Local AI Hardware Just Got Way More Expensive Than You Think

The DGX Spark went from $3,999 to $6,950 in a year and RAM prices went through the roof, but the model it was built to run got 57% cheaper to rent. This is the local AI hardware math almost nobody is doing. A year ago I had a spreadsheet that said buy the hardware and own your AI. I didn't buy, and that turned out to be the smartest thing I did all year. Everyone is comparing the new DGX Spark price to the old one. The real comparison is the box against the token it replaces. In October 2025, gpt-oss-120b cost 40 cents per million output tokens on OpenRouter. Today it's about 17 cents. At the 58.72 tokens per second llama.cpp measures on a Spark, that moves the payback from about five and a half years of nonstop use to over 22. The chip didn't change. The memory did. A 64GB DDR5 kit went from $191 to $1,118 in a year. The worst 128GB kits went from $329 to $3,399. Framework's 128GB Strix Halo desktop went from $1,999 to $2,459, and Framework blamed the memory. A used RTX 3090 with 24GB sells for about three times what a 10GB 3080 does. Whether you're weighing DGX Spark vs Mac Studio, a Strix Halo mini PC, or a 128GB build of your own, you're mostly paying for RAM in the middle of a shortage. So why didn't cloud token prices go up too? Batching. A datacenter H100 serves dozens of people from every trip through memory, while the box on your desk serves one. Even when a vendor raises prices, open weights keep it honest: when DeepSeek's V4.1 Flash came in at $1.20 per million at peak, dozens of other endpoints kept serving the same weights for far less. There's still one place where local AI hardware wins: dense models. Qwen3.8-27B costs at least $1.25 per million tokens hosted, and an RTX 3090 runs it locally at around 66 tokens per second. I also cover the Strata workaround (and why it moves the bill onto RAM), who should still buy hardware anyway, and a 5-minute check to run before you spend anything. CHAPTERS 0:00 I almost bought one 1:21 $3,999 to $6,950 in one year 2:35 The comparison everyone gets wrong 3:28 Payback went from 5.5 years to 22 4:34 Memory, not the chip 6:09 Why tokens got cheaper (batching) 7:51 What if the cloud raises prices? 9:10 What about Strata? 10:30 Where an RTX 3090 still wins 11:51 Who should buy anyway 13:01 Do this before you buy SOURCES - The Register: 128GB DGX Spark jumps to $6,950: https://www.theregister.com/systems/2026/10/02/nvidia-debuts-4999-dgx-spark-with-half-the-ram-and-storage-amid-memory-crunch/5300622 - NVIDIA DGX Spark hardware specs: https://docs.nvidia.com/dgx/dgx-spark/hardware.html - llama.cpp DGX Spark benchmarks: https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md - OpenRouter gpt-oss-120b (today): https://openrouter.ai/openai/gpt-oss-120b - OpenRouter gpt-oss-120b (October 2025, Internet Archive): https://web.archive.org/web/20251008144428/https://openrouter.ai/openai/gpt-oss-120b - Hardware Busters: DDR5 up 500% in a year: https://hwbusters.com/news/ddr5-memory-prices-are-up-500-in-a-year-and-a-128gb-kit-now-lists-at-3399/ - PCWorld: Framework raises prices as memory costs go up: https://www.pcworld.com/article/3031562/framework-raises-prices-yet-again-as-memory-costs-go-up.html - SemiAnalysis InferenceX, gpt-oss 120B on H100: https://inferencex.semianalysis.com/compare-per-dollar/gptoss-120b-h100-vs-mi300x - DeepSeek V4.1 Flash release: https://api-docs.deepseek.com/news/news260910 - DeepSeek API pricing: https://api-docs.deepseek.com/quick_start/pricing - OpenRouter DeepSeek V4.1 Flash endpoints: https://openrouter.ai/deepseek/deepseek-v4.1-flash - OpenRouter Qwen3.8-27B endpoints: https://openrouter.ai/qwen/qwen3.8-27b - Qwen3.8-27B on one RTX 3090 with MTP: https://hackmd.io/@thc1006/qwen3-8-27b-speculative-decoding-rtx-3090 - Strata: https://github.com/Niko1221/Strata Do the math yourself on OpenRouter's model pages: https://openrouter.ai/models Payback figures assume nonstop generation and exclude electricity, which only makes local look better than it is. Used-card prices are individual listings and eBay medians, not averages. #DGXSpark #LocalAI #LocalLLM