Introduction: What kind of historical turning point are we standing at?
In 2025, artificial intelligence reached an unprecedented threshold. If the explosion of ChatGPT in 2023 made the world aware of the knowledge-emergence capabilities of large language models, the central story of 2025 shifted to a deeper proposition: the marginal cost of agency is approaching zero.
The history of information technology reveals three structural inflection points at which marginal costs collapsed. Around 1995, the spread of the internet drove the marginal cost of obtaining information sharply downward. In 2023, large language models led by the GPT family pushed the cost of accessing public knowledge close to zero: anyone could obtain expert-level answers through natural-language conversation. In 2025, as the agent ecosystem began to take shape, we started to witness a third turning point—the ability to act itself becoming available at scale.
Here, “agency” means the capacity of a system to understand intent autonomously, make plans, use tools, execute tasks, and adjust in response to feedback. As the cost of deploying this capacity continues to fall, the productive structure of human society will be fundamentally reshaped.
Understanding the full scope of this transformation requires a full-stack analytical framework. This essay proceeds along three dimensions: the strategic landscape, development trends, and R&D models. They are not parallel topics, but a progression from macro to micro and from “why” to “how.”
Part I: The strategic landscape—the civilisational contest in the age of general intelligence
1. The full stack: a systemic map of the AI industry
Understanding the strategic landscape begins with a full-stack map.
At the bottom, the infrastructure layer encompasses energy consumption, mineral resources, materials science, and equipment manufacturing. AI’s deepest dependencies are physical. Training a frontier model consumes electricity measured in gigawatt-hours, while continuous inference demands an equally formidable energy supply. The location and construction of compute infrastructure are therefore constrained by energy and geography. The logic of America’s Stargate project is to bind compute and energy infrastructure into a single complex. China, meanwhile, is seeking differentiated advantages through its global lead in photovoltaics and continued progress in nuclear technology, including small modular reactors and research into controlled fusion.
The intelligent-compute layer includes AI computing centres, intelligent clouds, and data centres. Competition is moving beyond raw scale toward the efficiency of compute-model co-design. Whereas traditional cloud computing centred on general-purpose CPUs, the AI era requires floating-point matrix operations on GPU tensor cores and is evolving toward heterogeneous multi-chip architectures.
The model-training layer is the most fiercely contested. It comprises cognitive, contextual, embodied, spatial, and scientific intelligence—not isolated directions, but a spectrum of capabilities that together constitute general intelligence.
The intelligence-extension layer includes agents, Software 3.0, applications, and devices: the interfaces through which AI capabilities diffuse into the real world. Above it, the industrial-structure and diffusion layers concern how AI reshapes markets and spreads across consumer, business, and government settings.
This map matters because AI competition is never confined to one technological dimension. The strength of a country or region depends on its accumulated capabilities at every layer and on how efficiently those layers work together.
2. New sources of innovation: who drives the frontier?
At the top of the stack, the organisational form of the leading force deserves special attention.
The United States and China display sharply different structures. Frontier development in the United States is driven largely by research-intensive, closed-model start-ups. OpenAI, Anthropic, xAI, Thinking Machines Lab, and SSI form a distinctive cluster combining the depth of academic research with the speed of product engineering. Typically founded or led by elite researchers, these companies tightly couple basic research with product development and sustain exceptional “research density” through enormous financing. NVIDIA and Google participate as infrastructure and platform leaders.
China’s landscape is different. Major technology companies—ByteDance through Doubao, Tencent, and Alibaba—lead model development and deployment, while an open-source ecosystem led by DeepSeek provides an important complementary force. DeepSeek R1 marked a notable advance in reasoning, yet purely research-driven start-ups remain relatively scarce, reflecting deeper differences in talent systems, capital markets, and organisational culture.
The researcher-founder is therefore crucial. Under the fourth paradigm of science—computation-driven discovery—the path from basic research to industrial application has shortened dramatically. The conventional “industry–university–research” model assumed a linear transfer from university research to corporate application. AI research companies break that assumption: one organisation performs basic research, engineering implementation, and value creation simultaneously. OpenAI’s transformation from a non-profit laboratory into one of the world’s most highly valued AI companies is the most dramatic example.
At least four conditions sustain this model: a deep talent system, mature capital markets, compute at scale, and a closed loop from research to product. OpenAI and Anthropic lead because they possess all four.
3. A new globalisation: the geopolitics of AI diffusion
AI is rewriting the foundations of globalisation. America’s diffusion strategy exports its entire technology stack—from chip architecture and model systems to application ecosystems—and has already captured a dominant share of the global stack. Its strategic aim is long-term ecosystem control through technical standards. The Genesis Mission extends this logic by using AI to accelerate research in strategic fields such as nuclear energy, biotechnology, and advanced materials.
China’s response appears in its “AI+” industrial policy and the idea of a “new Digital Silk Road.” AI has been elevated further in the Fifteenth Five-Year Plan, while a digital version of the Belt and Road seeks adoption of Chinese AI stacks across the Global South. At heart, this is a contest between two paths for diffusing technological civilisation.
The underlying variable is changing, however. Traditional technological dominance relied on control of proprietary closed systems; open source is weakening that foundation. When DeepSeek released a high-performance reasoning model openly, it created both a technical and geopolitical shock: it showed that closed barriers could be bypassed and gave developers worldwide an alternative to the American stack.
4. New infrastructure: the 80-year cycle and bubble risk
Historically, the infrastructure phase of a major technological revolution brings massive capital expenditure, a cyclical bubble, and a “new dawn” after the bubble bursts. Railways, electricity, and the internet all followed this pattern.
The current AI build-out looks similar. Stargate represents an infrastructure-economy strategy: secure the commanding heights of the next decade through investment in compute and energy on a vast scale. Such concentration also carries bubble risk, especially when near-term application revenue cannot match infrastructure spending.
Carlota Perez’s Technological Revolutions and Financial Capital describes four phases: eruption, frenzy, synergy, and maturity. The frenzy phase features an excessive influx of financial capital, inflated asset prices, and bubbles. A crash does not mean that the technology has failed; it is more like a forced reallocation of resources that prepares the synergy phase. Seen through this lens, the current AI boom may be moving from eruption into frenzy.
China offers a contrast: government-guided, more balanced construction of computing and training centres aims to avoid purely market-driven overinvestment. Its strength in photovoltaics and continued nuclear research provide a longer-term answer to AI’s energy demand. If the rough 80-year rhythm of paradigm-level revolutions still holds, we are at the beginning of a new cycle.
5. New digitalisation: platform ecosystems and inflection effects
The digital economy is driven by platforms and ecosystems. Every change in computing paradigm produces a new platform and an ecosystem around it: Microsoft and Intel in the PC era, Google and Amazon on the internet, Apple and WeChat in the mobile era.
New AI platforms are now forming. The United States is building a global ecosystem on its lead in frontier models and cloud infrastructure. China has a distinct consumer advantage: its huge mobile-internet user base offers fertile ground for rapid AI diffusion. In enterprise software, China may have an unusual leapfrogging opportunity. Because many companies remain behind their American counterparts in conventional digitalisation, they may skip parts of the traditional SaaS stage and move directly to AI-native enterprise services.
6. New value and new industries
AI’s industrial impact goes far beyond efficiency. China’s Fifteenth Five-Year Plan treats semiconductors, AI, intelligent manufacturing, quantum computing, and the low-altitude economy as strategic emerging industries. These sectors are tightly coupled: AI chips depend on semiconductor advances; intelligent manufacturing depends on embodied intelligence; the low-altitude economy is a natural arena for spatial intelligence.
Another often-overlooked dimension is America’s need to reindustrialise. After decades of manufacturing hollowing-out, AI—especially embodied intelligence and industrial robotics—is viewed as a critical lever, and that need is shaping US policy priorities.
7. Civilisational systems: a change in the paradigm of science
At civilisational scale, AI is driving a fourth scientific paradigm.
The first paradigm was empirical: observation and thought accumulated knowledge slowly through oral transmission, from Egyptian irrigation rules to China’s twenty-four solar terms. The second, marked by Galileo’s systematic experiments, introduced controlled variables and reproducibility. The third was theoretical, reaching its heights in Newtonian mechanics, Maxwell’s equations, and relativity; mathematical modelling and peer review became standard.
The fourth paradigm, proposed by Jim Gray in 2007, is computational. It combines algorithms, large-scale data collection, human-machine collaboration, and iterative verification. AlphaFold’s protein-structure predictions, end-to-end autonomous driving, and neural decoding in brain-computer interfaces are representative. Discovery can shrink from years and months to weeks and days, while disciplinary boundaries expand.
This changes the organisation of innovation. Universities dominated the third paradigm, with research entering industry through technology-transfer mechanisms such as the Bayh–Dole Act. In the fourth, research-intensive start-ups become central because they combine research, engineering, and markets to accelerate the whole journey from minus one to one. OpenAI, DeepMind, and DeepSeek embody the model.
Vannevar Bush’s 1944 report Science, the Endless Frontier laid the strategic foundation of America’s post-war research system. Nearly eighty years later, AI is driving another reconstruction. America’s Genesis Mission targets advanced manufacturing, biotechnology, modern nuclear energy, fusion, and the electrical grid. China has elevated AI to the highest strategic level and is exploring a new research system. The new industry–university–research combination is, in essence, a rearrangement of innovative resources during the transition from the third paradigm to the fourth.
8. Technology, production relations, environment, and population
AI is accelerating four technological frontiers: new energy, life sciences, materials, and space. All rely heavily on simulation, data, and AI-assisted experimental design. In materials science, for example, AI can screen millions of possible compounds for target properties and compress years of discovery into weeks.
The financial system is becoming a key production variable. A developed chain from venture capital to IPO exits supports research-intensive companies. At the same time, structural pressure on the dollar’s reserve status, the internationalisation of the renminbi, and digital currencies including stablecoins are changing global capital flows. AI is also transforming finance itself, from quantitative trading and risk assessment to compliance monitoring.
The expansion of human activity from the ground to low altitude, high altitude, and space is creating new industrial ecosystems. Drone logistics, urban air mobility, and airspace management have become Chinese policy priorities. Integrated aerospace transport and information systems are moving from science fiction toward engineering reality. Their enabling technology is spatial intelligence: AI’s ability to understand, reason about, and plan in three-dimensional space.
China also faces a deep demographic challenge: more than 310 million people are over sixty and the birth rate has fallen to 6.77 per thousand. This creates both demand and constraint. Education requires AI-driven reform for new forms of talent; healthcare faces enormous demand in imaging, remote surgery, and intelligent elder care. In this sense, AI is a strategic tool for responding to ageing.
Part II: Development trends—from general intelligence to agents and new digitalisation
1. Five stages: from learning knowledge to learning organisation
The development of general intelligence can be divided into five qualitative stages.
Stage one: learning knowledge. Pre-training absorbs public knowledge from vast internet corpora. GPT-3 demonstrated surprising knowledge emergence, and Doubao also performed strongly at this stage. ChatGPT’s explosive adoption marked its maturity: for the first time, people could access knowledge from almost any field through natural-language dialogue. Scaling here centred on context length and parameter count.
Stage two: learning to reason. Since 2024, post-training has sharply improved reasoning. OpenAI’s o-series and DeepSeek R1 showed how reinforcement learning and self-reflection can strengthen logic and problem solving. DeepSeek’s open-source strategy accelerated the global diffusion of reasoning, while products such as Deep Research began to create practical value. Scaling shifted toward the iterative depth of self-consistent reflection.
Stage three: learning to act. This was the central battlefield of 2025. Continual training must combine intent understanding, instruction following, long-horizon planning, and tool use. Agency has two branches: digital-world agency, represented by Claude Code across product design, research, development, marketing, and operations; and physical-world agency in robots, autonomous vehicles, and other embodied systems. Anthropic’s dramatic valuation growth reflected the market’s pricing of agency. Claude Code showed that value grows non-linearly once a model can complete complex work autonomously inside a real development environment. In the physical world, multiple paths toward world models are beginning to provide embodied systems with physical intuition. Scaling becomes a trinity of environment, tools, and education.
Stage four: learning to innovate. Still early, this stage seeks genuine invention beyond existing knowledge. AI-assisted AI development and AI Scientists have entered experimentation, while mathematical reasoning at IMO gold-medal level hints at the potential. Scaling will depend on accumulated time spent innovating.
Stage five: learning to organise. The ultimate form is a multi-agent system capable of complex decisions, organisational coordination, and strategy. Research on multi-agent coordination remains nascent, but its long-term importance may equal that of single-agent breakthroughs. When highly autonomous agents can divide labour like a human organisation, a new social form will emerge. Scaling will depend on the complexity of adaptation to the environment.
2. A new path for scaling laws: from models to agents
Scaling laws have expanded profoundly. Early work by Kaplan et al. (2020) and Hoffmann et al. (2022) focused on power-law relationships among parameters, data, and compute in pre-training. As capabilities and use cases deepen, new scaling dimensions are appearing.
One useful analogy compares AI with the social development of civilisation: genetic inheritance, language, education, industrialisation, digitalisation, and finally intelligence each improved the transmission and amplification of capability. AI follows a similar path—from data-driven “inheritance” in pre-training, to reasoning-driven “cognition” in post-training, to interaction-driven “agency” in continual training, and ultimately to higher-order scaling in innovation and organisation.
The speed and path of this new scaling depend heavily on the maturity of digital environments and infrastructure. Competition in the agent era is therefore not only about the model, but about the quality and completeness of its environment: tools, data, cockpit, and context engineering.
3. The technical frontier of general intelligence
At the model foundation, sparse attention is improving efficiency and enabling extremely long contexts; unified architectures seek high-quality understanding and multimodal generation in one model; diffusion and flow-matching methods are advancing image, video, and audio generation; long-context and memory breakthroughs support million-token inputs; evaluation is moving from standard benchmarks to real-world, agency-oriented tests; and continual learning seeks new capabilities without catastrophic forgetting.
Cognitive intelligence pursues difficult problem solving and deep reflection. Its frontiers include AI-assisted AI development, end-to-end AI Scientists that propose hypotheses and design and analyse experiments, and a deeper union of cognition and agency so that models can both think and act. Superintelligence remains distant, but leading laboratories are already researching its safety and alignment.
Contextual intelligence concerns perception and understanding of real situations. Omnimodal integration across audio, video, touch, and other channels is a central challenge. “Deep Context” means understanding not only language but the full semantic structure of tasks and rewards. As agents operate in increasingly complex environments, demand for native contextual data is exploding.
Embodied intelligence has accelerated. Humanoid robots such as Figure and Unitree, quadrupeds for education, entertainment, and inspection, industrial robots, and autonomous driving form four major settings. Architectures are moving from modular pipelines to end-to-end systems, with world models and vision-language-action models as key directions. Work such as Physical Intelligence’s π0.5 suggests that language-model reasoning can transfer into physical manipulation. Data—embodiment, process, and annotation—remains the central bottleneck to scale.
Spatial intelligence is rapidly emerging. In digital space, systems such as World Labs and Marble build three-dimensional world models; in physical space, companies such as Qidaisong focus on contextual understanding and spatial-intelligence machines. Better spatial reasoning and attention will directly accelerate embodied deployment.
Scientific intelligence, or AI for Science, is among the most strategically valuable applications. It requires dedicated scientific data and model systems, agents participating end to end in research, and AI Scientists capable of proposing and testing hypotheses. The Genesis Mission represents national-level US investment. The ultimate aim is to accelerate every step of discovery, from data collection and hypothesis generation to experimental validation.
4. Agents: the full release of agency
2025 was widely called “the year of the agent.” Four capabilities define an agent: understanding the user’s real intent, following instructions faithfully, decomposing complex goals into executable long-range plans, and interacting with tools such as APIs, software interfaces, and file systems.
Releasing this agency depends on three conditions.
Real environments. Agents must learn through practice in real settings. The shift from static internet text to the complete context of valuable real-world situations is essential. Agents need to complete long, multi-turn tasks independently and improve through interaction.
Real evaluation. Static benchmarks no longer measure real capability effectively. Inputs and outputs must come from live data and authentic user needs; evaluations must change dynamically and introduce recent queries to prevent reward hacking and test-set overfitting.
Real data. Data collection is reorganising around environmental context. The decisive move from generic internet context to complete, multidimensional, high-value contexts is what turns an agent from merely usable into genuinely good.
5. The cockpit: how humans control agency
In Lu Qi’s framework, the cockpit is the interface through which people use and control an agent’s agency.
History clarifies the idea. The internal-combustion engine gave humanity powerful physical agency, but releasing it required a long evolution of cockpits: horse, carriage, Ford Model T, and highway system. Each iteration expanded the range and efficiency of mechanical power. The internet likewise provided informational agency, while its cockpit evolved from servers to browsers, search engines, and social platforms.
AI follows the same law. Models possess enormous latent agency, but cockpit iteration and engineering systems are needed to release it. OpenAI Codex was a first-generation programming cockpit; Claude Code marked a major engineering advance; the ecosystem forming around it signals further maturity.
Claude Code’s core innovation is a shared context between silicon and carbon. Earlier cloud-first tools asked developers to move their environment to the cloud, yet most real work remained local—in files, terminals, company databases, and internal resources. Claude Code instead installs on the developer’s machine and can work directly in that environment. Directory-level Markdown instructions describe local context so AI can see and use the same world as the human.
The deeper principle is that AI and humans should share one working context. Rather than building an isolated digital environment for AI, this design embeds AI into the environment people already inhabit. It reduces integration complexity and makes human-agent collaboration natural.
Cockpit evolution resembles levels of autonomous driving. In the assistance stage, humans lead and systems help; early conversational models paired with workflow frameworks such as Coze and LangChain belong here. In the conditional-autonomy stage, the system leads but humans can intervene; agentic reasoning models with long-horizon planning and environmental interaction are moving into this phase. The long-term destination is full autonomy: system-led, zero intervention, and high agency.
Each stage has distinct engineering problems. Assistance requires workflow orchestration to compensate for weak multi-turn and long-horizon reasoning. Conditional autonomy requires safe execution—sandboxes, permission audits, behavioural guardrails, and humans in the loop—plus context engineering. Full autonomy requires mature quality assurance: delivery contracts, evaluation rubrics, self-checks, and transparent processes.
6. Context engineering: the operating system for agents
If the cockpit is the hardware interface between human and agent, context engineering is the agent’s operating system. It has three core layers.
System prompts carry organisational purpose and personal context. They define who an agent is, whom it serves, and which principles it follows.
Structure prompts carry static environmental context. Like a physical office, they define the workspace, the function of each area, and the responsibilities of different organisational roles.
Task prompts carry dynamic context, from simple workflows to sophisticated cognitive scaffolding. Scaffolding does more than say what to do; it supplies a way to think—problem definition, construction of the solution space, decomposition into cognitive dimensions, and a systematic reasoning process.
The complete system also includes the cockpit beneath it—operating system, files, and hardware—and a context-processing loop: perceive the environment, analyse and plan, execute, remember state, and improve through feedback.
7. New digitalisation: the arrival of Software 3.0
Software 1.0 was the symbolic era: code as input, compiler as executor. Software 2.0 was the data era: data-driven neural inference. Software 3.0 is the language era: natural-language description, understood and executed by large models.
Its stack has five layers:
- Hardware compute: floating-point matrix operations on GPU tensor cores, with resources allocated across inference tasks through time-sharing.
- System core: language and world models form a new “context computer,” where prompts are input, inference is computation, and generated tokens are output.
- Commercial OS: cockpit plus context engineering. In the physical world, a spatial cockpit plus physical context forms a “spatial-intelligence machine.”
- Development extensions: plugins, skills, hooks, and the Model Context Protocol form an application ecosystem. MCP is becoming a standard way for agents to access tools and services.
- Application ecosystem: digital agents—agentic, customised, sometimes disposable apps—and embodied agents such as robots, autonomous vehicles, and new devices serve end users.
8. Reshaping industrial structure
Software 3.0 affects the entire industrial structure. In consumer markets, AI is rebuilding gateways across communication, social media, content, games, education, healthcare, finance, and housing. The change reaches beyond smarter interfaces to the underlying service logic: from standardised products to personalised intelligent services.
In enterprise markets, new IaaS for domestic compute and training, new PaaS for model-compute coordination and context efficiency, and new SaaS for agentic applications form a three-tier architecture. Cars, robots, drones, autonomous ships, AI glasses, wearables, and phones become physical extensions of AI.
At the macro level, an “AI factory” is taking shape. Supply chains, sales, customer support, employee experience, finance, legal work, administration, and government relations will all connect through AI cockpits to a unified intelligent operating system.
Part III: R&D models—from the laboratory to the battlefield
1. Six elements of a new R&D model
The new model has six pillars: a compute system built around the agent cloud; a model system with closed-loop agency; a data system based on context engineering; a cockpit that releases agency; a development and delivery model based on efficient, customised co-creation; and an application model based on systematic deployment.
2. A new model system: open agentic models and an evaluation ecosystem
Several directions matter.
Open agentic models. On open foundations such as Qwen3 and DeepSeek, developers must build intent understanding, goal adherence, planning and decomposition, and interactive reasoning. These capabilities should generalise across tool use, web interaction, deep research, and coding.
Real-world evaluation through an Open Agent Arena. Static benchmarks are insufficient. Dynamic evaluations must draw on authentic user needs and refresh regularly to prevent overfitting.
Training as a Service. High-value industrial scenarios should form a closed loop with model training: real settings supply training data and evaluation standards; trained models return to those settings for validation and continued iteration.
Next-generation agency systems. Agent-development frameworks such as NexAU can support end-to-end agentic data synthesis and reinforcement-learning infrastructure such as NexRL and NexVenusCL, moving agentic models from experiments to industrial deployment.
3. A new compute system: from compute cloud to agent cloud
The value logic of compute is changing. The old model rented GPU time slices, with thin margins, limited barriers, and severe commoditisation. “Token as a Service” raises value density by charging for inference but remains dependent on replaceable public internet context.
The new core is the agent cloud, a compute service organised around agent execution. Its proposition is “context as a service plus agent operation”: serve agents rather than people, move from internet context to valuable enterprise context, and replace usage-based billing with outcome-oriented pricing. Enterprise productivity context compounds—the longer a system is used, the richer its accumulated context and the more valuable the service.
The architecture has three layers. New IaaS provides intelligent-cloud infrastructure, advanced and domestic chips, training compute, and sandboxes. New PaaS provides agent runtimes, context engineering as a service, and permission and security management. New SaaS provides enterprise-productivity services, agent engineering, workspaces, agency configuration, and agent cockpits.
Across all three lies model-compute co-design: coordinating heterogeneous multi-chip systems with model architecture, plus Training as a Service under continual learning through automatic optimisation, reinforcement learning, and asynchronous parallelism.
4. A new data system: from internet data to environmental interaction
Data requirements change with agency. Pre-training relied on internet text and books; post-training turned to conversations and reasoning traces; agency training needs context and feedback from environmental interaction. Such data comes from real work settings, business processes, and decisions, and is highly situation-specific and time-sensitive.
American model and data companies are systematically building occupational data assets. Anthropic and Stanford use the US Department of Labor’s O*NET database to study AI’s impact on occupations and answer a strategic question: which jobs should AI learn first?
Surge AI is recreating full enterprise software environments and organisational structures, generating tens of thousands of logically consistent synthetic records. These “synthetic enterprises” offer agents safe, controllable, scalable places to practise.
OpenAI’s reported Mercury project has recruited Wall Street professionals at $150 an hour to annotate financial data, illustrating the scarcity and cost of high-value domain knowledge. Together these cases point to one conclusion: competition over agentic data is a contest to construct the most complete, authentic, and valuable situational context.
5. Engineering the cockpit
The central design principle is to harness agency while compensating for its weaknesses: behavioural drift over long tasks, insufficient business understanding, and unstable delivery quality.
The new cockpit builds four kinds of protection and enhancement:
- Safe execution: sandboxes, permission auditing, behavioural guardrails, and human-in-the-loop controls.
- Context engineering: contextualising raw data and tools, dynamically organising relevant context, and packaging reusable capabilities as skills.
- Delivery quality: explicit input, output, and quality contracts; multidimensional evaluation rubrics; model self-checks; and transparent delivery processes.
The goal is to integrate agentic models into work in a trustworthy and reliable way.
6. A paradigm shift in enterprise digitalisation
Enterprise digitalisation can be described in five stages.
System of Information: desktop digitalisation led by Microsoft Office, centred on documents.
System of Record: ERP and CRM record business processes and decisions.
System of Engagement: internet-era connections with customers, channels, suppliers, and employees through email and collaboration tools.
System of Insight: big data and BI extract insights for data-driven decisions, supported by cloud-native infrastructure, data stacks, and low-code platforms.
System of Intelligence: multimodal perception, complete data-to-model pipelines, and partially or fully automated decisions and execution across every function, industry, device, process, and role.
The present shift can be summarised in three movements: from data-driven to semantics-driven; from IT tools to business language; and from decision support to intelligent decision-making. Software 3.0 and context engineering make it possible to integrate enterprise digitalisation through a shared natural language. Like the move from paper to electronic documents, this changes not merely efficiency but the foundations of organisation and operation.
7. FDE: the key role in agent deployment
The Forward Deployed Engineer, a role pioneered by Palantir, is becoming central to the agent era.
FDEs are needed because of the application gap. Large models are general; every business is specific, with its own data formats, processes, organisational structures, and tacit domain knowledge. In traditional software, standard products plus custom implementation bridged the gap. AI adds uncertain requirements, rapid model improvement, and a shift from fixed functionality to continuously measurable value. The new role must understand technology and business, combine research and engineering, explore independently, and co-create with customers.
The workflow is iterative: understand needs with domain experts; translate them into AI task definitions, prompts, and workflows; implement output formats, tuning, and API orchestration; verify and correct quality; then deploy, monitor, and improve in production.
Palantir’s work with Airbus shows the path from zero to one to many. In late 2015, FDEs entered Airbus factories, found data silos, built an SDDI integration solution, used an ontology to create a semantic business layer, and developed an MVP for A350 production efficiency. In 2016, the approach expanded across production, supply chains, scheduling, finance, and quality. In 2017, these lessons became the Skywise platform—Apollo infrastructure, Foundry platform, and applications—and then opened to OEMs, airlines, suppliers, and MRO providers. By 2024, more than 150 airlines had joined, with an organisation added roughly every two weeks.
The deeper lesson is that FDE scaling begins with a custom problem, abstracts a general solution, and ultimately produces a product. Every step depends on immersion at the customer site.
OpenAI’s work with John Deere offers an agricultural example. Deere’s See & Spray system uses 36 cameras and edge compute to identify weeds and spray precisely, reportedly reducing chemical use by 60–70 percent while improving outcomes. The FDE workflow spans initial scoping, system and data review, and KPI definition; offline validation, edge simulation, and field trials; and finally production integration, monitoring, operations, and data feedback.
8. The challenge of FDE in China
China presents distinct obstacles.
Willingness to pay. Chinese enterprises have traditionally paid less readily for SaaS and knowledge services than for hardware and tangible assets. The knowledge-intensive blend of consulting and engineering must be redefined and explained.
Fragmented data infrastructure. Years of in-house and customised systems have produced silos. Integration that may take two weeks in the United States can stretch to two or three months, raising labour costs and reducing reuse.
Cultural differences. Palantir-style FDEs cross the boundaries of engineer, consultant, and researcher. Chinese organisations often favour clearly separated responsibilities, making a hybrid role harder to place.
Scarce talent. The role demands business understanding, technical depth, product judgement, and comfort with ambiguity. AI start-ups need such teams to deploy models, while traditional service companies need them for transformation; supply falls far short of demand.
Model capability gaps. American FDEs can often concentrate on integration and trust while using frontier models. Chinese FDEs may need extra fine-tuning and engineering work to compensate for weaker base capabilities, increasing complexity and cost.
9. From FDE to FDR and OPC
The role itself is evolving.
The Forward Deployed Researcher moves beyond configuring existing technology. OpenAI and Anthropic are putting researchers in customer environments so that customer problems become research questions. Successful methods are productised and fed back into the platform, creating a two-way flow between frontier research and industrial deployment.
Consider an enterprise knowledge assistant. A request to search historical project documents may lead an FDR to study how models understand industry terminology, develop new context-compression and multimodal-document techniques, and eventually integrate those innovations into a broader API platform.
The more radical evolution is the One Person Company. FDEs and solo founders share full-stack technical ability, deep business understanding, problem definition under ambiguity, cross-domain communication, results orientation, and adaptability. When AI tools become powerful enough, one person with FDE capabilities can build and run a company.
The Base44 case is a striking illustration. Founder Maor Shlomo reportedly operated essentially alone until hiring a first employee shortly before acquisition, used AI to build MVPs in two to four weeks, deployed directly to production, and grew through content and viral distribution. Wix acquired the company six months later for $80 million in cash. Reported annual recurring revenue was $3.5 million, with more than 250,000 users and monthly profit of $189,000.
The point is not merely the headline numbers. Agentic tools such as Claude Code are democratising the complex engineering capacity once available only to large organisations.
10. A systemic path for FDE-led industrial transformation
Embedded in industrial upgrading, FDE can support a path from point solutions to systemic transformation.
In government services, FDEs can help build centralised intelligent systems across public-service scenarios. Existing patterns—enterprise profiling, business support, and intelligent matching of policy subsidies—suggest a model combining agents, a unified operational map, and FDE delivery.
In enterprise services, FDEs can act as both “new consultants” and “new integrators.” Consulting helps organisations update their understanding and strategy for the AI era; integration uses reusable agent solutions to drive wider transformation. The aspiration is a kind of AI-native McKinsey for China’s new productive forces.
The relevant industries align closely with the Fifteenth Five-Year Plan and “AI+” policy: integrated circuits, artificial intelligence, biomedicine, advanced equipment, new-energy vehicles, and other priority sectors.
Conclusion: a new Long March toward intelligent civilisation
We return to the opening proposition: we stand at a historical turning point where the marginal cost of agency is approaching zero. Its meaning extends far beyond technology.
For individuals, access to agents raises the upper bound of what one person can execute. For organisations, advantage shifts from how many people they employ to the quality of their context and cockpit. For countries, AI competition is a contest not only of technology and capital but of organisational and civilisational paradigms. For civilisation, the arrival of the fourth scientific paradigm moves the speed of cognition and innovation to a new order of magnitude.
History does not repeat itself simply, but its rhythms are often strikingly similar. The steam engine gave rise to industrial civilisation, electricity to the electrical age, and computers to digital civilisation. General intelligence is giving rise to another form, in which knowledge, cognition, and agency become ubiquitous infrastructure like electricity and the internet.
China’s opportunity is to seize every critical link in this civilisational transformation—from infrastructure to industrial deployment and from model research to talent development—and build an autonomous, controllable full-stack AI system. FDE, Software 3.0, context engineering, and cockpit systems are not merely technical frameworks; they are components of the infrastructure of a new civilisation.
This is a new Long March toward intelligent civilisation. The road is long, but the direction is clear.