China Securities Co., Ltd.: Software sector deeply retraces, seeking opportunities for mispriced stocks.

China Securities Co.,Ltd. released a research report stating that the current window of opportunity is for deep differentiation pricing of the software sector. The current panic selling in the software sector provides a window for differentiation pricing based on deep barriers. The company recommends configuring along the "barrier offense and defense properties": 1) Overweight core "offensive barrier" targets, where the barrier itself becomes a company with a new growth foundation in the AI era; 2) Pay attention to the "defense to offense transformation" window, where traditional barriers are solid but companies need to validate AI revenue conversion; 3) Avoid targets with short barrier attenuation time constants, such as companies with low complexity and shallow data. Focus on progress in cooperation between OpenAI, Anthropic, and consulting companies, as well as the inflection point in pricing and ARR growth for enterprise software companies in AI. China Securities Co.,Ltd.'s main points are as follows: Model homogenization is accelerating at the layer, but not completely homogenizing Based on empirical research data from MIT and other teams on OpenRouter, closed-source model calls still account for 75% of demand, with short-term demand price elasticity of only -1.11 (close to unit elasticity), indicating that the big model market is presenting a "brand-centered competition" pattern, with extremely low internal conversion costs and very high costs for cross-ecosystem conversion. However, the lead time of top models has shortened from 7.5 months in early 2025 to within 4 months, the gap between the top and tenth place in GPQA continues to narrow, and the premium window for model manufacturers is narrowing. OpenAI's gross profit margin in 2025 has decreased from 40% to 33%, Anthropic's gross profit margin is 40% (lower than expected 50%), and the inference cost is 23% higher than internal expectations. The profit space at the model layer is being squeezed by the computational power consumption and API price deflation of test-time compute, forcing model manufacturers to penetrate the application layer. The "cost of errors" framework proves that high-value vertical fields still require the strongest models, and AI cannot uniformly replace all software Based on data from medical malpractice litigation in the United States, the expected economic loss for each mistake made in medical diagnostic scenarios by AI is $45-63 (adjusted), far exceeding the cost threshold of initially adopting high-performance inference models. In credit approval scenarios, considering false negative credit losses and false positive opportunity costs, the weighted cost per single error ranges from $27 to 125 (depending on the balance caliber). This means that in high-cost scenarios, the economic value supported by increasing model accuracy from 90% to 95% can reach tens or even hundreds of times the price premium, and the real competition focus is not on whose API is cheaper, but on who can approach expert-level accuracy in vertical fields. Scaling continues to advance in three parallel directions, and the "singularity path" in vertical fields is clear Empirical evidence from 2025-2026 shows that model performance improvement comes from: 1) RL/RLVR algorithm improvements (GRPODAPODr.GRPO-GRPO), shifting from human-annotated rewards to automatic validation; 2) Inference-time Scaling (Deep Think parallel thinking, Agent Swarm parallel sampling, thinking efficiency optimization), DeepSeek R1-0528 increased AIME accuracy from 70% to 87.5% by adding post-training computational power; 3) Architecture efficiency improvements (MoE sparse activation, linear attention, sparse attention), Kimi K2.5's PARL training reduced end-to-end runtime by 80%. In the background where the marginal benefits in the three directions have not quickly converged, accuracy in vertical fields will continue to improve, and the real moat path is "Mid-training injecting industry knowledge constructing a verifiable reward environment RL stimulating deep inference Test-time thorough thought". Deep differentiation in software barriers, AI impact differentiation The value of enterprise software does not lie in the code itself, as 96% of commercial software includes open-source code, but companies still pay for security, compliance, integration, and SLAs. In the AI era, barriers are differentiated along two dimensions: "workflow complexity depth of data moats": 1) High complexity + deep data (such as ServiceNow), where the value of workflow orchestration, context management, and compliance auditing is enhanced by AI; 2) Low complexity + deep data (such as HubSpot), where data is valuable but the seat-based billing logic is facing structural compression after AI improves efficiency; 3) Low complexity + shallow data (such as Five9, Freshworks), where core functions are directly covered by AI agents, with extremely thin moats. The failure of BloombergGPT proves that the "self-built model" route is not feasible, as GPT-4 surpassed proprietary models trained from scratch on 3.63 billion token financial data in less than a year; whereas successes like Hebbia ($13 billion valuation) and Harvey (ARR > $100 million) prove that the correct value capture method is "proprietary data + workflow + cutting-edge general models". Strong barriers come from the "abstraction encoding of business practices and legal regulations", and the replication cost of AI-native software is extremely high Taking SAP as an example (77% of global transaction revenue involves its systems), its barriers come from three layers of nesting: 1) Business rule encoding (executable logic of various national tax/labor laws/compliance regulations), 2) Irreversibility of customized configurations (tens of thousands of configuration parameters, hundreds of custom reports, decades of organizational memory accumulation), 3) Ecosystem lock-in (hundreds of thousands of certified consultants, S/4HANA mandatory migration turning into a re-locking event). Former SAP engineer Thomas Otter pointed out that many functions are not "business choices" but legal requirements (e.g., German payroll calculations involving church tax, social security apportionment, and dozens of interdependent variables), where a 0.01% error in payroll could lead to legal action. The threat of AI to core ERPs is "layered penetration" rather than "replacement", as UI/interaction layer and process automation layer (L1-L2) AI has enhanced these systems (such as SAPJoule), but in the foreseeable future, core business logic layer (L4, 2028+) and system record layer (L5) remain tools for enhancement rather than replacements. At the same time, AI is compressing the product development value chain, with internal practices at Anthropic showing a compression from weeks to hours in the cycle from idea to prototype, resulting in a decrease in the value of UI/UX as intermediate products and design tools like Figma facing the risk of "overall compression of the design phase", but barriers in visual collaboration and design system management still exist in the short term. Investment advice The current panic selling in the software sector provides a window of opportunity for pricing based on deep barriers. The company recommends configuring along the "barrier offense and defense properties": 1) Overweight core "offensive barrier" targets, where the barrier itself becomes a company with a new growth foundation in the AI era (such as Palantir, Ontology-driven AI platform, US commercial revenue +109%; ServiceNow, AI Control Tower positioned as an "AI agent governance platform", Now Assist ACV > $600 million targeting $1 billion); 2) Focus on the "defense to offense transformation" window, where traditional barriers are solid but companies need to validate AI revenue conversion (such as SAP, cloud revenue +26%, Cloud ERP Suite +33%, $64 billion cloud backlog; Salesforce, Agentforce ARR > $800 million but needs to shift from price-driven to AI consumption-driven); 3) Avoid targets with short barrier attenuation time constants, such as companies with low complexity and shallow data (such as Five9, revenue growth declining from 12% to 8%, the company has acknowledged the threat of AI replacement in its risk factors; UiPath, growth declining to 5-6%, RPA core value being directly replaced by AI agents). Follow-up on progress in cooperation between OpenAI, Anthropic, and consulting companies, as well as the inflection point in pricing and ARR growth for enterprise software companies in AI. Risk analysis If the pace of Fed rate cuts is slower than expected or the risk of global economic recession intensifies, it may lead to a comprehensive contraction in enterprise IT spending, further putting pressure on software sector valuations; The rate of improvement in model capabilities may exceed expectations, leading to AI agents penetrating medium to high complexity workflows faster than expected, putting targets currently considered "solid defensive barriers" at risk of erosion ahead of schedule; Failure to meet AI revenue growth expectations could pose significant risks of valuation retraction.