CITIC Construction Investment: AI computing power boom continues, holdings reach new highs, and cloud vendor capital investment support continues to increase

The Zhitong Finance App learned that CITIC Construction Investment Securities released a research report saying that test instruments, optical modules, and GPU/ASIC sectors continued to perform well in the first half of 2026, reflecting strong demand for AI computing power. Net profit from the 2026Q2 cable sector increased 205% year-on-year, and the fiber-optic cable boom continued to increase. 2026Q2, the share of public fund communications industry holdings reached a new high of 11.56%, reaching a new high. The main holdings are concentrated in optical modules, optical devices, optical chips and optical fiber cables. Four North American CSP manufacturers announced their second quarterly reports and continued to increase capital expenditure. The total capital expenditure of the four major North American cloud vendors in 2026Q2 was about US$171.2 billion, an increase of 78.6% over the previous year. The total capital expenditure is expected to be about US$73.25 billion in 2026 (median estimate), and it is expected to maintain a relatively rapid growth rate in 2027. Continue to be optimistic about AI development, and maintain long-term optimism about the North American computing power industry chain and the domestic AI computing power industry chain.

1. The revenue of the communications sector grew steadily, and profitability remained at a high level.

2026H1, the A-share communications sector achieved revenue of 1401,074 billion yuan, an increase of 5.83% year on year; net profit to mother was 146.363 billion yuan, up 6.14% year on year, and profit growth rate was slightly faster than revenue growth. The comprehensive gross margin of the 2026H1 communications sector was 28.10%, down 1.24pct year on year, and the net margin was 10.99%, down 0.11pct year on year. Looking at a single quarter, the 2026Q2 communications sector achieved revenue of 726.1 billion yuan, up 5.8% year on year, and net profit to mother of 91.6 billion yuan, up 6.89% year on year; comprehensive gross margin was 29.87%, down 1.63 pct year on year, and overall profitability remained steady.

2. Telecom operators were affected by VAT adjustments, and profits declined.

The three major operators were affected by value-added tax adjustments. Combined, the personal terminal business faced some pressure. Revenue and profit declined. 2026H1 achieved total revenue of 999.9 billion yuan, a year-on-year decrease of 1.53%, and realized net profit of 108 billion yuan to mother, a year-on-year decrease of 6.75%.

3. The 2026H1 test instrument, GPU/ASIC, and optical module/optical device sectors showed outstanding performance, and the Q2 cable sector saw a significant increase in profits.

Outstanding net profit performance of 2026H1 includes testing instruments (593 million yuan, up 549% year on year), GPU/ASIC (2,912 billion yuan, up 401% year on year), and optical modules/optical devices (24.060 billion yuan, up 147% year on year). The top three net profit growth sectors in 2026Q2 were test instruments (489 million yuan, up 505% year on year), GPU/ASIC (1,968 billion yuan, up 245% year on year), and cable (6.958 billion yuan, up 205% year on year).

4. Public fund communications industry holdings are further concentrated in the optical communications sector.

2026Q2, the market value of public fund communications industry holdings reached 794.492 billion yuan, accounting for 11.56% of the market value of holdings, a record high. The main holdings are concentrated in the optical communications sector. As of September 1, 2026, the Shenwan Communications Index PE-TTM was 57.04, at the 92.91% quantile in 5 years and 87.84% in 10 years.

North American CSP manufacturers continue to raise capital, and global models enter rapid commercialization

Nvidia's FY2028 guidelines exceed expectations as North American CSP vendor capital investment support continues to increase

Four North American CSP manufacturers announced their second quarterly reports and continued to increase capital expenditure. The total capital expenditure of the four major North American cloud vendors in 2026Q2 was about US$171.2 billion, up 78.6% year on year and about 30% month on month. Investment in AI infrastructure continued to accelerate markedly. Amazon, Google, and Meta have all further raised their 2026 capital expenditure guidelines. Although Microsoft adjusted its book CapEx forecast to about US$175 billion due to changes in data center rental accounting classification, the company emphasized that its actual infrastructure investment plans have not been lowered. According to estimates of Amazon's 220 billion US dollars, Microsoft's investment scale of about 175 billion US dollars, Google Guidelines median value of 200 billion US dollars, and Meta Guidelines median value of 137.5 billion US dollars, the total capital expenditure of the four companies in 2026 is about 73.25 billion US dollars, which is a further increase from the forecast of about 71 billion US dollars at the time of the first quarter report. We continue to be optimistic about the computing power demand brought about by AI development, and maintain long-term optimism about the North American computing power industry chain and the domestic AI computing power industry chain.

Amazon's capital expenditure for the second quarter of 2026 was about US$54.2 billion, up 68.4% year on year, and continues to be mainly used for AWS and generative AI infrastructure construction; AWS revenue in the second quarter reached US$42.2 billion, up 37% year on year, the fastest growth rate in 18 quarters. The company raised its forecast for cash capital expenditure in 2026 from about 200 billion US dollars to 220 billion US dollars, mainly reflecting rising prices of memory and other components and continued strong demand for AI infrastructure. Management said that even with a further increase in capital expenditure, it will not be able to meet all demand in 2026, and the tight supply and demand situation is expected to continue in 2027.

Microsoft's capital expenditure in the second quarter of 2026 (FY2026Q4) reached 41 billion US dollars, an increase of 69.4% over the previous year. Of this, about two-thirds were used for short-term assets such as GPUs and CPUs, and the rest was mainly for long-term assets such as data centers; customer demand continued to exceed current production capacity. Since Microsoft has extended the expected service life of data centers and office buildings from 15 to 25 years since FY2027, and some future data center leases will be converted from financial leases to operating leases, the company adjusted the estimated book capital expenditure forecast for the 2026 calendar year from about 190 billion US dollars to about 175 billion US dollars, but it was clearly stated that the actual scale of infrastructure investment has not been reduced, mainly about 15 billion US dollars of expenditure has moved out of CapEx. Furthermore, the company expects FY2027 capital expenditure to continue to increase year over year.

Google's capital expenditure in the second quarter of 2026 reached 44.9 billion US dollars, about double the previous year, and a further 26% increase from 35.7 billion US dollars in the first quarter. Of these, the investment in technical infrastructure was about 60% for servers and 40% for data centers and network equipment. Driven by continued strong demand for Google Cloud and AI computing power, the company once again raised its 2026 capital expenditure guidelines from the previous $180 to 190 billion US dollars to 1950 billion US dollars, with a median value of 200 billion US dollars. The company said the increase is mainly due to speeding up delivery of computing power production capacity to meet demand, while capital expenditure is still expected to increase significantly in 2027.

Meta's capital expenditure in the second quarter of 2026 reached US$31.1 billion, up 82.7% year on year, and 57% month-on-month increase from US$19.8 billion in the first quarter, mainly for server, data center and network infrastructure construction. The company further narrowed its 2026 capital expenditure guidance from the previous $125 to $145 billion to $130-145 billion, with a median value of about $137.5 billion, indicating that investment in AI infrastructure continues to increase. At the same time, Meta is cooperating with external capital such as BlackRock to build data centers. Some AI infrastructure investment may not be fully reflected in the traditional CapEx scale, so the actual scale of computing power expansion may be higher than the capital expenditure in the table.

Nvidia's FY2027Q2 performance fully exceeded expectations, with FY2028 revenue leading to a year-on-year growth rate of 70%, and global demand for AI computing power continues to be high. Nvidia's FY2027Q2 (as of July 26, 2026) achieved revenue of US$96.22 billion, up 106% year on year, of which data center revenue reached US$89 billion, up 117% year on year and 18% month on month, which is the core growth engine; GAAP net profit was about US$59.69 billion, up 126% year on year, and GAAP gross margin was about 75.0%. Looking ahead to FY2027Q3, the company expects revenue to reach US$108 billion ± 2%, corresponding to a year-on-year increase of about 89%. Furthermore, the guideline does not include data center computing revenue from the Chinese market, indicating that global demand for AI computing power remains high. In addition, the company's management first gave preliminary revenue guidance for the next fiscal year during the earnings call. The company's FY2028 revenue is expected to grow by about 70% year-on-year, far exceeding market expectations, and stated that this growth rate is at a level that can be achieved under supply constraints. Without supply constraints, the potential demand growth rate is expected to be higher. This guideline has significantly increased the certainty and visibility of the mid-term growth of global AI computing power.

Strong customer demand has led to a generational leap in product value. AWS and Nvidia recently announced plans to deploy 2 million additional Blackwell Ultra, Rubin and Rubin Ultra GPUs from 2027-2028. Previously, AWS only planned to add more than 1 million GPUs in GTC this year starting in 2026, and 2 million more in just 5 months. Officials clearly stated that the reason was that demand exceeded previous expectations. The two sides will also continue to expand cooperation in fields such as Spectrum Networks, Vera CPU, and AI Factory. Nvidia management said that mass production and shipment of Vera Rubin began in August, and each GW of computing power brings revenue opportunities (revenue opportunities) to Nvidia. It is expected to jump from about $18 billion for the Hopper architecture and about $25 billion for the Blackwell architecture to about $40 billion for the Vera Rubin architecture. The value of computing power platforms continues to rise, fully reflecting the demand for terminals with strong AI computing power and the industry trend of computing power infrastructure evolving to full-stack system solutions.

Demand for AI is spreading from hyperscale cloud vendors to more diverse customers. Nvidia's financial report for FY2027Q2 shows that within the data center business, the ACIE (AI Cloud, Industrial and Enterprise) sector's revenue reached US$40.3 billion, an increase of 138% year over year, which is faster than the Hyperscale (hyperscale customer) sector's 102% growth rate. ACIE's revenue share has increased to 45% of the data center business, showing that demand for AI computing power is rapidly penetrating from leading cloud vendors such as Amazon, Microsoft, and Google to a wider range of emerging AI cloud service providers (NeoCloud), industry, enterprises, and sovereign AI entities. Among them, Nvidia management said during the earnings call that their NeoCloud partner is rapidly deploying online computing power at a lower token cost. It is expected that by the end of 2026, its total installed capacity will increase from about 3 gigawatts at the end of 2025 to 8 gigawatts, and diversified customers are becoming an important source of growth for this round of AI computing power expansion.

Overall, as the Blackwell series of products continues to be released on a large scale, the next-generation Vera Rubin platform has officially entered the mass production and climbing stage. The diverse computing power requirements of traditional hyperscale cloud vendors, NeoCloud, industry, enterprises, and sovereign AI entities form a synergy to jointly drive Nvidia's FY2027Q2 revenue growth. Combined with Nvidia's history for the first time, it provides preliminary revenue guidance for FY2028 for the whole year. Multiple signals such as continuous improvement in downstream customer purchase orders and intergenerational increases in revenue opportunities corresponding to a single GW computing power confirm each other fully. Global AI computing power capital expenditure is still in a strong upward cycle. Demand for computing power is not only booming in the short term, but the certainty and visibility of medium-term growth has also been further strengthened.

Global big models are rapidly commercialized. As of the end of July 2026, Anthropic's annualized revenue exceeded 65 billion US dollars, an increase of about 622% from about 9 billion US dollars at the end of 2025, and an increase of 7.2 times the original in seven months; a two-month increase of about 38% from the 47 billion US dollars in May. Initial revenue for 2026Q2 exceeded US$11.5 billion, an increase of about 14 times over the previous year. Claude Code is still an important growth engine, but the revenue structure is expanding from a single coding to a wider range of enterprise AI as enterprise-grade agents and generic workflows are rapidly scaling up. The company recently signed a $45 billion Nscale and $35 billion Lambda cloud computing power cooperation, and commercial demand is being transformed into a definitive expansion of computing power.

OpenAI accelerates the expansion of chat to coding, co-work, and multiple business models. As of July 2026, OpenAI's annualized revenue exceeded 40 billion US dollars, nearly doubling from the end of 2025. The monthly annualized revenue in July increased by more than 20% month-on-month. The total number of active users of Codex and ChatGPT Work has exceeded 10 million. Of these, more than 1 million users have already used Codex for non-programming work, and agents are expanding from developer tools to knowledge work scenarios such as finance, law, recruitment, and operation. At the end of August, the annual revenue of ChatGPT Ads reached a further $1 billion, and the business model continued to expand from subscriptions and APIs to agents, enterprise services, and advertising.

The AI revenue of Chinese internet vendors has entered the large-scale cashout stage. ByteDance's AI-related annual revenue in July 2026 was about 4 billion US dollars, and the Feishu product team was integrated into the Doubao system, sales and customer service team into Volcano Engine to further open up AI applications, MaaS, SaaS and cloud computing. In the second quarter of 2026, Ali's AI cloud and computing service revenue reached 48.44 billion yuan, an increase of 45% over the previous year; of these, the large model MaaS business ARR exceeded 16 billion yuan, or about 2.4 billion US dollars, making it one of the largest enterprise-level large-scale model services in China.

The revenue of domestic independent model manufacturers increased simultaneously. At the end of August, Zhipu's MaaS platform ARR reached an annualized monthly rate of 1.6 billion US dollars, according to the latest weekly annualization; 2026H1 open platform and API revenue of 825 million yuan, an increase of about 27 times over the previous year, accounting for 86.5% of revenue, the API gross margin increased to 24.6%, the inference cost per token fell 80% from the beginning of the year, and a 1GW domestic computing power center has been built and partially put into operation. MiniMax's August ARR market caliber exceeded 800 million US dollars, and H1 corporate service revenue increased 703% year over year, accounting for 63.4%; Kimi had an ARR of about 300 million US dollars in June and officially launched the Hong Kong IPO process in September; DeepSeek's latest annualized revenue also saw a significant increase.

The acceleration of commercialization is becoming an important part of the closed loop of AI business logic. Coding took the lead in verifying the willingness of large models to pay, and the next stage of growth will come more from corporate agents, co-work, and long-term tasks. These tasks require longer context, multiple rounds of reasoning, tool calls, and repeated execution, and the single-task token consumption is significantly higher than traditional Chat. Therefore, even if the price per token continues to decline, total call volume and demand for inference computing power may still accelerate, and commercialization of large models and computing power capital expenditure are forming mutually reinforcing positive feedback.

The number of token calls on the OpenRouter platform accelerated. It nearly doubled in August, and the global AI penetration rate accelerated. Call volume continued to grow from May to June, adjusted briefly in July, then accelerated again in August. The OpenRouter platform's weekly token call volume increased from 28.9t on May 18 to 24 to 46.7t on June 15 to 21, an increase of about 62% in less than a month; it continued to hit a phased high in mid-July, then declined slightly for two consecutive weeks, and dropped to 56.8T from July 27 to August 2, reflecting short-term fluctuations caused by the release of new models and free traffic. Weekly calls in August continued to reach new highs. The platform's weekly token call volume increased from 56.8T from July 27 to August 2 to 69.0T, about 75.3T, 93.4T, and 113T; from August 24 to 30, it reached 113T, up 21.1% month-on-month, with a cumulative increase of about 99% over four weeks, nearly doubling in one month. Compared with 22.8T at the end of March 2026, the current call volume has increased to about 4.96 times; compared to 5.5T at the end of December 2025, it has increased to about 20.5 times in nine months.

Coding, agents, and cost-effective models are jointly driving the growth in call volume. According to OpenRouter data, a single request for an agent task consumes about 15 times more tokens than normal manual interaction, and the amount of tokens generated by agents has exceeded that of normal manual calls since February 2026. From August 24 to 30, the weekly call volume of the Chinese model reached 55.16T, up 36.3% month-on-month, surpassing the US model for 18 consecutive weeks; GLM-5.3-Flash, DeepSeek V4 Flash, and Xiaomi MIMO-v2.5 took the top three, reflecting that domestic open source models are entering the real workflow of overseas developers with power, price, and open ecology. The increase in call volume is being translated into commercial value. Of the task consumption amount on the OpenRouter platform, generic tasks, codes, and agents account for 32.4%, 30.3%, and 28.7%, respectively. Of these, Workflow Execution accounts for 18.7% of the total consumption amount of the platform, making it one of the segments with the highest paid value.

Although OpenRouter data is not directly equivalent to the global model market size or model manufacturer's revenue, it can be an important high-frequency indicator of developers' and enterprises' API needs. Weekly token calls continue to rise, and growth is shifting from Chat and model testing to coding, agents, and complex workflows, indicating that large models are being embedded deeper into production environments. As the running time of a single task, number of tool calls, and context length continue to increase, the total token consumption and demand for inference computing power are expected to continue to grow rapidly even if the price per token falls.

The capabilities of the global big model are still being rapidly improved, and they have not yet entered the competency platform period. According to the Artificial Analysis Intelligence Index, the comprehensive score of the latest generation head model has been raised to the 60-66 partition range, and the ability boundaries for complex reasoning, programming, mathematics, and long-range agent tasks continue to expand. The model iteration cycle remains at the level of several months, and each round of upgrades has achieved significant breakthroughs in the complexity, length, and reliability of tasks that can be completed. The overseas closed source model still leads the world at the cutting edge of intelligence. The Claude series is at the top of the list, and the latest models from manufacturers such as OpenAI, xAI, and Meta are also in the first tier. Overseas leading laboratories still have leading advantages in pre-training scale, reinforcement learning, inference calculation, and agent system engineering. In particular, in terms of complex programming, long-range tasks, tool calls, and task success rates, their overall capabilities are still ahead of domestic models.

Domestic open source models have caught up markedly faster. Kimi K3 Max and GLM-5.3 Max both reached 60 points, a difference of only 1 point from overseas first-line models such as GPT-5.6 and Grok 4.6; Qwen3.8 and GLM-5.3 Flash reached 58 points and 57 points respectively. The domestic model is moving from “low cost and basic usability” in the past to global cutting-edge capability competition, and with advantages such as open source, low price, and domestic computing power adaptation, it is gradually approaching the leading overseas programming agent represented by Codex in the coding and agent scenarios.

In addition, OpenAI launched GPT-6 Astra on September 3, and officially revealed that it has achieved a new round of leaps in complex reasoning, programming, computer use, research, and multi-step tasks. Cybersecurity and long-range agent capabilities have improved significantly compared to GPT-5.6 Sol. Currently, GPT-6 Astra has not been included in the unified rating by AA; however, its launch once again shows that overseas models are still driving the global intelligence frontier rapidly upward. Although the domestic open source model is catching up rapidly, the cutting-edge model capability competition continues to accelerate.

AI capabilities are rapidly shifting from answering questions to completing long-term tasks independently. The task span that Claude can handle has been increased from about 4 minutes in 2024 to at least 16 hours in Mythos Preview, and the success rate of open coding tasks has increased by 50 percentage points to 76% within six months. Increased capacity has been translated into real productivity: Claude has generated more than 80% of Anthropic's production code, and typical engineer code throughput is about 8 times higher than in 2024. The big model is no longer just about improving the speed of writing a single code, but begins to independently complete a complete closed loop of engineering to understand the codebase, develop solutions, call tools, run tests, troubleshoot errors, and deliver results. Coding is becoming the first realistic path for AI to RSI. The code has low-cost verifiers such as compilers, unit tests, and performance indicators, so that the model can be “build-run-evaluated-improved” over and over again. Claude's ability to optimize small model training programs increased from about 3 times that of Opus 4 to about 52 times that of Mythos; in an open research project, the multi-agent system recovered 97% of the target performance gap, while the two human researchers only recovered 23% in a week. This means that models are not only writing product code, but are also beginning to improve training procedures, design and run experiments, analyze results, and participate in the R&D process needed to create stronger AI.

The core of the next phase is moving from coding agents to AI R&D agents, and gradually forming an RSI closed loop. The model first generates better code and algorithms, then modifies Prompt, Memory, Tools, and Agent Harness, and then further optimizes training data, reward functions, training frameworks, and even model parameters; if the improved model is also better at developing next-generation models, it will form a recursive cycle of “AI improves AI - the successor model is stronger - R&D speed is further improved”. Currently, Claude is close to or surpassing skilled humans in targeted R&D execution, but research question selection, evaluation criteria, and final verification are still mainly controlled by humans, and the strong RSI for completely independent design and training of successor models is still a critical step.

RSI is recursive self-improvement, or “recursive self-improvement”: AI not only completes tasks, but also begins to participate in improving the R&D process for manufacturing the next generation of AI. “Recursive” means that one round of improvement can help the next. Current models participate in AI research and development to help train stronger models; stronger models will participate in the next round of research and development with better coding, experimentation, and research capabilities. Through repeated iterations in this way, AI's ability to improve AI will continue to improve. RSI can occur at different points. This includes revising current answers, remembering past experiences, and accumulating new skills. It can also generate training data, design reward rules, optimize training codes, and propose research plans. The key is whether these improvements can be saved and continue to be used in future tasks or model training. Anthropic divides AI participation in AI research and development into five stages: independent human development, chatbot assistance, coding agents, autonomous agents, and closed loop. Anthropic believes it is currently in the fourth stage — autonomous intelligence; the next stage is a “closed cycle” where AI will independently design, train, and improve the next generation of models, which is the state indicated by the complete RSI.