Kakao posted its best quarter ever, then announced what it won't build

On the morning of August 6, Kakao filed its Q2 2026 results. Consolidated revenue of 2.0985 trillion won (roughly $1.47 billion) and operating profit of 277 billion won (roughly $194 million). Both are all-time quarterly records. Operating profit rose 36% year over year and beat the analyst consensus by 22%. On the numbers alone, this was a quarter worth applauding.

But the heaviest sentence spoken on the earnings call that morning had nothing to do with the numbers. CEO Shina Chung announced that Kakao will not enter the AI data center or GPU cloud business. Her reasoning: "AI infrastructure businesses require massive upfront investment and continued expenditures on facilities," and because "the generation changes of advanced chips and AI servers are increasingly fast" alongside competitive supply expansion, "it will be difficult to guarantee a sustainable profitability." Translated: the very reason that market is printing money right now is the reason it won't last.

Here's why the timing makes this news. Exactly one day later, on August 7, Naver reported its own Q2. Revenue of 3.3888 trillion won, up 16.2%. Operating profit of 520.3 billion won, down 0.2%. The cause of that squeeze was AI infrastructure spending, and Naver has committed to scaling the Nvidia AI factory at its Sejong data center toward a long-term target of one gigawatt, with Nvidia and Brookfield as partners. Korea's two dominant platforms answered the same question in opposite directions in the same week. One is building the power plant. The other is refusing to build one and selling the wall socket instead.

And Kakao showed the first socket. It's food delivery. Kakao is entering an agent-to-agent (A2A) partnership with Coupang Eats, letting users get a recommendation, place an order, and complete payment entirely inside a KakaoTalk chat window. No app switching. Chung framed the thesis this way: competitiveness in the agentic AI era doesn't come only from building a better model, but from understanding user intent and connecting it to actual action.

The market's first reaction was lukewarm. Kakao shares slipped roughly 2% to around 37,500 won on the day of the release, and the seven brokerages that published notes afterward cut their average target price from about 64,000 won to about 56,000 won — a 13.1% reduction. A record quarter that got its price target trimmed. Explaining that gap is what the rest of this piece is for.

An app 49.6 million Koreans open every day, and a CEO 29 months into the job

Start with scale. KakaoTalk's domestic monthly active users hit 49.631 million in Q2. South Korea's population is around 51 million, so this is effectively every smartphone owner in the country. Including overseas users, global MAU was 55.346 million. Over the past two years that domestic figure has crept from 48.93 million to 49.63 million — never declining, never really accelerating. That's a number with almost no room to grow and almost no reason to fall.

Shina Chung became Kakao's sole CEO on March 28, 2024, confirmed by shareholder meeting and board vote. She came from six years running Kakao Ventures and is the company's first female chief executive. Her reappointment was confirmed in March 2026, starting a second term. What's striking is that most of what she has done since taking over is subtraction: shedding affiliates, freezing hiring, simplifying the governance structure.

You can read that directly off the headcount. Consolidated employees stood at 14,046 in Q2, down 2,338 from 16,384 a year earlier. Most of that drop is the Kakao Games deconsolidation rather than layoffs — headquarters headcount fell only from 3,986 to 3,918, a difference of 68 people. The IR deck's note on personnel expense is a single phrase: maintaining a conservative hiring stance. Personnel cost came in at 448 billion won, up just 2% year over year.

Kakao Games was the largest piece of that cleanup. In June 2026, an investment vehicle affiliated with LY Corporation completed payment on a 240 billion won rights issue and a 60 billion won convertible bond, taking a 33.43% stake. Kakao's holding fell from 37.93% to 14.68%, dropping it to second-largest shareholder. Kakao Games was classified as an asset held for sale in Q1 and left the consolidation scope in Q2. Kakao Healthcare was divested in the same cleanup.

That cleanup left one mark on the income statement that headlines mostly skipped. Net profit was 18 billion won, down 90% year over year. Operating profit up 36% while net profit falls 90% means something large happened below the operating line, and it did: a 180.8 billion won loss from discontinued operations tied to the deconsolidation and stake sales, plus 165.1 billion won in corporate tax expense, both recognized in the same quarter. The company called these one-time items, which is accurate accounting. It is also accurate that the cash and the equity are genuinely gone.

Now the AI org, which is the actual subject here. Kakao's in-house model brand is Kanana. On July 28 the company open-sourced four lightweight language models designed to run on phones — Kanana-2-1.3B and Kanana-2-3B, each in base and instruct variants — and it has also shipped Kanana Safeguard, which it describes as Korea's first open-source AI guardrail model. Kakao signed a partnership with Google DeepMind to apply SynthID watermarking and detection to its own models. And since 2025 it has worked with OpenAI on ChatGPT for Kakao, a ChatGPT tab pinned at the top of the KakaoTalk chat list, which has now passed 13 million cumulative subscribers. Usage runs at more than six messages per user per day and roughly eight minutes of daily engagement.

Put that together and the strategy is legible. Kakao is not trying to build a frontier model. It rents the big brain from OpenAI, builds the small fast brain itself and pushes it onto the device, and plugs both into a screen that 49.6 million people already open. This earnings call was the first time that plan came with numbers, partner names, and dates attached.

The quarter in numbers, and the 58 billion won that sits outside the table

Segment by segment, the character of this quarter is obvious. Platform earned all of the money.

Consolidated 2Q25 1Q26 2Q26 YoY QoQ
Revenue ₩1,917bn ₩1,942bn ₩2,098bn +9% +8%
Platform ₩1,055bn ₩1,153bn ₩1,230bn +17% +7%
├ TalkBiz ₩573bn ₩630bn ₩643bn +12% +2%
└ Platform Other ₩482bn ₩523bn ₩587bn +22% +12%
Content ₩862bn ₩789bn ₩868bn +1% +10%
├ Music ₩518bn ₩485bn ₩558bn +8% +15%
├ Story ₩250bn ₩211bn ₩211bn -16% -0.2%
└ Media ₩94bn ₩92bn ₩99bn +5% +7%
Operating expense ₩1,714bn ₩1,731bn ₩1,821bn +6% +5%
Operating profit ₩204bn ₩211bn ₩277bn +36% +31%
Operating margin 10.6% 10.9% 13.2% +2.6pp +2.3pp
Net profit ₩172bn ₩227bn ₩18bn -90% -92%

Open up TalkBiz and the advertising-and-subscription line grew 14% to 399.9 billion won. Business messaging rose 20%, driven mainly by financial-sector advertisers; KakaoTalk display advertising rose 28% after new feed-format ad products pulled in new advertisers. Commerce was 243.2 billion won, up 10%, with combined gift-and-deal transaction value of 2.7 trillion won, up 9%. Kakao's gifting service recorded its highest-ever May transaction volume.

The Platform Other line growing 22% to 587 billion won deserves attention too. That bucket holds Mobility and Pay. Kakao Pay posted record quarterly results with revenue of 335 billion won (up 41%) and operating profit of 59 billion won (up 528%); Mobility was helped by strength in last-mile logistics. That's evidence against the oldest criticism of this company — that everything depends on one chat app.

Content, on the other hand, grew 1%, which is a rounding error. Music carried the segment with 8% growth on anchor-IP concerts plus merchandise and licensing. Story fell 16%. Piccoma revenue declined 9% in yen terms (though it held the number-one spot for total app-market revenue in Japan), and the entertainment sub-line dropped 22% as user traffic slowed. Roughly half the reason analysts cut their targets lives in this paragraph.

And now the table that matters most. Kakao's IR deck breaks operating profit into three pieces, and one of them is the AI services division.

Operating profit split (₩bn) 2Q25 3Q25 4Q25 1Q26 2Q26
Kakao standalone (ex-AI) 150 146 174 174 184
Subsidiaries 96 123 88 92 151
AI services division -42 -46 -46 -55 -58
Total 204 224 216 211 277

The AI division lost 58 billion won in Q2, widening from a 42 billion won loss a year earlier. Add the five quarters and you get 247 billion won of cumulative losses. Kakao does not disclose AI revenue separately. In other words, AI is still purely a cost line on this income statement, and the cost is being absorbed by growth in TalkBiz advertising and Kakao Pay. Kukinews summed up the call with a headline that fits exactly: Kakao drew the AI blueprint, but TalkBiz earned the money.

So how big is the infrastructure Kakao is declining to build? Q2 capex was 188 billion won — 165 billion in tangible assets, 24 billion in intangibles. That's up 88.5 billion won year over year and equals 9% of revenue, with the IR deck attributing the increase to server purchases concentrated in the quarter. The more decisive line is buried in the operating expense notes. Outsourcing and infrastructure cost came in at 223 billion won, up only 2%, and the company's own explanation reads: AI-related infrastructure demand increased, but efficient investment and operation kept the cost increase limited. A company that says AI is its future grew its infrastructure spend by 2%. That single footnote is more honest financial evidence than any quote from the call.

Q2 2026 Kakao Naver
Revenue ₩2,098.5bn (+9%) ₩3,388.8bn (+16.2%)
Operating profit ₩277bn (+36%) ₩520.3bn (-0.2%)
Operating margin 13.2% (rising) 15.4% (falling)
AI infrastructure stance No data centers, no GPU cloud Sejong AI factory to 200MW, long-term 1GW
Share move on results day about -2% -7.08%

Line the two up and something interesting appears. Naver is still 1.6 times larger on revenue and 1.9 times on operating profit, but the margin trends point in opposite directions: Kakao's is climbing, Naver's is slipping. And both stocks fell. The market is asking the builder when it will earn the money back, and asking the non-builder what exactly it plans to win with. Neither question has an answer yet.

Who actually gets paid in this trade

The clearest winner is Coupang Eats. Korea's delivery app market consolidated into a two-horse race during 2026. As of March 2026, monthly active users ran roughly 22.49 million for Baedal Minjok (Baemin), about 12.49 million for Coupang Eats, and about 3.97 million for Yogiyo, with the top two holding a combined share in the high 80s. Coupang Eats is the perpetual number two. What it now gets is an ordering counter inside the chat window of an app 49.6 million people already have open. An entry path that requires no new app launch is the most expensive asset in delivery competition. It also explains why Kakao partnered with number two instead of number one: the market leader has no reason to walk into someone else's chat window and split a fee.

Kakao takes three things. First, time. Skipping data centers means depreciation and power costs don't balloon, which is why the operating margin can keep climbing from 13.2%. Kakao has targeted a standalone operating margin above 20% by 2028, and that target is inseparable from the decision not to build. Second, a seat at the transaction. When ordering and payment complete inside the chat, Kakao sits where a cut of gross transaction value gets taken. Third, a change in the nature of its ad inventory. Chung said the company will introduce, in the second half of 2026, "new AI-based advertising experiences that go beyond keyword-centric search ads." Korean search advertising has long been Naver's territory; Kakao is attempting a flanking entry by reading intent from conversation context rather than from a query box.

OpenAI quietly benefits too. Thirteen million cumulative ChatGPT for Kakao subscribers is a scale OpenAI would have struggled to reach alone in Korea. Eight minutes of daily engagement isn't much, but existing as a permanent tab inside the country's default messenger is very good for habit formation. As long as Kakao maintains its rent-the-big-model posture, OpenAI is effectively leasing a distribution channel for the Korean market.

Shareholders landed somewhere awkward. Results beat consensus by 22%, and target prices came down an average of 13%. Meritz Securities made the largest cut, from 75,000 won to 52,000 won, a 30.7% reduction. Hana Securities went from 75,000 to 58,000, Yuanta from 80,000 to 62,000. Samsung Securities was the lone raise, from 44,000 to 49,000. Mirae Asset's Lim Hee-seok put the shift plainly: where advertising growth used to set the multiple, platform valuations are now determined by the speed and profitability of AI monetization. Korea Investment & Securities' Jung Ho-yoon argued that AI service monetization requires user behavior to change, which takes considerable time, and that shortening that window is the real task. Hana's Kim So-hye made the sharpest observation — market-beating results have been arriving since last year, yet the stock has tracked expectations for the new AI business rather than the performance of the existing one.

The party most exposed to downside is merchants and small business owners. Fee structures haven't been published, so nothing can be asserted yet. Structurally, though, this adds one more intermediating layer to a delivery transaction. In a country where delivery app commissions are already a live political issue, the question of where the agent's cut comes from will be asked the moment the service actually opens.

KakaoTalk Order in 2017, Facebook M in 2018

This is not the first time Kakao has tried to make people order food inside KakaoTalk. On March 21, 2017, the company launched KakaoTalk Order, starting with 14 franchise brands across chicken, pizza, burgers and Korean food, handling ordering, payment and delivery notifications entirely inside the app with no separate download or signup. It ran the service with CNT Tech, a portfolio company. The concept was, functionally, what Kakao announced this week. The outcome? In October 2023 Kakao gave up operating it and handed the service to the company behind Yogiyo. The sign was changed to "Order by Yogiyo."

So the same idea has come back wearing a third face in nine years. Version one had Kakao running delivery itself. Version two leased the surface to Yogiyo. Version three has an AI agent reading conversation context and passing the order to Coupang Eats. If there's a lesson taken from two failures, it's that Kakao is running neither logistics nor operations this time. It keeps only the intent-reading layer. Owning no assets and collecting a toll is precisely the same logic as refusing to build AI infrastructure.

The closest international failure is Facebook M. Launched inside Messenger in 2015, this assistant promised to carry out real actions — booking appointments, ordering flowers, dealing with customer service. The problem was who was behind it. Complicated requests were handled by contract human trainers, with the AI learning from their responses, and the learning never advanced far enough to replace the humans. Facebook shut M down on January 19, 2018. The lesson is unkind: in-chat agents that connect conversation to action didn't fail from a shortage of ideas, they failed from a shortage of execution accuracy. Kakao's advantage in 2026 is that model capability isn't remotely comparable to 2015. Its disadvantage is that the bar for failure has risen just as much. An agent that orders the wrong meal and completes payment isn't a convenience, it's an incident.

The success case people usually cite is WeChat mini programs. Tencent laid a layer inside WeChat that runs services without opening an app, effectively turning a messenger into an operating system. That is exactly what Kakao wants. But there's a decisive difference. Tencent controlled both payment (WeChat Pay) and mini-program approval, and China's app store ecosystem was weak. Korea is not that. Kakao owns Kakao Pay, but the actual supply chains for delivery, commerce and reservations belong to other companies, and Korean users already hold accounts, saved cards and coupons inside each of those apps. Mini programs won in a market where apps were thin. Kakao's agent has to win in a market where the apps work fine.

There's also a domestic precedent worth remembering: the smart speaker wave. Around 2017, Kakao Mini, Naver Clova and SKT's Nugu all shipped with the same promise that speaking would place an order. All of them ended up playing music and reading the weather. The failure wasn't speech recognition, it was payment trust — people would not tolerate money leaving their account with no screen involved. What distinguishes Kakao's A2A from a speaker is that a chat window has a screen and can insert a confirmation step. Whether that actually changes behavior is the central variable in this experiment.

What Baemin, Naver, and OpenAI can do about it

The fastest mover will be Woowa Brothers, which runs Baemin. It is the market leader and the only large player excluded from this deal. It has two options: belatedly join the same kind of A2A integration to secure the surface, or double down on its own in-app agent and argue that ordering in Baemin is simply more accurate. If it picks the second, Kakao's agent becomes a half-service carrying only the number-two operator, and users will start hitting the experience of asking the chat window for a restaurant that isn't listed. Baemin's MAU being roughly 1.8 times Coupang Eats' makes that scenario quite plausible.

Naver's counterplay is already running, in the opposite direction. Naver started monetizing AI inside search first with products like AI Briefing ads, and said AI-driven ad optimization contributed more than 60% of its advertising revenue growth in Q2. Naver is collecting; Kakao hasn't started. The tradeoff is that Naver accepted a heavy fixed cost in the form of gigawatt-class infrastructure. The real verdict lands in 2027. If Kakao monetizes agents without owning infrastructure, Naver's capex looks like overbuilding. If model quality and large-scale inference capacity turn out to determine service quality, Kakao ends up splitting margin with whoever owns the compute. The weakness of selling only the socket is that the power plant sets the tariff.

OpenAI is simultaneously the partner and the most dangerous competitor. Right now it sits politely inside a KakaoTalk tab, but ChatGPT is adding payment and commerce capability of its own. The moment users start ordering delivery directly in the ChatGPT app rather than in a KakaoTalk chat, the intent-reading layer Kakao worked to build gets routed around. That is the most natural explanation for why Kakao is pushing its own Kanana lightweight models on-device. If conversation context is processed on the handset, the data and the judgment made in that moment stay with Kakao.

Coupang itself faces a more complicated calculation than it looks. Coupang Eats gets more orders inside KakaoTalk, but Coupang's biggest asset — WOW membership lock-in — weakens if users open the Coupang app less often. Agreeing to this deal makes sense for a number-two operator that needs growth more than it needs purity. Whether Coupang keeps the same terms after its delivery share climbs is a separate question.

Finally, telecom carriers and Samsung Electronics compete on exactly the same layer. Carrier assistants like SKT's A. and Galaxy's on-device AI are fighting for the same prize: being the first thing that reads user intent. Whoever owns the operating system and the handset has a structural advantage, and Kakao is an app sitting on top of both. That context is also how to read Kakao being selected on August 6 as an implementing organization for the Ministry of Science and ICT's AI Agent Marketplace development support project. Setting the standard early helps in a layer war.

So what actually changes

For people who use KakaoTalk, the change starts in the second half of this year. Kakao said Kanana in KakaoTalk arrives in H2, with a year-end target of 10 million monthly active users for AI services inside KakaoTalk. Ten million out of 49.6 million is one in five. The experience is meant to work like this: you're texting a friend about what to eat, and without an explicit request, an on-device model reads the context, suggests a menu, and the order and payment finish right there. If it works, it becomes habit. If it misfires, it becomes irritation. And note the precondition — this only functions if an AI reads the conversation. Half the reason Kakao keeps stressing on-device processing is performance; the other half is this. How the privacy disclosure is written is worth checking on launch day.

For advertisers and commerce teams, new inventory is opening. KakaoTalk display advertising is already growing 28%, and AI-based ad products land in H2. Early inventory usually performs well because competition for it is thin, and Kakao's flanking entry into search advertising happens to coincide with that window. That said, conversation-context advertising cannot be measured the same way keyword search ads are. When the product opens, the first question to ask isn't click-through rate — it's how the conversion path is attributed.

For investors, this quarter is a clearly conditional message. The existing business is genuinely improving: operating margin has risen five quarters in a row, results beat consensus by 22%, and management said it expects to exceed its 2026 guidance of 10%-plus revenue growth and a 10% operating margin, describing Q2 as a new baseline rather than a one-off peak. And target prices still came down. What the market is watching isn't the core business, it's when that -58 billion won AI line flips positive. The company's stated schedule is monetization beginning in 2027 through transaction fees, subscriptions and new ad formats, with AI-related revenue reaching a double-digit share of TalkBiz revenue by 2028. Two metrics matter more than revenue: whether the 10 million year-end AI MAU target is hit, and how many partners follow Coupang Eats — and in which verticals.

For developers and people working in AI, there's a different angle. Kakao open-sourcing the Kanana lightweight models while trying to standardize A2A partnerships means Korea is about to get its first commercially validated case of agents transacting with other agents. Put the Ministry of Science and ICT marketplace project alongside it and the shape of that protocol may well determine the interface for every Korean service that wants to attach an agent later.

Compressed into one sentence: Kakao didn't abandon infrastructure in order to win the AI race. It abandoned infrastructure because that's the only way to protect today's margin, and it intends to use that margin to buy time to change the habits of 49.6 million people. If it works, Kakao becomes the most capital-efficient toll collector in Korean tech. If it doesn't, it ends up a thin layer brokering someone else's products on someone else's models running on someone else's compute. The window in which that resolves runs through 2027, and right now the only two hard numbers in hand are a 58 billion won quarterly loss and a 10 million-user target.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? If you use KakaoTalk, starting in the second half of this year you'll meet an AI in your chat window suggesting food orders. The convenience is real, but it only works if the AI reads conversation context. Kakao says that processing happens on-device, so the disclosure language at launch is the thing worth actually reading.

— Is skipping infrastructure the right call? On today's P&L, yes. That's precisely why operating margin climbed to 13.2% while quarterly capex stayed at 188 billion won. The catch is that this strategy means permanently renting models and inference capacity from someone else, and Kakao's negotiating position weakens if those suppliers raise prices or start selling directly to consumers. It's a question that gets graded years from now, so calling it early would be premature.

— When does the AI actually make money? The company's own schedule is monetization starting in 2027 via transaction fees, subscriptions and new advertising formats, reaching a double-digit share of TalkBiz revenue by 2028. But the AI division still lost 58 billion won this quarter, and seven brokerages cut their average target price by 13% after the release. The market is waiting for the first commission to land, not the blueprint.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!