Surface Laptop Ultra: buy now or wait?

Microsoft's Surface Laptop Ultra pairs Windows with an NVIDIA RTX Spark chip, up to 128GB of unified memory, and full CUDA support inside a 15-inch touchscreen laptop under 2kg. No price. No confirmed release date beyond later 2026. No refurbished stock for years, since that only builds up once real units are out in the world.

The honest answer for most people: buy a proven refurbished machine now. Wait only if a narrow set of needs, such as huge local AI models, CUDA, Windows, touch, and portability all at once, applies to how you work.

Microsoft's own preview and product pages confirm the hardware: NVIDIA RTX Spark graphics with full CUDA support, up to 128GB of unified memory, and a claimed one petaflop of sparse FP4 compute aimed at local AI models up to 120 billion parameters. The screen is a 15-inch, 3:2 mini-LED touchscreen at 262 pixels per inch and up to 2,000-nit HDR peak, in a body under 18mm thick and under 2kg. Ports cover USB-C, USB-A, HDMI, a full-size SD reader, and a headphone jack, alongside a user-serviceable SSD and a haptic touchpad.

What's still missing: price, exact memory and storage tiers, battery capacity, port speeds, and regional service terms. Availability is later in 2026, subject to certification and region, with no fixed date.

Strip away the platform story and the decision comes down to five checks.

Capability. Does the job need 70 to 120 billion parameter local models, or a shared CUDA and memory pool bigger than a normal laptop GPU offers? Most work doesn't.

Compatibility. Windows on Arm emulates most software, but kernel-mode drivers, VPN clients, capture cards, and specialist audio or CAD tools often need native Arm versions that may not exist yet. Check every driver and plug-in the workflow depends on, not just whether the main application launches.

Governance. Large local memory can run more AI agents for longer, which raises the stakes if one misbehaves: reading another agent's files, following instructions hidden inside a document, or continuing to run after network access should have been cut. Microsoft's containment and identity controls for that are still rolling out, not fully proven.

Economics. With no price announced, there's nothing to weigh yet against a workstation, a refurbished laptop, or renting cloud GPU time for occasional jobs.

Service. A replaceable SSD is a promise, not a finished service network. Confirm parts and repair turnaround exist in your country before counting on it.

Two or more unclear answers here mean the safer move is to buy something proven now instead of waiting on all five.

Choose the machine that fits the work. If your tools already run on macOS and work with Metal, buy a high-memory Mac Studio 2025 M4 Max now. For Windows and a large shared memory pool without a CUDA requirement, the HP ZBook Ultra 14 G1a comes closest to the Surface Laptop Ultra's promise without the first-generation risk. If certified Windows applications, VPN clients or specialist drivers are non-negotiable, choose the mature x86 CUDA Dell Precision 7680. Driver maturity matters more than a new platform's promise.

Working with models under roughly 27 billion parameters? You need none of those three. Buy a mature Windows laptop instead; a modern RTX laptop handles that work comfortably.

Wait and pilot the Surface Laptop Ultra only if you regularly need to run 70 to 120 billion parameter local models on Windows, with data that can't leave the device. Audit every driver and application, benchmark real workloads on a few units, then integrate them with existing device management before any wider rollout.

Just want the most powerful Surface? Without a workload behind it, wait for price and evidence rather than paying a platform premium. For a similar buy-or-wait decision on a very different machine, see the Acer Aspire Go 15.

Surface devices already fit established Autopilot, Intune, Entra, firmware and service workflows. That's a head start, not proof that the Surface Laptop Ultra will arrive ready for your organisation. Validate the exact retail configuration before a wider rollout.

Autopilot: On the real device, test hardware-hash registration, zero-touch deployment, the right user role and enrolment timing. Don't rely on results from an earlier Surface.

Intune and Entra: Confirm policy application, compliance state, certificates and conditional access before the device receives corporate data.

Firmware and drivers: Verify staged rollout, rollback, reporting and recovery media for this exact model.

Data protection: Test full-disk encryption, key escrow and a working secure-erase process for the removable SSD. A swappable drive changes how you reassign a device.

Local AI agents: Identity, containment, policy and audit logs on the device are the newest and least-proven part of the stack. Treat them as unconfirmed until Microsoft's documentation for this build says otherwise.

A small pilot fleet can answer these questions; a broad refresh before they're answered cannot.

Every option above assumes regular use. If the 70 to 120 billion parameter job comes up rarely, renting cloud GPU time can cost less than owning hardware that sits idle most days.

Independently published cloud pricing (Runpod, August 2026) runs from around $1.39 an hour for an NVIDIA A100 with 80GB of memory, up to $1.99 for an RTX Pro 6000 with 96GB, and $4.39 for an H200 with 141GB. Storage keeps billing even while the compute is switched off, so that adds up for anything left sitting between sessions.

Buying refurbished instead of a new, unproven platform frees up real money for exactly this. A few hundred pounds saved on the daily machine covers hundreds of hours of rented GPU time, for the rare week that needs it.

Manufacturing any new laptop carries a real cost before it ever gets switched on: mining raw materials, manufacturing components, and shipping the finished device around the world. A refurbished purchase skips that step entirely by extending a machine that's already in circulation.

Buying refurbished instead of new saves, on average, around 355kg of CO2 and roughly 90,000 litres of water per laptop, weighted across real sales data for refurbished laptops sold on refurbed. Larger workstation-class machines, like the alternatives above, tend to sit toward the higher end of that range.

It isn't a reason to buy on its own. It's a real number attached to a real choice: the same class of machine, tested and warrantied, with a smaller footprint attached.

A big unified memory number is not the same as fast responses. Whether a large local model feels usable in practice depends on memory bandwidth, how the model is compressed, how much context it's holding, and whether the software drivers are mature enough to use the hardware well.

Independent memory-bandwidth testing across this class of hardware found no single architecture wins every time: strong raw bandwidth doesn't automatically beat a well-optimised software stack, and the reverse holds too. A model fitting in memory is the first check, not proof it runs at a usable speed.

Ask for load time, prompt processing speed, and sustained speed after the laptop heats up, not just the maximum model size a spec sheet claims.

How much does the Surface Laptop Ultra cost?

Unannounced. Microsoft has confirmed the hardware but not pricing, exact memory and storage tiers, or a fixed release date, only later in 2026 subject to certification and region.

When can I buy one?

Not yet, and there's no fixed date. Microsoft's own timeline is later in 2026, subject to certification and regional availability, which can slip.

Will refurbed sell a refurbished Surface Laptop Ultra?

Not for a long while. Refurbished stock builds up from real trade-ins and returns, which takes years for a brand-new platform. Browse current refurbished Microsoft laptops instead.

Is 128GB of memory overkill for most people?

Yes. Unless the work involves large local AI models or professional CUDA workloads, that much unified memory goes mostly unused.

What should I buy right now instead?

It depends on the workload: a high-memory Mac for macOS AI work, a Windows laptop with a large shared memory pool for a similar setup without CUDA, or an x86 workstation for certified apps and conventional CUDA. All three are covered above, in stock, and refurbished.