Francesco Di Costanzo
Back to shorts

(7) The Price of Private Intelligence

/5 min read

Intelligence Now Has a Purchase Price

For under ten thousand dollars of hardware, a single person can own and run a model that scores 52 on the Artificial Analysis Intelligence Index, against a frontier peak of 61 held by Claude Opus 5. Two NVIDIA DGX Spark units cost $7,998 to $9,398 at current pricing, following NVIDIA's increase from a $3,999 launch price to $4,699 per unit in February 2026, and that pair reportedly serves DeepSeek-V4-Flash-0731 at around 41 tokens per second across its full one-million-token context. The weights are MIT licensed. Nothing about that arrangement depends on a vendor's continued goodwill, pricing decisions, or terms of service.

The reason it fits on a desk is the precision format rather than any breakthrough in compression. DeepSeek ships Flash in mixed FP4 and FP8, which puts the full checkpoint at roughly 146GB, and its Mixture-of-Experts design activates only 13 billion of its 284 billion parameters per token. Two units of unified memory are enough. The interesting question has moved from whether frontier-adjacent intelligence can run privately to how much of it you are willing to buy.

The Gap Closed Through Training, Not Silicon

DeepSeek's model card for the 31 July release is explicit that 0731 "keeps the same model architecture and size" as the April preview and "was only re-post-trained." Identical 284 billion total parameters, identical 13 billion active per token, identical one-million-token context. On DeepSeek's own reported figures, Terminal Bench 2.1 rose from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4, both now ahead of the far larger V4-Pro-Preview that activates 49 billion parameters per token.

That matters for anyone deciding what hardware to buy. When capability arrives through post-training rather than parameter growth, the machine you buy this year keeps getting better without getting bigger, because the next checkpoint lands in the same memory footprint. The honest caveat is that DeepSeek's agentic scores come from an evaluation suite it has described but not released, and the one independent re-run, Artificial Analysis on Terminal Bench 2.1, came in at 79 against the claimed 82.7. DeepSeek's own April technical report puts its reasoning three to six months behind state of the art. A lag measured in months is a different proposition from a gap measured in tiers.

Privacy Is What You Are Actually Buying

The cost argument for local inference is the weakest one available. DeepSeek serves the same model at $0.14 per million cache-miss input tokens and $0.28 per million output, and independent modelling puts hardware breakeven against cloud APIs at three to six months only under sustained heavy load, stretching to years at ordinary developer volumes. On pure economics, almost everyone should rent.

What the purchase buys is a boundary. Nothing crosses the network, so there is no processor agreement to negotiate, no cross-border transfer to justify, no retention policy to audit at a third party, and no prospect of a prompt containing customer data reaching someone else's training corpus. For a regulated lender or a healthcare provider, that boundary is the difference between a model that can be pointed at real customer records and one confined to synthetic data. The token price is a rounding error next to the compliance work it removes, and DeepSeek has already said its API pricing will rise, which is precisely the kind of decision owned weights make irrelevant.

The Threshold Keeps Falling

The number worth tracking is the hardware price at which useful private intelligence becomes available, rather than the benchmark score at the top of the market. It currently sits at roughly two desktop units for a model within single digits of the frontier. The local runtimes have not caught up either: stock llama.cpp on Blackwell cards reaches only 36 tokens per second at 30 to 40 percent GPU utilisation, because Flash Attention is not yet implemented for this model's attention graph. That headroom is free performance waiting on software. When the next open checkpoint arrives in the same footprint with better post-training and a matured runtime beneath it, the machine already on the desk collects the upgrade at no cost.

What local ownership does not buy is speed. Forty-one tokens per second is comfortably below hosted inference, which pushes the sensible use toward background and batch work rather than interactive chat. DeepSeek appears to expect exactly that, having shipped the 0731 checkpoint with a speculative-decoding module attached and native Responses API support for agent frameworks. An agent that runs unattended overnight against a million tokens of your own documents does not care about latency, and it is the one workload where owning the machine, the weights and the data at once actually pays.

Sources

Model and Benchmark Sources

  1. DeepSeek, "DeepSeek-V4-Flash-0731 model card" https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

  2. DeepSeek, "DeepSeek-V4 Technical Report, arXiv:2606.19348" https://arxiv.org/abs/2606.19348

  3. Artificial Analysis, "DeepSeek V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index" https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash

  4. Artificial Analysis, "Intelligence Index leaderboard" https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index

Hardware, Cost and Local Deployment Sources

  1. Engadget, "NVIDIA starts selling its $3,999 DGX Spark AI developer PC" https://www.engadget.com/ai/nvidia-starts-selling-its-3999-dgx-spark-ai-developer-pc-120034479.html

  2. GPUSmith, "NVIDIA DGX Spark review: price, specs and performance" https://gpusmith.com/articles/en/nvidia-dgx-spark-review-specs-performance

  3. Flowtivity, "DeepSeek V4 Flash at 1M context on dual DGX Spark" https://flowtivity.ai/blog/deepseek-v4-flash-1m-context-dual-dgx-spark/

  4. Loft LLC, "DeepSeek V4 Flash on llama.cpp, Blackwell local inference" https://loftllc.dev/en/docs/tech/llm-research/deepseek-v4-flash-llama-cpp-blackwell-local-inference/

  5. DeepSeek, "API pricing" https://api-docs.deepseek.com/quick_start/pricing

  6. Presenc AI, "Local LLM vs cloud API cost comparison 2026" https://presenc.ai/research/local-llm-vs-cloud-api-cost-2026