Cloud gaming works — but it's not winning

David Pierce reports that SpaceX is considering buying Boost Mobile as part of an effort to become a major fourth wireless carrier. The plan is complicated because Boost currently relies on T-Mobile and AT&T networks, potentially requiring cooperation from future competitors; SpaceX is also buying spectrum and reportedly discussing access to 80,000 U.S. cell sites. Anthropic’s updated IPO filing says it earned $33.24 billion through September 30, up 1,206% year over year, including $17 billion in one quarter, while losing $23.29 billion and committing nearly $900 billion to compute and infrastructure. European grocery stores are also reporting hacked freezers, including one raised to 122°F.

Pierce says he has recently had strong experiences playing 007 First Light on GeForce Now and Star Wars Galactic Racer on Amazon Luna, despite generally being a casual gamer. Sean Hollister explains that cloud gaming compresses and streams game frames like Netflix or YouTube, but also sends the player’s button presses back to a remote server. The key measurement is end-to-end latency: the time between pressing a button and seeing the action on-screen. Hollister has even beaten Sekiro’s Genichiro boss through cloud gaming, but says a sudden bandwidth or connection problem can ruin an otherwise playable session.

Hollister says cloud gaming has worked since the Gaikai and OnLive era around 2010 when users had a robust connection and were close to a server. Dedicated cloud servers could theoretically give competitive players a more even footing than peer-to-peer multiplayer, where one player may have a host advantage. Single-player games, FIFA, and other titles that do not demand perfect real-time reactions have been good enough for years, especially on GeForce Now’s expensive Ultimate tier. The larger obstacle is friction: users must consider their connection, data-center distance, subscription tier, game availability, and whether a license might disappear later.

The hosts contrast cloud gaming with locally installed games, which require hardware, downloads, patches, discs in earlier console eras, and sometimes separate logins. Gaikai once let users click a browser button and play Mass Effect, while Google Stadia imagined letting someone click an advertisement and receive an hour-long trial immediately. Hollister says online games such as Fortnite are not already fully streamed: the local device still renders the graphics while the cloud tracks other players and game state. Microsoft Flight Simulator comes closer to a hybrid model by streaming enormous world environments that cannot conveniently fit on a local drive.

Pierce suggests faster internet should gradually solve cloud gaming, but Hollister says raw bandwidth is usually sufficient: roughly 25–35 Mbps can support a 4K stream, and about 5 Mbps can support basic 1080p. The real problem is reliability. Every wireless or wired hop between the player and the game server creates another chance for a hiccup, including traffic inside the home. Netflix can buffer and cache predictable video content near users, but a game cannot know in advance what the player will do. Cellular connections are especially inconsistent; Hollister found cloud gaming worked under a nearby 5G tower but degraded at home.

Hollister says a cloud provider needs roughly the same GPU power to run a game remotely that a player would need locally, so cloud gaming is expensive infrastructure rather than cheap video delivery. A powerful GPU might be divided between two users running less demanding games, but current GPUs are costly because AI data centers are consuming them. NVIDIA can sustain GeForce Now more easily because it makes GPUs and can absorb experiments; Google effectively abandoned Stadia, while Amazon has repeatedly changed Luna after a poor launch. Hollister speculates that aging AI servers could eventually be repurposed to run the enormous long tail of older games at lower subscription prices.

Pierce argues Microsoft should be able to force game streaming into the mainstream because it owns Azure, Xbox, major publishers, distribution, and Activision Blizzard King. Hollister says Azure was not automatically the right kind of gaming compute, and Microsoft’s acquisition created legal commitments, including bringing Call of Duty to more platforms. Apple and Google also limit Microsoft’s ability to turn phones into Xboxes: Apple blocks an in-store Xbox game store from bypassing its fees, while Google only recently became subject to rules allowing alternative stores after Epic’s lawsuit. Microsoft additionally still has reasons to sell physical Xbox hardware and exclusive games.

GeForce Now originally presented itself as a rented remote GeForce PC: users would log into Steam or Epic Games and install games they already owned. Publishers objected that NVIDIA did not automatically have streaming rights to their games, turning licensing into another barrier. Hollister says Valve’s newer Steam publisher agreement apparently grants Valve rights needed for users to stream games to themselves, including within their homes. Pierce compares the issue to placing a computer a mile away and connecting it with a very long HDMI cable, but notes that copyright law treats remote streaming differently. The result is that users’ own hardware and games do not guarantee unrestricted cloud access.

Pierce asks whether Grand Theft Auto VI could become the watershed event that finally moves gaming toward the cloud, especially as hardware grows more expensive. Hollister says GTA is likely to make consumers spend unusually freely: people are already buying $700–$900 PlayStation 5 Pros for up to $1,200 on eBay, while $1,800 handhelds are selling out despite warnings that they are poor purchases. Cloud services will pitch themselves as a way to play GTA VI, but the game may push many users toward consoles instead. Even if GTA VI Online creates enormous demand, Hollister doubts players will keep expensive subscriptions once the initial excitement fades.

Pierce reads from a recent address by Pope Leo, who wrote that AI’s future is not predetermined because humans develop it and can guide or reject its applications. The decisive question, he says, is not simply what AI can do but what purpose it serves. Pierce endorses that framing against people at frontier AI labs who describe AI as something grown, alive, or possessing its own needs and desires. He also highlights the sentence “machines are tools, not persons,” arguing that people should treat software and AI as tools unless there is evidence to the contrary. He recommends the full address as a useful way to think about personhood in an AI-filled world.