Woter AI detection.Hurry - ends Jul 16th

New Year Sales :up to 80% OFF

AI Humanize AI Translator Bypass AI AI Rewriter AI Detector

PRICING

TRY FOR FREE

An investor’s hard-core observation: AI investment paradigm rewritten by DeepSeek and Manus

Written by

Caleb Hayes

Updated on:July-12th-2025

In the past two years, Zhou Wei, founding managing partner of Genesis Partners Venture Capital, has been intensively investigating AI projects in China and the United States. He has always been puzzled by one question: "Except for the underlying large model companies, application layer companies rarely focus on providing truly complete solutions. Every product can bring surprises, but few products can truly and thoroughly solve users' problems."

Coincidentally, at the beginning of March, Zhou Wei’s discussion with Peking University Professor Hou Hong on AI also focused on this: “When will there be products that truly break through the application threshold—that is, applications that can completely replace humans and completely solve problems in specific scenarios?” The next day, Manus came out.

Zhou Wei believes that the window of opportunity for startups at the bottom of the big model has basically closed, both in China and the United States. The real opportunity lies in accurately capturing user needs like Manus, breaking through the threshold of user experience, and truly providing a complete "end-to-end" solution. " Deepseek sounded the alarm for the world: big models can still be done this way; Manus set a benchmark for global application developers: applications should be done this way. Both brought epiphany-like revelations, but at different levels: the bottom layer and the application layer. There is no need to use the standards of the bottom-level technology to measure the application-level products."

Obviously, the competition logic of the AI industry is undergoing profound changes starting from 2025. The arms race of technical parameters is gradually cooling down, and the industry's focus is shifting from "competing on parameters" to "competing on ecology."

"Without the support of the application layer, the underlying technology can only be self-entertaining; only by truly moving towards the application layer can AI achieve universal value." Zhou Wei's judgment reveals the key to the development of the industry - from the fantasy of simply pursuing technological breakthroughs to the efficiency revolution of ecological synergy, this transformation has just begun.

Ecological reconstruction

From "three-layer pyramid" to "network symbiosis"

The current AI ecosystem has formed a clear "three-layer architecture": Foundation LM, AI infrastructure (MLOps/AIOps/data orchestration) and application layer (2P/2C/2B). However, this architecture is evolving towards "network symbiosis" - the basic model vendors integrate computing power and data downward, the application layer optimizes interactions and scenarios upward, and the open source ecosystem becomes the glue connecting the various layers.

This evolution is not only a technological upgrade, but also a reshaping of the business model. The winners in the future will no longer be pure technology leaders, but ecological players who can deeply integrate technology, scenarios and business logic. When investing, we pay more attention to how companies integrate into and reshape this ecological "network" rather than simply leading in technology for a few months.

Basic big model: open source and proprietary models coexist

There will not be a winner-takes-all situation in the field of large models, and open source and proprietary models will coexist for a long time. Open source models are becoming the "water, electricity and gas" of infrastructure, while proprietary models are customized tools for high-value scenarios.

For example, DeepSeek is rapidly gaining the attention of developers with its inference cost advantage (only 1/20 of GPT-4), and Meta has to accelerate the development of Llama 4. However, the pattern of this field has been initially determined, and the opportunity window for early venture capital has basically closed.

AI infrastructure: new opportunities for edge computing and energy optimization

The field of AI infrastructure is experiencing "ice and fire": the market for pure computing infrastructure is shrinking, while edge computing and energy optimization technologies are experiencing explosive growth.

Taking NVIDIA's Jetson Orin Nano chip as an example, its energy efficiency ratio has increased by 5 times, providing technical support for edge scenarios such as XR devices and in-vehicle AI. Yopao Technology, which we invested in, has reduced the cost of collecting road data for dynamic pricing systems by 80% through edge computing technology.

The key to future infrastructure is no longer just computing power, but the comprehensive optimization of energy consumption and cost.

Application layer: competition logic in vertical fields

The value of AI lies not in what it can do, but in what problems it solves. The competition logic at the application layer is clearer: scene depth and data moat are the real competitive advantages.

2P applications (for professionals): such as the legal contract generation tool Harvey and the medical image analysis platform Arterys, which fine-tune models through vertical data and become the "super co-pilot" of doctors and lawyers.
2C applications: OpenAI and Microsoft have occupied general scenarios with Copilot, but there are still opportunities in vertical fields. For example, the financial technology company Upstart used AI to reconstruct the credit assessment model, reducing the bad debt rate by 30%; the Chinese startup "Shenyan Technology" launched the AI psychological assistant "Heart Mirror", with a user retention rate of over 70%, confirming the deep dependence of Generation Z on AI.
2B application: The core lies in "data moat + scenario closed loop". Retail giant Walmart uses RAG technology to link inventory data with consumer behavior, implements a dynamic pricing system, and increases inventory turnover by 18%.

Five major trends predicted for the AI industry in 2025

Trend 1: Edge Intelligence Revolution - SLM and TinyLM usher in the era of "device awakening"

We believe that "edge intelligence will become the next technological explosion point." Thanks to distillation training and hardware adaptation technology, the TinyLM can run smoothly on devices with 50 TOPS computing power, bringing revolutionary changes to scenarios such as smart glasses and industrial sensors.

Case: The case of Silicon Valley startup Nexa AI is quite representative: its 7B model realizes multi-round dialogue on Raspberry Pi, with a response delay of less than 0.5 seconds and a power consumption of only 2W.

Investment logic: The end-cloud collaborative architecture is the key to edge intelligence. For example, Tesla's Optimus robot uses the end-side SLM to process real-time environmental perception and the cloud-based large model to optimize the decision path. The collaboration between the two increases the training efficiency by 3 times.

Trend 2: Cost cliff effect, DeepSeek reshapes AI industry economics

When the cost of inference drops by an order of magnitude, the rules of the game for the entire industry will change. DeepSeek's technological breakthrough is a watershed for the AI industry. We have observed that this cost cliff effect directly triggers three major chain reactions. The valuation logic of the capital market has also begun to shift from "competing for computing power" to "competing for scenarios", and the valuation premium of AI applications that can run through the unit economic model exceeds 30%.

Reduced hardware dependence: The low-cost model reduces the dependence on high-end GPUs, and some scenarios can be replaced by CPUs.
Reconstruction of the open source ecosystem: More than 50% of Llama developers worldwide turned to DeepSeek, and Meta was forced to accelerate the development of Llama 4.
The valuation logic has been overturned: capital has shifted from "competing in computing power" to "competing in scenarios", and AI applications that can run unit economic models have a valuation premium of 30%.

Trend 3: The rise of AI Agent ecosystem: a qualitative change from “tool empowerment” to “organizational reconstruction”

Product power leap: 2025 marks the evolution of AI Agent from "functional module" to "organizational unit". The latest test of Microsoft Research shows that Agent system equipped with multimodal perception and memory engine has demonstrated collaboration efficiency that exceeds that of human teams in complex tasks (test case: a multinational law firm completed cross-border M&A due diligence through LegalMind Agent cluster, with efficiency increased by 12 times and risk point discovery rate increased by 83%).

Innovation team map:

Breakthrough at the architectural level: Anthropic’s “Constitutional AI 2.0” framework uses ethical decision-making chain visualization technology to increase the accuracy of value alignment to 89%, clearing ethical barriers in highly regulated fields such as finance and healthcare
Interaction layer innovation: Devin Agent, an engineer at Silicon Valley startup Cognition Lab, has achieved a complete closed loop from demand analysis to code deployment, with an independent completion rate of 72% in GitHub real project tests
Business evolution: The "Tiangong" Agent system developed by the Chinese team through in-depth exploration realizes a dynamic decision-making-execution-feedback closed loop in the supply chain optimization scenario, and increases the inventory turnover rate of a manufacturing customer by 34%.

Investment logic upgrade:

Scene penetration: Prioritize the deployment of teams with complete closed-loop capabilities of "perception-decision-execution" (such as Yushu Technology embedding motion control agents into robot joint-level decision-making)
Data flywheel effect: Pay attention to agent platforms with unique interactive data accumulation (such as the emotion recognition model formed by the psychological assistant "Heart Mirror" with 30 million cumulative conversations)
Organizational adaptability: Focus on the design capabilities of the human-machine division of responsibilities system (refer to the "three-level fuse mechanism" of Amazon warehouses to ensure the safety of agent decision-making)

Trend 4: The awakening of embodied intelligence – the quantum leap of physical interaction

According to the latest report from Morgan Stanley, the financing scale of the global embodied intelligence market increased by 240% year-on-year in Q3 2024, among which the humanoid robot track received $5.8 billion in investment in a single quarter. Boston Dynamics' Atlas robot has achieved autonomous navigation in complex terrains, and Tesla's Optimus Gen-2 has broken the 0.1mm accuracy threshold in fine operation, marking a qualitative leap in physical interaction capabilities.

The biggest pain point at present comes from the "double shortage" of training data:

Multimodal data gap: A single robot needs to integrate heterogeneous data streams such as vision (2TB/day), mechanics (800GB/day), and spatial positioning (500GB/day).
Simulation-reality gap: There is a 17-23% deviation rate between the virtual training environment and the real physical world (MIT 2024 laboratory data)

Path to breakthrough:

NVIDIA Omniverse platform launches "Physical Accuracy Certification" system to control simulation training error within 3%
OpenAI and Boston Dynamics collaborate to develop a "tactile feedback reinforcement learning" framework to increase the efficiency of robotic arm operation learning by 40 times
The Ministry of Industry and Information Technology of China's Embodied Intelligence Data Collection Standard 1.0 will be implemented in Q1 2025 to standardize industrial scene data annotation

Investment trend:

China's innovative power is rewriting the industrial landscape:

Unitree B2-W quadruped robot released by Yushu Technology in January 2025, equipped with the self-developed "Chitu 2.0" motion control system, can achieve adaptive gait adjustment in complex terrain
Its Spring Festival Gala special edition robot integrates joint modules with a localization rate of 92%, and the peak torque of a single leg reaches 360N·m (surpassing the 320N·m of Boston Dynamics Spot 2.0)
The FZMotion motion capture system developed jointly with Lingyun Optoelectronics has increased the efficiency of training data collection by 7 times, solving the problem of "data hunger" in embodied intelligence.

Investment trend:

This track is creating a financing legend: Yushu Technology completed a B2 round of financing of 1 billion yuan in 2024, setting a record for the highest single financing in the field of humanoid robots. Investors include top institutions such as Hillhouse Capital and Sequoia Capital, and the valuation has soared 150 times compared with the angel round. In its 2024 trend report, a16z listed "physical intelligence" as the annual keyword, predicting that embodied smart hardware will have an "iPhone moment" in 2025, and the penetration rate of service robots is expected to jump from the current 4.7% to 19%.

Trend 5: Value alignment engineering: from ethical debate to technology standards competition

Paradigm Shift:

The 2024 EU AI Act requires high-risk systems to pass ISO/IEC 23894 certification, giving rise to a new business model of "Ethics-as-a-Service". Microsoft Azure launched the world's first AI ethics audit platform, which can detect 178 value deviation indicators.

Emerging Focus:

Three-dimensional ethical challenges brought by embodied intelligence: Quantification of safety margins in human-machine collaboration scenarios (such as the force control threshold of service robots)

Anthropic proposes a "Constitutional AI 2.0" framework to increase the visualization of ethical decision-making chains to 89%

China Academy of Information and Communications Technology takes the lead in formulating the "Generative AI Value Alignment Test Method", covering localization indicators such as cultural adaptability

Investment focus and industry impact

Our investment logic has always focused on vertical AI applications and dedicated AI fields with clear scenario requirements and data moats. Our current and future investment focuses are on vertical AI applications (including robots) and dedicated AI (such as SLM applications on the edge or on devices and intelligent AI applications). Our investments also include use cases for AI customer support and AI work scenarios (such as unstructured data search, etc.).

The impact of AI on industry

Highly digitalized industries: AI will bring profound empowerment rather than disruption. For example, in the financial services sector, AI has brought significant improvements in risk management and Alpha asset management, including fraud detection, automated credit scoring, market sentiment analysis, etc.
Industries lacking big data: The impact of AI is relatively small, but local optimization can still be achieved through vertical data fine-tuning
Retail and e-commerce: AI uses big models to optimize customer service and inventory management, while developing intelligent systems to analyze consumer data. This application improves personalized shopping experience, optimizes inventory levels, and provides actionable insights to drive sales.

Conclusion
Finding certainty in an “accelerating world”

In this era of accelerating technological change, we believe that the future of AI belongs not only to technology leaders, but also to those who truly understand the essence of business. In 2025, the key to AI investment will no longer be simple technological breakthroughs, but the synergy and win-win of the ecosystem.

In this ecological revolution, only those companies that can closely integrate technology, scenarios and business logic can find certainty in the fog and create real and lasting value.