Flitto Deep Dive: One Big Tech Data Order Grew 5.6x. The Stock Is 56% Below Its High. The November 10 Filing Gives the First Answer.

Fri, 30 Oct 2026 12:03:00 +0900

In October 2025, a US big tech company sent Flitto an order for AI training data worth 4.23 billion won. Over the next year the customer extended that order once and raised it five times. It now stands at 23.78 billion won ($16.9 million), 5.6 times the original. Flitto’s stock went the other way. It closed at a high of 19,750 won on December 2, 2025, and traded at 8,740 won on the morning of October 30. That is 56% below the high.

The filings show what the market is worried about. The contract’s end date is November 10. But the filing says delivery was due by September 26, and November 10 is the deadline for payment, 45 days after delivery. The 12.63 billion won of data added on June 29 also had to be delivered by September 26. Third-quarter revenue rose 9.68 billion won from the second quarter, and operating profit of 8.97 billion won was a record. That jump overlaps the delivery window for the June 29 increase. The company does not report revenue by contract, so contract A’s exact contribution cannot be confirmed. As of October 30, no new order big enough to repeat it in the fourth quarter had been filed.

Our view: the third-quarter profit will not repeat in the fourth quarter. But the filing record does not support calling the contract a one-off windfall either. This customer has ordered about every six months since September 2024. In December 2025, instead of closing this order, it extended the order by nine months and then kept adding volume. The AI training-data bottleneck supports demand for human-made multilingual data. So we think a next order is likely. We think it is unlikely to be as large as 2026. The same customer’s other contract was cut 55% in October, and no autumn order has appeared yet. The size of the next order will move the stock more than whether there is one.

The first answer will likely come from the amendment filing around November 10. This customer’s earlier orders each drew an amendment filing on their end date. The second comes from the third-quarter report, due by November 16, which will show the backlog at the end of September.

Written October 30, 2026. This is filing analysis, not a recommendation. The customer’s name is withheld at its request to protect trade secrets. We assume that the “global IT company” in the contract filings and “US-based global big tech company A” in the half-year report are the same customer. In our October 30 Korea small-cap discovery screen, Flitto had the best earnings of the six names, but domestic institutions bought it mostly on one day, the day of its earnings release.

Key points: the third quarter was a delivery quarter; the order book decides the next one

All of Flitto’s disclosed 2026 data orders were increases to 2025 orders. The customer added volume to its October 3, 2025 order five times, for a total of 19.55 billion won ($13.9 million).
November 10 is a payment deadline. Delivery was due September 26. So the filing around November 10 is likely to fix the final contract amount.
Final amounts have changed before. The customer’s March 2025 order was cut 16% at its end date. In December 2025, contract A was extended instead, and it grew in 2026.
Profit follows data deliveries. In 1Q26, revenue of about 5.2 billion won left operating profit near zero, and 78% of the revenue added from the second to the third quarter became operating profit.
Fourth-quarter revenue is about 10.8 billion won from filed contracts alone, and up to about 16.5 billion won by backlog arithmetic (our estimates). The won has strengthened since September, which also shrinks the won value of the remaining dollar contracts.
The demand case is strong. Public text is projected to run out between 2026 and 2032, and key web sources have started to block AI crawlers. Korean is 0.88% of Common Crawl pages.
Books are a stock that you scan once. Flitto sells a flow that it makes to order. A US court held that scanning purchased books for training was fair use, and a pirated library led to a settlement of about $1.5 billion.

1. One purchase order grew 5.6x in a year

Contract A is an order received on October 3, 2025. The filing describes it as “data supply for the research and development of AI-based language models.” The counterparty is a “global IT company.” Payment is due within 45 days of delivery.

Filing date	Change	Contract value	In dollars
2025-10-10	First filing (order received Oct 3)	4.23bn won	$3.01M
2025-12-22	End date moved from 2026-02-10 to 2026-11-10	4.23bn won	$3.01M
2026-04-03	Increase	5.68bn won	-
2026-04-15	Increase	7.46bn won	-
2026-06-26	Increase	10.04bn won	$7.14M
2026-06-29	Increase to 111.65% of prior-year revenue (trading halt)	22.67bn won	$16.12M
2026-09-02	Increase	23.78bn won	$16.92M

All won values use the October 2, 2025 exchange rate of 1,406 won per dollar. The filing says the terms can change with the quantity, type and timing of the data the customer requests. The contract value is the order ceiling, and revenue is booked on delivery. Sources: DART Oct 10, 2025, Dec 22, 2025, Jun 29, 2026, Sep 2, 2026.

No new data order from this customer was filed in 2026. The ceiling on an existing order kept rising instead. That suits the customer: it can add languages and volume as it needs them without a new contract. It does not suit investors: nobody can tell in advance whether the next tranche will appear as a new contract filing or as an amendment.

2. November 10 is a payment deadline

The September 2 amendment explains how the contract period was set. The end date reflects “the contract term that payment is made within 45 days of delivery completion (September 26, 2026).” The June 29 filing gave the same delivery date. So the volume added on June 29 had to be delivered within three months. That timing fits the 9.68 billion won jump in third-quarter revenue over the second quarter.

What will the November 10 filing show? The customer’s earlier orders show how its contracts end. Each of them drew an amendment filing on its end date, under continued confidential treatment that discloses only summary terms at the customer’s request.

Order	First filed value	Last value	What happened at the end
September 2024 order	3.91bn won	4.26bn won	Raised before the end, held at the end date
March 2025 order	9.04bn won	6.68bn won	Cut 16% at the end-date amendment, “final supply amount changed” (-26% vs first filing)
July 2025 order	6.37bn won	6.37bn won	Extended, value held
October 2025 contract A	4.23bn won	23.78bn won	End date pushed out nine months, raised five times, November 10 pending
October 2025 contract B	3.88bn won	4.03bn won	Raised to 8.89bn won in June, cut 55% on October 6, end date moved to 2027-02-09

Sources: DART Aug 12, 2025, Nov 11, 2025, Nov 28, 2025, Oct 6, 2026.

Based on those precedents, the filing around November 10 can be read like this:

Extension and increase: the December 2025 precedent. The customer reopened the order instead of closing it. That would be the first signal of 2027 volume.
Settlement cut: the March 2025 precedent. The final supply amount is fixed below the ceiling, so less volume is left in the contract. The filing alone cannot show why, or whether revenue already booked is revised.
Value held: the July 2025 precedent. The contract ends at 23.78 billion won, and the next volume has to come from a new order.

Contract B has already sent a signal. On October 6 its value fell from 8.89 billion won to 4.03 billion won ($2.65 million), and its end date moved from November 10 to February 9, 2027. The new delivery date is December 26. The filing gives the reason only as “change in contract amount and period.” We cannot tell whether the customer cut volume or moved it to a later contract.

3. Profit follows data deliveries

Quarter	Revenue	Operating profit	Margin
1Q25	4.61bn won	0.17bn won	3.8%
2Q25	9.38bn won	2.43bn won	25.9%
3Q25	11.78bn won	2.83bn won	24.0%
4Q25	10.21bn won	0.72bn won	7.1%
1Q26	5.21bn won	-0.06bn won	-1.2%
2Q26	8.11bn won	1.40bn won	17.2%
3Q26 (preliminary)	17.79bn won	8.97bn won	50.4%

Consolidated. 3Q26 is from the October 20 preliminary earnings filing. 4Q25 is the full year minus the first nine months. Nine-month operating profit of 10.31 billion won already exceeds the 6.15 billion won for all of 2025.

In 1Q26, when data deliveries were light, revenue of about 5.2 billion won left operating profit near zero. We read that as a sign that quarterly fixed costs are in the 5 billion won range. From 2Q26 to 3Q26, revenue rose 9.68 billion won and operating profit rose 7.57 billion won: 78% of the added revenue became profit. If the same structure works in reverse, profit falls back to the 1Q26 level when data deliveries stop.

Data sales revenue grew from 8.90 billion won in 2023 to 13.99 billion won in 2024 and 21.78 billion won in 2025. Exports were 18.75 billion won of the 2025 total. In the first half of 2026, data sales were 6.59 billion won, or 49.5% of revenue. In its 2025 annual report, the company says average prices are rising as contracts get larger and the price per language goes up. Sources: 2025 annual report, 2026 half-year report.

4. The fourth quarter: between filed contracts and backlog arithmetic

The half-year report puts the data backlog at 28.31 billion won on June 30. The report defines it as the amount of contracts not yet ended at the report date, converted at the report-date exchange rate. At the end of June, contracts A and B together came to $21.97 million, or about 34 billion won at the June-end rate of about 1,548 won per dollar. The backlog is smaller than that, so we read it as the amount not yet delivered.

The exchange rate matters too. The dollar fell from about 1,548 won on June 30 to about 1,358 won on September 30 and about 1,337 won on October 30, as the won strengthened. The won value of the remaining dollar contracts fell with it. So we did the fourth-quarter arithmetic in dollars. All of it is our estimate.

Case	4Q revenue	Calculation
Filed contracts only	about 10.8bn won	Base business about 3.4bn won (first-half non-data revenue averaged 3.36bn won a quarter) + contract B up to $2.65M (about 3.5bn won) + contract C 3.87bn won. Assumes the contract B remainder and contract C are all booked in the fourth quarter
Backlog arithmetic	up to about 16.5bn won	June-end backlog about $18.3M + September contract A increase $0.8M - estimated 3Q data deliveries of $9.4M to $10.1M - October contract B cut $3.2M = about $5.8M to $6.5M (about 7.8 to 8.7bn won). Add 3.5 to 4.0bn won of base business and contract C

The third-quarter data deliveries are third-quarter revenue minus 3.5 to 4.5 billion won of non-data revenue, converted at the third-quarter average rate of about 1,420 won per dollar. If part of the data revenue comes from won contracts, the numbers change.

The gap between the two is about 4.2 to 5.2 billion won ($3.2 million to $3.9 million) of data backlog that contract B cannot explain. It may be contract A volume that missed the September 26 delivery date. It may also be another customer’s data contract. This figure comes from our arithmetic; it is not a confirmed backlog. The September-end backlog in the third-quarter report is the first thing to check it against.

Contract C is a solution build, not a data sale. On October 8, Flitto signed a 3.87 billion won contract with a “global semiconductor company” to build an AI real-time interpretation system. It runs to December 30, and the value combines a solution license with system-integration labor. The filing says Flitto had no similar contract with this counterparty in the past three years. Because it is a build project with labor costs, its margin may be lower than data sales. Source: DART Oct 12, 2026.

If operating profit is (revenue - 5.2 billion won) x 0.7, fourth-quarter operating profit is about 3.9 to 8.0 billion won. The 5.2 billion won is 1Q26 breakeven revenue, and the 0.7 is the third-quarter incremental margin of 78%, cut for caution. Neither is a tested model. On these assumptions, 2026 operating profit is about 14.2 to 18.3 billion won, and the market value of about 144.5 billion won (October 30, intraday) is 8 to 10 times that. At the end of June, cash and financial assets were about 18.8 billion won against about 0.1 billion won of debt (half-year report, our calculation). No filed contract yet carries volume for delivery in 2027. Even contract B finishes delivery on December 26. The stock sits between 2026 profit and a 2027 that is still empty.

5. Demand: the human writing that AI learns from is running short

Why the customer buys this data is the case for a repeat order.

Public text is finite. Epoch AI estimates the stock of quality- and repetition-adjusted public human text at about 300 trillion tokens (90% interval: 100 to 1,000 trillion). At current trends it will be fully used between 2026 and 2032 (80% interval), or earlier with heavy overtraining. Epoch AI

Usable text is shrinking faster. Researchers at the Data Provenance Initiative reported that in one year, from 2023 to 2024, more than 5% of all tokens in the C4 dataset, and more than 28% of its most actively maintained critical sources, became fully restricted for AI crawling. By terms of service, 45% of C4 was restricted. Consent in Crisis, arXiv 2407.14933

AI-made text is a poor substitute. Shumailov and co-authors showed that when each model generation trains on the previous generation’s output, the tails of the original distribution disappear. In runs that kept 10% of the original data, performance degraded only slightly. Human-made data slows the collapse. Nature, 2024

Korean was scarce from the start. In Common Crawl’s latest crawl (CC-MAIN-2026-43), Korean is 0.88% of pages and English is 42.6%. Common Crawl statistics In Meta’s Llama 2 pretraining data, Korean was 0.06%. Llama 2 paper, Table 10 The rare language pairs Flitto highlights, such as English-Arabic, specialist fields such as medicine and law, and generational slang are even harder to find on the web.

6. Books are a stock; Flitto sells a flow

The same bottleneck explains why AI companies buy and scan print books. On June 23, 2025, Judge William Alsup of the US District Court for the Northern District of California split his ruling in Bartz v. Anthropic. Training a model on lawfully purchased books was fair use. Destroying purchased print originals and scanning them into digital copies was also fair use, because the number of copies did not grow. Keeping millions of books downloaded from pirate sites in a central library was not fair use. Anthropic agreed to a settlement of about $1.5 billion (about $3,000 per work), and the court granted preliminary approval on September 25, 2025. Goodwin analysis of the ruling, Publishers Weekly It is a district court ruling and does not bind other courts.

The ruling shows the economics of data supply. Books are a stock of text written in the past. You scan them once, and no new ones appear. Flitto sells something else. On its own platform, Arcade, it pays participants in points to complete language missions, and so produces new data to order in the languages and formats a customer wants. The half-year report lists the data types as text and voice. The company names multilingual data checked by both machine translation and professional translators as its edge.

That difference is the logic of repeat orders. You buy a stock once. You buy a flow again each time a model needs a new language, a new feature or a new evaluation standard. The customer’s behavior over the past year, raising the ceiling instead of closing the order, fits that logic.

7. The other side: a bottleneck is not the same as Flitto’s volume

There is one customer. The half-year report says US big tech company A is more than 10% of revenue. Both 2026 data contracts, A and B, name a “global IT company” as the counterparty, and we take it to be company A. If its data roadmap (languages, formats, timing) changes, volume falls whatever the bottleneck does.
The size signals have weakened. Contract B was cut 55% from its peak on October 6. A new order arrived each autumn in 2024 and 2025, but none for autumn 2026 had been filed by October 30.
Small orders show up late. The disclosure threshold is 10% of the prior year’s revenue, now about 3.6 billion won. Contract B was first filed six months after the order, when an increase pushed it over the threshold. No filing does not mean no order. But when a filing comes, it can lag the order by months.
$16.9 million is small money for big tech. We think the next order depends on the customer’s data needs more than on its budget. The customer can also split volume with other data suppliers.

8. What to check next

Date	Event	How to read it
Around November 10	Contract A amendment	Extension plus increase is the first signal of 2027 volume. A settlement cut means less volume left for the fourth quarter. A held value means the next volume needs a new order
By November 16	Third-quarter report	Our arithmetic puts the September-end data backlog at about 12 to 13 billion won (before the contract B cut). A higher figure may mean new volume arrived; a lower one may mean larger deliveries or cuts. Check the data share of revenue too
December 26 to 30	Contract B delivery, contract C completion	Watch for delay or cut amendments
February 2027	2026 results	Full-year operating profit below about 14.2 billion won means the fourth quarter fell short of our low end. Whether deliveries, exchange rates or costs caused it needs a separate check
Any time	A new order from the same customer	Filed if it is 3.6 billion won or more

Foreign ownership rose from 3.8% on July 1 to 11.3% on October 29. Domestic investment trusts bought around the earnings release (October 20 to 22), and pensions, trusts, private funds and insurers together sold a net 1.2 billion won on October 26 to 29. The stock fell 24% from an October 21 close of 11,550 won to 8,740 won intraday on October 30.

When AI training data runs short, who keeps the profit longer: the companies that own data, or the companies that make it? And will the November 10 filing be an extension or a settlement? Tell us your call in the comments.

Research and information only; not investment advice. Not a trade instruction or a personalized portfolio decision. The fourth-quarter and 2026 profit ranges are our estimates built from filed numbers. Figures are date-bound, and the October 30 price is intraday and provisional. Disclosure: AI tools, including Anthropic’s, assisted research and editing. Details of the Anthropic case cite only published legal analysis and reporting.

AI Training Data on Korea Invest Insights