Ghost Core, a San Francisco startup led by a 19-year-old founder, sold out the first production run of its $3,499 screenless AI computer—a hardware debut timed to coincide with an $11 million seed round led by Andreessen Horowitz. The bet: that developers and privacy-conscious users will pay a premium for a turnkey local-inference machine rather than cobble together their own rigs or feed their data to cloud APIs.
The timing is bold, if not slightly reckless. NVIDIA rolled out its RTX Spark PC platform the same week, alongside a wave of AI-optimized Windows laptops from ASUS, Dell, HP, and others. Yet Ghost founder Zain Javaid appears unbothered by the competitive pressure, pitching Core as something different—what the company calls a "brain in a box" that runs AI models entirely on-device, no screen required, no subscription fees attached.
Core connects to a user's apps, files, and smart-home devices over Wi-Fi, processing queries locally rather than routing them through the cloud. Ghost ships four models preinstalled: Qwen 3.8-Next, Qwen 3.8-27B, Gemma 4-31B, and Muse-Glimmer-30B. Users can add others from Hugging Face or load custom-trained weights if they prefer. The machine's interface is compatible with the OpenAI Responses API, meaning third-party apps and scripts can call local models without per-token fees—a potentially significant draw for developers tired of metered cloud costs.
Ghost claims Core delivers "over 87 tokens per second" on Qwen-3.8-27B and 32 tokens per second on Qwen-3.8-Next, speeds Javaid attributes to kernel-level optimizations and multi-token prediction. No independent lab has tested throughput, thermal behavior, or power draw ahead of shipment, though the company has set an initial delivery window. Ethereum co-founder Vitalik Buterin once remarked that 20 to 30 tokens per second "feels slow" while roughly 100 tokens per second "feels fast"—a threshold Ghost claims to approach, if not quite cross, on its 27-billion-parameter model.
The hardware itself pairs an NVIDIA RTX PRO 4000 Blackwell SFF Edition GPU with an AMD Ryzen 5 7600 processor, 64 GB of DDR5 memory, and a 1 TB NVMe SSD. The GPU carries 24 GB of GDDR7 ECC memory and 432 GB/s bandwidth. Ghost rates the system at 770 AI TOPS. Pricing the RTX PRO 4000 is tricky; reports cite a launch price closer to $1,500, while current retail pricing appears to be around $3,200. Javaid wrote on X that Ghost "locked down supply of this GPU when it was widely available" at roughly $1,700 per unit a few months back and is "happy to pass on our savings to the customer." He hinted that later batches may carry higher prices if GPU supply tightens.
Privacy is the other pillar of Ghost's pitch. "All of the source code, logic, and model weights live on the device, not the services," Javaid told TechCrunch. "The user owns all of it." The company's privacy policy offers a more granular view: while enabled, Core's companion app captures screen images and recognized text—complete with app, window, time, and URL metadata—and sends them to the paired Core device for on-device search and recall. Ghost says it does not keep cloud copies of on-device conversations, files, connected-app content, or memories, nor store content forwarded through its relay. The product page shows integration icons for email, calendar, finances, files, recordings, browser, smart lights, refrigerators, speakers, cameras, health records, and wearables including Whoop, Oura, and Eight Sleep.
Whether that architecture delivers on its privacy promise is a question best answered by security researchers once units ship. For now, it's a compelling narrative in a climate where data-hungry cloud models dominate the AI landscape.
Javaid founded Ghost with Nicholas Chua, 19, Yifei Chen, 24, and Gautam Sharda, 23. The company operates as ZMJ, Inc., doing business as Ghost. Abstract Ventures, Audacious Ventures, SV Angel, and Nova joined Andreessen Horowitz in the seed round, with angels including Cory Levy, Max Mullen, Vincent Weisser, Zach Sims, and Maria Davidson. A third-party LinkedIn post claimed Javaid "spent $3.4 million of Ghost's $11 million total funding on GPUs in one call," though neither the company nor its founders have confirmed that figure. Ghost is reportedly assembling units at a San Francisco facility.
The broader market context complicates Ghost's position. NVIDIA's RTX Spark platform offers unified memory up to 128 GB and 1 petaFLOP AI performance across a range of form factors, while the company's DGX Spark 64 GB compact "personal AI supercomputer" targets local development workflows. Some analysts have positioned Core as an alternative to maxed-out Mac mini or Mac Studio builds, though the price-to-performance comparison gets murky depending on configuration.
Ghost's pitch centers on preinstalled models, no-subscription pricing, proactive agent behavior, and turnkey setup for users who find DIY local inference—in Javaid's words to TechCrunch—"a terrible experience on many levels." That's a bet on convenience over raw horsepower, and on privacy over the network effects of cloud platforms.
Core ships with a one-year warranty and free shipping for early buyers. Ghost has indicated it will push over-the-air updates for new models and improvements "similar to how Tesla works," Javaid said. A refundable deposit of $499 was reportedly available to reserve interest in the second production batch.
"The most powerful technology humanity has built should belong to the people whose lives it changes," Javaid wrote on LinkedIn when the funding round was announced. It's the kind of sweeping declaration that sounds either visionary or naive depending on whether the hardware delivers. For a 19-year-old founder with a sold-out first run and $11 million in the bank, the answer will come soon enough.
