What Bott actually built
Bott describes a deliberately ordinary open-source stack:
- Inference runtime: llama.cpp.
- Model: the open-weights Qwen3.8-27B.
- Coding agent: Cline, running as an extension in Visual Studio Code.
He says Copilot helped him configure the environment, including PowerShell scripts that got the project running within about half an hour. He estimates that doing it by hand would have taken weeks of reading and trial and error. After that he wrote a detailed spec, handed it to Cline, and mostly watched. The agent wrote code, ran tests, read the results and revised, and occasionally asked him to approve a step. He calls this "human-in-the-loop agentic coding".
The task was to turn a two-year-old spreadsheet of links to his ZDNET articles into a searchable archive. According to Bott:
- The agent retrieved nearly every listed article.
- It wrote a separate routine to pull about 90 dead links (404 errors) from the Internet Archive's Wayback Machine.
- It then listed roughly another 90 pages it could not recover. He spent about 30 minutes resolving those.
- The final output was two folders of 1,951 files each, one in raw HTML and one in Markdown.
- It also built a script for adding new posts to the archive.
He also reports that the fan was audible from across the room but the machine never got uncomfortably hot. He still recommends a desk over a lap for this kind of load.
Caveat: nobody has independently audited the archive's completeness, the code quality, or how much human correction it needed. Treat these as one writer's observations.
The "$0.00" claim, read carefully
Bott reports 479 inference calls and nearly 50 million logical tokens, for a net cost of zero. That is zero in per-token charges. It is not zero total cost:
- The hardware is the cost. Bott calls his review unit a $7,000 notebook.
- Electricity, setup time and maintenance still count.
- Copilot assisted with setup, so the workflow was not entirely free of cloud help.
His argument against hosted tools is that agentic coding makes token consumption hard to predict. Context size, iteration counts and model choice all change the bill. He puts entry-level plans for ChatGPT Plus with Codex, Claude Code Pro and GitHub Copilot at $10 to $20 a month, and higher-limit tiers around $100. He says he would have hit rate limits on the cheap plans. He also quotes analyst Joey Brookhart on the Decoder podcast saying the top 5 percent of Claude Code users spend about $18,000 a year on average.
The piece does not price his actual workload against any provider. Plan names, prices and quotas change often, so the comparison is directional rather than a calculated bill.
Privacy: real, but conditional
Bott's bigger argument is privacy. With a fully local stack, prompts, documents and inference need not leave the PC. His archive project used already-published pages, so privacy didn't matter much there. He says he would feed financial documents, card statements and invoices to a local model but not to a cloud coding service.
That benefit depends on how the stack is configured. Before pointing a local agent at sensitive files, check:
- That the model runtime is actually serving locally and not forwarding to a hosted endpoint.
- What each connected extension or tool sends over the network.
- Whether any retrieval or web-fetch routine reaches outside the machine.
- Whether cloud assistance, such as Copilot in setup, touches the data.
Nvidia's launch announcement points the same way. It says locally processed code and context stay on the device. It also describes OpenShell and Microsoft Execution Containers as ways to control agent permissions.
The hardware, and a price discrepancy
Nvidia describes RTX Spark as a superchip pairing a Blackwell RTX GPU with a Grace CPU. HP's 16-inch Ultra has a 2880x1800 120Hz OLED display and three memory configurations from 32GB to 128GB. The 128 GB and 1 petaflop FP4 figures are "up to" platform ceilings, not what every configuration delivers.
On pricing, Bott's $7,000 figure doesn't match other published information:
- Tom's Guide says the Ultra 16 starts at $3,199 at Best Buy with 32GB of RAM and a 1TB SSD.
- VideoCardz reported HP US listings of $3,199.99 for the Ultra 16 base model and $4,999.99 for a 64GB version.
- HotHardware lists $3,199 for the 32GB model and $6,229 for its tested configuration. That is the top-end unit with 128GB of LPDDR5X.
- Windows Central cited a $2,799 sale price for the 32GB model.
- Techaeris noted that no HP price was published for the 128 GB custom configuration.
I could not find a published SKU that confirms $7,000, so read it as Bott's own figure. The storage label is similarly unconfirmed. Bott says Gen5 NVMe, and a retailer-tracking source lists a 1TB PCIe Gen5 SSD on the base model, so that part is plausible. Nvidia's range for RTX Spark laptops starts at $2,599, with pricing varying by region and configuration.
Memory matters more than the logo
Other reviewers stress that RAM is the deciding factor for local AI:
- How-To Geek says 32GB leaves almost nothing for local AI and recommends the 64GB or 128GB versions.
- Notebookcheck notes the 32 GB entry model has non-upgradeable memory shared between the OS, user tasks and any local models.
- Notebookcheck also says buyers who don't prioritize AI may find significantly more powerful x64 laptops for the same price.
Bott's result came from a 128 GB machine running a 27B-parameter model. It should not be extrapolated to the $3,199 base model. Unified memory is soldered, so you have to pick the right configuration at purchase.
Who this is for
My read, from this reporting and general industry knowledge:
- Good fit: developers who run long agent sessions, have sensitive code or documents, and want predictable costs.
- Poor fit: general office users, and anyone expecting a gaming or all-round performance buy. HotHardware lists a price premium, mediocre battery life, and underwhelming web and office app performance as downsides.
- Worth watching: Microsoft's "hybrid intelligence" framing, with AI running locally when it makes sense and in the cloud when needed. The pitch is hybrid, not a wholesale replacement for cloud services.
Bott's own conclusion is cautious. His accountant would not let him justify the purchase, and he expects configurations like this to become mainstream business machines in about three years. That is a forecast, not a finding.
Bottom line
This is a credible, concrete demonstration that a 27B-class open model can drive an agentic coding workflow on a Windows laptop with no per-token billing. The privacy upside is real but depends on how you set the stack up. The "free" part holds only if you ignore a hardware bill in the thousands. If you plan to try this, size the memory first and audit the data flows second.