Last week a friend called me to share about their new Windows laptop: ARM-based with some embedded “AI chip”. That came as a surprise for a consumer-grade PC. So I decided to research and was indeed shocked to learn about the recent (2020-2025) industry shift towards ARM.
My friend bought a Lenovo ThinkPad T14s with a Snapdragon X Elite X1E78100. This chip is a SoC (System on Chip): CPU + GPU + NPU (Neural Processing Unit) all-in-one, just like Apple M-series.
The NPU (the “AI chip”) is a specialized block for neural network inference, optimized for matrix operations at low precision (INT8, INT4). It consumes much less power than a GPU for targeted AI tasks, can run 13B+ parameter LLMs on-device, and delivers 45 TOPS, significantly ahead of Intel Core Ultra “Meteor Lake” at 10 TOPS and AMD Ryzen 8040 at 16 TOPS. source
ARM on the desktop
Windows on ARM technically existed since 2012 (Surface RT), but it is now mature for wide consumer adoption: much better x86-64 emulation, Snapdragon X Elite confirmed as broadly competitive with Apple M3, significantly better battery life than Intel equivalents. source
One major hindrance compared to x86: ARM on the laptop has no standardized hardware abstraction layer (ACPI). Instead, each platform uses Device Trees1, specific to each board, requiring dedicated porting work for each new device. source
For servers ARM actually defined an ACPI profile (“SBSA” — Server Base System Architecture), which is already supported by AWS Graviton, Ampere Altra etc. Which is why Linux “just works” on ARM cloud instances without any Device Tree fiddling. The server world already has its IBM-PC-moment equivalent.
On laptops and consumer devices, SoC vendors (Qualcomm especially) historically come from the mobile world where Device Tree is the norm and nobody cared about OS portability. They provide the closed proprietary firmware and drivers for Windows.
But Linux has to port everything from scratch. And so Linux is simply not ready yet for ARM laptops. Ubuntu has decent support for the ThinkPad T14s Gen 6 though source, but forget about Arch or FreeBSD. 😢 Windows is simply more mature for ARM right now.
So what happened? Apple M1 in 2020 is what happened. It proved ARM could be faster and more efficient than x86. That was a hard-to-ignore signal for the industry.
ARM on the server
ARM server market share grew from ~15% in 2024 to ~21-25% in 2025 source source, driven largely by NVIDIA’s Grace-Blackwell AI racks. In 2025 ARM claimed targeting 50% new compute among top hyperscalers (AWS, Google, Microsoft). This was partly recently confirmed by Amazon’s SVP who stated that the majority of compute capacity AWS added to its fleet in 2025 was Graviton-powered source. The growth seems real and significant.
This includes internal infrastructure invisible to customers, but also customer-facing services. Cloud providers actively encourage ARM adoption (EC2 Graviton instances, ARM-based RDS and Lambda) through cheaper pricing, typically 20-40% below equivalent x86 instances. source
Major platforms are actively migrating to ARM at scale: Uber is moving more than 5,000 services to ARM hosts source, Datadog has migrated 70% of its workloads to AWS Graviton source source.
Same enthusiasm for Paramount+ or Spotify on Google Cloud’s Axion ARM-based processors source.
Note: some figures come straight from the ARM newsroom and should be taken with a pinch of salt. Independent analysts (IDC, Dell’Oro) generally put the numbers lower but confirm the trend.
Docker
If Linux laptops and servers are becoming ARM, what does that mean for Docker? Nothing much actually: docker continues to use OS building blocks (cgroups, namespaces, overlayfs), just on a different architecture.
On Mac, Docker Desktop runs all containers inside a small Linux VM. arm64
images run natively without emulation. Legacy amd64/x86 images go through
Rosetta 2 — Apple’s binary translator, hardware-accelerated on M chips —
making it surprisingly fast in practice.
All popular images (postgres, nginx, redis, node, python, …) now publish multi-arch manifests. The compatibility pain of 2020-2021 is largely behind us.
On Linux ARM (Graviton servers, ARM laptops), there is no VM at all: containers run directly on the kernel, just as Docker was always meant to work.
Conclusion
I thought I kept myself reasonably informed about tech trends. Yet despite Linus Torvalds predicting “2016 will be the year of the ARM laptop” at LinuxCon 2015 source — before later arguing against ARM source — I failed to connect the dots from disparate signals into a single narrative.
The takeover didn’t come with an announcement. It came in pieces: a chip here, a cloud instance there, a laptop nobody in my circle bought. Until last week.