Qualcomm sold me a number at Computex 2024. Forty-five TOPS, x86-class performance, all of it sipping power inside a fanless chassis. Two years later the sequel has not just been announced, it has shipped and been reviewed, and I keep asking the same question I asked back then. Qualcomm unveiled the Snapdragon X2 Elite and the higher X2 Elite Extreme at its Snapdragon Summit in September 2025, the laptops reached retail this spring, and the review embargo lifted on April 7, 2026. So this is no longer a spec sheet I am squinting at. It is a platform with benchmarks, shipping hardware, and a verdict. The question is whether the second generation moved the things holding the first one back, or whether Qualcomm just threw more silicon at a software ecosystem that has not moved fast enough to deserve it.

Answering that means being clear about where the original X Elite landed, because the marketing version and the reality diverged almost immediately. The first platform, Elite, Plus, and the entry X across the 85-plus device designs Qualcomm counted, came out of 2024 and 2025 with a strong core, real battery life, and a market share that ABI Research pinned at roughly 4% of Windows PC shipments. Four percent. After two years of Microsoft flag-waving, after Epic added Easy Anti-Cheat ARM64 support in March 2025, after $600 Copilot+ machines finally hit shelves. The hardware got out ahead of everything else and then sat there. That is the backdrop the X2 walks into.

Qualcomm paid $1.4 billion for Nuvia in January 2021 to get a custom ARM64 core built by people who had done it before at Apple, and what Oryon is matters for judging whether the X2 core changes anything. On the top first-gen SKU, the X1E-84-100, you got twelve Oryon cores in a single homogeneous cluster with no big.LITTLE split, boosting to 4.3 GHz on a dual-core turbo. The twelve cores sat in three groups of four, each sharing 12 MB of L2, with a 42 MB System Level Cache behind it feeding the CPU, the Adreno, and the memory controller from one shared pool. That homogeneous layout was a deliberate bet, and it was the right one for Windows the first time around.

Here is where the X2 surprised me, because Qualcomm walked away from that bet. The Windows scheduler grew up on x86 symmetric multiprocessing and tends to make a mess of the asymmetric core topologies that mobile chips lean on, so a flat cluster of identical cores was the safe call in 2024. The X2 Elite Extreme does not do that. It runs eighteen cores split twelve Prime plus six Performance, two clusters of different core types, the kind of heterogeneous layout the first design pointedly avoided. Qualcomm decided the scheduler had matured enough, or that the multi-core and efficiency gains were worth the risk. Early reviews suggest the gamble paid: the Zenbook A16 with the Extreme posted a Geekbench 6 single-core near 3,800 and multi-core past 23,000, the top of the thin-and-light category. But the clean homogeneous story I told about the first generation is not the X2’s story. The bet changed.

The Extreme, model X2E-96-100, is built on a 3nm node and is the first Arm-based PC chip to reach 5.0 GHz, boosting there on one or two cores while holding around 4.4 GHz all-core across the twelve Prime cores, with the six Performance cores at 3.6 GHz, and 53 MB of total cache behind the lot. Those are not leaked or hedged numbers anymore. They sit on Qualcomm’s product page and in a dozen reviews.

The NPU is the spec Qualcomm wants you to read, and it is a real jump: 80 TOPS at INT8, up from 45, measured the same rough way across the Hexagon’s sub-units. What has not changed since 2024 is that the NPU only matters if software targets it. On the first platform the Windows inference stack split into at least three paths, with Qualcomm’s QNN SDK routing to the Hexagon, DirectML routing to the Adreno, and everything else falling back to the CPU, so an arbitrary AI app from the Microsoft Store could leave the NPU idle and the 45 TOPS would be real and irrelevant at the same time. Qualcomm and Microsoft have been pushing deeper Windows ML integration meant to let apps reach the NPU through Microsoft’s abstraction layer without wiring up QNN directly. If that ships and works, it chips at the fragmentation. I have been waiting for that fix since 2024, and I want to see it land in shipping apps before I call it solved.

Quieter than the NPU number, and more important, is the memory pipe, and this is where the X2 changes the calculus. The first X Elite ran unified LPDDR5X on a 128-bit bus at a theoretical 135.8 GB/s, soldered, capped at 64 GB. The X2 Elite moves to 152 GB/s, the Elite Extreme to 228 GB/s, and the maximum addressable memory doubles to 128 GB. That last figure is the quiet headline, because a Windows-on-Arm laptop can now hold a 70B-class model in memory, something only Apple could do before. The bandwidth jump moves the inference ceiling the first-gen NPU was never the real bottleneck against. A 7B model at INT8 is about 7 GB of weights per forward pass. On the original platform, call it 100 GB/s sustained after controller overhead, you were looking at roughly 14 tokens per second as a memory-bandwidth ceiling, not a compute one. Push the Extreme’s 228 GB/s through the same back-of-envelope and you clear 25 tokens per second on the same model, before any INT4 quantization. The TOPS headline grabbed the slide. The bandwidth is what changes what you can run.

The Adreno was the weak leg of the first platform, and the X2-90 is a real rebuild rather than a clock bump. Qualcomm quotes 2.3x the performance per watt of the old Adreno 740 and around 80% more raw GPU throughput, and the reviews back the leap: in 3DMark Steel Nomad the Extreme landed ahead of both the Apple M5 MacBook Air and Intel’s Core Ultra 7 355. But the architecture is still tile-based deferred rendering, and TBDR is exactly what broke the DX11 and DX9 back catalog the first time. TBDR shades geometry in screen-space tiles to save bandwidth on a unified memory pool, and a lot of older DirectX titles assume immediate-mode rendering, so a TBDR part either enforces ordering at a performance cost or risks artifacts. Making the GPU bigger and faster does not dissolve that friction with a legacy x86 catalog. A faster Adreno renders the games that already run well even faster. It does nothing for the ones that were never going to run right.

Prism, the x86-to-ARM translator, carried roughly 5-15% overhead on browsers and Office and climbed to 30-50% on AVX2-heavy compute, because AVX and AVX2 had no native ARM64 mapping and had to be emulated in full. By late 2025 the native column was solid: Chrome, Edge, Firefox, Microsoft 365, VS Code, Python, .NET, Photoshop. Premiere was still x86. And underneath all of it sits the wall nobody at Qualcomm can move, because Windows 11 on Arm does not run 32-bit x86 binaries at all, which takes out legacy enterprise tools, a chunk of the Steam back catalog, and older games with no fix on the horizon. None of that changes with an eighteen-core die. It is a Microsoft OS decision, not a silicon one.

The anti-cheat problem is the same story told in kernel drivers. Epic shipped a native ARM64 Easy Anti-Cheat developer preview in March 2025 with Fortnite first, but writing a kernel security driver for a new architecture is not a recompile, and getting a thousand-plus EAC titles to each opt in and rebuild plays out over 12-24 months. BattlEye still has not announced ARM64 support, which strands PUBG, Rainbow Six Siege, and Destiny 2. VAC runs under Prism with no native build. Eighteen cores at 5 GHz do not write a single one of those drivers.

None of those barriers is the thing I keep coming back to, though. What I keep thinking about is Windows RT. In 2012 Microsoft shipped an Arm version of Windows that could not run legacy desktop apps and locked the desktop to a bundled copy of Office. It told developers to rewrite everything for a Store nobody wanted to build for, then watched nobody show up. It wrote off most of a billion dollars on unsold Surface RT inventory and buried the idea. The hardware was never the problem. The platform was empty, and Microsoft expected the world to fill it for free. The Snapdragon X era is the inverse of that mistake. For two years Qualcomm and Microsoft did the slow, unglamorous work RT skipped: turning Prism into an emulator that runs x64 and now AVX2, shipping ARM64EC so native and x86 code can share a process, standing up Windows ML so the NPU is reachable without bespoke wiring, and grinding through it app by app to get native Arm builds of the things people live in. Not just the consumer staples like Chrome and Spotify, but the professional stack that decides whether a laptop is a toy or a tool, the Photoshop and the dev environments a paid user opens every morning. That groundwork is why the X2 ships into a Windows that has software. RT shipped into one that had a tech demo.

Here is the part that makes me laugh a little. Qualcomm did not do that work as charity. It did it as the sole Microsoft partner, the only company allowed to ship an Arm chip in a Windows PC from 2016 onward. That exclusivity lapsed in 2024, and the door Qualcomm spent eight years walking through is now open to everyone. NVIDIA showed up first, its RTX Spark and N1X silicon landing in Dell and Lenovo machines and a Surface this year, twenty Arm cores and a Blackwell GPU stapled to the exact ecosystem Qualcomm built. MediaTek and AMD are right behind. And the detail I cannot get over is that NVIDIA already failed at this once, because its Tegra 3 was the chip inside that doomed Surface RT in 2012. The company that helped Microsoft faceplant on Arm the first time gets to walk back into a Windows where the heavy lifting is done, the apps ported and the emulator working, none of which it had to pay for. Qualcomm cleared the minefield and the competition is strolling across it. The same software ceiling still caps everyone, because a native CUDA stack says nothing about the long tail of legacy Windows nobody has recompiled. But two years of quiet platform work, and the biggest beneficiaries might be the people who never did it.

And the competition the X2 faces is not the 2024 lineup the first-gen story was built against. Intel’s Lunar Lake is a generation old now. Its successor, Panther Lake, the Core Ultra 300 series, launched at CES 2026 on Intel’s own 18A node, with an NPU 5 rated around 50 TOPS standalone and a much stronger Xe3 Arc GPU that closes a lot of the gaming gap Qualcomm has always lost. AMD’s Strix Point gave way to Gorgon Point, the Ryzen AI 400 line. The X2 reviews benchmark straight against Panther Lake, and on CPU the Extreme wins clearly, single-core roughly 24% ahead of Intel’s Core Ultra X9 388H. But Intel’s GPU answer is real, and the most quoted finding from the whole review wave is that x86 laptops shed close to half their performance unplugged while the Snapdragon holds nearly all of it. That battery-bound gap is the X2’s most durable advantage, and the one thing none of the x86 successors have closed.

The most strategically interesting Qualcomm move is not the Extreme at all. At CES 2026 Qualcomm showed the X2 Plus and signaled a budget Snapdragon C class aimed at the part of the market where the volume lives. The first-gen entry X already proved Qualcomm could hit $600 while keeping the 45 TOPS NPU intact, because the NPU was the one block it would not cut, so every Snapdragon device at every price cleared the Copilot+ floor. If Qualcomm can get ARM64 Windows into $400-500 machines in volume, that is how you move off 4%, not by stacking Oryon cores into Extreme SKUs that start at $1,699 on the Asus Zenbook A16. The premium parts win benchmarks. The cheap parts win share. Those are different fights.

So what did the X2 fix? The core is faster and there are more of them, the GPU stopped being an embarrassment, the NPU doubled, the memory ceiling doubled, and on battery it humiliates x86. Those gains are real, and for the traveling professional who lives in browsers, Microsoft 365, Teams, and native Arm tooling, the Extreme is a fast, absurdly efficient machine that reviewers came away impressed by. What it did not fix is the part that was never made of transistors. The AVX2 Prism overhead is structurally the same. BattlEye still has no ARM64 driver. The 32-bit block is a Microsoft decision nobody is revisiting. The TBDR friction with the back catalog is a property of the GPU family, not a bug. Enterprise procurement on 3-5 year cycles, the kind that will not approve a fleet without ARM64 builds for every deployed app, is untouched by anything Qualcomm shipped.

The Nuvia acquisition bought Qualcomm an excellent CPU, and the second generation proves the team can build. But a better chip does not rewrite a kernel anti-cheat driver, and it does not talk enterprise IT into recertifying an app stack. Two years and a full generation in, the reviews say what the spec sheet said: the silicon is no longer the question. Qualcomm keeps building a better answer to a problem that was never going to be solved in the chip, and at some point the fair read is that the ceiling was never made of transistors. It is made of everyone else’s software, and that is not Qualcomm’s to ship.

Sources & related reading