Microsoft's $2,599 Laptop Sold Out Its $5,899 Top Model

Nvidia CEO Jensen Huang shared the stage with Satya Nadella at Microsoft's Oct 7 event. RTX Spark is Nvidia's Arm chip entering Windows laptops. Photo: Peter Dasilva / European Union, CC BY 4.0, via Wikimedia Commons.

Microsoft wants you to run AI on your laptop for free. The catch: the laptop costs $2,599, the configurations that actually make the pitch work cost up to $5,899 — and the most expensive one was already sold out when they announced it.

That was the scene at Dogpatch Studios in San Francisco on Wednesday, where CEO Satya Nadella and Nvidia's Jensen Huang took the stage together during the city's Tech Week, with Sriram Krishnan moderating their fireside chat. Microsoft's big idea is "hybrid intelligence" — local models handle what they can, the cloud picks up the rest — and the hardware at the center of it is the Surface Laptop Ultra, the first flagship Windows machine built around Nvidia's RTX Spark Arm chip. Preorders opened the same day; the laptops ship October 16.

What launched on Oct 7

Windows and Surface chief Pavan Davuluri announced the pricing on stage. The Surface Laptop Ultra comes in at $2,599 for the base model (TechCrunch rounds it to $2,600), a more powerful chip configuration starts at $3,699.99, and the high-end build — 128GB of unified RAM, a 20-core CPU, 1TB of storage — tops out at $5,899.99. A companion workstation, the Surface RTX Spark Dev Box, starts at $5,999, ships with VS Code, GitHub Copilot CLI, WSL and PowerShell 7 preloaded, and lands in November.

It's not just Microsoft's silicon show. Dell had its XPS 16 Creator Edition on US preorder the same day through Best Buy at $3,799.99 for a 32GB/512GB configuration, also arriving Oct 16. Asus (ProArt P14 at $2,599.99), HP (OmniBook X 14 at $2,999.99), Lenovo, MSI and others have their own RTX Spark machines queued — some priced, some not. The same chip is suddenly the default platform for the Windows AI PC.

One detail that barely made the headlines: the highest-end Surface Laptop Ultra configuration was already out of stock at announcement, per TechCrunch. Whether that's real demand or a deliberately thin launch SKU, it's the kind of line that tells you Microsoft wants this machine to feel scarce.

The machine and the chip

RTX Spark is an Nvidia Arm-based SoC: a Blackwell RTX GPU with up to 6,144 CUDA cores and fifth-gen Tensor Cores, a Grace CPU with up to 20 cores, and up to 128GB of unified LPDDR5x memory fused over NVLink. Nvidia's claim is up to 1 petaflop of FP4 AI performance — enough, per Ars Technica, to run models up to 120B parameters locally.

The machine around it is properly flagship: a 15-inch mini-LED HDR touchscreen pushing up to 2,000 nits peak, which Microsoft says is 25% brighter than the M5 MacBook Pro's peak. A large haptic trackpad. Under 18mm thick and under 4.5 pounds. Three USB-C ports, USB-A, HDMI, a full-size SD reader, a headphone jack — and built-in magnetic USB-C charging, which Microsoft claims is a world first for a laptop, replacing the old proprietary Surface Connect port.

The comparisons are deliberate and direct. Microsoft is benchmarking against the 16-inch M5 Pro MacBook Pro: AI image generation up to 4.3x faster, AI responses 2.1x quicker, per Microsoft's own commissioned tests — not independently verified, and you should read them as marketing until someone with no stake in the outcome runs the numbers. Same with the gaming demos: Gears of War: E-Day at 60fps natively, Easy Anti-Cheat and BattlEye supported, Riot Games and KRAFTON working on native ports. This is Microsoft telling developers, in the most expensive way possible, that Windows on Arm is done apologizing.

One honest caveat buried in the spec sheets: the entry $2,599 machine doesn't get the headline silicon. It's a reduced part — 8-core CPU, 24GB of RAM, 512GB SSD. The 128GB memory figure, the 20 cores, the 1 petaflop — that all lives at the top of the stack. Windows still runs x86 apps through Prism emulation, which MacRumors confirms, so the compatibility story holds. But the "run big models locally" dream is a $5,899 dream.

The $1,000 MacBook bribe

Then there's the trade-in. Microsoft will give you up to $1,000 off for a returned MacBook Pro. TechCrunch noted the obvious: Macs are popular with developers and AI-agent enthusiasts, "particularly in Silicon Valley." This isn't a discount. It's a defection program.

Do the arithmetic and the price war gets real. Apple's first touchscreen Mac era is just beginning — and Microsoft is pricing the Surface Laptop Ultra to sit right under Apple's flagship while paying you to cross the aisle. A $2,599 machine with a $1,000 trade-in is a $1,599 machine, which is a number that changes conversations. The fine print on which MacBooks qualify isn't public yet, so treat the headline figure as the best-case, not the promise.

Microsoft CEO Satya Nadella, who led the Oct 7 keynote on hybrid intelligence and agent security. Photo: Brian Smale and Microsoft, CC BY-SA 4.0, via Wikimedia Commons.

"Free" AI has a price tag

The pitch, as TechCrunch relayed it from the event, is that these PCs are "designed to run AI models on-device locally (and for free)." That parenthetical is Microsoft's own framing, and it's doing a lot of work. The inference is free in the same way the coffee is free at a $2,599 hotel: the marginal cost is zero because you've already paid the entry fee.

There is a real economic argument underneath. Microsoft says local computing for suitable tasks helps customers "stretch their AI token budgets" — GitHub's HydraFusion routing now picks between device and cloud models depending on the job. And Microsoft's scale figures are striking: more than 2 trillion inferences a month already run locally across Copilot+ PCs, with over 40% of business-built laptops now Copilot+ machines. The company is also making Copilot more autonomous — local file context, OS-wide actions with permission — and shipping local/cloud model routing in GitHub Copilot later this month.

But "free" local AI on the base $2,599 config means running modest models on 24GB of RAM. The configs where local inference genuinely replaces cloud spend are the $5,899 ones. Microsoft is selling you the idea of escaping the token meter by handing you a five-figure invoice first. The economics can still work for heavy users — they just aren't the economics of the sticker price.

The other headline: Windows learns to cage its agents

The hardware got the photos; the more consequential news may be the plumbing. Microsoft Execution Containers (MXC) went generally available on October 7 — an OS-level sandbox for AI agents on Windows 11, built on three pillars: containment (limit what an agent touches), identity (distinguish agent activity from yours), and manageability (tools to govern and monitor).

The design principle is the good part: the policy sits outside the agent workload, so neither the agent nor the code it generates can grant itself more access. Four backends — process, session, WSL and MicroVM containers — and three operating modes: Enforcement, Learning, Permissive. Live at launch: GitHub Copilot, OpenAI Codex, Replit, LM Studio, Unsloth AI, OpenClaw and Nvidia's OpenShell. Coming soon: Anthropic's Claude Code, Nous Research's Hermes Agent, Manus, Perplexity, Raycast and others. Meta's Muse is planned as a native Windows app with MXC integration.

As Nadella put it on stage: "not just for our apps but apps for anyone's agent. That's what the Windows platform is all about going forward." And on the security framing: "We needed to make the desktop the most secure place for agents to execute" — with Huang replying, "What Satya just said is going to be the foundation of the next generation of IT as we know it."

One honest caveat, because coverage is glossing over it: MXC at GA is the containment layer only. The Entra-based agent identity separation and Intune management pieces are still ahead. "Containment today, governance tomorrow" is the real story — and it's a good one, given the week the industry has had. When tens of thousands of rogue AI incidents are on the books and hackers used Claude to break into systems, an OS that treats every agent as untrusted by default is the right instinct.

Our take: the agent PC is here before the agents deserve it

Step back and the strategy is brazen and coherent: Microsoft is building the developer desktop of the agent era — local inference, agent sandboxing, model routing — before the agents themselves are fully trustworthy. That's either visionary or a bet that containment will arrive just in time for the containment to matter. The MXC launch suggests Microsoft knows exactly which one it is: it spent its biggest Windows keynote in years talking about caging the things it's also trying to sell you.

The hardware play is the riskier half. $2,599 to $5,899 is MacBook Pro territory, and Microsoft has lost that fight before — repeatedly. What's different this time is the alliance: Nvidia's best silicon, a genuine ecosystem of OEMs, and a trade-in program aimed squarely at the developers who never left macOS. The $1,000 bribe tells you Microsoft has done the market research and knows exactly whose loyalty it's buying.

Will it work? The out-of-stock top config is either a signal or a stunt, the performance claims are vendor arithmetic, and the "free inference" pitch has a $5,899-shaped hole in it. But the platform bet — Windows as the secure substrate for agents, local and cloud stitched together — is the most interesting thing Microsoft has said about its future in years. Huang's line lands: "the computer has to be revolutionised." On Wednesday, for the first time in a long time, it looked like Microsoft was the one doing the revolutionizing.

Frequently asked questions

How much does the Surface Laptop Ultra cost?
Starts at $2,599 (base, 8-core, 24GB RAM), $3,699.99 for the higher chip config, up to $5,899.99 fully loaded. The Dev Box workstation starts at $5,999.

When does it ship?
October 16, 2026. Preorders opened Oct 7. Dell's XPS 16 Creator Edition ($3,799.99 at Best Buy) ships the same day.

What is RTX Spark?
Nvidia's Arm-based SoC: Blackwell RTX GPU, Grace CPU (up to 20 cores), up to 128GB unified memory, up to 1 petaflop of FP4 AI performance. It's designed to run large AI models locally on the device.

Is local AI really free?
The inference itself costs no tokens — but the machine does. The base config's 24GB of RAM limits what you can run locally; the configs built for serious local AI are the expensive ones.

What are Microsoft Execution Containers?
An OS-level sandbox (generally available Oct 7 on Windows 11) that limits what AI agents can touch on your machine — containment, identity and manageability, with policy the agent can't override.

Sources

TechCrunch's launch report by Julie Bort (Oct 7, 2026); The Straits Times wire report from San Francisco (Oct 7); MacRumors (Oct 7); Computer Weekly's MXC analysis by Cliff Saran (Oct 9); Notebookcheck on the Dell XPS 16 (Oct 9); GSMArena, theagenttimes, developmentstoday, WebProNews, Unite.ai and intelligibberish launch coverage. Performance figures are vendor claims, not independently verified.

By Daniel Rao · October 10, 2026

Daniel Rao covers AI funding, frontier labs and the economics of machine intelligence for TechPulse USA. He has previously reported on the Isomorphic Labs raise, the Firmus IPO withdrawal, and the decision-model benchmark wars.

Search This Blog

TechPulse USA