expensive AI shares will remain vulnerable whenever another industry reports improving profits.

The direction of margins matters as much as revenue growth. Hardware companies need enough pricing power to offset manufacturing and capacity costs. Application providers need revenue growth that exceeds incremental inference expenses.

The second signal is a credible change in AI commercialization. That change could involve recursive self-improvement, effective anti-distillation systems, or another mechanism that strengthens model economics.

For RSI, the important evidence is not a laboratory saying its models help researchers. The stronger test is whether AI systems shorten development cycles while producing independently validated improvements in successor models.

For anti-distillation, investors should look for deployment evidence. A laboratory must show that its defenses deter extraction without degrading ordinary customer output or falsely restricting legitimate use.

Either result would strengthen the distant valuation case. RSI could expand demand and accelerate capability growth, while effective anti-distillation could protect the returns earned by frontier model developers.

Failure would weaken the most speculative part of the AI trade. Compute demand could remain high, but model providers would still struggle to convert technical leadership into a durable commercial advantage.

The third signal is whether energy and chemical profit gains persist through subsequent monthly data. Investors should compare profit growth with production, pricing, utilization, inventories, and cash flow.

Continued improvement across several measures would support CITIC’s PB-ROE logic. It would suggest that sector earnings reflect better industry economics rather than a favorable base comparison.

A rapid slowdown would weaken the rotation thesis. If chemical profits lose momentum while producers restart capacity, valuation repair could end before a durable new market leader forms.

Policy and trade developments will influence all three signals. Additional trade restrictions can reduce export valuations, alter technology supply chains, and move energy or feedstock costs.

Currency changes can also shift reported results across sectors. Investors should separate translation effects from operating improvements, particularly for companies with large overseas revenue or imported inputs.

The original report remains a brokerage strategy call, not an official forecast or investment guarantee. Public summaries provide the central conclusions but do not disclose every model assumption behind them.

That verification gap deserves attention. The expanded report account confirms the core claims about rotation, PB-ROE, AI computing demand, and the energy-chemicals barbell. It does not make the strategy immune to changing data.

For technology buyers and developers, the analysis has a practical message beyond equity markets. Infrastructure demand can grow rapidly while the economic value of individual AI products remains unsettled.

Teams should therefore track outcomes, not only model capability. Useful measures include completed work, error rates, customer retention, inference expense, and the human review required for production use.

A searchable AI knowledge base can help teams preserve those evaluations alongside vendor claims, meeting notes, and deployment results. That record becomes valuable when models and prices change quickly.

CITIC’s framework ultimately asks investors to hold two ideas at once. AI computing demand remains a real growth force, but familiar demand evidence no longer guarantees additional valuation expansion.

Energy and chemicals offer a different route through earnings recovery and lower starting valuations. They are not risk-free substitutes for technology, and their improving data still require confirmation.

Over the next three months, watch whether earnings upgrades broaden, whether AI laboratories establish stronger commercial moats, and whether industrial profits remain durable. Those signals will reveal whether the barbell is becoming a sustainable allocation or merely the latest turn in a fast market.