The cloud gaming latency floor: where physics ends the argument

Every cloud gaming press release for the last four years has included some version of the same claim: latency is improving, the gap with local hardware is closing, soon it'll be indistinguishable. Marketing teams point at codec improvements (AV1), encoder optimisations (NVENC generations), and data centre expansion as the three levers.

All three are real and all three matter. But there's a fourth lever the marketing never mentions: the speed of light. And it sets a hard floor below which no amount of engineering can take you.

Light in a vacuum travels 300,000 km/s. Light in fibre optics travels about 200,000 km/s — slower because the photons bounce off the cable walls. That's 5 microseconds per kilometre.

A round trip from London to a data centre in Frankfurt (~700 km of cable, not as the crow flies) is about 7 ms just in fibre transit. Add the router hops, the queueing, the last-mile copper or coax, and you're at 10–15 ms before the cloud server has done a single thing.

Our methodology lab measures GeForce Now Ultimate at ~23 ms of overhead on top of the network round-trip. Of that 23 ms, roughly 6 ms is the server-side render at high frame rates, 7 ms is the encoder, 2 ms is the transit between the cloud GPU and its NIC, 5 ms is the client decoder, and 3 ms is display pipeline (depending on the screen).

Each of these has a theoretical lower bound. The render time can't go below 1/framerate (8.3 ms at 120 Hz). Hardware encoders have a per-frame minimum measured in low single-digit ms. Display pipelines have similar floors.

Add the speed-of-light minimum for a typical residential-to-DC route (~10 ms) and the lab-floor numbers (~16 ms even on perfect hardware), and the absolute physical minimum for end-to-end cloud gaming latency in the most favourable geography is around 26 ms.

A gaming PC with a 240 Hz display has its own latency budget — the game's render loop, the input device's poll rate, the display's pixel response. Modern competitive setups can hit ~6 ms total click-to-photon latency.

So the gap between the best possible cloud gaming setup and the best possible local setup, in 2026, is approximately 20 ms. Twenty ms is two frames at 100 Hz. It's measurable, it's perceptible to competitive players, and it can't be engineered away.

The places where cloud will keep closing the gap with local: at 60 Hz and below (where local's per-frame budget is bigger than the cloud overhead), in regions where the user previously had bad local hardware (most of the developing world), and on devices that can't run local rendering at all (phones, low-end laptops).

These are also the places where cloud gaming's growth will continue to be strong, and where most non-enthusiast players live anyway.

Competitive shooters at the top of the skill ladder. Sim racing on direct-drive wheels. Rhythm games at high charts. Any genre where ~15–20 ms of input lag is the difference between a viable session and a frustrating one.

These are also the places where the marketing keeps insisting cloud is closing the gap. It isn't. The physics doesn't allow it. The honest answer is: cloud gaming will be excellent for 95% of the gaming population in 2030, and the remaining 5% — the competitive top end — will still be on local hardware for the foreseeable future.

If you're choosing between cloud and a new gaming PC for casual play, MMO grinding, single-player RPGs, story games, co-op shooters, or anything where 30 ms vs 50 ms is invisible to you: cloud is the better economic choice and likely the better experiential one too.

If you're choosing between cloud and a new gaming PC because you want to climb a ranked ladder in CS or Valorant or fighting games: get the PC. The 20 ms gap is not closing and your competition assumes you're not paying it.

If you're already at the latency floor in your current setup, the only direction left is to get physically closer to a data centre — which is, increasingly, a real factor in where people choose to live. We're not joking. We've seen it.