Cheap AI is killing AI companies. That's the point.

There is a simple way to start an AI company, and for a while it was a pretty good way. You rent access to a powerful model, put a cleaner interface on top of it, pick a job that people dislike doing, and charge more than the model costs you. This is not a stupid business. Many useful businesses begin as wrappers around someone else's infrastructure. The awkward part comes later, when the infrastructure company notices the same customer, lowers the price, improves the product, and starts looking less like your supplier and more like your competitor.

I first saw that problem as a price tag. Not in a pitch deck, not in a technical architecture diagram, but in a subscription plan: twenty dollars a month. An American startup had sold AI-written marketing copy, raised heavily, and reached a valuation of about 1.5 billion US dollars in late 2022. Then ChatGPT arrived at 20 dollars a month and did enough of the same job. The startup had been charging roughly three times that, which is a fine price when the customer cannot get the thing anywhere else and a difficult price when the customer can get it from the company making the thing.

That is the funny part, though not funny if it is your revenue line. The issue was not that the startup forgot to add a button, or hired the wrong sales team, or failed to say "AI-native" with sufficient conviction. The thing it sold had become cheap. According to research firm Contrary, the company lost roughly one in six customers within eighteen months. Trade-press estimates, which no auditor has checked, put revenue down by about half from the peak. Sacra had described the business model plainly in 2022: a reseller of someone else's AI. That sentence is not a footnote. It is the plot.

I spend a lot of time looking at AI companies before they become obvious. Founders show growth, customers praise demos, and decks contain diagrams in which all the arrows point toward inevitability. But the price curve is less impressed. Cheap AI is not killing the AI industry. It is killing companies that confused access to intelligence with ownership of a business.

Thailand's late start just got cheaper

The Nation reported in late May that Thailand ranked 89th in AI use, with 10.7 percent of the population using AI against a world average of 16.3 percent. Vietnam stood at 23.5 percent. Singapore stood at 60.9 percent. The natural reading is that Thailand is late, and late is usually bad. Late means the good seats are gone, the winners have been picked, and someone else has already named the category with an English word that no one wants to translate.

But technology markets have a strange habit. Being late to a bad version of a product is not always a loss. A country buying AI in 2026 is not buying the 2023 version. It is buying after the price has collapsed, after the first weak products have been exposed, and after every serious buyer has had three years to learn that a good demo and a good workflow are different things. That does not make delay a strategy. It does make the bill smaller.

The constraint is no longer access, because access is rapidly becoming the cheap part. The National Board of Digital Economy and Society projects Thailand's digital economy at about 5.6 trillion baht this year. The surface area is already here: accounting, logistics, clinics, schools, factories, retail, call centers, law offices, and banks. The interesting question is not whether Thailand can buy AI. Of course it can. The interesting question is what Thailand attaches cheap intelligence to.

The engine is not the business

Here is the price curve, because this is where the story stops being philosophy and starts being accounting. In March 2023, the most capable AI engines charged about 60 US dollars for a standard unit of work, roughly the text of seven novels processed. By early 2026, public price lists tracked by Epoch AI showed the same quality of work selling for about one dollar. Sixty became one. Stanford University's AI Index found a steeper version one tier down: the cost of producing a 2022-grade AI answer fell more than 280-fold between late 2022 and late 2024. Epoch AI's data, as of June 2026, shows the price of a fixed quality of answer now halving roughly every two months.

This is wonderful news if AI is an input to your business. It is less wonderful if AI is your business. Imagine a restaurant that became famous because flour was hard to buy. The flour gets cheaper, which is good for restaurants, but bad for the restaurant whose competitive advantage was "we have flour." AI models are not flour exactly, because flour does not occasionally write Python or summarize a contract, but the economic point is the same. When the raw material gets cheap, the value has to be somewhere else.

For AI products, the raw material is the engine: the model built by a handful of global laboratories and rented through APIs. The engine keeps improving, and the rent keeps falling. Around that engine sits the actual business: your files, your history, the systems it connects to, the routines it changes, the data it accumulates, and the small operational frictions that make leaving inconvenient. The engine is what makes the product impressive on day one. The workflow is what makes it hard to remove on day one hundred.

The marketing-copy startup owned the engine badly and the workflow barely at all. When the landlord started selling directly at a lower price, the customer's decision stopped being strategic and became arithmetic. Another American writing tool followed the same pattern. It rented its models, raised 13.9 million dollars according to Crunchbase, and was acquired in October 2025 at an undisclosed price, according to the buyer's announcement. That is not a moral failure. It is a layer problem.

Now compare the companies that look less fragile. Cursor, a programming tool, became the place many users actually write software. It grew from 100 million US dollars of yearly revenue in January 2025 to two billion by February 2026, according to TechCrunch and Bloomberg reporting in March 2026. Harvey, a legal AI company, reports 98 percent customer retention and says existing customers grow spending by about two-thirds. Those are company-published figures, not audited ones, so keep the usual investor discount handy. Still, the shape matters. Both rent the engine. Both try to own the work.

Cheap does not mean small

One tempting conclusion is that if AI gets cheaper, the AI market gets smaller. This is a reasonable instinct and also, so far, wrong. IDC estimates spending on AI infrastructure at 153 billion US dollars in 2024, 318 billion in 2025, and 487 billion projected for 2026.

The price went down, usage went up, and the money moved. That is the useful distinction. A collapse in the price of a unit of intelligence is not the same as a collapse in demand for intelligence. It is closer to what happens when a useful input becomes cheap enough to use everywhere. The old margin disappears in one place and reappears, if it reappears at all, in the systems that can use much more of the input.

This is the part that matters for Thailand. The failures on the front pages are not proof that AI is a bubble. They are proof that the market is sorting. One group sold intelligence as if intelligence would stay scarce. The other used cheap intelligence to make a workflow harder to leave. The first group had a product demo. The second group had a place in the customer's day.

The Thai founder's constraint

A Thai founder does not need to win the global model race, which is useful because the global model race is being run by companies with data centers, model teams, and balance sheets measured in tens of billions of dollars. There are easier ways to spend a decade than competing with that directly. The local constraint is different, and more promising: Thai accounting is not generic, Thai logistics is not generic, Thai clinics are not generic, and Thai procurement, language service work, education paperwork, restaurant operations, and bank workflows are not empty boxes waiting for a chatbot.

Those are places where cheap intelligence has to become part of work. A founder who sells a Thai interface over rented intelligence is standing on the falling curve. A founder who connects cheap intelligence to a painful local workflow is standing somewhere more defensible. The product may still fail, because most products do. But at least it fails because the workflow was not valuable enough, not because the raw material became cheaper on someone else's pricing page.

The non-obvious opportunity is that falling model prices make local specialization more attractive, not less. When the engine was expensive, everyone had to spend attention on conserving it. When the engine becomes cheap, the scarce work shifts to knowing where to put it. That is a better game for Thai founders than pretending to be a global model lab with a smaller office and a more optimistic spreadsheet.

The Thai buyer's constraint

A Thai buyer has the opposite problem. The cost of experimenting is falling, but the cost of installing the wrong workflow is not. A bad AI subscription is mildly annoying. A bad AI workflow, wired into documents, approvals, customer records, and staff habits, is more like a renovation carried out by someone who only saw the building over Zoom.

That changes the test. The demo is not the test. The first output is not the test. The smart answer is not the test, because every major model will keep producing smarter answers at lower prices. The test is what remains after ChatGPT, Gemini, Claude, and the next model all get cheaper again. Does the product know the company's documents? Does it connect to its systems? Does it remember work in a useful way? Does tomorrow's task get easier because it saw yesterday's? If not, the vendor is asking the buyer to pay a premium for a falling input.

Asst. Dr. Tanwa Arpornthip is Senior Advisor to the Venture Capital Team at SCB 10X, where he provides insights on emerging technologies, innovation ecosystems, and startup development. He is also a lecturer at the Faculty of Technology and Environment, Prince of Songkla University, Phuket Campus.

About SCB 10X

SCB 10X is the disruptive technology investment arm of SCBX Group. With an investment track record spanning since 2016, SCB 10X has deployed over USD 500 million globally into startups in AI, blockchain, and fintech. SCB 10X has backed exceptional companies such as Together AI, Pagaya, Ripple, Fireblocks, Anchorage Digital.

Beyond capital, SCB 10X partners with our portfolio founders to test, grow and scale their solutions through SCBX’s network, unlocking commercial opportunities into Thailand and Southeast Asia. Mandated as the group’s speedboat, we discover and ship state-of-the-art technologies and solutions into SCBX group.

For more information, please visit https://scb10x.com/