The latest generation of AI models and a new intelligent agent have been launched, signaling a profound shift in the industry's competitive landscape and reinforcing the long-term demand for computing infrastructure.
AI application leader OpenAI has officially rolled out its GPT-5.6 series of models and simultaneously introduced a significant new tool: the ChatGPT Work intelligent agent. This AI agent is designed to operate continuously across applications for multiple hours, handling broader and more complex AI workloads. This move advances OpenAI's strategy to attract more business professionals. The joint release of the GPT-5.6 series and ChatGPT Work signifies a core shift beyond merely increasing model parameters. It aims to propel artificial intelligence from a "content generation tool" to a transformative "infrastructure for autonomous, high-efficiency productivity."
While market focus in recent years centered on parameter scale, reasoning capability, and benchmark rankings, GPT-5.6 emphasizes token efficiency, task completion, and the automation of enterprise workflows. This indicates that AI's commercial value is migrating from "answering questions" to "autonomously completing tasks." The key highlight of ChatGPT Work is its ability to connect with enterprise systems like files, applications, email, calendars, and CRMs. This evolution allows AI to transition from an assistive tool into an agentic, workflow-driven "digital employee" capable of executing complex processes and large-scale projects like engineering designs.
From a technological competition standpoint, GPT-5.6's release sends a strong signal that the AI industry is entering a phase of intense dual competition focused on both "intelligence level" and "unit cost efficiency." The newly launched GPT-5.6 series includes three versions: the flagship Sol, the balanced Terra, and the cost-effective Luna. Sol targets complex reasoning, agentic programming, cybersecurity, and scientific research tasks. Terra emphasizes balanced capabilities for everyday enterprise applications, while Luna is aimed at large-scale, low-cost deployment.
OpenAI's CEO, Sam Altman, highlighted that GPT-5.6 achieves a 54% improvement in token efficiency for agentic programming tasks. This means enterprises can accomplish more agentic AI workflow tasks under the same computational budget. The launch of GPT-5.6 and ChatGPT Work, coupled with the latest analysis from research firm SemiAnalysis showing Anthropic's transition from long-term losses to a period of explosive profit growth, sends a crucial signal to global equity markets.
The AI computing power supply chain is gradually moving from the "super-cycle of AI capital expenditure for training large models" into a new phase of "exponential expansion in AI inference computing demand driven by the large-scale application of agents." These developments directly counter recent pessimistic narratives about "computing power oversupply," which had contributed to declines in the AI computing theme, particularly within the AI semiconductor sector.
Wall Street investment firm Nomura has published a report refuting "peak semiconductor" theories. Meanwhile, a recent report from Bank of America indicates that, driven by the massive wave of AI agents and the continued surge in AI inference computing power, global capital expenditure on cloud computing and AI-related infrastructure is projected to reach $1.5 trillion by 2027. The report suggests the current summer pullback in AI semiconductors, including memory chip stocks, represents a healthy reset trajectory rather than any structural change in underlying AI computing demand.
From Goldman Sachs' perspective, the AI bull market is far from over. It is transitioning from a first phase of "AI chip purchasing frenzy" into a second phase of "large-scale construction of AI factories." Consequently, the next wave of excess alpha returns will not be confined solely to the strongest leaders in AI GPUs or AI ASICs. Instead, it will systematically spread across the full stack of AI computing infrastructure for "AI factories," including data center high-performance CPUs, DRAM/NAND/HBM memory, AI PCBs, liquid cooling systems, data center optical interconnect systems, ABF substrates/glass substrates, MLCCs, electronic fabric, and a broad range of wafer foundries.
GPT-5.6 Ignites the AI Agent Era: OpenAI Evolves from 'Chatbot' to Enterprise Digital Employee Platform
Test data released by OpenAI shows GPT-5.6 Sol achieving industry-leading levels in multiple evaluations for agentic coding, long-cycle tasks, and security. It enhances complex task execution through its "Sol max" deep reasoning mode and "Sol ultra" multi-agent collaboration mode. Concurrently, its core commercial logic revolves around reducing inference costs: fewer token consumption, fewer tool calls, and shorter execution times mean lower marginal costs for enterprises adopting AI, which will directly accelerate the commercial penetration of AI software.
Compared to competitors like Anthropic, OpenAI is attempting to capture the complete market chain—from large enterprises to developers to ordinary office users—through a combined strategy of "flagship model performance leadership" and "price reduction for mid-to-low-end models." The newly launched tool, ChatGPT Work, is designed to help users create documents, spreadsheets, presentations, and web applications. It is fully powered by the GPT-5.6 large model, which was officially launched on Thursday after a delay due to regulatory intervention.
OpenAI and its long-time strongest competitor in the AI application space, Anthropic PBC, have been racing to develop more advanced AI agents to streamline workflows across broader domains. Both companies have previously achieved notable success with AI development tools capable of automating code writing and the complete debugging and deployment process. Earlier this year, Anthropic launched a similar product named Claude Cowork, aiming to attract a wider user base into the unprecedented super-wave of AI agents.
Both OpenAI and Anthropic have confidentially filed for public listings. Reports indicate Anthropic could go public in the US stock market as early as this autumn, while OpenAI is considering a listing next year. From the perspective of the coding benchmark Terminal-Bench 2.1 test metrics, GPT-5.6 Sol Ultra ranks first with a score of 91.9%, followed by GPT-5.6 Sol at 88.8%. The competing product Claude Mythos 5 ranks third at 88.0%, a gap of about 0.8 percentage points. Gemini 3.1 Pro Preview ranks last at 70.7%, showing a significant gap from the top tier.
For enterprise users and developers, the core impact of this release is a comprehensive improvement in cost-performance. In the Agents' Last Exam test for professional workflows, GPT-5.6 Sol scored 53.6, exceeding Claude Fable 5 by 13.1 percentage points. Even with medium reasoning settings, its cost is approximately one-quarter that of Fable 5. In terms of cost efficiency, GPT-5.6 Sol consistently achieves the highest scores under the same API cost, giving it the best cost-performance in the series. Some internal tests even show its cost-performance far exceeds that of Claude Mythos 5.
GPT-5.5 and GPT-5.6 Luna show a clear "cost bottleneck"—increased investment yields very limited performance gains. Regarding reasoning capability, as the number of output tokens increases, GPT-5.6 Sol's score improvement slope is the steepest, sufficiently demonstrating its ability to most effectively utilize complex reasoning processes to enhance output quality. The flagship Sol version strengthens competitiveness in complex programming, agentic tasks, and scientific reasoning while lowering actual deployment costs through higher token utilization efficiency. This directly challenges the premium space established by competitors like Anthropic, which rely on high-performance models.
The future core competition in the AI industry may no longer be about who possesses the highest single-point capability, but rather who can replicate model capabilities into hundreds of millions of workflows at the lowest inference cost. OpenAI CEO Sam Altman stated in a media interview on Thursday that the company's latest AI large model, GPT-5.6 Sol, has improved token efficiency for agentic programming tasks by 54%, and its performance is "as good as, or better than, competing models in the market."
"Every company is now thinking about spending and the value they get from AI, and that's what we really want to achieve," Altman said. He described the company's collaboration with the government as a "collaborative back-and-forth process" involving testing and problem-solving. "If you want to achieve widespread use—which is our goal—and you have a powerful model, you really want to be confident in your safety claims, because otherwise the world will become uneasy very quickly," he added.
Altman expressed hope for a global approach to regulation where people can use AI without constantly worrying about safety. "Everyone will have access," he said. "It's not that the US will get a disproportionate advantage here." If AI agents can continuously reduce corporate costs in R&D, operations, sales, financial analysis, and administration, they could drive a new wave of labor productivity improvement and further expand the contribution of cutting-edge AI technology to global economic growth.
The competition among OpenAI, Anthropic, Google, Microsoft, Amazon, Meta, and Elon Musk's AI ecosystem is essentially evolving into a war for the entry point to the next generation of enterprise digital infrastructure. OpenAI currently holds a private valuation of approximately $852 billion from investors, and Anthropic is also preparing for a potential IPO. This indicates capital markets are beginning to price in the long-term cash flow value brought by the commercialization of AI agents.
OpenAI is attempting to establish an ecosystem position akin to an operating system-level entry point from the mobile internet era: models provide intelligent capabilities, agents are responsible for executing tasks, and enterprise data becomes the continuous optimization fuel. If this closed loop forms, AI competition will no longer be just about models, but about "who can become the global enterprise digital labor infrastructure."
Overall, the significant importance of GPT-5.6 lies in its further confirmation that the AI industry is transitioning from a "model competition" phase into a new stage of "AI employees, AI software platforms, and enterprise productivity revolution." Companies that master the agent ecosystem entry point are poised to become core assets in the next technological capital cycle.
Capital May Continue Flocking to the 'Silicon-Based Inflation' Theme! The AI Semiconductor Summer Pullback Presents a Buying Opportunity
Past market concerns centered on the potential premature exhaustion of AI infrastructure investment. However, GPT-5.6's demonstrated 54% improvement in token efficiency for agentic programming, its cross-application autonomous task execution capability, and ChatGPT Work's penetration into enterprise office scenarios suggest that improvements in AI model capabilities will not simply reduce computing demand. Instead, by lowering the cost per task, expanding the user base, and increasing enterprise call frequency, they are likely to create even more substantial long-term inference demand.
OpenAI's release of three models at different price points—Sol, Terra, and Luna—essentially aims to expand token consumption scale by lowering the barrier to AI use, providing stronger demand certainty for future AI computing infrastructure. The latest analysis from research firm SemiAnalysis reveals that Anthropic is reshaping the AI commercialization landscape with profitability and growth rates far exceeding its competitors. Leveraging a high-margin, API-centric business model, Anthropic has become a leader in the B2B AI market.
SemiAnalysis's in-depth report forecasts that Anthropic will achieve $10 billion in GAAP EBIT by the third quarter of 2026, corresponding to a roughly 6% margin. Concurrently, its Annual Recurring Revenue (ARR) has skyrocketed from $9 billion at the end of 2025 to over $60 billion currently. The firm predicts that if Anthropic maintains a net new ARR (NNARR) pace of approximately $15 billion per month, its ARR could reach $3 trillion by the end of 2027, corresponding to a $60 trillion enterprise value, making it the world's highest-valued company.
Anthropic's inflection point stems from the explosive adoption of Claude Code. Statistics exclusively compiled by SemiAnalysis show that Claude Code currently accounts for over 7% of all code commits on GitHub, directly driving the company's monthly ARR additions to surge from $3 billion in January to $11 billion in March. Furthermore, SemiAnalysis estimates show Anthropic's current consolidated gross margin has risen to the mid-60% range, compared to negative 94% in 2024, with its API business gross margin exceeding 80%.
The grand investment narrative of "seeking silicon-based inflation, weakening carbon-based" assets pursued by global capital this year essentially represents a shift from traditional "carbon-based assets" reliant on population, resources, and linear economic growth—such as manufacturing, autos, consumer goods, real estate, and energy—towards the high-end manufacturing supply chain centered on silicon wafers for AI computing infrastructure.
Therefore, the arrival of GPT-5.6 and ChatGPT Work, combined with Anthropic's commercial data, reinforces a core investment judgment: the unprecedented demand cycle for AI computing infrastructure is not over. It is transitioning from a phase driven by AI large model training to one driven by AI inference-side applications. The true super-cycle for AI computing infrastructure may stem from the global, large-scale deployment of AI agents as a new generation of digital employees by enterprises. This also implies that the current pullback in the AI semiconductor sector is a healthy adjustment, not a bear market crash driven by "computing power oversupply."
The unparalleled preliminary Q2 results just disclosed by South Korea-based memory chip giant Samsung Electronics serve as the most直观的 profit sample for this memory chip super-cycle. Operating profit for April to June is estimated to have soared approximately 19-fold year-over-year to 89.4 trillion won (about $584 billion), setting another quarterly record and representing 56% sequential growth from a strong base. Revenue for the period is expected to reach 171 trillion won, exceeding market estimates of 169.2 trillion won and marking about 129% year-over-year growth. The company plans to announce full results on July 30, disclosing net profit and segment data.
Samsung Electronics' quarterly operating profit has surpassed NVIDIA's (NVDA) last quarter operating profit of $53.536 billion (approximately 82 trillion won), making it the company with the highest quarterly operating profit globally. In Goldman Sachs' view, the global bull market surrounding the AI computing chain is far from finished. The market's main theme has upgraded from the long-standing "software valuation expansion driven by programming/code and light-asset software" to a "re-pricing of a series of physical assets for AI computing infrastructure."
Goldman Sachs' latest estimates suggest that hyperscale cloud providers' total investment in AI-related infrastructure could exceed $6 trillion by 2030. The global AI capital expenditure base model is projected to grow from $765 billion annually in 2026 to $1.65 trillion annually in 2031, with cumulative capital expenditure from 2026 to 2031 estimated at around $7.6 trillion. US data center power demand is expected to rise from 31 GW in 2025 to 66 GW in 2027.
Wall Street firm Nomura recently published a report refuting "peak semiconductor" theories. The key to Nomura's rebuttal is not simply stating that AI chips will continue to rise, but pointing out that AI cloud infrastructure demand is diffusing from a single-point GPU shortage into a systemic component mismatch. According to Nomura's research framework, AI server revenue is projected to grow 78% and 76% in 2026 and 2027, respectively. The number of global data center projects is expected to increase from 240 to 280, with about 50 being gigawatt-scale projects. New computing power deployment in 2027 is forecast at 32 GW, with 23 GW already visible for 2028. However, the real bottleneck is spreading from NVIDIA AI GPU and Google TPU capacity, as well as TSMC's CoWoS advanced packaging, to memory chips, wafer-level substrates, AI PCBs, copper-clad laminate (CCL), electronic fabric, MLCCs, glass substrates/ABF substrates, IC substrates, high-end capacitors, power management chips, and data center optical high-speed interconnect components.
A team of Bank of America analysts led by veteran strategist Vivek Arya stated: "The AI semiconductor-driven market surge is not over. After a record 88% surge in Q2, the Philadelphia Semiconductor Index (SOX) pulled back 11% in Q3, consistent with its historically weakest seasonal period. We view the current pullback as a healthy reset, not any structural change in AI demand. This reset is expected for the summer, with a potential rebound in the fall. Short-term leadership may favor lower-beta stocks like NVIDIA (NVDA), Texas Instruments (TXN), Analog Devices (ADI), and the two leading chip design EDA firms—Cadence Design Systems (CDNS) and Synopsys (SNPS). However, historical experience shows that following consolidation periods, new momentum often emerges as investors regain strong confidence in the next cycle of earnings and capital expenditure growth."
Arya and his team added: "We expect global cloud and AI computing infrastructure capital expenditure to approach $1.5 trillion by 2027, implying potential for another 40% to 50% year-over-year growth. This is strongly supported by continued token scale growth, a surge in enterprise AI agent adoption, and constrained infrastructure supply. Importantly, the focus for hyperscale cloud providers remains on maximizing utilization and AI-driven performance growth trajectories, not on optimizing depreciation."
Comments