A new research report indicates that the AI computing power demand cycle remains robust, with key segments such as test instruments, optical modules, and GPU/ASIC continuing to deliver strong performance in the first half of 2026. The fiber optic cable sector also posted a remarkable 205% year-on-year increase in attributable net profit during the second quarter.

By the end of Q2 2026, public fund holdings in the communications sector reached 11.56% of total market value—a historic peak—with positions heavily concentrated in optical modules, optical components, optical chips, and fiber optic cables. Meanwhile, the four major North American cloud service providers have all reported second-quarter results showing continued and substantial increases in capital expenditure.

Where the growth is coming from

The combined capital expenditure of the four dominant North American cloud providers hit approximately $171.2 billion in Q2 2026, representing a 78.6% year-on-year surge. Projections suggest total spending for 2026 could reach around $732.5 billion at the median estimate, with expectations of continued strong growth heading into 2027. The outlook for both the North American and domestic Chinese AI computing supply chains remains firmly bullish.

Steady revenue growth and resilient profitability

In the first half of 2026, A-share communications companies generated total revenue of 1.401 trillion yuan, up 5.83% year-on-year, with attributable net profit reaching 146.36 billion yuan, an increase of 6.14%. The sector's gross margin stood at 28.10%, down 1.24 percentage points year-on-year, while the net margin was 10.99%. For the second quarter alone, revenue reached 726.1 billion yuan (up 5.8%) with net profit of 91.6 billion yuan (up 6.89%), reflecting stable overall profitability despite slight margin compression.

Telecom operators feel the pinch

The three major telecom operators experienced a decline in both revenue and profit during H1 2026, primarily due to value-added tax adjustments and pressure on consumer-facing businesses. Their combined revenue fell 1.53% to 999.9 billion yuan, while attributable net profit dropped 6.75% to 108 billion yuan.

Standout performers across segments

Test instruments led the pack with attributable net profit of 593 million yuan, a remarkable 549% year-on-year increase, followed by GPU/ASIC at 2.912 billion yuan (up 401%) and optical modules/components at 24.06 billion yuan (up 147%). In Q2 2026, the top three growth segments were test instruments (up 505%), GPU/ASIC (up 245%), and cables (up 205% to 6.958 billion yuan).

Concentration in optical communications

Public fund holdings in the communications sector reached 794.492 billion yuan in Q2 2026, accounting for 11.56% of total fund assets—another record high. As of September 1, 2026, the Shenwan Communications Index traded at a PE-TTM of 57.04, placing it in the 92.91st percentile over the past five years and the 87.84th percentile over the past decade.

Cloud giants double down on infrastructure

The four largest North American cloud providers collectively spent approximately $171.2 billion on capital expenditures in Q2 2026, a 78.6% year-on-year increase and roughly 30% quarter-on-quarter growth. Amazon, Google, and Meta all raised their full-year 2026 capex guidance, while Microsoft adjusted its reported figure due to an accounting change for data center leases—though the company emphasized its actual infrastructure investment plans remain unchanged.

Based on updated guidance—Amazon at $220 billion, Microsoft at roughly $175 billion, Google at a mid-point of $200 billion, and Meta at a mid-point of $137.5 billion—the four combined are expected to spend approximately $732.5 billion in 2026, up from the roughly $710 billion projected after the first quarter.

Amazon scales up AI investment

Amazon's Q2 2026 capital expenditure reached approximately $54.2 billion, up 68.4% year-on-year, primarily directed toward AWS and generative AI infrastructure. AWS revenue hit $42.2 billion for the quarter, a 37% year-on-year increase—the fastest growth in 18 quarters. The company raised its 2026 cash capex guidance from roughly $200 billion to $220 billion, citing rising memory component prices and sustained AI infrastructure demand. Management indicated that even with increased spending, supply will not fully meet demand in 2026, with tight conditions expected to persist into 2027.

Microsoft's accounting adjustment

Microsoft's Q2 2026 (FY2026 Q4) capital expenditure reached $41 billion, up 69.4% year-on-year, with about two-thirds allocated to short-cycle assets like GPUs and CPUs. Due to a change in the estimated useful life of data centers and office buildings from 15 to 25 years starting in FY2027, the company adjusted its 2026 calendar year reported capex guidance from approximately $190 billion to $175 billion. However, management clarified that actual infrastructure investment has not been reduced—roughly $15 billion has simply moved off the capex ledger. Microsoft also guided that FY2027 capital expenditure will continue to grow year-on-year.

Google accelerates capacity delivery

Google's Q2 2026 capital expenditure reached $44.9 billion, approximately doubling year-on-year and up 26% from Q1's $35.7 billion. About 60% of technology infrastructure investment went to servers, with 40% for data centers and networking. Driven by strong demand from Google Cloud and AI computing, the company raised its 2026 capex guidance from $180-190 billion to $195-205 billion (mid-point: $200 billion), citing the need to accelerate computing capacity delivery. Google also expects significant capex growth in 2027.

Meta expands through external partnerships

Meta's Q2 2026 capital expenditure reached $31.1 billion, up 82.7% year-on-year and 57% quarter-on-quarter, allocated to servers, data centers, and network infrastructure. The company narrowed its 2026 capex guidance to $130-145 billion (mid-point: approximately $137.5 billion). Notably, Meta is partnering with external capital sources like BlackRock for data center construction, meaning some AI infrastructure investment may not appear fully in traditional capex figures—suggesting actual compute expansion could exceed reported spending.

A pivotal guidance moment for Nvidia

Nvidia's FY2027 Q2 results exceeded expectations across the board, with the company issuing an extraordinary ~70% year-on-year revenue growth guidance for FY2028. Revenue reached $96.22 billion, up 106% year-on-year, with data center revenue alone hitting $89 billion—a 117% increase. GAAP net profit was approximately $59.69 billion, up 126%, with GAAP gross margins around 75%.

For FY2027 Q3, Nvidia guided revenue to approximately $108 billion (±2%), representing about 89% year-on-year growth—and notably excluding data center computing revenue from China. The unprecedented FY2028 preliminary guidance suggests this growth rate is achievable under current supply constraints, implying potential demand is even higher. This significantly enhances the certainty and visibility of medium-term AI computing growth.

Customer demand drives generational value leaps

AWS and Nvidia recently announced plans to deploy an additional 2 million Blackwell Ultra, Rubin, and Rubin Ultra GPUs in 2027-2028. This follows AWS's earlier plan at GTC this year to add over 1 million GPUs from 2026 onward—a decision revised upward in just five months due to demand exceeding expectations. The two companies are also expanding cooperation in Spectrum networking, Vera CPU, and AI Factory initiatives.

Nvidia management noted that Vera Rubin began volume shipping in August, with revenue opportunity per gigawatt of compute rising from approximately $18 billion for Hopper, to $25 billion for Blackwell, and now expected to reach about $40 billion for Vera Rubin. This escalating platform value clearly reflects robust end-user AI demand and the industry-wide shift toward full-stack system solutions.

Demand diversifying beyond hyperscalers

Nvidia's FY2027 Q2 results show that the ACIE (AI cloud, industrial, and enterprise) segment generated $40.3 billion in revenue, up 138% year-on-year—outpacing the 102% growth seen in the Hyperscale segment. ACIE now accounts for 45% of data center revenue, indicating AI computing demand is rapidly spreading from major cloud providers to emerging AI cloud services (NeoCloud), industrial, enterprise, and sovereign AI entities.

Nvidia management reported that NeoCloud partners are deploying compute capacity at lower token costs, with total installed capacity expected to grow from approximately 3 gigawatts at end-2025 to 8 gigawatts by end-2026. Diversified customers are becoming a significant growth source in this AI computing expansion.

Overall, with Blackwell products scaling massively, the next-generation Vera Rubin platform entering production ramp-up, and demand converging across hyperscalers, NeoCloud, industrial, enterprise, and sovereign AI entities, Nvidia's FY2027 Q2 revenue surge—combined with its first-ever FY2028 guidance—confirms that global AI computing capital expenditure remains in a powerful upcycle. Short-term momentum is strong, and medium-term growth certainty has been significantly reinforced.

Global large models hit commercial acceleration

Anthropic's annualized revenue exceeded $65 billion by end-July 2026, representing a 622% increase from approximately $9 billion at end-2025—a 7.2x expansion in seven months—and a 38% jump from $47 billion in May. Q2 2026 preliminary revenue surpassed $11.5 billion, up roughly 14x year-on-year. While Claude Code remains a key growth engine, the revenue mix is shifting from coding-focused offerings to broader enterprise AI solutions as agents and general workflows scale rapidly. The company recently signed $45 billion and $35 billion cloud computing partnerships with Nscale and Lambda, respectively, converting commercial demand into concrete compute expansion.

OpenAI is extending from ChatGPT toward coding, collaborative work, and diverse business models. As of July 2026, annualized revenue exceeded $40 billion, nearly doubling from end-2025, with July alone seeing over 20% month-on-month growth. Codex and ChatGPT Work combined surpassed 10 million weekly active users, with over 1 million using Codex for non-programming tasks—agents are expanding from developer tools into finance, legal, recruiting, and operations. By end-August, ChatGPT Ads reached $1 billion in annualized revenue, diversifying beyond subscriptions and APIs into agents, enterprise services, and advertising.

Chinese internet players enter monetization phase

ByteDance generated approximately $4 billion in annualized AI-related revenue in July 2026, and has integrated its Feishu product team into the Doubao ecosystem while merging sales and customer service teams into Volcano Engine to streamline AI applications, MaaS, SaaS, and cloud services. Alibaba's Q2 2026 AI cloud and computing services revenue reached 48.44 billion yuan, up 45% year-on-year, with its large model MaaS business achieving an ARR exceeding 16 billion yuan (approximately $2.4 billion)—making it one of China's largest enterprise-grade model services.

Independent domestic model providers are also scaling revenue. Zhipu's MaaS platform ARR reached $1.6 billion on a monthly annualized basis by end-August, exceeding $2 billion on a weekly annualized basis. H1 2026 open platform and API revenue grew approximately 27x year-on-year to 825 million yuan, accounting for 86.5% of total revenue, with API gross margins improving to 24.6%. Unit token inference costs have dropped 80% since the start of the year, and a 1GW-scale domestic computing center has been completed and partially operational. MiniMax's August ARR exceeded $800 million per market estimates, with H1 enterprise service revenue up 703% year-on-year. Kimi achieved approximately $300 million in June ARR and launched its Hong Kong IPO process in September. DeepSeek's latest annualized revenue has also grown substantially.

The positive feedback loop of monetization and compute

Coding has validated willingness to pay for large models, and the next growth phase is expected from enterprise agents, collaborative work, and long-horizon tasks. These require extended contexts, multi-turn reasoning, tool invocation, and repeated execution, consuming significantly more tokens per task than traditional chat interactions. Consequently, even as unit token prices decline, total call volume and inference computing demand are likely to accelerate, creating a self-reinforcing cycle between model commercialization and compute capex.

Token usage accelerates sharply

OpenRouter platform token calls surged dramatically, with August monthly volumes nearly doubling. Weekly token usage grew from 28.9T (May 18-24) to 46.7T (June 15-21)—a 62% increase in under a month—before briefly plateauing in mid-July. From late July, weekly volumes hit successive records: 69.0T, 75.3T, 93.4T, and 113T by August 24-30, representing a 21.1% week-on-week increase and roughly 99% growth over four weeks. Compared to 22.8T at end-March 2026, current usage has expanded about 4.96x; versus 5.5T at end-December 2025, that's approximately 20.5x growth in nine months.

Coding, agents, and cost-effective models are jointly driving this growth. OpenRouter data shows agent tasks consume approximately 15x more tokens per request than human interactions, with agent-generated token volume surpassing human calls since February 2026. During August 24-30, Chinese models reached 55.16T weekly calls, up 36.3% week-on-week and exceeding US models for the eighteenth consecutive week. GLM-5.3-Flash, DeepSeek V4 Flash, and Xiaomi's MiMo-V2.5 led the rankings, indicating that domestic open-source models are entering overseas developers' real workflows through capability, pricing, and open ecosystems.

The usage growth is converting to commercial value. On OpenRouter, general tasks, code, and agents accounted for 32.4%, 30.3%, and 28.7% of spending respectively, with Workflow Execution alone representing 18.7% of platform consumption—the highest-value scenario category. While OpenRouter data cannot directly equate to global large model market size or vendor revenue, it serves as a key high-frequency indicator of developer and enterprise API demand. The shift from chat and model testing toward coding, agents, and complex workflows demonstrates that large models are becoming more deeply embedded in production environments.

Capability frontier continues expanding

According to the Artificial Analysis Intelligence Index, the latest leading models now score 60-66 points, with capability boundaries continuously expanding in complex reasoning, programming, mathematics, and long-horizon agent tasks. Model iteration cycles remain in the multi-month range, with each upgrade achieving significant breakthroughs in task complexity, length, and reliability.

Overseas closed-source models still lead the global intelligence frontier. Claude series dominates the rankings, with OpenAI, xAI, and Meta's latest models in the first tier. Overseas labs maintain advantages in pretraining scale, reinforcement learning, inference computation, and agent systems engineering, particularly in complex programming, long-horizon tasks, tool invocation, and task success rates.

Domestic open-source models are catching up rapidly. Kimi K3 Max and GLM-5.3 Max both achieved AA composite scores of 60, just one point behind overseas leaders like GPT-5.6 and Grok 4.6. Qwen3.8 and GLM-5.3 Flash scored 58 and 57 respectively. Chinese models are transitioning from "low-cost, basically usable" to globally competitive, gradually closing the gap with Codex in coding and agent scenarios through open-source advantages, competitive pricing, and domestic compute adaptation.

OpenAI launched GPT-6 Astra on September 3, with official disclosures indicating a new leap forward in complex reasoning, programming, computer use, research, and multi-step tasks, with significantly improved cybersecurity and long-horizon agent capabilities compared to GPT-5.6 Sol. While GPT-6 Astra has not yet been included in AA's unified scoring, its release underscores that overseas models continue to push the global intelligence frontier upward—and the frontier capability race remains intense even as domestic open-source models advance quickly.

From answering questions to autonomous work

Claude's task-span capability has expanded from approximately 4 minutes in 2024 to at least 16 hours with Mythos Preview, with open-ended coding task success rates improving by 50 percentage points to 76% in six months. This capability has translated into real productivity: Claude generates over 80% of Anthropic's production code, and typical engineer code throughput has increased roughly 8x compared to 2024. Large models are no longer just accelerating single code-writing tasks—they are independently completing the full engineering loop of understanding codebases, devising plans, invoking tools, running tests, debugging, and delivering results.

Coding as the first practical path to RSI

Code provides low-cost validators like compilers, unit tests, and performance metrics, enabling models to repeatedly cycle through "generate, run, evaluate, improve." Claude's optimization capability for small model training programs improved from roughly 3x with Opus 4 to about 52x with Mythos. In open-ended research projects, multi-agent systems recovered 97% of target performance gaps, whereas two human researchers recovered only 23% in a week. Models are no longer just writing production code—they are improving training programs, designing and running experiments, analyzing results, and participating in the R&D process required to create stronger AI systems.

The next phase involves progressing from coding agents to AI research agents, gradually forming an RSI loop. Models first generate better code and algorithms, then modify prompts, memory, tools, and agent harnesses, before optimizing training data, reward functions, training frameworks, and even model parameters. When improved models become better at developing subsequent models, a recursive cycle emerges: "AI improves AI—successor models grow stronger—R&D accelerates further."

Currently, Claude approaches or exceeds skilled humans in well-defined research execution tasks, but research question selection, evaluation criteria, and final validation remain primarily human-controlled. The path to fully autonomous design and training of successor models—strong RSI—still requires a critical step. RSI refers to recursive self-improvement: AI not only completes tasks but also participates in improving the R&D process for building the next generation of AI. "Recursive" means each improvement round enables the next round of improvement, creating a compounding effect across iterations.

RSI can occur at various levels: modifying current responses, retaining past experience, accumulating new skills, generating training data, designing reward rules, optimizing training code, and proposing research approaches. The key is whether these improvements can be preserved and applied in subsequent tasks or model training.

Anthropic categorizes AI participation in AI R&D into five stages: human-independent research, chatbot assistance, coding agents, autonomous agents, and closed loops. Anthropic assesses itself as currently in the fourth stage—autonomous agents. The next stage is a "closed loop" where AI autonomously designs, trains, and improves successor models—the state that complete RSI aspires to achieve.

As global large models enter a phase of rapid commercialization, with token consumption doubling monthly, Nvidia delivering unprecedented forward guidance, cloud providers raising capex targets, and diversity in demand spreading across hyperscalers, NeoCloud, enterprises, and sovereign entities, the AI computing power cycle shows no signs of abating. The convergence of short-term momentum and strengthened medium-term visibility confirms that AI infrastructure investment remains on a robust upward trajectory, with implications for supply chains both in North America and domestically in China.