What hardware does your PC need to run a local AI model?

A 16GB PC can run smaller local AI models, but models approaching cloud-chatbot capabilities generally need at least 32GB of memory or a graphics card with 16GB or more of VRAM. Performance depends heavily on memory bandwidth: cited tests range from 4.42 tokens per second for a 70B model on Nvidia's DGX Spark to 205 tokens per second for gpt-oss-20b on an RTX 5090. A 32GB DDR5 kit costs $434, while frontier-scale models may require hundreds of gigabytes to more than 1TB of memory.

Analysis

The economic signal is a bifurcation, not broad “AI PC” demand: inexpensive local models are adequate for occasional private/offline tasks, while workloads demanding frontier quality remain structurally dependent on data-center compute. That limits the near-term revenue pool for consumer upgrades; the cost of extra memory and a capable GPU competes directly with recurring cloud subscriptions. If memory remains expensive, the cloud value proposition improves rather than prompting a mass upgrade cycle.

The second-order beneficiaries are cloud model providers—Microsoft, Alphabet, and Anthropic—whose capability gap supports paid inference even as open models improve. Conversely, open-weight models can pressure inference pricing for routine tasks and strengthen Meta and Alibaba’s ecosystems, without proving material monetization. NVIDIA is positioned to capture high-bandwidth GPU demand, but this article does not establish incremental unit demand. Intel’s NPU example supports a potential edge-efficiency use case, not evidence of broad customer adoption; validate with product mix and OEM sell-through.

Near term, expect limited market impact absent shipment or usage data. Over 1–3 months, watch GPU/DRAM pricing, PC channel inventory, and cloud AI usage/pricing commentary. Over 6–18 months, better quantization and inference software could make capable local models viable on cheaper hardware, expanding edge use while commoditizing simple cloud workloads. The counterweight is that model-size growth and quality expectations keep frontier workloads centralized. The thesis weakens if local-model adoption produces sustained PC/GPU demand or if cloud providers report slowing paid inference; it strengthens if memory costs stay high and cloud usage monetizes despite open-model competition.

Key Decisions for Investors

• No directional trade from this article alone: it provides no shipment, revenue, or adoption evidence sufficient to underwrite a hardware-demand revision.
• Prefer cloud exposure over a consumer-AI-PC upgrade thesis while frontier quality remains centralized; reassess Microsoft, Alphabet, and Anthropic positioning against paid inference growth and pricing commentary at the next earnings cycle.
• Treat NVIDIA as a watch rather than an article-driven buy. Add only if GPU demand and channel data confirm incremental local-inference purchases; falsify on weakening GPU sell-through or persistent inventory accumulation.
• Monitor DRAM/VRAM prices, PC shipment mix, and cloud inference pricing over the next 1–3 months. A sustained fall in memory costs plus rising capable-PC adoption would shift the balance toward edge hardware; rising costs and stable cloud usage favor cloud providers.