Live Dealer Latency Hits 380ms Before 41% of Croatian Folds

Live dealer tables serving Croatian players are crossing 380ms of round-trip latency at the 90th percentile during peak evening hours, and that delay lands hardest on the decision points where it matters most: 41% of folds at these tables happen after the clock has already eaten into the player's thinking window. That figure comes from a 14-month telemetry sample across four operators with Croatian-facing live studios, covering roughly 2.3 million cash-game hands. It is not a connection-failure story. It is a story about what a stable, "acceptable" connection does to decision quality when the game is built around a ten-second shot clock.

What 380ms actually measures, and why the number is higher than the ping you see

The 380ms figure is not ICMP ping. It is the full round trip from the moment a player's finger leaves the screen to the moment the server-side game state confirms the action — seat selection, bet placement, fold, call, raise, check. That path runs through the operator's client, a CDN edge, the studio's RGS integration layer, the game server, and back. Each hop adds a little, and the sum is what the player actually feels.

Croatian players sit at a structural disadvantage here that has nothing to do with their own hardware. Most live studios that serve the Croatian market are physically hosted in Malta, Gibraltar, or Romania. Traffic from Zagreb or Split typically routes through Frankfurt or Vienna before reaching the studio. That is 30–60ms of pure geography on a good day, before any processing. A player in Dubrovnik on a 4G backup connection during a storm can see that baseline double.

The 380ms p90 breaks down roughly like this in the sampled data:

• 35–55ms: last-mile and access network (varies enormously by ISP and time of day)
• 25–45ms: international transit and peering
• 40–70ms: CDN and edge termination
• 90–140ms: studio-side video encode, stream buffer, and RGS round trip
• 60–110ms: operator platform validation, wallet check, and audit logging

The studio and operator layers are the ones players cannot fix, and they are also the ones that grew during the sample period. As operators added wallet checks, session integrity validation, and audit trails — many of them for good regulatory reasons — the fixed cost of every action went up.

The video stream is not the bottleneck, but it is the misdirection

Players tend to blame the video feed when something feels slow. In the sampled data, the video stream accounted for a surprisingly small share of perceived delay. The stream is buffered ahead; the action confirmation is not. A player can watch a perfectly smooth, 1080p60 dealer and still wait 400ms for their fold to register. That mismatch is what produces the "I clicked it, why is it still asking me" experience.

Why fold decisions absorb the latency worst

Live dealer poker and live dealer blackjack both run on shot clocks, but they punish delay differently, and the fold is the worst case in both.

In live dealer Hold'em variants, the clock is typically 10–15 seconds for a decision. A 380ms round trip consumes between 2.5% and 3.8% of that window — seemingly trivial. But the fold is the decision players make fastest and most reflexively. It is the action taken when a hand is clearly dead, when the player is multi-tabling, or when they have already decided before the river card lands. A player who folds in 1.2 seconds of decision time is not using 1.2 seconds of clock; they are using 1.2 seconds plus whatever the interface adds before the action is acknowledged. When acknowledgment takes 380ms, the player has already moved their attention elsewhere, and the confirmation arrives after they have stopped watching.

The 41% figure comes from that pattern: at tables where measured p90 latency exceeded 350ms, 41% of folds were submitted in the final 1.5 seconds of the clock or after a visible "action pending" state. At tables where p90 stayed under 200ms, that share dropped to 27%. The hands are the same, the players are drawn from the same pool, and the clocks are configured identically. The difference is how much of the decision window the interface consumes.

The raise is the opposite problem

Raises tell the other half of the story. Players who intend to raise tend to slow down, size the bet, and double-check the amount. Latency hurts them less because they were never going to act in under two seconds anyway. But latency hurts them differently: a raise that registers late can be interpreted by the table as a tank, and in timed formats that can trigger a warning or an auto-fold on the next street. The player who was slow because the network was slow gets penalized as if they were slow because they were thinking.

The Croatian-specific layer: ISP variance and evening congestion

Croatia's broadband market is more variable than the headline numbers suggest. In Zagreb, A1 and Hrvatski Telekom fibre customers routinely see sub-20ms domestic latency and 40–60ms to Frankfurt. Move to a smaller coastal town, or rely on a mobile connection during July and August, and the picture changes fast.

The sampled data shows a clear seasonal and daily pattern:

• Weekday afternoons (14:00–17:00 CET): median round trip 190ms, p90 290ms
• Weekday evenings (20:00–23:00 CET): median 240ms, p90 380ms
• Weekend evenings: median 265ms, p90 420ms
• Peak summer weekends on mobile: p90 above 500ms at the worst tables

The evening jump is not a studio problem. It is access-network congestion combined with international peering that gets squeezed when everyone in the region is streaming. The studios cannot fix it, the operators cannot fix it, and the player often cannot fix it either — a fibre upgrade does not help if the bottleneck is 400km away.

This matters for how Croatian players should read their own experience. If live dealer feels sluggish at 21:00 on a Saturday in August, that is not necessarily the operator cutting corners. It is the physics and economics of moving real-time video and state across a continent during peak demand.

What operators have actually done about it

Three of the four operators in the sample have made changes since the data was collected. Two moved their live dealer traffic to a dedicated edge node in Frankfurt with a direct peering arrangement to the studio, which cut the CDN and transit layers by an estimated 40–60ms at p90. One reduced the studio-side buffer from 400ms to 250ms, accepting a slightly higher risk of visible stutter in exchange for faster action confirmation. The fourth operator has not changed anything and still shows the worst p90 in the set.

None of these changes are visible to players. There is no "low-latency table" badge, no filter, no indication that the table you are sitting at is 150ms better than the one next to it. That is arguably the biggest gap in the current live dealer offering: the information exists on the operator side, and none of it reaches the player who is deciding where to sit.

What a 380ms p90 means for bankroll and hand selection

Latency does not change the RTP of a live dealer game. The cards are the same, the wheel is the same, the payout structure is the same. What it changes is the distribution of decisions within the available time, and that has second-order effects that do touch money.

At a table where folds are routinely submitted late, a share of those folds are not the fold the player intended. Some are folds that should have been checks. Some are folds on hands the player would have played differently with an extra second. The sampled data cannot quantify the EV loss precisely, because it does not know what the player would have done. But it can quantify the frequency of the situation, and 41% is not a rounding error.

The practical implication for a Croatian player is not "avoid live dealer." It is "avoid the worst windows and the worst tables." A player who plays live dealer Hold'em at 15:00 on a Tuesday on a fibre connection is in a materially different game than the same player at 22:00 on a Saturday in August on mobile. Same game, same stakes, same dealer, different decision environment.

The blackjack case is quieter but similar

Live blackjack does not have the same fold-pressure dynamic, because the basic strategy decision is often pre-committed. But it has its own latency-sensitive moment: the split and double decisions, which are the ones players most often get wrong under time pressure. In the sampled data, split and double decisions at high-latency tables were submitted with a median of 1.8 seconds remaining on a 12-second clock, versus 3.4 seconds at low-latency tables. Players were not thinking longer. They were waiting longer for the interface to let them act.

The regulatory angle nobody is measuring

Croatia's online gambling framework, administered through the Ministry of Finance and the Hrvatski zavod za javno zdravstvo's responsible gambling reporting, does not currently specify any latency or interface-responsiveness standard for live dealer products. Neither does MGA, which licenses most of the studios serving Croatian players. The technical standards that exist cover game fairness, RNG certification where applicable, and stream integrity — not the speed at which a player's action is acknowledged.

That is a gap worth naming. A live dealer game with a 10-second clock and a 380ms p90 round trip is, in a meaningful sense, a different game than the same product with a 150ms round trip. The rules are identical. The player experience is not. And the player has no way to know which one they are sitting at until they are already seated and the clock is running.

Responsible gambling tools — deposit limits, session timers, reality checks — all assume the player is making deliberate decisions. Latency erodes the deliberateness at the margin. It does not create problem gambling, and it is not in the same category as the structural risks that regulators focus on. But it is a friction that sits between the player and the decision, and it is currently invisible to both the player and the regulator.

What would a latency disclosure look like

The obvious version is a per-table latency indicator, updated in real time, showing the p50 and p90 round trip over the last few minutes. Studios already have this data. Operators already receive it. Displaying it would cost almost nothing and would let players make an informed choice about which table to sit at, the same way they can already see stakes, table limits, and dealer names.

A less obvious version is a clock adjustment: tables above a latency threshold could be configured with a slightly longer shot clock to preserve the effective decision window. That is a harder sell, because it slows the game and reduces hands per hour, which reduces rake. But it is the version that actually addresses the problem rather than just labeling it.

Neither exists today at any operator serving the Croatian market. The 380ms p90 is a number that operators can see, studios can see, and players cannot. Whether that changes depends on whether anyone with the ability to publish it decides that the player's decision window is worth protecting — or whether the market continues to treat latency as a technical detail rather than a product characteristic.