Upbit And Bithumb List New AI Tokens In Korea

I refreshed the won books twice before the clocks hit mid-afternoon in Seoul, mostly because same-day listings from the two biggest local venues rarely land this cleanly. One exchange opened a Korean-won market for an AI-agent token. The other, an hour later, opened three pairs for a token tied to distributed GPU inference. Same headline theme. Very different plumbing. If you only read the ticker tape, both names look like “AI coins that just got a Korea bid.” Sit with the notices for ten minutes and that shortcut starts to fray.

Two Korean Venues, Two Very Different AI Tokens

On October 2, 2026, Bithumb added Talus, ticker US, against the Korean won, with trading scheduled for 3 p.m. Korea Standard Time. The reference price printed at KRW 37.85 before the book went live. Deposits and withdrawals were limited to the Sui network, and the exchange said those rails would open within about two hours of its early-afternoon notice, after the usual twenty network confirmations.

Upbit took a wider cut. Dolphin, ticker POD, was set to trade from 4 p.m. across Korean won, Bitcoin, and Tether. Pre-listing reference marks, taken around 1:20 p.m. local time, sat near KRW 428, 0.00000367 BTC, and 0.3147 USDT. Deposits and withdrawals were expected on a similar two-hour clock, but only through Base, and only at the contract address the venue published. Wrong chain, and the coins do not arrive. That part is not a footnote. It is the whole transfer rule.

I’ve found that traders treat a dual listing day as a single event. It is not. One book is a direct won market with a narrow deposit path. The other is a three-pair launch with a different chain and a token that had already been trading offshore for months. The overlap is the theme, not the product.

What Actually Opened, And When

Timing matters more than the press line. Bithumb’s notice landed around 12:14 p.m. local time. Trading for US/KRW was penciled in for 3 p.m. Upbit’s Dolphin markets were penciled in for 4 p.m. That one-hour gap is long enough for the first book to set a mood, and short enough that both names can still trade inside the same Seoul afternoon session.

Both venues also kept a familiar escape hatch. If liquidity is not there, the listing can be pushed. That sentence gets skipped when prices are green. It is the sentence that matters if the opening book is thin.

A listing is a door, not a promise. The door can open late, open narrow, or open into a room that empties by the close.

Perhaps the most interesting aspect is how ordinary the controls looked. New coins in this market do not get a free-for-all open. They get a short leash.

Opening Rules That Quietly Shape The First Print

Bithumb blocked buy orders for the first five minutes. Sell orders in that window could not sit below 90 percent or above 200 percent of the KRW 37.85 reference. For roughly the first two hours, only limit orders were allowed. Upbit used a similar cage for POD: purchases restricted for about five minutes, an early cap on low-priced sells, and limit-only trading for about two hours.

Those rules do not kill volatility. They change its shape. You do not get a market-order stampede in minute one. You get a queue of limits, a reference price acting like a fence, and a five-minute window where sellers can act and buyers cannot. I have watched that pattern turn a “listing pump” into a choppy open more than once. Sometimes the fence holds. Sometimes the first real print still gaps once the buy side is unlocked.

• Buys paused for about five minutes on both names
• Early sells boxed around the reference price
• Limit orders only for roughly the first two hours
• Deposits tied to one network each: Sui for US, Base for POD
• Either listing can slip if the book is too thin

Read that list again if you plan to move coins in. The network rule is the one that burns people. Talus deposits ride Sui. Dolphin deposits ride Base. A token sent on the wrong rail is not a delayed deposit. It is often a lost one.

Talus, The Agent Network On Sui

Talus is pitched as an onchain platform for AI agents. The idea is simple to say and fiddly to build. Agent identity, permissions, and actions get recorded where they can be checked. Heavy computation, the part that does not fit inside a smart contract, runs off the chain. The project’s Nexus framework is meant to stitch those pieces into workflows, so a task can hop from a verifiable step to an offchain service and back.

The first version of the protocol reached Sui mainnet in April. Protocol v2.0 followed in August, aimed at the infrastructure agents use when they execute transactions and leave a trail that can be inspected later. In September the team released an API so developers could pull structured data from those agent runs. That sequence matters. It is not a white paper with a countdown. It is a stack that has already been switched on, then revised, then given a developer door.

US is described as the network token for service pricing, staking, and governance-related participation. Built on the Sui Move stack, the project pairs its agent framework with the Nexus protocol. In plain language, the token is supposed to sit where agents pay for work, where operators lock value, and where decisions about the system get a vote. Whether that loop is busy yet is a separate question from whether the listing exists.

Bithumb’s won pair gives local users a direct route. No need to hop through a stablecoin or Bitcoin just to hold US. That is the practical gift of a KRW market. It is also why these listings punch above their global volume. A buyer who lives in won does not have to solve a second currency problem before taking a position.

Dolphin, The Inference Pool On Base

Dolphin sits closer to machines than to agents. The project is framed as an AI and decentralized physical infrastructure effort that spreads inference across unused GPUs. Nodes running the same model are grouped into pools. A user request gets assigned to whatever compute is free. Owners of graphics cards can offer capacity. The network, in theory, becomes a distributed alternative to sending every job to a centralized cloud.

POD is a Base token. Total supply has been cited at 500 million. The token is described as the unit for staking and bonding inside the network. Earlier work includes a Dolphin Mistral 24B Venice Edition built with Venice, which is a concrete model release rather than a slogan. Offshore venues had already listed the coin. A major international exchange opened a POD/USDT market on May 12, months before the Korean books.

That history changes how you read the Upbit print. POD was not born on October 2. It already had a price, a range, and a set of holders. The Korean listing can still move it. It cannot be the whole story of the move.

Upbit’s three pairs also change the texture of the open. A won pair pulls local retail. A Bitcoin pair pulls traders who already sit in BTC. A Tether pair pulls anyone who thinks in dollars. Three doors, one token, one chain for deposits. Convenient for trading. Strict for transfers.

Same Theme, Different Job

Both tokens get filed under AI infrastructure. That filing is fair and also lazy. Talus is about agents that can act, and about leaving a record those actions can be checked against. Dolphin is about compute. One asks whether a software agent did what it was allowed to do. The other asks whether a spare GPU can run the model without a hyperscale bill.

I keep coming back to that split because the market keeps erasing it. “AI token” is a label that travels well on social feeds and poorly in a portfolio. An agent network can fail if developers never route real tasks through it. A GPU pool can fail if the hardware never shows up, or if the jobs are cheaper on a normal cloud. Those are different failure modes. They deserve different questions.

The table is the clean version. Live markets are messier. Reference prices are not promises. They are fences for the open.

Did The Listings Move The Price?

Both names had loud sessions. Market trackers showed Talus near $0.0376, up roughly 40 percent over 24 hours, with a daily range from about $0.0182 to $0.0383. Reported daily volume sat around $20.9 million. Dolphin traded near $0.6765, up roughly 115 percent, with a range from about $0.3046 to $0.7234.

Those numbers are real enough to respect and too tangled to worship. POD’s move started near the level Upbit had cited before its own open, and other venues were already trading it. US had a sharper local catalyst in the sense that a direct won book is new demand, but global books were not empty either. A clean “listing effect” cannot be isolated from the rest of the tape. Anyone who claims otherwise is selling a story, not a measurement.

Still, the pattern is familiar. Smaller tokens have jumped when these two venues open Korean markets. One meme-style name rose nearly 10 percent when both exchanges listed it. A spring listing of another project on both venues came with a 48 percent rally in the related token. Past jumps are not a formula. They are a reminder that local access can reprice a thin book in a single session.

Price can confirm attention. It cannot confirm that the network is being used.

A habit worth keeping on listing day

In my experience, the first green candle answers the wrong question. The better question shows up a week later. Is volume still local, or did it leak back offshore? Did deposits actually arrive on the supported chain? Did the book stay wide enough that a modest sell does not gap the price?

Why A Won Market Still Punches Hard

South Korea’s retail crypto culture is intense, and also smaller than the folklore. The five main won exchanges recorded a 54.6 percent year-over-year drop in first-half trading volume. Profits at local venues have been under pressure. Upbit still holds most of the domestic liquidity. That mix, less volume overall and one venue still dominant, is the backdrop for today’s listings.

A weaker year does not mean a listing cannot rip. It means the rip is more dependent on a narrow set of buyers. When the local book is the marginal bid, a few busy hours can look like a regime change. Then the session ends, overseas books reassert themselves, and the chart gives some of it back. I would rather assume that path than assume a new floor.

There is another wrinkle. Listing two different AI names on the same day spreads attention. In August, one venue listed one token while the other added two more, and local traders suddenly had several fresh won markets at once. Today rhymes with that. Attention is finite. Two stories compete for the same afternoon.

How To Read An AI Token Without The Costume

Strip the costume and you are left with a few plain tests. Does the token pay for something people already want? Can that something be faked by a cheaper service? Is the chain you must use the chain you can actually withdraw on? Are the early trading rules going to distort the first hour?

For Talus, the tests sit around agents. Are workflows being built? Are actions leaving records anyone checks? Is US used for pricing and staking, or mostly for listing-day speculation? The August upgrade and the September API are evidence of a team shipping. They are not evidence of demand.

For Dolphin, the tests sit around hardware and jobs. Are GPU owners bonding capacity? Are inference requests landing in the pools? Is a 500 million supply being absorbed by use, or by a new exchange logo? A model release is a proof of taste. A busy pool is a proof of need.

1. Confirm the deposit network before you move a single coin
2. Treat the reference price as a fence, not a target
3. Separate agent infrastructure from GPU infrastructure
4. Watch whether volume stays after the first Seoul session
5. Assume a postponement clause can still be used

None of that is glamorous. Glamour is how thin books get expensive.

The First Hours, In Practice

Picture the open as a room with the lights coming on in stages. For five minutes, sellers can post inside a band and buyers cannot. Limits only. Then the buy side unlocks, still inside a limit-only window that lasts about two hours. Deposits may not even be live yet, depending on when the notice went out and how fast confirmations land.

That sequence creates a weird market. Early price can be set by people who already hold the coin elsewhere, not by fresh deposits. If offshore sellers are faster than local buyers, the first won prints can lean heavy. If local buyers are faster, the band gets tested from above the moment buys unlock. Both outcomes have happened on other names. Today does not get a special exemption because the theme is AI.

A small practical note I wish more notices shouted: the contract address for POD on Base is specific. Copy it from the venue, not from a reply under a social post. US only moves on Sui for this listing. Shortcuts here are how people donate coins to nowhere.

Opening checklist:
  Network match before transfer
  Limit orders only in the first window
  Reference band is a rule, not a view
  Liquidity clause can delay the open
  Offshore price is already in the chart

What The Wider AI Trade Is Actually Pricing

Public markets have spent two years paying up for anything that sounds like compute or agents. Crypto does the same thing with less earnings and more narrative. That is not an insult. It is the asset class. The risk is category error. A token that meters agent permissions is not a token that rents a GPU, even if both slides say “AI.”

Talus is betting that verifiable action becomes a feature people pay for. If agents start moving money, signing steps, and calling tools, someone will want a record that is hard to rewrite. Putting that record on a chain, and pushing the heavy math off it, is a reasonable architecture. It only becomes a business if builders route real workflows through Nexus instead of a private server.

Dolphin is betting that idle GPUs are a market, not a hobby. Inference is the boring, repeated part of AI. Training gets the headlines. Serving the model, over and over, gets the bill. If a pool can match a request to a card that is already on, and do it reliably, the pitch is economic. If latency, uptime, or model consistency slip, users go back to the cloud they already trust.

I do not think either bet is silly. I do think listing day is a terrible moment to decide which bet is working. The tape is loud. The usage data is quiet. Those two clocks rarely agree.

Liquidity, Postponement, And The Quiet Risk

Both exchanges reserved the right to delay if liquidity is not sufficient. That clause is easy to ignore when a social post is already circulating. It exists because a won book with no depth is a hazard, not a product. A postponement would not mean the project failed. It would mean the venue refused to open a hollow market.

There is a second quiet risk, and it is operational. Supported networks are not suggestions. Sui for US. Base for POD. A trader who bridges to the wrong environment, or who withdraws to an exchange address on an unsupported chain, is outside the listing. Customer support cannot invent a rail the notice did not offer.

A third risk is narrative compression. Once both tickers trend, commentary collapses them into one trade. Then one name fades and the other holds, and people act surprised. They should not be. Different chains, different utilities, different prior trading histories, different pair counts. Divergence is the base case, not the twist.

A Slower Year, A Fast Afternoon

The domestic backdrop is not euphoric. First-half volume on the main won venues was down more than half from a year earlier. Exchange profits have been squeezed. Upbit still concentrates local liquidity, which means a listing there can matter even in a softer year, and a listing only on the second venue can matter less than the headline implies. Today splits that map. Talus gets Bithumb’s won book. Dolphin gets Upbit’s three pairs. Neither gets both, at least not in the notices that opened the day.

That split is easy to miss if you only track “Korea listed AI.” It is the detail I would keep. Access is not uniform. A buyer who uses one app may see one token and not the other. Liquidity will not be shared just because the theme is shared.

Softer volumes can also make percentage moves look heroic. A 40 percent day on roughly $21 million of turnover is a real session for a small name. It is not evidence of a deep two-way market. A 115 percent day on a token that already traded offshore can be a mix of catch-up, local urgency, and short covering. Untangling that mix takes more than one close.

What I Would Watch After The Open

First, whether deposits actually opened on schedule. A market that trades before transfers settle is a market of existing holders. Fresh supply arrives later, if it arrives.

Second, the spread after the limit-only window ends. If the book stays wide, the listing is a logo more than a market. If it tightens, someone is willing to warehouse risk.

Third, whether POD’s three pairs agree with each other. A won premium that does not show up in the Tether pair is a local squeeze. A move that shows up in all three is harder to dismiss as one currency’s mood.

Fourth, US against its own reference. KRW 37.85 was a fence, not a forecast. A settle well above it, with volume that does not vanish at the Seoul close, is more interesting than a spike that round-trips before dinner.

Fifth, the thing almost nobody charts on day one: whether either network’s stated use shows a pulse. Staking, bonding, agent calls, inference jobs. Without that, you are trading an access event. Access events fade.

A Fair Way To Hold Both Ideas At Once

You can believe AI infrastructure is a real spend category and still be picky about tokens. You can like the idea of checked agent actions and still want to see workflows. You can like idle-GPU markets and still want uptime numbers. Those positions are not in conflict. The conflict is only with the feed, which prefers one adjective and a green candle.

Talus has a clearer “software did a thing, here is the record” story, anchored on Sui, with a token aimed at pricing, staking, and governance participation. Dolphin has a clearer “hardware did a job, here is the pool” story, anchored on Base, with a token aimed at staking and bonding, and a supply figure of 500 million already in circulation narratives offshore. Korea just gave each story a local microphone. Microphones do not write the script.

Questions Worth Asking Before The Next Session

Who is the marginal buyer after the first afternoon? If it is only listing-day curiosity, the chart has a habit of mean-reverting toward the offshore price. If it is won savings looking for a new AI sleeve, the local premium can linger. You will not know which one you have until the second and third sessions.

Is the product even comparable to what people think they bought? An agent framework is not an inference marketplace. Mixing them in one basket is how portfolios get accidental bets. I would rather own a view on agents, or a view on distributed compute, than a view on a label.

And the dull one, which keeps being the important one: can you move the coin the way the venue requires? Sui for US. Base for POD, at the published contract. Everything else is noise until that part is right.

Simple filter: network match + real use + volume after day one. Miss two, and the listing is mostly a headline.

Korea’s two largest venues just widened local access to a pair of AI-linked projects on the same day. Bithumb gave Talus a direct won market and a Sui-only rail. Upbit gave Dolphin won, Bitcoin, and Tether pairs, with deposits stuck to Base. Prices moved hard. Rules around the open were tight. The year underneath those moves has been a softer one for domestic turnover.

That is the whole shape of the day. Not a single AI trade. Two doors, two chains, two jobs, one crowded afternoon. The part still open is whether either door stays busy once the notices stop circulating.

A Longer Look At The Agent Bet

Agent tokens have a storytelling advantage. People can picture a piece of software booking, paying, and reporting back. The harder picture is the audit trail. Talus leans on that harder picture. Identity, permissions, and actions are supposed to land where they can be checked, while the bulky computation stays offchain. Nexus is the organizing layer, the place workflows get assembled so a task is not a loose script on someone’s laptop.

Protocol v2.0 on Sui mainnet, deployed in August after an April launch, is the project’s way of saying the rails for those agents are no longer a demo. A September API for structured execution data is the way of saying outsiders can read what the agents did. I like that sequence more than a token launch with no software behind it. Liking the sequence is not the same as underwriting the token.

Service pricing is the economic hinge. If agents pay US to call a workflow, the token has a reason to move that is not a listing. Staking is the second hinge, a way to lock supply against some duty inside the network. Governance participation is the third, and the easiest to overrate. Votes do not create demand. Usage does. A won market can pull the price toward local attention for a session. It cannot invent a workflow.

There is also a chain bet tucked inside the token bet. Sui, and the Move stack, is the home turf. Anyone who wanted US exposure before today could already think in that ecosystem. Bithumb’s contribution is currency access and a supervised open, not a new chain. Keep those layers separate when the chart gets loud.

A Longer Look At The Compute Bet

Inference is unglamorous on purpose. A model that cannot be served is a paperweight. Dolphin’s design groups nodes that run the same model, then assigns requests to free capacity. GPU owners supply the machines. Users supply the jobs. POD is meant to sit in the middle as the staking and bonding asset, the thing that ties an operator to the pool.

The 24B Venice Edition is a useful marker because it is specific. A named model, built with a named partner, is harder to fake than a roadmap slide. It still does not tell you whether the pool is cheaper or more reliable than renting a normal cloud GPU for an afternoon. That comparison is the whole business. Crypto distribution does not win by being distributed. It wins if the job gets done at a price and a latency someone will accept twice.

Supply sits at 500 million in the figures venues have used. A fixed supply is not tightness. Tightness is how much of that supply is bonded, staked, or otherwise unwilling to meet a market sell. Upbit’s listing increases the number of people who can meet that sell with won, bitcoin, or tether. More doors can mean more demand. They can also mean a faster exit. Both readings are live until the book shows its hand.

Offshore trading since at least mid-May is the other anchor. A May 12 USDT pair on a large international venue means POD already had a discovery process. Korean reference marks near KRW 428, a sliver of bitcoin, and about 0.31 tether were snapshots, not origins. If the local open diverges from those snapshots and then snaps back, you watched a venue premium. If it diverges and stays, you watched a new marginal buyer. Day one rarely finishes that argument.

How Listing Premiums Usually Age

I have seen three rough ages for this kind of premium. The first is the fenced hour, when rules do more work than opinions. The second is the evening handover, when Seoul liquidity thins and other time zones decide whether they care. The third is the week after, when the notice is no longer new and the token has to trade on whatever use, or lack of use, it had before the logo appeared.

Names that jumped on earlier Korean dual opens did not all keep the jump. A near-10 percent pop is a different animal from a 48 percent pop, and neither is a destiny. Today’s 40 percent and 115 percent sessions belong in that family of abrupt moves. They do not belong in a model that assumes the close is a floor.

Perhaps the healthiest habit is to write down the reference and the offshore price before you look at the live book. KRW 37.85 for US. The trio of POD references from early afternoon. If you cannot remember where the fence was, every tick feels like information. A lot of ticks are just the fence doing its job.

Retail Access Versus Project Quality

Local listings change who can buy. They do not change what was built. That sentence sounds obvious until a feed treats the listing as due diligence. It is distribution. Distribution can be valuable. A direct won pair lowers friction for a large retail base. Three pairs lower it further. Friction is not quality.

Quality, for these two, lives in unfashionable places. For Talus, in whether agent runs produce records developers actually query. For Dolphin, in whether pools fill with cards and requests. A trader can ignore both and still trade the access event. That trader should label the position honestly. It is a bet on attention and market structure, not a bet that idle GPUs or checked agents have won.

I would rather see that honesty in more write-ups. The afternoon was genuinely notable. Two major Korean venues, two AI-linked assets, staggered opens, strict transfer rules, large intraday ranges. Notable is enough. It does not need a victory lap.

What A Cautious Buyer Might Actually Do

Nothing here is advice in the personal sense. It is a posture. Wait until deposits are confirmed open. Match the network. Prefer limit orders even after the venue stops requiring them, because spreads on new books lie. Compare the won price with the tether or bitcoin price before calling a move “the” price. Size for the chance that liquidity was the condition, not the guarantee.

A cautious reader might also split research time unevenly. An hour on the notices. An hour on what each token is for. Almost no time on posts that treat US and POD as the same coin in different clothes. They are not. One is an agent and verification story on Sui. The other is a compute and bonding story on Base. Korea listed both. Korea did not merge them.

If the books stay orderly into the next session, the listings will have done the modest thing venues are for. They will have turned a foreign ticker into something a won account can reach, under rules that slow the first minutes down. If the books gap and fade, the listings will have done the other familiar thing. They will have dated a chart. Either result can happen without telling you whether the underlying networks are any good. That verdict takes longer than an afternoon in Seoul.

So the day closes on a narrower claim than the headlines want. Access expanded. Rules were tight. Prices were not calm. The products remain different. Anyone trading the label instead of the rail is trading a costume. Costumes photograph well. They do not settle.