In March 2026, Steam reported Linux at 5.33% of its user base. StatCounter reported 3.16% of global desktop share the same month. Estimates aggregating across tracking methods for year-end 2025 landed near 4.7%.
All three are correct. None of them is measuring the same thing.
Anyone citing a Linux market share figure is citing a population definition and a detection method, usually without saying which.
The gap between these numbers is not noise or error, and understanding where it comes from is more useful than picking a favourite.
The PC gaming operating system breakdown reports the Steam-derived figures; this piece is about why they differ from everything else.
The Tell Is in the Windows Number
Start with the figure nobody quotes, because it exposes the mechanism immediately.
Steam’s March 2026 survey put Windows at 92.33%, Linux at 5.33% and macOS at 2.35%. StatCounter’s April 2026 desktop data put Windows at 66.6%, macOS at 7.34% and Linux at 2.99%.
Windows differs by nearly 26 points between the two. macOS differs by a factor of three, in the opposite direction from Linux.
That is not a Linux measurement problem, it is two entirely different populations: Steam surveys people who play PC games, StatCounter samples general web traffic, and Linux users are distributed differently across those groups.
The Discipline Has a Name Somewhere Else
Reading figures like these is inference rather than arithmetic, and the field that has formalised that skill most thoroughly is not statistics.
Poker is played entirely on incomplete information. You never see an opponent’s cards. You observe their actions and reason backwards to the holdings consistent with them.
Free strategy material on that process has been published since 2004, as covered by pokerology, and three of its standard concepts map directly onto reading platform-share data.
- Hand reading is range construction, not guessing. A competent player never concludes “he has aces.” They narrow the set of hands consistent with the betting so far, then keep narrowing as more information arrives. The equivalent here is declining to treat 5.33% as a fact about the world, and instead asking which populations and detection methods would produce that reading.
- Sample size decides whether a statistic exists at all. An opponent’s aggression frequency across fifty hands is noise. Across ten thousand it is a read worth acting on. That is precisely why a single month of an opt-in survey is a weaker instrument than the same survey’s trend across a year, and it is the reason the February 2025 figure below should not have surprised anyone.
- Showdown bias is the hard one, and it is the same problem as an Unknown bucket. A hand-history database records only the hands that reached a showdown, and those are systematically the hands where somebody held something worth showing. Everything folded earlier is invisible, and the invisible portion is not a random sample of the visible one. Any detection method that counts only the browsers volunteering their operating system has exactly that shape.
The transferable principle underneath all three: an observation is evidence about the process that produced it before it is evidence about the world.
Steam’s Survey: Opt-In, and Hardware Does the Voting
The Steam Hardware Survey prompts a rotating subset of users each month and asks permission to report their configuration. Two structural features follow from that.
It is opt-in, so participation itself is a variable. The clearest demonstration is February 2025, when Linux on Steam dropped to 1.45%. Linux users did not vanish. Increased survey participation from Simplified Chinese language users temporarily shifted the regional composition of the sample, and Linux share fell as an arithmetic consequence. It recovered to 3.05% by October 2025 and 3.20% by November.
If a figure can move by more than half its value because of who answered that month, the monthly number is a weaker signal than the trend. Fifty hands, not ten thousand.
Hardware sales register as OS adoption. Every Steam Deck reports SteamOS. Valve shipped an estimated 5.6 million units by mid-2025, and each one enters the survey as a Linux system whether the owner thinks of themselves as a Linux user or not.
That is a definition question rather than a distortion. The devices genuinely run Linux. Whether they represent desktop Linux adoption is separate, and Valve’s own Linux-only view helps: SteamOS Holo held 22.24% of Linux Steam systems in July 2026, with CachyOS at 14.31% and Arch Linux at 8.29%. Source: Valve, Steam Hardware & Software Survey, Linux-only view, July 2026.
Roughly three-quarters of Linux on Steam is therefore not Steam Deck.
StatCounter: User Agents and the Unknown Bucket
StatCounter works from browser user-agent strings collected across tracked page views. Different population, different failure mode.
The failure mode is undercounting, and it is structural. Privacy-hardened browsers, user-agent reduction, VPN traffic and anti-fingerprinting configurations all produce requests that do not resolve cleanly to an operating system.
Those land in an Unknown category, and that category skews toward exactly the users most likely to be running Linux.
This is the showdown problem in its purest form. The hands you never see are not a random selection of the hands that were played.
So the StatCounter Linux figure should be read as a floor rather than an estimate. Analyses attempting to correct for the Unknown bucket put year-end 2025 nearer 4.7%, against monthly readings around 3%. Source: StatCounter monthly desktop data, with adjusted estimates via aggregating analyses.
Worth noting the StatCounter series is also not monotonic. Linux touched 4.6% in August 2024 before falling back, a reminder that detection methodology changes can move a series independently of anything happening in the world.
Which Number to Use?
Depends entirely on the question, and the questions are genuinely different.
- “Should I ship a Linux build?” Use Steam. Your customers are gamers, Steam surveys gamers, and 5.33% of a platform that set a concurrent-user record above 42.3 million in March 2026 is a substantial absolute audience regardless of the percentage.
- “Should we support Linux desktops in the enterprise?” Neither, really. Use developer surveys. Stack Overflow’s 2025 data had Ubuntu at 27.8% personal and 27.7% professional use, which describes the population being supported far better than global browser traffic does.
- “Is Linux desktop adoption growing?” Use StatCounter for the trend and ignore the level. The direction is reliable even though the absolute number is a floor.
- “How many people run Linux?” Nobody knows, and anyone offering a confident single figure is quoting one method without saying so.
The Practical Version
Three rules, and they generalise well past this topic.
- Ask what population a figure describes before asking whether it is high or low. Steam and StatCounter disagree about Windows by 26 points, which says the populations differ more than the measurements do.
- Treat opt-in surveys as trend instruments, not levels. February 2025 moved Linux on Steam by more than half its value on participation alone.
- Assume undercounting where detection is passive. A method that infers the OS from what a browser volunteers will systematically miss the users who configure their browser to volunteer less.
None of which makes these sources unreliable. It makes them specific, which is a different property and a considerably more useful one once you know which specificity you needed.
The Steam Deck figures are a good worked example, since the same underlying data supports both “the Deck drove Linux growth” and “desktop Linux grew faster than the Deck” depending on which slice you take.
