According to a research report from China Securities Co., Ltd., the first-half earnings forecasts from A-share computer companies and overseas model iterations jointly confirm the sustained high景气 of the AI industry chain. On one hand, hardware segments such as AI servers and intelligent computing infrastructure have shown outstanding earnings elasticity, while software and AI application companies have begun to demonstrate operational improvements and revenue realization. On the other hand, overseas models from OpenAI, xAI, and Meta continue to strengthen capabilities in agents, coding, multimodality, and office entry points. Model competition has shifted from capability verification to high-frequency scenario implementation, and the consumption of inference computing power, infrastructure investment, and the commercialization of AI applications are expected to resonate further.
The first-half earnings forecasts from A-share computer companies are being disclosed gradually, with the AI computing hardware chain maintaining strong performance and software companies showing earnings recovery. The AI computing chain remains the area with the most definitive earnings elasticity. Inspur Electronic Information Industry Co., Ltd. is expected to report a net profit attributable to the parent company of 2.600 to 3.100 billion yuan in the first half of 2026, representing a year-over-year increase of 226% to 288%. Unisplendour Corporation Limited is expected to report a net profit of 1.910 to 2.320 billion yuan, up 83.50% to 122.89% year-over-year. Against the backdrop of expanding demand for AI servers, intelligent computing infrastructure, and network equipment, the revenue and profit elasticity of the hardware side are still being realized, maintaining a high industry sentiment. The operational quality of software and security companies is marginally improving, with an increasing number of signals pointing to earnings recovery. SIE Consulting Co., Ltd. is expected to report a net profit of 0.053 to 0.065 billion yuan in the first half of 2026, a year-over-year increase of 188.36% to 257.01%, mainly benefiting from improved delivery efficiency, cost control, and overseas business expansion. Venustech Group Inc. is expected to report a net profit of 0.026 to 0.038 billion yuan, turning a profit year-over-year, with revenue resuming growth and a significant improvement in operating cash flow. Qihoo 360 Technology Co., Ltd. is expected to report a net profit of 0.180 to 0.260 billion yuan, also turning a profit year-over-year. Overall, while performance divergence within the computer sector remains significant, the high景气 of computing hardware, cost reduction and efficiency improvement by software companies, and the advancement of AI productization are supporting the fundamentals of the sector's first-half earnings.
Overseas models are undergoing intensive iterations, with agents, multimodality, and office entry points becoming the core of this round of competition. This week, OpenAI, xAI, and Meta successively released models or products such as GPT-5.6/GPT-Live, Grok 4.5, Muse Spark 1.1, and Muse Image. In terms of performance direction, the updates focus on long-horizon tasks, multi-agent collaboration, coding, computer use, tool invocation, real-time voice, and image/video generation. Model competition is shifting from single-turn Q&A and static benchmarks to task completion rates, token efficiency, response speed, and product entry points in real-world workflows. OpenAI's frontier models have been further upgraded, with GPT-5.6 enhancing capabilities in long-horizon agents, coding, and computer use. On July 9, 2026, OpenAI released the GPT-5.6 series, including the Sol, Terra, and Luna models, priced at $5/$30, $2.5/$15, and $1/$6 per million tokens for input/output, respectively. In terms of performance, GPT-5.6 Sol has shown significant improvements in coding agents, computer use, network security, and long contexts. Official disclosures indicate its Artificial Analysis Coding Agent Index reached 80, SWE-Bench Pro reached 64.6%, and OSWorld 2.0 reached 62.6%, showing continued improvement over GPT-5.5 in most engineering and computer use tasks. The GPT-Live, released on July 8, enhances full-duplex voice capabilities, allowing simultaneous listening and responding, improving real-time conversations, natural interruptions, and continuous interaction experiences. The core of OpenAI's update is to translate frontier reasoning capabilities into tasks within Codex, voice, and office domains, expanding inference consumption from text-based Q&A to long-horizon tasks and real-time interactions.
xAI released Grok 4.5, with coding agent capabilities essentially reaching the level of GPT-5.5, and a low-price strategy intensifying competition for developer entry points. On July 8, 2026, xAI released Grok 4.5, targeting coding, agentic tasks, and knowledge work. Its API pricing is $2 per million tokens for input and $6 for output, significantly lower than GPT-5.5's $5/$30, representing a 60% lower input price and an 80% lower output price. In terms of performance, Artificial Analysis shows Grok 4.5 has a comprehensive intelligence index of 54, ranking behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. However, it achieved a score of 76 on the Coding Agent Index, essentially matching GPT-5.5's performance in Codex. Additionally, its single-task cost for coding agents is approximately $2.49, lower than GPT-5.5 Codex's $5.07, and its token usage is also significantly lower. Although Grok 4.5 does not fully surpass GPT-5.5, it has entered the same capability band in developer workflows, forming a differentiated competition through lower pricing and higher token efficiency.
Meta is addressing its shortcomings in agents and multimodality, using low-cost APIs and social traffic entry points to accelerate AI productization. On July 9, 2026, Meta released Muse Spark 1.1, a multimodal reasoning model for agentic tasks, supporting a 1M token context and focusing on improving tool use, computer use, coding, and multimodal understanding. According to media and developer sources, the Muse Spark 1.1 API is priced at $1.25 per million tokens for input and $4.25 for output, lower than Grok 4.5 and OpenAI's GPT-5.6 flagship models. On July 7, Meta released Muse Image and previewed Muse Video. Muse Image has been integrated into entry points like Meta AI, Instagram Stories, and WhatsApp, ranking second in arenas for text-to-image, single-image editing, and multi-image editing. On the infrastructure front, computing power leasing, high capital expenditure, and data center expansion collectively reflect Meta's continued investment in AI infrastructure. Citing Bloomberg from July 1, TechCrunch reported that Meta is planning a "Meta Compute" cloud infrastructure business, intending to sell managed model services and AI computing power to external customers. Potential models include opening access to Meta's model hosting or renting out raw GPU computing power similar to CoreWeave. In its first-quarter report, the company raised its 2026 capital expenditure guidance to $125-145 billion, primarily for AI infrastructure and data center capacity expansion. On July 8, Meta announced an investment of approximately $9.1 billion in Canada to build its first AI data center in the country, also one of its largest outside the United States. Meta is catching up with frontier model companies through low-cost models, social traffic entry points, computing power leasing, and high-intensity infrastructure investment. Its future demand for GPUs, CPUs, networks, storage, and data center resources is expected to remain strong, further validating the commercial potential of AI computing power leasing services.
Domestic model companies continue to strengthen their investment in AGI, with financing, team incentives, and open-source ecosystems sending long-term signals. On July 11, according to an internal letter from Zhipu AI founder Tang Jie titled "The Giant Wave Has Arrived," the company will strategically invest in the "Touch High Plan" over the next two years, focusing on long-horizon task capabilities, autonomous agent systems, and self-evolution. It will also continue its direction of open-sourcing GLM-5.2 and security governance. According to public reports, Zhipu AI recently raised approximately 31.4 billion Hong Kong dollars through a placement, with funds primarily used for foundational model R&D, computing infrastructure construction, commercial expansion, and global ecosystem layout. For MiniMax, CEO Yan Junjie issued a company-wide letter on July 10, announcing he would forgo his salary and allocate personal shares equivalent to 4% of the company's total equity for team incentives and 1% to support the open-source community. On the same day, the company completed a 16 billion Hong Kong dollar financing round, intending to use 80% of the net proceeds for AI infrastructure and model R&D. On one hand, leading domestic model companies are strengthening their teams and long-term commitments amid capital market fluctuations; on the other hand, they are increasing investment in AGI, coding, agents, and infrastructure expansion. Their subsequent pull on computing power, data, developer ecosystems, and AI application commercialization remains noteworthy.
Comments