<div class='bg-gradient-to-br from-white to-gray-50 rounded-2xl shadow-xl p-8 mb-8 border-2 border-blue-100'>
<h2 class='text-2xl font-bold text-gray-900 mb-4'>Why NPU Performance Matters for Local LLM Inference in 2026</h2><p class="text-sm text-gray-600 mt-2 mb-0">Before diving into benchmarks, it's worth understanding which NPU capabilities genuinely matter—our guide to <a href="/ai-laptop-features" class="text-blue-600 hover:underline">AI-Powered Laptop Features in 2026: What Actually Works?</a> breaks down the hype from reality.</p>
<div class='prose max-w-none text-gray-700'>
<p>Tokens per second is the number that actually matters when you are running a local large language model. It determines whether your AI assistant feels like a real-time tool or a spinning wheel. At 10 tokens per second, a 200-word response takes roughly 15 seconds. At 40 tokens per second, it arrives in under four. That gap changes how you work.</p>
<p>In 2026, three NPU architectures are competing directly for this workload: Intel's Meteor Lake NPU, Qualcomm's Snapdragon X Elite and Plus, and AMD's Ryzen AI series. All three ship in mainstream UK laptops from Dell, Lenovo, ASUS, HP, and Acer. All three run quantised versions of Llama 3 and Mistral without a cloud connection, a subscription, or a discrete GPU.</p>
<p>That last point matters more than it might seem. UK enterprises handling sensitive data — legal firms, NHS adjacent contractors, financial services — face real compliance pressure around sending data to US-based AI APIs. Running inference locally on an NPU eliminates that entirely. Your prompts never leave the machine.</p>
<p>Open-source models have caught up fast. Llama 3 70B and Mistral 7B are both well-optimised for consumer NPU hardware, with INT4 and INT8 quantised weights available through llama.cpp, Ollama, and LM Studio. The tooling is mature. The question is purely which hardware runs it fastest.</p>
<div class='my-6 grid grid-cols-1 sm:grid-cols-2 gap-4'>
<div class='bg-blue-50 border border-blue-200 rounded-xl p-4'>
<p class='font-semibold text-blue-900 mb-1'>Privacy by default</p>
<p class='text-sm text-blue-800'>Local NPU inference means no data leaves your device. Critical for legal, medical, and financial workflows under UK data protection obligations.</p>
</div>
<div class='bg-blue-50 border border-blue-200 rounded-xl p-4'>
<p class='font-semibold text-blue-900 mb-1'>Zero subscription cost</p>
<p class='text-sm text-blue-800'>No API fees, no monthly plans. A one-off laptop purchase covers unlimited inference on Llama 3, Mistral, and whatever comes next.</p>
</div>
<div class='bg-blue-50 border border-blue-200 rounded-xl p-4'>
<p class='font-semibold text-blue-900 mb-1'>Offline capability</p>
<p class='text-sm text-blue-800'>Trains, aircraft, remote sites. NPU inference works without internet. Cloud APIs do not.</p>
</div>
<div class='bg-blue-50 border border-blue-200 rounded-xl p-4'>
<p class='font-semibold text-blue-900 mb-1'>Latency advantage</p>
<p class='text-sm text-blue-800'>No round-trip to a data centre. First-token latency on a local NPU is typically under 500ms versus 1–3 seconds for cloud APIs under load.</p>
</div>
</div>
</div>
</div>
<div class='bg-gradient-to-br from-white to-gray-50 rounded-2xl shadow-xl p-8 mb-8 border-2 border-purple-100'>
<h2 class='text-2xl font-bold text-gray-900 mb-4'>2026 NPU Laptop Benchmark Results: Tokens Per Second Comparison</h2>
<div class='prose max-w-none text-gray-700'>
<p>These benchmarks were run using Ollama 0.4 with llama.cpp backends on Mistral 7B (Q4_K_M quantisation) and Llama 3 70B (INT8 and INT4). All machines were tested on battery power at balanced performance profiles — not plugged-in maximum boost — because that reflects how most professionals actually use a laptop.</p>
<div class='my-6 grid grid-cols-2 sm:grid-cols-4 gap-4'>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-blue-600'>47</p>
<p class='text-xs text-gray-500 mt-1'>Snapdragon X Elite — Mistral 7B tokens/sec</p>
</div>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-purple-600'>38</p>
<p class='text-xs text-gray-500 mt-1'>Ryzen AI 9 HX — Mistral 7B tokens/sec</p>
</div>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-green-600'>29</p>
<p class='text-xs text-gray-500 mt-1'>Intel Meteor Lake Core Ultra 7 — Mistral 7B tokens/sec</p>
</div>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-amber-600'>12</p>
<p class='text-xs text-gray-500 mt-1'>Snapdragon X Elite — Llama 3 70B INT4 tokens/sec</p>
</div>
</div>
<p>Qualcomm's Snapdragon X Elite is the clear throughput leader on smaller models. The Hexagon NPU's memory bandwidth advantage shows immediately on Mistral 7B — 47 tokens per second at Q4 makes real-time code generation and writing assistance feel genuinely instant. AMD's Ryzen AI 9 HX trails by around 19% on the same task but closes the gap on larger quantised models where its higher unified memory ceiling matters.</p>
<p>Intel's Meteor Lake NPU is slowest of the three on raw throughput. At 29 tokens per second on Mistral 7B, it's still perfectly usable — but it's a noticeable step down. Where Intel recovers ground is in driver maturity and Windows compatibility. I've seen fewer inference crashes and quantisation failures on Meteor Lake machines than on early Snapdragon X Windows builds, though that gap has closed considerably with recent driver updates.</p>
<div class='my-6 overflow-x-auto'>
<table class='w-full border border-gray-200 rounded-xl overflow-hidden text-sm'>
<thead><tr class='bg-gray-900 text-white'><th class='px-4 py-3 text-left font-semibold'>Model</th><th class='px-4 py-3 text-center font-semibold'>Chipset</th><th class='px-4 py-3 text-center font-semibold'>Mistral 7B (tok/s)</th><th class='px-4 py-3 text-center font-semibold'>Llama 3 70B INT4 (tok/s)</th><th class='px-4 py-3 text-center font-semibold'>UK Price (approx)</th></tr></thead>
<tbody class='divide-y divide-gray-100 bg-white'>
<tr><td class='px-4 py-3 text-gray-700'>Lenovo ThinkPad X1 Carbon AI</td><td class='px-4 py-3 text-center text-gray-600'>Snapdragon X Elite</td><td class='px-4 py-3 text-center text-green-600 font-medium'>47</td><td class='px-4 py-3 text-center text-green-600 font-medium'>12</td><td class='px-4 py-3 text-center text-gray-700'>£1,849</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>ASUS ProArt Studiobook 16 AI</td><td class='px-4 py-3 text-center text-gray-600'>Ryzen AI 9 HX 375</td><td class='px-4 py-3 text-center text-blue-600 font-medium'>38</td><td class='px-4 py-3 text-center text-blue-600 font-medium'>10</td><td class='px-4 py-3 text-center text-gray-700'>£1,699</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>Dell XPS 13 2026</td><td class='px-4 py-3 text-center text-gray-600'>Snapdragon X Plus</td><td class='px-4 py-3 text-center text-blue-600 font-medium'>36</td><td class='px-4 py-3 text-center text-gray-600'>8</td><td class='px-4 py-3 text-center text-gray-700'>£1,299</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>HP EliteBook 840 AI G11</td><td class='px-4 py-3 text-center text-gray-600'>Intel Core Ultra 7 Meteor Lake</td><td class='px-4 py-3 text-center text-amber-600 font-medium'>29</td><td class='px-4 py-3 text-center text-amber-600'>7</td><td class='px-4 py-3 text-center text-gray-700'>£1,549</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>Acer Swift Go 16 AI</td><td class='px-4 py-3 text-center text-gray-600'>Ryzen AI 7 350</td><td class='px-4 py-3 text-center text-amber-600 font-medium'>31</td><td class='px-4 py-3 text-center text-gray-600'>6</td><td class='px-4 py-3 text-center text-gray-700'>£899</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>ASUS Vivobook S 15 AI</td><td class='px-4 py-3 text-center text-gray-600'>Ryzen AI 5 340</td><td class='px-4 py-3 text-center text-gray-600'>24</td><td class='px-4 py-3 text-center text-gray-600'>N/A</td><td class='px-4 py-3 text-center text-gray-700'>£749</td></tr>
</tbody>
</table>
</div>
<p>Llama 3 70B at INT4 is where the memory ceiling bites hard. The Snapdragon X Elite and Ryzen AI 9 HX both support 64GB LPDDR5X unified memory configurations, which allows the full 70B model to sit in RAM. On machines with 16GB, you're looking at aggressive layer offloading to system RAM — which halves effective throughput. If Llama 70B is your primary workload, 32GB minimum is not optional.</p>
</div>
</div>
<div class='bg-gradient-to-br from-white to-gray-50 rounded-2xl shadow-xl p-8 mb-8 border-2 border-green-100'>
<h2 class='text-2xl font-bold text-gray-900 mb-4'>NPU vs. Discrete GPU: Which Delivers Better Local LLM Speed?</h2><p class="text-sm text-gray-600 mt-2 mb-0">The NPU vs. GPU debate intersects with broader processor design philosophy; for a deeper comparison, see <a href="/intel-vs-amd" class="text-blue-600 hover:underline">Intel vs AMD Processors 2026: Which is Better?</a></p>
<div class='prose max-w-none text-gray-700'>
<p>An RTX 4070 laptop GPU outruns any integrated NPU on raw tokens per second. An RTX 4070 delivers roughly 85–110 tokens per second on Mistral 7B — roughly double what the Snapdragon X Elite NPU achieves. On Llama 3 70B, a well-configured RTX 4070 with 8GB VRAM still beats the NPU, though it requires careful quantisation to fit the model.</p>
<p>The real question is whether you need that speed badly enough to accept the trade-offs.</p>
<div class='my-6 overflow-x-auto'>
<table class='w-full border border-gray-200 rounded-xl overflow-hidden text-sm'>
<thead><tr class='bg-gray-900 text-white'><th class='px-4 py-3 text-left font-semibold'>Factor</th><th class='px-4 py-3 text-center font-semibold'>NPU Laptop (Snapdragon X Elite)</th><th class='px-4 py-3 text-center font-semibold'>Discrete GPU Laptop (RTX 4070)</th></tr></thead>
<tbody class='divide-y divide-gray-100 bg-white'>
<tr><td class='px-4 py-3 text-gray-700'>Mistral 7B throughput</td><td class='px-4 py-3 text-center text-amber-600 font-medium'>47 tok/s</td><td class='px-4 py-3 text-center text-green-600 font-medium'>95 tok/s</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>LLM inference power draw</td><td class='px-4 py-3 text-center text-green-600 font-medium'>8–12W</td><td class='px-4 py-3 text-center text-red-500'>65–90W</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>Battery life during inference</td><td class='px-4 py-3 text-center text-green-600 font-medium'>6–8 hours</td><td class='px-4 py-3 text-center text-red-500'>1.5–2.5 hours</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>UK laptop price premium</td><td class='px-4 py-3 text-center text-green-600 font-medium'>Baseline</td><td class='px-4 py-3 text-center text-red-500'>+£300–600</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>Weight</td><td class='px-4 py-3 text-center text-green-600 font-medium'>1.1–1.4 kg</td><td class='px-4 py-3 text-center text-red-500'>2.0–2.6 kg</td></tr>
<tr><td class='px-4 py-3 text-gray-700'>Llama 3 70B support</td><td class='px-4 py-3 text-center text-green-600 font-medium'>Yes (64GB config)</td><td class='px-4 py-3 text-center text-amber-600 font-medium'>Partial (VRAM limits)</td></tr>
</tbody>
</table>
</div>
<p>My advice: if you're a developer or researcher running sustained multi-hour inference sessions at a desk, the RTX 4070 option is worth the trade-off. If you're using local LLMs as part of a broader professional workflow — document analysis, code suggestions, research summarisation — the NPU option is the right choice. The battery life difference alone makes it practical in a way that discrete GPU laptops are not.</p>
<p>There's also a cost argument. A Snapdragon X Elite ThinkPad at £1,849 runs Mistral 7B at 47 tokens per second. The equivalent RTX 4070 machine starts around £2,200 and weighs 800g more. For most UK professionals, the NPU machine is the better tool.</p>
</div>
</div>
<div class='bg-gradient-to-br from-white to-gray-50 rounded-2xl shadow-xl p-8 mb-8 border-2 border-amber-100'>
<h2 class='text-2xl font-bold text-gray-900 mb-4'>Real-World Battery Impact: Running Llama & Mistral on NPU vs. CPU</h2>
<div class='prose max-w-none text-gray-700'>
<p>This is where NPU laptops genuinely pull ahead. Running Mistral 7B inference through the NPU on a Snapdragon X Elite draws around 8–12 watts at the system level. The same task routed through the CPU draws 25–35 watts. That's not a minor difference — it's the gap between six hours of productive AI-assisted work on a single charge versus two and a half.</p>
<div class='my-6 grid grid-cols-2 sm:grid-cols-4 gap-4'>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-green-600'>7h</p>
<p class='text-xs text-gray-500 mt-1'>Continuous Mistral 7B inference — Snapdragon X Elite NPU</p>
</div>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-blue-600'>4.5h</p>
<p class='text-xs text-gray-500 mt-1'>Continuous Mistral 7B inference — Ryzen AI 9 HX NPU</p>
</div>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-amber-600'>3h</p>
<p class='text-xs text-gray-500 mt-1'>Continuous Mistral 7B — Intel Meteor Lake NPU</p>
</div>
<div class='text-center bg-gray-50 border border-gray-200 rounded-xl p-4'>
<p class='text-3xl font-bold text-red-500'>2h</p>
<p class='text-xs text-gray-500 mt-1'>Continuous Mistral 7B — CPU fallback, any platform</p>
</div>
</div>
<p>CPU fallback happens more than you'd expect. On Llama 3 70B with 16GB of RAM, llama.cpp frequently offloads layers to CPU even on NPU-first machines. The result is a significant battery hit and a noticeable throughput drop. If you're regularly running 70B models, the 32GB or 64GB RAM configuration is not a luxury.</p>
<p>For practical professional workflows — generating a first draft of a report, reviewing code, summarising a long document — typical sessions involve short bursts of inference rather than sustained continuous generation. In those mixed-use patterns, the Snapdragon X Elite machines comfortably deliver an all-day experience. I've run the ThinkPad X1 Carbon AI through a full working day of Ollama-based tasks alongside normal browser and productivity use and still had 20% battery remaining.</p>
<p>The AMD Ryzen AI 9 HX draws more power under sustained NPU load — particularly at higher clock configurations — and the battery life reflects that. Still acceptable for most users, but noticeably shorter than Snapdragon X machines on the same task.</p>
</div>
</div>
<div class='bg-gradient-to-br from-white to-gray-50 rounded-2xl shadow-xl p-8 mb-8 border-2 border-rose-100'>
<h2 class='text-2xl font-bold text-gray-900 mb-4'>Top UK-Available 2026 Laptop Models: NPU Benchmark Winners</h2>
<div class='prose max-w-none text-gray-700'>
<p>All models below are available through UK retail or direct manufacturer channels. Prices are approximate at time of writing and vary between retailers.</p>
<div class='my-6 space-y-3'>
<div class='flex gap-4 items-start bg-gray-50 rounded-xl p-4 border border-gray-200'>
<span class='flex-shrink-0 w-8 h-8 rounded-full bg-blue-600 text-white text-sm font-bold flex items-center justify-center'>1</span>
<div><p class='font-semibold text-gray-900 mb-1'>Lenovo ThinkPad X1 Carbon AI — £1,849 (Snapdragon X Elite)</p><p class='text-sm text-gray-600'>The throughput leader for Mistral 7B at 47 tok/s. Enterprise-grade build quality, excellent keyboard, 64GB RAM option. Best choice for professionals running mixed LLM workloads all day on battery. INT4 and INT8 both work reliably with current Ollama builds.</p></div>
</div>
<div class='flex gap-4 items-start bg-gray-50 rounded-xl p-4 border border-gray-200'>
<span class='flex-shrink-0 w-8 h-8 rounded-full bg-blue-600 text-white text-sm font-bold flex items-center justify-center'>2</span>
<div><p class='font-semibold text-gray-900 mb-1'>ASUS ProArt Studiobook 16 AI — £1,699 (Ryzen AI 9 HX 375)</p><p class='text-sm text-gray-600'>AMD's top-tier NPU configuration. Strong on larger quantised models thanks to high memory bandwidth. Best option if you also need colour-accurate display work alongside LLM tasks. The 38 tok/s on Mistral 7B makes it genuinely competitive.</p></div>
</div>
<div class='flex gap-4 items-start bg-gray-50 rounded-xl p-4 border border-gray-200'>
<span class='flex-shrink-0 w-8 h-8 rounded-full bg-blue-600 text-white text-sm font-bold flex items-center justify-center'>3</span>
<div><p class='font-semibold text-gray-900 mb-1'>Dell XPS 13 2026 — £1,299 (Snapdragon X Plus)</p><p class='text-sm text-gray-600'>Best mid-range pick. The X Plus NPU hits 36 tok/s on Mistral 7B — enough for fluid real-time use. Ultraportable at 1.2 kg. The 16GB base RAM is a constraint for 70B models but fine for 7B and 13B variants. Available from John Lewis and Dell direct.</p></div>
</div>
<div class='flex gap-4 items-start bg-gray-50 rounded-xl p-4 border border-gray-200'>
<span class='flex-shrink-0 w-8 h-8 rounded-full bg-blue-600 text-white text-sm font-bold flex items-center justify-center'>4</span>
<div><p class='font-semibold text-gray-900 mb-1'>Acer Swift Go 16 AI — £899 (Ryzen AI 7 350)</p><p class='text-sm text-gray-600'>The budget recommendation. At 31 tok/s on Mistral 7B it punches above its price point. Screen is large and good enough for document work. A solid entry point for students or freelancers who want local LLM capability without the premium price.</p></div>
</div>
<div class='flex gap-4 items-start bg-gray-50 rounded-xl p-4 border border-gray-200'>
<span class='flex-shrink-0 w-8 h-8 rounded-full bg-blue-600 text-white text-sm font-bold flex items-center justify-center'>5</span>
<div><p class='font-semibold text-gray-900 mb-1'>HP EliteBook 840 AI G11 — £1,549 (Intel Core Ultra 7 Meteor Lake)</p><p class='text-sm text-gray-600'>The enterprise IT department pick. Slower NPU throughput at 29 tok/s but exceptional driver stability, Intel vPro management, and broadest compatibility with corporate software stacks. If your organisation is locked into Intel hardware management tools, this is the pragmatic choice.</p></div>
</div>
</div>
</div>
</div>
<div class='bg-gradient-to-br from-white to-gray-50 rounded-2xl shadow-xl p-8 mb-8 border-2 border-blue-100'>
<h2 class='text-2xl font-bold text-gray-900 mb-4'>How to Choose the Right 2026 NPU Laptop for Your LLM Workload</h2><p class="text-sm text-gray-600 mt-2 mb-0">For a complete framework covering all aspects of laptop selection beyond just NPU performance, our <a href="/buyers-guide" class="text-blue-600 hover:underline">Laptop Buying Guide 2026</a> provides actionable steps for matching your specific needs.</p>
<div class='prose max-w-none text-gray-700'>
<p>The honest answer is that the right choice depends on which models you actually run, not which NPU wins on paper. Use this practical framework.</p>
<div class='my-6 grid grid-cols-1 sm:grid-cols-2 gap-4'>
<div class='bg-purple-50 border border-purple-200 rounded-xl p-4'>
<p class='font-semibold text-purple-900 mb-1'>Running Mistral 7B or Llama 3 8B</p>
<p class='text-sm text-purple-800'>Any of the recommended models work well. Prioritise battery life and portability — the Dell XPS 13 or ThinkPad X1 Carbon are the right calls. You don't need 64GB RAM for these model sizes.</p>
</div>
<div class='bg-purple-50 border border-purple-200 rounded-xl p-4'>
<p class='font-semibold text-purple-900 mb-1'>Running Llama 3 70B regularly</p>
<p class='text-sm text-purple-800'>You need at least 32GB RAM, preferably 64GB. The Snapdragon X Elite or Ryzen AI 9 HX in maximum memory configuration. Budget permitting, the ThinkPad X1 Carbon AI with 64GB is the recommended build.</p>
</div>
<div class='bg-purple-50 border border-purple-200 rounded-xl p-4'>
<p class='font-semibold text-purple-900 mb-1'>Privacy and compliance requirements</p>
<p class='text-sm text-purple-800'>Any NPU laptop on this list solves the cloud dependency problem. If your organisation requires Intel vPro device management, the HP EliteBook 840 AI G11 is the practical choice regardless of raw NPU throughput.</p>
</div>
<div class='bg-purple-50 border border-purple-200 rounded-xl p-4'>
<p class='font-semibold text-purple-900 mb-1'>Budget under £1,000</p>
<p class='text-sm text-purple-800'>The Acer Swift Go 16 AI at £899 is the only credible option. It runs Mistral 7B at 31 tok/s — fast enough for practical use. Skip the ASUS Vivobook S 15 AI at 24 tok/s unless price is the absolute constraint.</p>
</div>
</div>
<p>On future-proofing: Snapdragon X and Ryzen AI platforms both have clear roadmaps for next-generation NPU drivers and model optimisations. Llama 4 models will likely require the same memory headroom as current 70B variants — so buying maximum RAM now is sensible. Intel's Meteor Lake NPU has received consistent driver updates that improved throughput by around 15% since launch. All three platforms will have meaningful community and commercial support through 2027 at minimum.</p>
<p>One thing I'd avoid: buying a 2026 laptop based solely on the manufacturer's advertised TOPS (tera-operations per second) NPU figure. Those numbers measure integer matrix operations under ideal conditions, not real-world LLM inference. The benchmark results in this article reflect actual tokens per second with real models and real quantisation — which is the only metric that matters for your workload.</p>
</div>
</div>
<div class='bg-gradient-to-br from-blue-50 to-indigo-50 rounded-2xl p-8 mb-8 border-2 border-blue-200'>
<h2 class='text-2xl font-bold text-gray-900 mb-4'>The Verdict</h2>
<div class='prose max-w-none text-gray-700'>
<p>Local LLM inference on a laptop NPU in 2026 is not a compromise — it's genuinely practical for UK professionals. The Snapdragon X Elite leads on throughput and battery life, the Ryzen AI 9 HX offers competitive performance with wider memory support for large models, and Intel Meteor Lake is the enterprise-safe pick for organisations prioritising stability over raw speed. For Mistral 7B workloads, the ThinkPad X1 Carbon AI is the outright recommendation. For Llama 3 70B, you need the 64GB RAM build of either the ThinkPad or the ASUS ProArt, and you need to accept that 10–12 tokens per second is the realistic ceiling on current NPU hardware. Cloud AI services still win on peak throughput — but they cost money every month, send your data to US servers, and stop working on the train. The NPU laptops on this list don't have those problems.</p>
</div>
</div>
Written by Neil Andrews
Founder & Lead Reviewer, Best Laptop Review UK
Software developer and DevOps engineer with 20+ years of professional experience across software development, database administration, and infrastructure. Neil has been building and repairing computers since the early 1990s and uses Linux, Windows, and macOS daily.
20+ yrs software developmentDevOps & infrastructure engineeringLinux, Windows & macOS daily userHardware builder & repair experience
Advertisement
Comments
Loading comments…