In one line: the invite narrows the theme to how local AI will shape the next chapter of the PC, but the useful question is which specs already belong in a comparison table and which lines are still press inference. The sections below split three threads: the session and attendees that are already confirmed, the Surface price and ship details that are still missing, and the build layer that on-device compute does not replace.
What is already confirmed
Around September 15, 2026, The Verge, Windows Central, 9to5Windows, and Windows Latest reported the same invite: October 7 in San Francisco, 10 a.m. Pacific, 1 p.m. Eastern. October 7 is still inside daylight saving time, so Pacific time is 15 hours behind Beijing. For teams in China that is 1 a.m. on October 8. There is no reason to stay up for an agenda that is not public yet. Write the question list during the day on October 7.
The line reporters quote from the invite is a conversation on “how local AI will shape the next chapter of the PC.” Three names show up consistently: Microsoft CEO Satya Nadella, Windows and Surface chief Pavan Davuluri, and NVIDIA CEO Jensen Huang. Huang on stage is why several outlets put RTX Spark at the center. That is an inference, not an agenda item Microsoft has published.
The gap since the last event is what tempts people to assume a new version number. The previous major Windows event was Copilot+ PCs in May 2024, built around Arm machines and the Qualcomm partnership. Before that, Windows 11 itself was the June 2021 livestream. The invite does not say this October session is a new operating system.
Do not paste a livestream URL into the calendar yet
As of September 29, 2026, the coverage above did not include an official stream link. It pointed readers at Microsoft’s own channels. This article does not invent one. If a URL appears before October 7, use that page.
Surface hardware that has been shown, not priced
The hardware October 7 is most likely to clarify is not a machine nobody has seen. It is the one Surface already introduced on May 31, 2026, and that still cannot be ordered. Surface corporate vice president Brett Ostrum called the Surface Laptop Ultra “the most powerful Surface Laptop ever built.” The figures that repeat across coverage are an NVIDIA Blackwell RTX GPU on the RTX Spark platform, up to 128 GB of unified memory, full CUDA support, and the ability to run models of up to about 120 billion parameters on the device. Microsoft’s stated window was fall 2026. October 7 sits inside that window.
The other machine is the Surface RTX Spark Dev Box, a compact desktop for people building local models, aimed at the same chip direction as the laptop. Ports, memory steps, power, and whether it shares the laptop’s software stack are not in a spec sheet anyone can cite yet.
| Item | Safe to write down on September 29 | Still waiting on October 7 |
|---|---|---|
| Surface Laptop Ultra | Introduced in May; Blackwell RTX / RTX Spark; up to 128 GB unified memory; CUDA; about 120B parameters on device; fall availability promised | Price, exact on-sale date, which memory steps you can buy, whether regulatory approval is done |
| RTX Spark Dev Box | Compact desktop for local-model work, same chip direction as the laptop | Full specs, regions, and whether pricing logic matches the laptop |
| Other OEMs | Dell, Lenovo, ASUS and others are also reported to be preparing RTX Spark laptops | Whether they share the stage or stay a footnote |
Some reports also mention up to about 1 petaflop of local AI compute, and that the laptop is still pre-release pending regulatory approval. Those two points do not appear as consistently as 128 GB and 120 billion parameters, so they stay in the “listen for whether the keynote repeats them” pile, not in the confirmed column. A dollar figure invented before the event becomes a budget rewrite on October 8, not a news update.
Windows AI: keep guesses out of the changelog
Windows on the invite is not Windows 12. 9to5Windows and Windows Latest both said not to wait for a new version number. Their basis is that when speculation rose in May 2026, Microsoft had already said Windows 12 was not on this hardware line, and the October theme is still local AI and the PC. Tom Warren’s expectation at The Verge is quality work and “what comes next,” not a surprise generation change.
Keep reported wishes separate from things you can already click in Windows 11:
- Changes you can already compare. Windows Central pointed at the Windows 11 K2 quality effort, including a taskbar you can resize and move, as an early visible fix. That is the current system, not an operating system that installs only on October 7.
- Still inference. Workload scheduling for RTX Spark, spreading background tasks across about 20 CPU cores, Prism emulation tweaks, and a developer-device class codenamed Project Zenith. Those lines live in expectation sections, not in the invite quote. If the stage does not say them, they are not shipped.
- How this sits next to Copilot+ PCs. May 2024 tied NPUs and on-device features to Arm laptops. Huang’s appearance points more at NVIDIA’s local CUDA path than at a rerun of the Qualcomm story. The two local-AI paths can coexist. Do not collapse them into “Windows AI means one kind of PC” before the event.
The feature list is whatever is said and documented that day
This article was published on September 29. If October 7 brings a new Windows API, a new Copilot entry point, or a new hardware badge, use Microsoft’s announcement and docs. Do not treat this preview as a changelog.
What local compute does and does not settle
“120 billion parameters on the device” means something concrete: weights, KV cache, and context do not have to leave that machine on every request. 128 GB of unified memory is the hardware condition for that claim. CUDA decides that the software stack is NVIDIA’s, not a handful of system features on an NPU. The workloads worth comparing are long-context local coding assistants, batched GPU inference, and experiments that should not ship a repo snapshot to an external API.
Three things it does not settle:
- A memory headline is not your current model. 120 billion parameters is Microsoft’s public ceiling for the machine, not a measurement of the quantized file you already run. Quantization, context length, and concurrency eat the margin. Write down today’s model size before the slogan. For how coding-assistant memory actually steps from 16 GB to 64 GB, see Copilot App hardware at 16, 32, and 64 GB.
- A Windows GPU does not sign an iOS app. Xcode, signing, the simulator, and App Store upload still need Apple’s environment. A stronger Surface is still a Windows inference and development machine. Teams already building iOS from Windows do not get a new boundary from this event. See which iOS steps still need a Mac.
- Local compute is not automatically a shared API. Running a model on a laptop and opening that port to a team network are different operations. If you also compare against MLX on a Mac, auth and exposure still need their own design. See how to close an MLX-LM local API.
Do not merge two machines into one budget line
RTX Spark answers whether a Windows desk needs a CUDA workstation for local models. iOS builds answer where signing and archives run. The two budgets can be discussed the same week. They do not cancel each other. A remote Mac can stay the build node, separate from the Windows model machine that does not have a price yet.
A checklist before the keynote
The keynote is 1 a.m. China Standard Time on October 8. Put these four lines in the team doc during the day so a price or a new API has something to sit next to.
- The local model you already run. Parameter count, quantization, whether it needs CUDA, and how much memory that machine has.
- Whether you are missing a price or a capability. If 128 GB and CUDA already cover the experiment, October 7 only decides when to buy. If the load is still 7B to 14B, the event is more likely a watch item than a replacement.
- Do not pre-write a Windows 12 migration. Until a version ships, app compatibility, drivers, and images have no upgrade ticket.
- Leave iOS steps on a Mac. Builds, signing, and the simulator do not belong in “we’ll move it when the Surface arrives.”
# 1. On-device model: size / quantization / CUDA or not # 2. Current RAM, and whether 128 GB actually changes the queue # 3. Windows features: only what is installed, not rumors # 4. iOS builds still on a separate macOS node runs-on: [self-hosted, macOS, cloud-mac]
Questions
What time does it start? October 7, 2026, San Francisco, 10 a.m. PT, 1 p.m. ET. That is 1 a.m. on October 8 in China Standard Time.
Who is on stage? Coverage confirms Satya Nadella, Pavan Davuluri, and Jensen Huang. Running order, and whether other OEMs join them, is not in the public invite reporting.
Is this Windows 12? The public theme is local AI and the next chapter of the PC. Windows reporters wrote that Microsoft took Windows 12 off this hardware line in May, and they do not expect a reversal. Until there is an announcement, do not treat it as scheduled.
Can you buy the Surface Laptop Ultra now? Not at a published retail price. It was introduced on May 31 with a fall window. Price and the exact date are still open. Those are the items October 7 is for.
Does 120 billion parameters on device retire cloud models? No. It is the public ceiling for this Surface, and neither price nor supply is set. External APIs, shared inference, and any step that has to finish on macOS stay where they are.
- Building iOS on Windows: the steps that still need a Mac
- How local coding-assistant memory moves from 16 GB to 64 GB
- Deploying and closing a local model API on a Mac
ZavCloud Cloud Mac
iOS builds do not have to wait for a Windows AI box
Surface and RTX Spark are about local models and CUDA on Windows. Xcode, signing, and the simulator still need macOS. A dedicated Mac mini cloud host can stay the build node, off the same purchase order as a Windows workstation that does not have a price yet.
View Cloud Mac plans