Development & Automation

A Chess Grandmaster Cut His $300 AI Subscriptions With a DGX Spark.

After 10 days running local AI on an Nvidia DGX Spark, chess grandmaster Vidit Gujrathi says he cut AI subscriptions of more than $300 a month to $40. He still uses frontier models for complex work.

· Huuush

A Chess Grandmaster Cut His $300 AI Subscriptions With a DGX Spark

Key takeaways

  1. 1Chess grandmaster Vidit Gujrathi says that 10 days after getting a DGX Spark, he cut AI subscriptions of more than $300 a month to $40.
  2. 2He still uses frontier models such as Grok 4.6 for complex work, so it is not a full switch to local.
  3. 3He can't use Qwen 3.8 27B yet because of speed, and is considering adding an RTX 5090 or expanding with another Spark.
A green and brown boxed Nvidia DGX Spark sitting on a desk, with the DGX Spark logo visible on top

From $300 a month in subscriptions to $40

Chess grandmaster Vidit Gujrathi posted a hands-on review of 10 days with the Nvidia DGX Spark on his X account on August 22. Gujrathi is an Indian player who earned the grandmaster title in 2013, won the 2023 FIDE Grand Swiss, and is a 2024 Chess Olympiad gold medalist with Team India. Away from the board, he is also known for building and experimenting with AI tools himself.

Here is how he describes his spending change. He used to spend more than $300 a month, with $200 on Claude Code and $100 on ChatGPT. After shifting more work to local models, he says he's down to $20 for ChatGPT and $20 for Cursor. These are figures he reported himself and haven't been verified by a third party, but it remains a concrete example of an individual cutting subscription costs with local AI.

Some things local models can't do

The model Gujrathi currently uses most in Hermes Agent, a local AI agent framework, is "Ornith 1.5." He says he likes its balance of speed and passable intelligence. But he drew a line: for complex tasks he still trusts frontier models more. He has recently been experimenting with Grok 4.6, and for work that has to happen in the mobile app he uses Luna Max. For tasks like updating his "second brain," where he wants to protect his privacy, he prefers local models.

In other words, going local isn't an all-or-nothing choice. It's closer to splitting work between local and cloud depending on the task. Cutting subscriptions doesn't mean he dropped frontier models entirely.

The next wall is speed

Gujrathi likes Qwen 3.8 27B but wrote that on the DGX Spark it runs at only about 20 tokens per second, which makes it "basically unusable." He says he is considering adding an RTX 5090 or buying a second Spark. In the replies, another user said Qwen 3.8 27B's benchmark scores "seem optimized more for benchmarks than for real use," and Gujrathi agreed that "perceived performance can vary with the quantization level."

Another open question is how much to trust benchmarks for quantized models. He gets excited about intelligence scores published by sites like Artificial Analysis, but what he actually runs locally is a quantized version, so whether those scores carry over has to be checked separately. There's no answer yet, and he says he plans to test further whether DeepSeek Flash 0731 is practical on a single Spark.

Frequently asked questions

What is the DGX Spark?
A personal AI computer from Nvidia with a GB10 Grace Blackwell superchip and 128GB of unified memory. It can run models of up to roughly 200 billion parameters locally at 35 to 80 tokens per second or more, and its price rose from $3,999 at launch to $4,699.
Did his subscription costs really fall by more than two-thirds?
These are the spending figures Vidit Gujrathi gave in his own post. He used to spend more than $300 a month on $200 for Claude Code plus $100 for ChatGPT, and after replacing much of that with local models he says he's down to $20 for ChatGPT and $20 for Cursor. They haven't been verified by a third party.
Did local models fully replace the cloud?
No. He says he still trusts frontier models more for complex work and uses cloud models like Grok 4.6 and Luna Max alongside local ones. He mainly uses local models for work that needs speed and for updates to his 'second brain' where he wants to protect his privacy.
Why can't he use Qwen 3.8 27B?
He says he likes the model, but at about 20 tokens per second on the DGX Spark it is too slow to be practical. So he uses Ornith 1.5 as his main model, and he's weighing adding an RTX 5090 or buying a second Spark.

Related articles