Developers who run AI agents throughout the day can accumulate substantial usage-based API charges. Ghost is betting that some will prefer paying $3,499 upfront for a computer that processes AI requests locally, potentially reducing their dependence on cloud services.
The startup emerged from stealth on October 5 with $11 million in seed funding led by Andreessen Horowitz. Its new product, Core, is a screenless personal AI computer designed to run language models and agents on dedicated hardware. Whether it can replace existing AI subscriptions depends on how much paid cloud usage users can transfer to local models.
What does $3,499 buy?
Core runs on an NVIDIA RTX PRO 4000 Blackwell SFF Edition GPU with 24GB of dedicated GDDR7 memory. It also includes an AMD Ryzen 5 7600 processor and 64GB of DDR5 system RAM, with a 1TB SSD for local storage. The GPU memory determines which models can run entirely on the graphics card, while larger models may require compromises in precision or processing speed.
According to Ghost’s setup guide, users pair Core with a computer through a companion app on their local network. From there, they can select an AI model and connect information sources such as email or documents. Ghost says its software uses this information to build persistent memory, so agents can recall previous interactions and carry out tasks using connected accounts.
For developers, compatibility with the OpenAI Responses API could allow existing applications to send requests to local models instead of cloud providers. However, API compatibility does not guarantee support for every function or application.
How fast is Ghost Core’s local AI?
Ghost claims Core generates more than 87 tokens per second using Qwen 3.8-27B, a 27-billion-parameter language model. Tokens are units of text, so the figure indicates how quickly the model produces an answer after generation begins.
The reported benchmark has yet to receive independent confirmation. Its published description also lacks important testing details, including model quantization and prompt length, both of which can affect performance.
Output speed measures only part of an agent’s work. A task that requires several reasoning steps or external tool calls may take considerably longer than a simple text response. Independent evaluations would need to measure task completion and accuracy to establish how Core compares with cloud-hosted alternatives.
Can local AI save money?
Local inference eliminates per-token API charges for requests processed entirely on Core. For developers who regularly pay for high-volume model usage, those avoided charges could eventually recover the computer’s purchase price.
Consider two hypothetical scenarios. If Core replaced $200 in monthly API charges, recovering the $3,499 purchase would take approximately 18 months. At $100 per month, the payback period would extend to roughly 35 months. Neither estimate includes electricity or maintenance, and both assume Core can replace the full amount of paid usage.
Fixed-price subscriptions present a different calculation. Canceling a hypothetical $20-per-month chatbot subscription would save $240 annually, so recovering Core’s purchase price would take nearly 15 years before operating expenses. Users who continue paying for the same cloud plan would see no reduction in that subscription bill, even if they processed some requests locally.
Model performance further complicates the economics. A local model may be sufficient for document summaries but require additional corrections when debugging complex code or completing multi-step analysis. If users must send difficult requests back to paid cloud models, their actual savings will be lower.
For frequent API users, Core could offer an alternative to recurring inference charges. For occasional chatbot users, the initial hardware expense may outweigh years of subscription payments.
Local AI offers privacy benefits, with limits
Running models locally can reduce the amount of personal information transmitted to external AI providers. According to Ghost’s privacy policy, conversations and memory processing remain on Core, and the company does not use remote AI models as a fallback for those functions.
However, the policy distinguishes local inference from online features that require external services. Web searches and phone calls, for example, can send task information through Ghost’s gateway to third-party providers. Ghost says it does not retain forwarded content, although recipients may process or store information under their own policies.
Core’s persistent memory also introduces security responsibilities. Imported browser sessions can provide access to connected accounts, while copies saved on companion devices may enter cloud backups. Ghost’s optional screen-history feature can capture information from open applications, including private browsing sessions when the feature is enabled.
Disconnecting an account stops future collection through that connection but does not automatically delete memories already created from its content. Users handling confidential data would need to review account permissions and stored memories, especially when agents have access to sensitive documents.
How does Ghost compare with other local AI computers?
Ghost enters a market where developers already have alternatives for running AI models locally. NVIDIA’s DGX Spark is designed for AI development and autonomous agents, with configurations offering up to 128GB of unified memory. Unlike Core’s separate system RAM and 24GB of GPU memory, DGX Spark’s shared memory architecture can accommodate larger model workloads. However, memory capacity alone does not establish which computer performs faster.
Existing laptops and desktop computers can also run local models, as The Neuron’s guide to local AI explains. Ghost therefore faces competition from hardware that potential customers may already own.
Its proposed advantage lies partly in the software. Preinstalled models and a companion app could reduce the technical work required to configure local AI. Persistent memory and connected-account access may also appeal to people who want an assistant that can use information from several applications. How consistently those agents complete tasks will be important to evaluating the overall product.
Early demand suggests interest in the concept. Founderland reports that Core’s first batch sold out, while Ghost’s product page lists October 31, 2026, as its scheduled shipping date. Reservations indicate willingness to purchase, although customer experiences and sustained demand remain unproven.
Core’s launch points to a developing category of personal AI computers designed around persistent agents rather than conventional desktop applications. If dedicated hardware makes local models easier to use, more developers may choose to process recurring workloads on devices they own.
The larger question is how many users need a separate AI computer when cloud services and existing PCs can already perform similar tasks. Ghost’s commercial prospects will depend on whether its software provides enough additional value to justify the hardware expense, especially for users with low AI expenses.
