An analytical report from Guolian Minsheng Securities Company Limited suggests that, based on current core metrics, the overall AI industry cycle remains supported without signs of a comprehensive deterioration. Going forward, it is crucial to continuously monitor the marginal changes in three key signposts to cross-verify demand strength and risk levels. Should a combination of signals emerge—including consecutive declines in Annual Recurring Revenue (ARR) growth, simultaneous drops in both Token consumption volume and price, an accelerated fading of cost-cutting benefits, and a persistent widening of credit spreads for NeoCloud companies—it would be prudent to be alert to the risk of a systemic correction in the AI investment theme. The main viewpoints of Guolian Minsheng Securities Company Limited are outlined below.
As the current AI market rally enters an intense phase, market expectations have gradually shifted from widespread optimism to divergence, with growing debate over the sustainability of the industry's expansion. In particular, recent news that Meta plans to commercially lease its surplus internal AI computing capacity to the market—a marginal change on the supply side—has triggered concentrated market-wide skepticism about the sustainability of the AI industry's capital-intensive expansion model, directly prompting a broad correction in global technology stocks. Against this backdrop, accurately judging the upside potential and cyclical inflection point of this AI rally, and constructing a practical monitoring framework, has become a central question for the market.
Key Monitoring Framework for the AI Cycle
Based on the complete transmission logic of the AI industry cycle, we have distilled three core observation signposts from three dimensions—industry momentum, profit support, and tail risks—to systematically track the market's trajectory and inflection point signals.
First Signpost: Tracking AI Industry Momentum
This involves anchoring the entire demand transmission chain. Following the path from "downstream commercial applications → midstream model subscription revenue → upstream cloud provider capital expenditure," we can verify the real strength and sustainability of AI industry momentum layer by layer using indicators such as Token consumption, Annual Recurring Revenue (ARR), and computing power capital expenditure.
Second Signpost: Monitoring the Inflection Point in Cost Efficiency for Tech Giants
This focuses on assessing the resilience of profit support. By tracking the pace at which cost-saving benefits from workforce optimization and expense reduction diminish, we can gauge sustainability. If cost efficiency improvements become unsustainable and AI commercialization fails to yield significant results, the maintenance of profit margins for major tech firms could face challenges.
Third Signpost: Monitoring Tail Risks in NeoCloud
This entails vigilance regarding the transmission effects from high-leverage segments. The focus is on the high-leverage weak links within the industrial chain—NeoCloud providers represented by companies like CoreWeave. Tracking leading indicators such as their bond credit spreads and CDS quotes helps monitor their tail risks and potential spillover effects on the entire sector.
Detailed Analysis of the First Signpost
At the industry level, the momentum and realization of the AI sector remain the core observation thread. The initial market enthusiasm was precisely triggered by the non-linear, rapid increase in ARR from algorithm companies. The AI industrial chain is structured from upstream computing infrastructure (chips, servers, data centers) to midstream model providers (the hub connecting supply and implementation) and downstream application scenarios (the endogenous driver for end-user payment and technological iteration). The upside and duration of this AI rally are, to some extent, determined by the willingness of downstream end-users to pay.
A recovery in downstream commercial demand directly boosts midstream model call volumes and subscription revenue, which in turn drives upstream cloud and computing power providers to increase capital expenditure for capacity expansion. In essence, demand signals travel upstream along the path: "downstream commercial applications → midstream model calls and subscriptions → upstream cloud provider capital expenditure." Based on this, we track the following indicators:
First, the growth rate and slope of capital expenditures, such as those for computing power, reflect the willingness of upstream cloud service providers to expand supply. In Q1 2026, the quarterly capital expenditure of the top five cloud providers remained above $140 billion, and market consensus expectations for their subsequent quarterly capital expenditures continue to trend upward, showing no significant slowdown. Future quarterly reports from these major firms regarding their Capex guidance for the next two quarters warrant close attention. If the growth rate of capital expenditure shifts from its previous steep ascent to a flattening trend, or even shows a sequential stall, it may indicate that this round of computing power supply expansion has reached a阶段性 ceiling, and order momentum in the upstream computing hardware chain could face marginal contraction pressure first.
Second, Annual Recurring Revenue (ARR) is used to verify the true realization level of midstream large model commercialization. Compared to quarterly revenue or one-time project income, ARR excludes the disturbance of non-recurring gains and can accurately reflect the sustainable growth capability of AI model and cloud service subscription businesses, serving as a core benchmark for midstream commercialization. Currently, leading large model providers, including OpenAI and Anthropic, are showing non-linear growth in ARR. Taking Anthropic as an example, its ARR surged from $9 billion at the end of 2025 to $47 billion in May this year, a more than fourfold increase in half a year, validating the logic of "model capability leap → developer call volume surge → revenue growth." This suggests the AI industry is not solely a "cash-burning narrative" but has entered a crucial phase of profit realization, providing mid-term fundamental support for the market. Going forward, continuous tracking of ARR growth rates for leading providers is essential. If growth shows consecutive months of marginal slowdown or weakening, it may signal a阶段性 softening in downstream付费 demand, potentially leading to a阶段性 adjustment in the AI theme.
Third, end-user Token consumption serves as a leading indicator for predicting changes in downstream demand. As the smallest unit for measuring AI calls and computing power consumption, the weighted spending price and total network consumption volume of Tokens are currently the most sensitive leading indicators of downstream demand. Among these, the Large Model Token Spend Index published by Silicon Data measures the average cost paid by the market per million Tokens consumed. This index more than doubled from December last year and continued to climb until May this year. However, the data for June showed a sudden decline,一度引发市场担忧. Nonetheless, weakening Token pricing does not necessarily equate to a peak in overall AI demand. Another core possibility is that users are actively shifting towards more cost-effective, lower-priced base models, thereby pulling down the overall average price.
According to global large model Token total consumption data disclosed by OpenRouter, since June, the weekly total Token call volume for global large models has surged significantly again, steadily exceeding the 45 trillion level, indicating no clear sign of an overall decline in industry usage demand. This suggests the industry's situation is not entirely due to demand contraction but rather price normalization driven by improved model efficiency and intensified industry competition, with real downstream computing power demand still expanding. If subsequent Token consumption persistently shows a combined signal of "decline in both volume and price," it would imply a short-term peak in downstream demand, warranting vigilance for systemic correction risks in the AI theme.
Examining the Second Signpost
Beyond industry cycle indicators and demand-side factors, the cost side—specifically the cost efficiency progress of tech giants—is the second core dimension for assessing the sustainability of the AI rally. While the performance of leading cloud providers has generally maintained high growth, the阶段性 profit improvement is not entirely attributable to business expansion and AI commercialization. A significant portion relies on cost-cutting measures such as workforce optimization and expense reduction, introducing structural concerns about profit resilience.
We selected five global tech leaders—Oracle, Google, Microsoft, Meta, and Amazon—for analysis. The sales and administrative expense ratio (average of the five companies), highly correlated with workforce optimization, has continuously declined from 15% in fiscal 2023 to around 12% in fiscal 2025, showing significant expense reduction effects. Correspondingly, the operating profit margin for these companies increased from 27% to over 32% during the same period. This implies that the decline in sales and administrative expenses contributed over 50% to the marginal increase in operating profit margins, making cost compression a key variable in the profit recovery of these tech giants this cycle.
A rough calculation based on不完全统计裁员 announcements from these five cloud providers in the first half of fiscal 2025-2026 suggests a裁员规模 of around 100,000 over the past two years. Assuming an average per-person cost of $200,000, this would释放至少 $20 billion in cash flow. In fiscal 2025, the total revenue of these five companies was $1.65 trillion, with total operating profit exceeding $400 billion. Roughly estimated, the人力成本 savings from裁员 account for about 1.2% of total revenue and 5% of total operating profit. This means that裁员 alone could potentially contribute over 1 percentage point to the overall operating profit margin of these five giants (though actual profit增厚 would be slightly lower considering one-time severance支出抵扣).
From an incremental perspective, the total revenue of these five companies in fiscal 2025 increased by $200 billion compared to fiscal 2024, with operating profit新增 $60 billion. The cost savings from裁员降本 are equivalent to about 10% of the fiscal 2025 revenue增量 and nearly one-third of the operating profit增量. Cost reduction has become one of the core drivers of profit improvement for major tech firms.
Looking ahead, as the红利 from workforce optimization gradually peaks and the room for further cost efficiency gains narrows, the marginal momentum for profit recovery driven by the cost side will significantly weaken. If incremental AI business profits cannot timely form接力 to对冲 the退坡 of cost-side红利, the overall profit growth rate of tech giants will face significant downward constraints, and the underlying profit support for the AI theme will consequently loosen.
Analyzing the Third Signpost
During this cycle of computing power expansion by tech companies, debt financing has become a core source of funding for capital expenditure. With expectations of rising market interest rate中枢强化, credit default risk has also become a tail risk that cannot be ignored in the AI industrial chain. Current bond market pricing shows that despite persistently high long-term U.S. Treasury yields, overall credit spreads for U.S. investment-grade and high-yield corporate bonds remain in historically low ranges, comparable only to levels around 1996. Credit spreads in the tech sector have also not shown a trend of widening, indicating that the market has not yet priced in systemic偿债风险 for the tech industry, and overall sector debt pressure appears manageable.
However, this平稳表象 at the aggregate level masks significant structural differentiation within the industrial chain. Credit risks in the vulnerable segments of the AI赛道 require focused identification. By entity, leading tech giants, supported by stable profit bases, ample operating cash flow, and diversified business buffers, possess sufficient debt抗风险能力. Credit脆弱点 are mainly concentrated in midstream, rapidly expanding新兴算力厂商 that rely on high leverage. Among these, NeoCloud (new AI cloud service providers) represented by CoreWeave and Nebius are one of the sectors with higher leverage and credit risk in this AI rally.
Judging from the pricing of Option-Adjusted Spread (OAS), the bond OAS range for a典型代表 like CoreWeave reaches 500–600 basis points, significantly higher than that of leading U.S. tech firms like Microsoft, Google, and Oracle,直观体现 the market's higher违约风险溢价 for it.
The leverage model of NeoCloud (emerging AI-specific cloud providers)本质上 is a debt-driven expansion model based on dual抵押 of "contractual cash flow + hardware assets." By securing long-term computing power orders from leading AI clients to obtain stable cash flow expectations, they use this as a core credit foundation,叠加 GPU实物抵押, to leverage large-scale, low-cost debt financing. This rapidly amplifies computing power capital expenditure, creating a self-reinforcing cycle of "order — financing — expansion," a金融化扩张 "玩法" born during the AI computing power爆发期.
The process typically involves: 1) Locking in anchor orders with high-credit clients like Microsoft and OpenAI, often with "take-or-pay" clauses, guaranteeing payment regardless of actual usage, creating确定性 future cash flow. 2) Dual抵押债务融资: Using this legally binding long-term contract (future revenue stream) as the core偿债担保, alongside purchased GPU clusters as实物抵押品, to secure debt financing. 3) Computing power expansion: Using the raised funds to prepay Nvidia for GPU capacity, rapidly building AI computing clusters, delivering them to clients, and starting to recognize revenue. 4)循环加杠杆: Using operational income from newly delivered computing power to repay部分本息, while leveraging the deployed computing scale to secure more leading orders, then using new orders as抵押 for the next round of融资扩张, continuously放大 the资产负债表.
Taking龙头 CoreWeave as an example, the core risk points for NeoCloud currently mainly集中在 the following aspects:
Risk One: High dependence on external融资循环, lacking自身盈利 and现金流支撑. Although CoreWeave's Q1 2026 revenue reached $2.078 billion,同比增长 112%, with积压订单 rising to nearly $99.4 billion, supporting future high revenue growth expectations, the core痛点掩盖在高成长预期下 is its极度滞后的内生盈利能力. Its Q1 net loss widened from $315 million in the same period last year to $740 million, with adjusted每股亏损达 $1.12, indicating the "revenue增长 without profit增长"模式尚未扭转.
A deeper财务隐患 lies in its extreme债务杠杆与过度拉伸的资产负债表. CoreWeave's重资产军备竞赛 is entirely driven by高频负债. As of Q1 2026, the company's total liabilities had risen to $50.814 billion (compared to total assets of $55.573 billion), resulting in a负债率高企至 91.4% and an equity ratio (net debt-to-equity) as high as 10.7 times. With the rapid rise in interest-bearing debt,财务成本 is加速侵蚀经营成果. Q1 net interest expense surged from $264 million a year ago to $536 million, more than doubling year-over-year, with interest expense接近 26% of quarterly revenue. The company's Q1 EBITDA/interest expense (interest coverage ratio) has滑落至 1.9 times,处于信用评级市场的投机级 (junk bond)危险边缘.
This重资产杠杆模式 has极低的现金流链条容错率. In the absence of a底层内生现金流安全垫,一旦下游大模型大客户出现履约延期 or算力需求边际放缓, the庞大的利息刚性支出 could瞬间引发流动性断裂风险.
Risk Two: High客户集中度, facing连锁反应 from单一客户违约. CoreWeave has high客户集中度,主要集中 in a few clients like Microsoft and OpenAI, with 67% of its fiscal 2025 revenue coming from Microsoft alone. This structure is a "优质现金流背书" during景气期, but一旦核心客户需求不及预期 or重新谈判价格, NeoCloud's收入确定性 could瞬间崩塌,面临严重的产能闲置和现金流断裂风险.
Furthermore, major clients随时可能转化为 "竞争对手." The recent news that Meta plans to commercially lease its own surplus internal AI infrastructure and computing power directly to the market directly导致股价大跌 for CoreWeave and Nebius. This indicates that while巨头 are currently NeoCloud's biggest "金主," the成熟 of巨头自研芯片 and转售 of surplus算力 are forming直接的 "供给端去中心化"冲击 on the NeoCloud model.
Risk Three: Asset折旧与技术的非对称风险. NeoCloud债务期限 are typically 5-7 years, but the technological折旧周期 for GPUs is about 3 years (the computing value of old cards大幅缩水 after新一代芯片推出). This means债务还没还完,核心资产 may already be outdated, with the asset side facing较大的减值压力.届时原有的资产担保比例可能会直接受挫,引发信用利差 (OAS)的惩罚性拉宽.
Therefore, the NeoCloud leverage model's成立高度依赖 external conditions and exhibits明显的顺周期脆弱性. These conditions include sustained high growth in AI computing demand, maintained scarcity in GPU supply, controllable GPU hardware technology iteration speed (avoiding rapid asset贬值),长期维持高位 utilization of computing clusters (covering折旧与资金成本), and low market利率 (reducing融资成本).一旦AI大模型投资退潮,头部客户缩减算力采购, utilization下滑等都将直接冲击NeoCloud公司现金流,叠加利率的上移,可能导致偿债能力迅速恶化,信用危机骤然上升.
Compared to滞后财务报表,信用溢价 and CDS利差 are核心先行指标 for预判NeoCloud风险 and前瞻AI行情拐点. CDS利差 directly reflect the market's pricing of违约概率 for computing power firms, often able to提前预判板块波动与产业风险. During this rally, the 5-year CDS利差 for core NeoCloud target CoreWeave一度飙升至 900bp,充分定价 the违约担忧 from高杠杆 and再融资压力. Subsequently,伴随行业订单落地 and融资环境边际宽松, the利差回落至 500-600bp左右, but the绝对水平仍处于历史高位,显著高于北美高收益债CDS指数以及其他云服务厂商,侧面印证板块整体风险偏好偏弱、波动韧性不足.
Therefore,后续跟踪NeoCloud板块债务风险 is also an important indicator for甄别AI需求和尾部风险. If CDS利差 for this板块持续走阔 and存量债券收益率大幅上行, it意味着市场对其资产负债表稳健性的质疑加剧,届时需高度警惕风险向整条AI产业链传导扩散.一旦AI行业估值泡沫出清,企业发展格局或将显著分化: tech巨头 with stable经营性现金流 capable of自主支撑AI资本开支 will possess stronger周期抵御能力凭借深厚信用护城河; while enterprises relying on高杠杆扩张 may迎来完全不同的发展结局.
Conclusion and Monitoring Summary
In summary, based on current core indicators, overall industry momentum remains supported without signals of comprehensive恶化. Going forward, it is essential to持续跟踪 the边际变化 of the three key signposts,交叉验证需求强度与风险水平. Should a combination of signals emerge—including consecutive declines in ARR growth, simultaneous drops in both Token consumption volume and price, an accelerated fading of cost-cutting红利, and a persistent widening of NeoCloud credit利差—it is necessary to警惕AI行情迎来系统性回调风险.