Preface|The annual benchmark for watching AI globally
Every autumn, the global technology and investment communities look to a long-form benchmark that condenses the year’s movement — the State of AI Report. Led by the British venture capitalist Nathan Benaich and his team at Air Street Capital, it has now been published for seven consecutive years, and is one of the few systematic yearbooks that can stretch at once across five dimensions: Research, Industry, Politics, Safety, and Survey. The key signal in the 2025 edition is clear: a shift from a “contest of parameters and scale” to a “contest of reasoning capability.” Energy and compute are beginning to form a hard boundary, and AI has fully entered geopolitics. That means AI is no longer only a list of technologies, or a commercialization checklist. It will rewrite the next round of global order and industrial division of labor at the level of institutional design, national strategy, and infrastructure. For Taiwan, this yearbook is not only material for upgrading ideas. It is also a working reference for laying out “green compute, trustworthy governance, and dual-ecosystem integration.”
I|From a model contest to a reasoning contest: AI enters a stage of “being able to think”
The frontier movement of 2025 clearly shifts from “whose model is larger” to “who can think and solve problems better.” OpenAI’s o1 series builds “reason first, then answer” into the product design: through longer-step thinking, planning, and self-correction, the model moves from fluent answers to a problem-solving process that can be verified. China’s DeepSeek has iterated quickly on R1-lite / R1, and has shown strong potential on several international benchmarks. What is driving the shift is the maturing of training mechanisms such as chain of thought, rubric-based rewards, and verifiable reasoning. The essential change is this: AI no longer stays at linguistic expression and knowledge recall. It begins to understand the structure of a problem, to unpack premises in context, and to generate a path — and so comes closer to “a cognitive system that can be audited.” On the enterprise side, the center of competition moves with it. Rather than competing over who can write a flashier prompt, it is better to raise “the ability to define the problem”: break a business question into a logic graph AI can compute (tasks, constraints, signals, feedback), and build an evaluation and gatekeeping mechanism that can be iterated. If Taiwan’s technology manufacturing, financial services, and content industries can turn requirements engineering and evaluation engineering (evals) into everyday capabilities early, AI’s output will not stop at the demo. It can enter the process, and the P&L, on a steady basis.
II|Open versus closed source: China rises as a second pole, and ecosystem power becomes the deciding point
The map of AI is moving from “a single technical point” to “ecosystem integration.” The report shows a collective burst in China’s open-source ecosystem in 2025: Alibaba’s Qwen, Moonshot’s Kimi, and the model series from Baidu, Zhipu, and others have gradually filled in from the base model to the toolchain, the community, and the application scenarios. Activity and contribution on platforms such as Hugging Face have clearly risen. Open source is therefore no longer only a research toy. It has become an industrial base — it has driven the speed of distributed innovation and a fall in cost, and at the same time has pushed “operations and governance” to the front of the stage. The global center of innovation is moving from a unipolar Silicon Valley toward a bipolar China–US pattern, with many nodes running in parallel. For Taiwanese companies, the real difficulty is not “which model to pick,” but “which ecosystem to enter”: APIs, permissions, data governance, compute supply, commercial terms, and even future compatibility will all produce path dependence from the choice of ecosystem. A more practical approach is to use a “dual-ecosystem strategy” to lower single-point risk: under the Personal Data Protection Act and cross-border data rules, adopt US-side cloud and models on one side to meet global compliance and partnership, and configure a Chinese or open-source stack on the other to gain an advantage in local cost and iteration speed; at the same time, leave interfaces for long-term maintenance and talent development (observability, evaluation, permissions, hot-swappable models), so that you are not locked in technically or commercially.
III|AI for Science: machines “walk into the lab” and become co-authors of research
AI’s role in scientific research has crossed from an assisting tool into a decision partner. Take DeepMind’s Co-Scientist as an example: the system can combine domain knowledge, data, and hypotheses into verifiable experiments, and iterate automatically in a closed loop. In materials, algorithms, and biomedicine, a “computer first, experiment as corroboration” workflow has also appeared — for instance, proposing a new kind of matrix multiplication, exploring generation paths for atomic-scale crystals, and shortening the cycle for screening targets in particular diseases. The core value is not “replacing the scientist.” It is tightening the experimental space down to a candidate set with a higher win rate, so that the value of research per unit of time rises. Taiwan has long had advantages in semiconductors, materials, and biomedicine. If it can pack “data governance (including master data, experimental records, and reproducibility),” “high-throughput simulation (simulation-first),” and “auditable evaluation (model cards / risk logs)” into R&D infrastructure, and put data ownership and open terms for academic–industry collaboration up front, AI for Science can move from scattered pilots to output at scale — so that R&D is not only faster, but more verifiable, and more transferable.
IV|Energy, compute, and NVIDIA’s “circular empire”
The burst of generative and reasoning workloads is running straight into the ceiling of energy and infrastructure. The report estimates that US data centers could face a 68 GW power gap before 2028. At the same time, NVIDIA still dominates the hardware and software stack for training and inference, with a very high share of the market and of the developer ecosystem. What is more worth watching is the closed loop of capital and supply: frontier labs spend their funds on cloud compute; cloud providers then use capital and cash flow to replenish GPU purchases and cabinet expansion, producing an expansion loop of continuing positive feedback. AI is therefore no longer only a contest of software and algorithms. It is a combined engineering problem of “energy × cooling × network × supply chain.” For Taiwan, the point is to advance at both ends. On the enterprise side, when bringing in AI, put “electricity, cooling, networking, and data movement” into TCO, rather than looking only at the API unit price. On the policy side, offshore wind, renewable energy certificates (T-REC), high-efficiency liquid cooling, and edge data centers can be combined to plan a “low-carbon compute corridor.” If industrial parks and power planning can reserve capacity for AI loads and network redundancy, and attract international supply chains to land through transparent carbon footprints, Taiwan’s position as a hub in the next geography of compute will be more secure.
V|Politics: AI becomes a front line of national strategy
The AI contest has been upgraded from market competition at the company level to a long-term game at the level of sovereignty. The United States, on its new-administration path, emphasizes “AI first” and is pushing ultra-large-scale compute projects, to secure a lead at the frontier and industrial dominance. China is accelerating its layout in technological self-reliance, supply chains, and cooperation with the Global South, seeking comprehensive influence in the export of models, cloud, and applications. The European Union has laid down a regulatory frame with the AI Act, but has been relatively conservative in pace on compute and the production of large-scale models. The Middle East (for example the UAE) is using energy and capital to build sovereign AI quickly, trying to become a new stronghold in the global distribution of compute. AI has therefore become a language of geopolitical technology: it concerns economic competitiveness, and it also defines the sovereign boundaries of data, algorithms, and infrastructure. Taiwan’s practical path in this structure is to take “trustworthy governance + multilateral collaboration” as the core: build norms and processes that international partners can trust in cybersecurity, cross-border data, industry–academia–research, and supply-chain transparency; and, under the principle of “compute at the place of production,” develop regional data centers and edge compute nodes, to lower cross-border uncertainty and political risk.
VI|Safety: the risks of AI are also evolving faster
As frontier capability grows at a multiple of speed, the map of attack and defense expands with it. The report notes that AI-driven cyber-attack capability is iterating quickly: from automated phishing and social engineering to emotionally driven manipulation launched with a coding assistant (so-called vibe hacking), the threshold and the efficiency of attacks are both falling. Large companies have already brought safety forward into research, development, and deployment, and have introduced higher grades of control and testing. But relative to the capital put into model capability, global funding for external safety research and public goods (tools, datasets, test platforms) is still visibly insufficient, creating an institutional tension of “development is fast, safety is slow.” For Taiwan’s critical infrastructure, finance, telecoms, and healthcare, practical near-term steps include: taking a zero-trust architecture as the network base; building a model red team (including scenarios such as privilege escalation, data leakage, adversarial examples, and prompt induction); introducing data masking and synthetic data to lower training risk; and writing incident reporting, trace-back, and the assignment of responsibility into everyday operations. Only when safety is treated as a product characteristic, not a make-up class after the fact, can AI services bear social use at a larger scale.
VII|Ten predictions for the future: the next year, as AI walks into the real world
The report uses ten predictions to sketch observable events in the next twelve months. First, more than 5% of the online revenue of large global retailers will come from agentic checkout, a sign that the shopping journey is led by an agent and that conversion is optimized through multi-step reasoning. Second, at least one leading AI lab may return to open source, in exchange for policy forbearance or market cooperation, and the boundary between open source and commercialization will be renegotiated. Third, AI-assisted scientific research will complete a breakthrough discovery in the real world, and the path of “simulation first, experimental verification” will be accepted by more disciplines. Fourth, the first major cybersecurity incident caused by AI forgery will trigger a global discussion of norms, and give rise to new tools for tracing and forensics. Fifth, an AI-generated game may become the year’s most-watched title on Twitch, a sign that content production can be automated from materials through to logic. Sixth, “AI neutrality” may become a new diplomatic stance for some countries, trying to keep more room for maneuver between technology and regulation. Seventh, an AI-generated film will be seen at a mainstream awards ceremony, and will trigger a public debate on copyright and ethics. Eighth, a Chinese lab may overtake a US team on a mainstream leaderboard (such as the LMSYS Chatbot Arena), strengthening the bipolar narrative. Ninth, NIMBYism in the United States may affect the siting of data centers and the supply curve, pushing up the cost of compute and a geographic reallocation. Tenth, the White House may use administrative and judicial paths to handle conflicts in state-level AI legislation. For Taiwanese companies and policy departments, these ten predictions can be turned directly into annual assumptions for scenario planning, to configure product roadmaps, compliance resources, and public-communication responses early.
VIII|A reflection: from chasing the wind to making the wind — turn a technology dividend into a governance dividend
This yearbook is not a compilation of technical gossip. It is a clear reminder: the boundary of AI is moving from capability to institutions and governance. For the past decade, we have relied on platform dividends and the allocation of traffic. For the next decade, the critical capabilities will be system design and logic governance — how to let AI understand a company’s objectives, read the real context, and act efficiently inside the boundaries of compliance and ethics. Taiwan’s advantages lie in deep supply-chain integration, an engineering culture, and the flexibility of small and mid-sized firms. The short boards are the standardization of data governance, cross-department evaluation capability, and the long-term development of talent. If “dual ecosystem, low-carbon compute, and trustworthy governance” can be taken as a triangular fulcrum, and requirements engineering, evaluation engineering, and safety engineering can be built into products and processes, AI will become a new kind of air — not a gust of wind, but a productivity that can be breathed every day, expanded, and trusted by society.
More deep-dive pieces
- ChatGPT will now show brands how often they are mentioned. That is not proof they were recommended well.
- AI will not only eliminate people: the functions being enhanced, and a manager’s handbook for this week
- AI unemployment ranking: the high-risk functions most likely to be eliminated or largely replaced
- Will AI cause unemployment? Start with the enhance / accelerate / eliminate axes
The full original report
State of AI Report – 2025 ONLINE · Download the PDF
Official site: https://www.stateof.ai/
The judgment, restated
My judgment is that after reasoning rises, the bottleneck turns to energy, safety, and geopolitics — corporate strategy has to go one layer deeper than “chasing the model.”
Three things this week
- Take stock of the real effect of compute / energy constraints on your product rhythm.
- Write one page of safety red lines: which decisions must never be fully automatic.
- Make one geopolitics / supply-chain scenario table (best / base / worst).
You can read this alongside the close read of McKinsey’s The State of AI in 2025: from “everyone is using it” to the few who actually know how. You are welcome to subscribe to Future Insights Lab, or reply to discuss where your strategy is stuck.
