Thai AI startups are now at a critical turning point. As foundation models like GPT and Claude continue to improve, startups that only build wrappers or UI layers will inevitably be overtaken.
The problem is that most of us still rely on business models dependent on OpenAI APIs. When they adjust pricing or add new features, we become mere middlemen with nothing special to offer. Customers can simply go directly to OpenAI instead.
I believe we have only 12 months left to build truly defensible moats. We need to find proprietary data, create deep domain expertise, or develop specialized models specifically for the Thai market. Otherwise, we’ll become just free add-ons that nobody wants to pay for.
Competition now isn’t about technology anymore - it’s about moats and timing.
Case study examples of Thai startups under threat
Take one startup that built Thai language chatbots for banks. Last year, they could still charge $60,000-$90,000 per project. But as Claude and ChatGPT improved their Thai language support, clients started asking “Why should we hire you when we can use OpenAI API directly?”
Now they’re cutting prices in half, but still can’t compete with foundation models that improve monthly. Revenue has dropped 60% over the past 6 months, and new competitors using GPT-4 as backend are selling even cheaper.
I think this case is a clear warning sign. Without finding new strengths quickly, startups like this will become unwanted middlemen.
Thai AI landscape: Where do we stand in the bigger game
Most Thai AI startups are stuck in the same trap: building on others’ foundation models and wrapping them into solutions. When OpenAI or Google adjusts pricing or releases new features, our businesses get immediately impacted.
Many previously strong-looking startups are struggling to adapt because what we can sell today might become a free built-in feature of foundation models tomorrow. Foreign competitors are also grabbing market share with cheaper and better APIs.
I think the Thai landscape is in a dangerous transition period. Those who don’t have defensible moats or irreplaceable unique value need to urgently find solutions within these 12 months. Otherwise, they’ll become mere add-ons in a bigger system.
Comparison: Before and after foundation models rule the world
| Factor | Before Foundation Models | After Foundation Models |
|---|---|---|
| Barrier to Entry | Very High - Need ML team | Very Low - Just call APIs |
| Development Time | 6-12 months | 2-4 weeks |
| Core Value | Algorithm & Model | Data + UX + Integration |
| Competition | Tech vs Tech | Execution vs Execution |
| Defensibility | Model Performance | Network Effect + Data Moat |
Before, AI startups had to build their own models, competing on algorithms and accuracy. But now everyone uses GPT-4 or Claude alike. The competition has shifted to who manages data better and who creates more engaging UX.
I think the game has truly changed. Now we’re not competing on AI, but on who builds moats that competitors can’t cross.
Four key strategies for building defensible moats
Proprietary data moats means accumulating data competitors can’t get. Like Krung Thai Bank with millions of transaction records, or CP Group with supply chain data from 7-Elevens nationwide.
Specialized domain expertise requires deep industry focus, like startups doing AI for shrimp farming since Thailand is the world’s 3rd largest shrimp exporter, or AI for rubber plantations.
Regulatory advantages uses laws as barriers, like banks with banking licenses or fintechs that passed BOT sandbox.
Network effects become more valuable with more users, like platforms connecting Thai SMEs with suppliers or marketplaces for farmers.
I think one strategy isn’t enough - you need to combine several approaches to survive.
Comparison of leading startup approaches
| Factor | Jasper AI | Character.AI | Hugging Face |
|---|---|---|---|
| Main Strategy | Brand + Content Marketing | User-generated Characters | Open Source Community |
| Moat Type | Customer Lock-in | Network Effects | Developer Ecosystem |
| Revenue Model | SaaS Subscription | Freemium + Premium | Enterprise + Cloud |
| Foundation Model Defense | Specialized Workflows | Social Features | Infrastructure Layer |
Jasper built a strong brand in the marketer segment by focusing on specialized templates and workflows. Character.AI uses social elements letting users create their own AI characters, building community.
Hugging Face went back to building an infrastructure layer for everyone to use. Instead of competing with foundation models, they became the platform for them.
I think Thai startups should look at Character.AI as a model because social features and user-generated content create the highest switching costs.
Pros and cons of rushing to build moats
Pros
- +Creates high switching costs making it hard for users to move to competitors
- +Provides pricing power since you're not just an AI model wrapper
- +Builds network effects from community or user data
- +Prevents becoming a commodity when foundation models improve
Cons
- −Requires heavy investment in time and money while still finding product-market fit
- −Might miss opportunities in fast-changing markets
- −Limited resources of Thai startups may not build strong enough moats
- −High risk if you choose wrong moat or market doesn't want it
I think a balanced approach is starting with simple wrappers first, then gradually building moats from user behavior data. Honestly, waiting 12 months might be too late because foundation models develop so fast.
Better to start with what’s possible today, then collect data to iterate further.
Hidden costs in building defensible positions
The invisible cost more serious than investment money is opportunity cost from not launching products quickly. While we spend 6-8 months building moats, competitors might grab market share.
Finding talent who understands both AI and domain expertise costs more than expected because most talented people went to work on foundation models. R&D budget needs at least 40-60% of first round funding.
Time-to-market delay is most dangerous because foundation models improve every 3-4 months. Instead of building perfect moats from the start, I think we should begin with MVPs having basic differentiation, then build moats from real user data.
Waiting until 100% ready might mean missing opportunities entirely.
Which Thai startup groups should rush, which don’t need to yet
Urgent groups: B2B automation, fintech AI, healthcare AI with natural regulatory barriers, and vertical SaaS requiring deep domain expertise, because general GPT can’t access sensitive or industry-specific data.
Groups not rushing yet: content generation, basic chatbots, translation services competing directly with foundation models. Instead of fighting high walls, I think they should pivot to narrower niches.
Middle groups: e-commerce personalization, education tech still have opportunities, but need to quickly build network effects or proprietary datasets because OpenAI is building a plugins ecosystem that might capture this market share.
It’s safer to fight in arenas where local data or regulations provide advantages.
The next 12 months will be the turning point
The countdown has begun. In the next 12 months, foundation models will commoditize AI features we use as selling points today. Those without defensible moats will become wrapper services that customers can switch from with one click.
Prepared startups will gain advantages from having proprietary data, network effects, or regulatory compliance that can’t be competed against. This group will find Series A easier and get higher valuations than before.
Groups still searching for moats will have to compete on price with everyone globally using the same ChatGPT API. I think by then, only 2-3 major players will remain in each vertical.