Foreword: The "Busy-Work" Predicament of Taiwan's Software Industry
Taiwan's software industry has long been trapped in an awkward cycle where a small market, insufficient talent, and chaotic processes feed into one another, creating an ever-tightening spiral. The main source of revenue for most software companies is customized government or enterprise projects, developed for acceptance testing and delivered for project closure. Under this model, earning more means hiring more people, resulting in extremely high replication costs and profit margins that can never be raised.
This "just get it working" project mentality makes it difficult for companies to accumulate true core technologies and product assets. Meanwhile, the semiconductor industry continues to siphon off STEM talent. The National Development Council predicts a labor shortage of 480,000 people in Taiwan by 2030, and the salaries the software industry can offer give it almost no chance in the war for talent. From my own observations in the industry, it's not that many teams don't want to do things well, but that structural limitations leave them willing but powerless.
This article aims to discuss the root causes of these difficulties and whether, in today's era of rapid AI adoption, Taiwan's software industry can transition from "project outsourcing" to being "product-driven" and forge a different path.
Step One: A Mindset Shift, from Project to Product
To talk about solutions, we must first define the problem clearly. The difference between a "Project" and a "Product" is not just semantic; it's a fundamental difference in business logic:
Aspect | Project Mindset | Product Mindset |
|---|---|---|
Goal | Responsible for a single client's acceptance | Responsible for market needs |
Revenue Model | Selling time (billed by man-months, linear growth) | Creating assets (subscription/licensing, exponential growth) |
Development Process | Dive in on receiving requirements; document after completion | User Research → Define MVP → Iterative Validation |
Success Metrics | On-time delivery, client sign-off | User Retention, MRR, NPS |
Knowledge Accumulation | Experience stays with individuals; lost upon turnover | Embedded in the product and processes; replicable |
It's not that Taiwan has no successful software product companies. Trend Micro focuses on cybersecurity, Perfect Corp. delves deep into beauty tech, and CyberLink specializes in video processing. The common thread in these cases is: they resisted the temptation to "do everything," chose a vertical, and built a moat with accumulated domain know-how.
In my experience, the problem with many small software companies isn't a lack of technical skill, but that they take on too many projects in different directions simultaneously. Each project is only developed to a "usable" state, with no spare capacity to polish any single one. Alice Chang, CEO of Perfect Corp., once said something very on-point: Taiwanese software companies "must define their products to be sold worldwide from day one." This isn't a slogan but a survival strategy; the local market cannot support the scaling of a product company.
The same goes for process. Survey data shows that 84% of software projects fail to be completed within the scheduled time and budget, with an average budget overrun of 189%. The root cause is often not technical issues, but unclear requirements, frequent specification changes, and low transparency in the development process. Adopting agile development and establishing a habit of user research isn't about following trends; it's about ensuring that limited resources are invested in directions with real market value.
The Core Solution: Building Small, Beautiful Niche Products with an AI-Native Mindset
Why AI, specifically? Because AI can solve some of the most painful problems in Taiwan's software industry.
Talent shortage → LLMs and AI development tools allow a 3-person team to achieve the output of a 10-person team in the past. Automated testing, document generation, and code assistance are no longer just concepts but tools used daily.
Small market → AI-driven SaaS products have inherent potential for globalization. Natural language processing has drastically reduced the cost of multi-language support, allowing a product to target overseas markets from day one.
Chaotic processes → AI can help analyze user behavior, predict changes in demand, and even automatically generate preliminary specification documents, reducing the waste caused by "starting work with unclear requirements." However, that doesn't mean a single sentence is enough for an AI to clarify all features.
But a key concept must be emphasized here: simply "adding an AI feature" to an existing product is not a transformation. The real opportunity lies in redesigning products with an AI-Native mindset, making AI the core engine of the product, not just a decorative accessory.
There's no need for Taiwan to go head-to-head with OpenAI and Google on general-purpose large language models; that's a money-burning arms race. The true late-mover advantage lies in combining Taiwan's industrial strengths (manufacturing, healthcare, finance, semiconductors) to develop Small Language Models (SLMs) for specific domains. These models don't need billions of parameters, but their performance on specific tasks can surpass general-purpose models, and they have lower deployment costs, faster inference speeds, and better data privacy control.
As a conceptual example, fine-tuning a pre-trained model with an open-source framework for a specific domain is not as difficult as one might imagine:
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer
# 載入基礎模型（以輕量級模型為例）
model_name = "microsoft/phi-2"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
# 使用 LoRA 進行參數高效微調，大幅降低訓練成本
lora_config = LoraConfig(
r=16, lora_alpha=32, lora_dropout=0.05,
target_modules=["q_proj", "v_proj"], # 只調整部分權重
)
model = get_peft_model(model, lora_config)
# 準備領域專用資料集（例如：法律條文、金融報告、製造 SOP）
# domain_dataset = load_dataset("your-domain-data")
# 微調訓練 — 用自己領域的資料教會模型「說行話」
trainer = SFTTrainer(
model=model,
train_dataset=domain_dataset,
args=TrainingArguments(
output_dir="./domain-slm",
num_train_epochs=3,
per_device_train_batch_size=4,
),
)
trainer.train()
The point is not the code itself, but the strategy behind it: use your own domain-specific data to train a proprietary model that is difficult for others to replicate. This is the moat.
A Practical Blueprint: How AI Can Reshape the Competitiveness of Taiwan's Software Industry
Putting the above ideas into practice, AI can reshape the competitiveness of Taiwan's software industry in three aspects:
Overcoming market limitations: In the past, creating a multi-language product involved significant costs for translation and localization alone. Now, the quality of AI translation has reached a "usable" or even "good" level. Combined with human review, a small team can manage multiple language markets simultaneously. Language is no longer the ceiling; product strength is. Former Google Taiwan Managing Director Lee-Feng Chien also pointed out that language barriers are actually greater than cultural barriers, and AI is rapidly eliminating this obstacle.
Alleviating the talent gap: 71% of employers in Taiwan report difficulty filling key positions. AI Copilots won't replace engineers, but they will enable existing talent to do more. Tasks that used to consume a lot of man-hours, such as automated testing, code review, and documentation, can now be handed over to AI for the first draft, allowing engineers to focus on architectural design and product innovation. In my own experience, after introducing AI-assisted development, many repetitive tasks have indeed been significantly reduced, allowing the team to spend time on more valuable things.
Optimizing the development process: Unclear requirements and frequent specification changes are the most common reasons for software project failures in Taiwan. AI can extract insights from user behavior data, predict feature priorities, and even evaluate the impact scope of a requirement change based on historical data. This is not meant to replace the judgment of a PM, but to provide more data to support decisions and reduce the risks associated with "making decisions on a whim."
Conclusion: Seizing the AI Opportunity to Ignite Taiwan's Software "Productization" Revolution
The predicament of Taiwan's software industry is structural. Market size, talent distribution, and development culture are problems that won't automatically disappear with mere effort. The only way out is a fundamental mindset shift: from the Project model of selling time to the Product model of creating assets, with AI as the core engine to achieve this transformation.
Now might be the best time. The small scale of Taiwan's software industry means it doesn't have the "slow to turn" baggage of large corporations. Focusing on vertical domains, designing products with an AI-native mindset, and developing domain-specific SLMs is a path that doesn't require burning huge amounts of money. However, it does require the determination to leave the comfort zone of "taking any project that comes along." This is not just a technological upgrade, but a business model choice crucial for future survival.
Comments