Re: [HBRobotics] ai dev machines

[Thomas Messerschmidt]

Nice! In order to run the AI models I need, currently I would need to spend over $5000. Hopefully these new machines will be significantly cheaper. 

Thomas Messerschmidt

-  

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

Contact me directly or through LinkedIn:   

https://www.linkedin.com/in/ai-robotics/

On Jul 15, 2026, at 12:08 PM, A J <aj48...@gmail.com> wrote:



Hey Folks,

It looks like vendors will be pushing out more laptops with shared memory.

For work stations LLM under 1 TB can power complex AI development tools for Bots.

Apple: Pushing the Upper Capacity Limits (2026-2027)

Apple continues to dominate the raw capacity threshold of unified memory, utilizing proprietary chiplet fusing. [1, 2]

• Current Lineup: Laptops running the M5 Max ship with up to 128GB of high-speed unified RAM.
• The 2026/2027 Horizon: Apple is testing the upcoming M5 Ultra chip (slated for the Mac Studio refresh), which fuses two M5 Max chips together. Reports show these workstations pushing up to 768GB of unified memory explicitly to allow local machine learning pipelines and multi-billion parameter model hosting. [1, 2, 3, 4]

NVIDIA: Re-entering the Consumer PC Space (2026-2027)

NVIDIA has bypassed traditional discrete desktop GPUs for AI-native workflows by introducing its own consumer "Superchip" platform. [1, 2]

• The Hardware: Announced at GTC Taipei, the NVIDIA RTX Spark is an ARM-based SoC developed alongside MediaTek to disrupt the thin-and-light laptop and small desktop space. [1, 2, 3]
• The Unified Memory: The RTX Spark chips feature up to 128GB of unified memory. It is structurally integrated to run 120-billion-parameter local AI models and render massive 90GB+ 3D scenes seamlessly. Original Equipment Manufacturers (OEMs) like ASUS, Dell, HP, Lenovo, and MSI are shipping these Windows-on-ARM systems. [1, 2]

AMD: The x86 "Super-APU" Strategy (2026-2027)

AMD is relying on its x86 inheritance to offer unified memory without breaking existing Windows app compatibility. [1]

• The Hardware: AMD's flagship UMA family is the Ryzen AI Max 400 Series (codenamed "Strix Halo"). It combines Zen 5 CPU cores with powerful RDNA 3.5 integrated graphics. [1, 2, 3]
• The Unified Memory: AMD upgraded the memory controllers on the Strix Halo SoC to support up to 192GB of unified LPDDR5X memory. This hardware gives developers an identical software footprint to AMD's enterprise datacenter chips, allowing them to train AI models in the cloud and run them natively on a laptop. [1, 2]

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/148cf2ed-ca25-4e8f-abd9-a29569786b98n%40googlegroups.com.

[A J]

Hey Thomas,

I did a search for RTX Spark Windows 11 machines and it seems like some models as less.

Was researching a Codex machine using gamer or workstation parts and it is about 2x or more.

Best!

[Chris Albertson]

This is the classic “Buy vs Rent” decision that applies to many things, not just AI.

As I see it, I can get a reasonable amount of AI computing for free.  But if I use a lot, then I have to pay for it.   This is OK up to the next limit, where it is cheaper to buy a system than to rent time.     But you must be a big user to justify the cost of buying.    Don’t buy anything until you are saying to yourself, “These AI Cloud subscriptions are killing me.”   Only then can you justify spending $5K or $10K for a local system.

But there was so much talk about humanoid robots here.  The good thing about humanoid robots is the model that controls locomotion is TINY compared to the LLMs that handle text.   The locomotion models can run on microcontrollers as they have only thousands of parameters, not billions.  

On Jul 17, 2026, at 12:10 AM, A J <aj48...@gmail.com> wrote:

Hey Thomas,

I did a search for RTX Spark Windows 11 machines and it seems like some models as less.

Was researching a Codex machine using gamer or workstation parts and it is about 2x or more.

Best!

On Wednesday, July 15, 2026 at 3:06:06 PM UTC-7 Thomas Messerschmidt wrote:

Nice! In order to run the AI models I need, currently I would need to spend over $5000. Hopefully these new machines will be significantly cheaper. 

No, that will never happen.    What will happen is your expectations will go up, and the new models will be larger.  The cost of the computer will ALWAYS be at that economic point where you make the Buy vs Rent calculation.  You might buy if there is a 1-year breakeven.  That 1-year breakeven price point might always stay at $10K, but the technology might move.     I think the cost of an “AI Computer” will always be about the cost of renting heavy-use AI for a straight year.  

If you are a hobbyist, you can control the price of the hobby by choosing to work on a different part of the problem.   For example, with humanoid robots, the first steps don’t require much cost; simulation and RL training is easy even on a midrange PC/Mac, and CAD software can be free.   You’d be years into the project before you needed to spend much.        But there are other approaches that are heavy up-front costs.  Just don't do that.    

[Albert Margolis]

All other things being equal, this is a valid way to evaluate the purchase. Some things that may tip the scale toward own:

- latency

- connectivity / reliability:

--- you may want to take your system away from good internet connectivity

--- they won't warn you before taking down systems for maintenance or throttling your usage which could be dangerous for your bot

- privacy: they use your data for training. there is nothing that I do that has particularly great value but the fact that they can aggregate small learnings from all of us at no cost and then charge for the benefit is an issue. I don't know how you could open source the high cost of training the large models but I would feel much better about that than the current environment. I doubt these models could code anything if all the free contributions to Stack Overflow were removed from the training data.

- Al Margolis

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/531D0372-0E7E-4889-8E4C-A9B37EB06244%40gmail.com.

[andy]

Hey Chris,

There are some good points when looking at cloud gpu. I think a sweet spot is the H200 which has over 160 GB vram.

With some many open source tools out there maybe providers could have tiers of ready made images. The local

Linux AI could suggest strategies and even negotiate prices for extra compute power.

Best!

On Fri, Jul 17, 2026 at 2:40 PM Chris Albertson <alberts...@gmail.com> wrote:

--
You received this message because you are subscribed to a topic in the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/hbrobotics/TxIGQV-se4c/unsubscribe.
To unsubscribe from this group and all its topics, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/531D0372-0E7E-4889-8E4C-A9B37EB06244%40gmail.com.