Welcome Avatar! We have a guest post this Monday with BowTiedStack
As with all Guest Posts, this is entirely Written by BowTiedStack. A CTO for & Engineer for an AI Startup
We’ll add our thoughts on Wednesday but think its a pretty good argument for local AI. Talk then!
BowTiedStack: Grok Sets Things in Motion
On the first of November in the year of our Lord 2026, Elon Musk admitted in a reply on X what many had suspected but couldn’t prove.
After 10 weeks of taking the AI twittersphere by a storm, Grok Bot had become totally unusable for the previous week.
Not from bugs or withdrawn capabilities, Grok Bot was still the most polished and powerful 24/7 agent harness that we’ve seen yet – a long way from the buggy and cult spawning OpenClaw which launched at the end of 2025.
No, Grok Bot had become unusable from new usage limits. Countless users were reporting that their usage quota, which used to last a week with room to spare, was now only lasting hours.
Rumours swirled as to why, a recent announcement of using more expensive Opus-5.5 instead of Grok-4.7 was the primary suspect, a potential result of Anthropic’s increasing compute spend at SpaceX potentially crowding out available compute for in-house cheaper models like Grok-4.7 or projects like Grok Bot.
Regardless of the reason, business owners had moved their entire businesses to Grok Bot, setup automated routines, painstakingly pasted in their account usernames and passwords when prompted, all to be rug pulled just weeks later.
Unfortunately, for anyone following the space, this is not really a surprise, though it still was disappointing.
Since the first “all you can eat” AI plans were launched, OpenAI, Anthropic, and even the Chinese AI companies like Kimi have consistently reduced usage quota for paying customers, raised prices, and silently degraded model performance weeks after new model releases. Consumer fraud is the business model of AI companies.
So, where does that leave business owners like you, anon? How can you leverage the efficiencies and power of AI without the risk of getting rug pulled again, and again?
This post will answer these questions, and put you miles ahead of business owners who just last week downloaded Claude Cowork for the first time, and are either deep into their first wave of AI psychosis, or have already cancelled their subscription because they can’t think of what to use it for.
There is a third way. Let’s begin.
The Rise of Agents
The launch of ChatGPT in November 2022 introduced the chat paradigm for interacting with AI models. For programming use cases, it was powerful but involved a lot of copy/pasting of written code, debugging errors, and actual output, so it in practice was not meaningfully different compared to the old days of searching Stack Overflow.
Claude Code in February 2025 introduced agents, but primarily limited to a programming use case.
“Agents” came to mean a harness wrapping AI chat.
The harness would expose available tools like read_file, write_file, run_command...etc. and loop as the conversation now included inputs from tool calls (not just human prompts) until the model reached a conclusion and terminal state. The agent harness paradigm unlocked full programming loops where the agent harness could iterate through reading files, writing, running test and compiler commands, parsing the output, fixing bugs automatically... all without human intervention.
OpenClaw in November 2025 was the first breakthrough use of agents beyond programming. Non-engineers for the first time could leverage 24/7 agents to control their computer, conduct research, manage social media accounts, or automate a variety of other business processes. All without much technical knowledge of programming.
Mac Minis were out of stock nationwide for a few weeks after the initial hype, but OpenClaw was plagued from the beginning by being the product of a single engineer’s vibe coding, not a seriously built product by a team or company. And when OpenClaw was acquired by OpenAI in February 2026, quality and level of polish continued to fall. Though, token inefficiency might have been a feature of this deal for Sam Altman – not a bug, as OpenClaw was one of the first times that customers started to max out cloud plans and face the actual API prices for model access.
Beyond some initially committed die hard users, OpenClaw adoption slowed and most returned to Claude Code (or Cowork) or OpenAI’s ChatGPT (or Codex) as their primary harness.
Others like myself discovered Hermes from Nous Research, a competitor to OpenClaw which launched weeks after, but with a much stronger team and community building it, and a limited but more robust set of capabilities, self-learning skill generation, and safety guardrails built-in. Hermes offered a much more refined product than OpenClaw, and has seen increased adoption throughout the year, even as OpenClaw usage has slowed as users burn out and give up after its many rough edges and vulnerabilities.
Grok Bot launched August 11, 2026 and quickly gained traction as the most polished and mobile-first agent experience yet. Combined with the usual AI chat interface and rich included tools (like Cursor cloud agents for programming), was the ability to name and create unique “bots” for different purposes (ie. chief of staff, researcher, engineer, reviewer, QA, marketer, pentester...) and a full Linux VM in the cloud to unlock full computer use by any of the bots.
Need to log into a website? Instead of getting blocked using a headless browser from Claude Code in the command line, the full Linux VM with Chrome pre-installed meant that anything you do on a computer, could now be easily automated with Grok Bot.
Taken to the extreme, I had Grok Bot reverse engineer multiple Windows applications for my current startup, and build a sidecar tray app to integrate with them. To do this, I had the Grok Bot Linux VM use remote desktop to connect into a third-party Windows VM I rented (to avoid paying for a Windows computer).
The layers of remote desktop seemed over the top and unlikely to work, until after one weekend of it churning away night and day, I woke up Monday morning and it was done. The sidecar could talk to the closed source Windows apps, and my startup now had working technology to start selling to customers. For context, this would have probably taken months for a team of three engineers pre-AI, or still weeks of work if simply using Claude Code single-threaded.
My QA bot could even log into Gmail, and do full signup and login workflows to test and catch bugs in our SaaS web app. And it caught many bugs, filed Github issues, asked Engineer bot to push fix PRs (pull request, an atomic unit of code change), asked Reviewer bot to review Engineer’s code changes, and then ping QA bot when the PR landed and the next round of testing could begin... all without any manual intervention. Bots would even argue back and forth about the best approach, and the final result was always much better than the initial approach.
By leveraging Grok Bot, the connected Cursor cloud agents, Linux VM, and the multiple adversarial agent bots model, I was able to automate the entire technical side of a new software startup.
What previously would have taken a full year with a staff of 6-10 senior engineers, I did in under two months on my own, with two $300/month SuperGrok Heavy subscriptions.
The inevitable shrinking usage quota rug pull
By late October, the entire tech side for my startup was running on Grok Bot (& Cloudflare Workers – but that’s another story).
And while usage quota had certainly gotten tighter (I now needed a second Grok Bot account), I figured my usage had legitimately gone up and it was only another $300/m, not the end of the world.
But then the great rug pull arrived. 8 hours into my week, all progress halted. My usage was at 100%. On a Monday. With half a dozen partially written PRs now stuck in limbo, waiting for me to pay for extra credits.
One PR was especially pertinent so I coughed up for the extra credits. It ended up taking $70 to finish that PR.
If this is the new status quo, then I’m screwed. Modern velocity simply can’t afford $140 per PR.
When I was a senior staff software engineer at a big tech job in 2022, I was expected to ship around 100 PRs per year.
By the end of 2025, with forced AI adoption in full swing, expectations had risen to 100 PRs per month.
When I joined a startup in early 2026, anything less than 100 PRs per week was suspect.
And now in late 2026 with Grok Bot, many days I’ve been merging 100 PRs in a single day.
At the current PR run rate, I’m on track to ship 25,000 PRs per year, up from 100 in 2022.
And if each PR now costs $140 in actual AI costs to complete, my non-VC-subsidized AI token spend is already over $3.5 million.
So much for only needing to spend $3,600/year on cloud AI plans.
Front running the cloud collapse
Subsidized cloud AI plans are fast coming to an end as the “AI bubble” money to subsidize usage runs out.
Despite how much they try to dog walk their users wth airdropped resets, included API credits, or claiming that a 20x plan doesn’t mean 20x the weekly usage, but only 20x in select 5 hour windows (Anthropic was forced to defend this recently in court), the open secret is that all of the cloud AI plans are coming to an end.
You didn’t think that $100/m for near-unlimited tokens was going to last forever? Did you?
For many heavy users, what used to take a $20/m or $100/m plan, now requires multiple $200/m and $300/m plans.
And quota shrinkage + price hikes are not an annual affair to keep up with inflation. They are now changing every 6-10 weeks, always worse for customers.
Grok Bot’s recent rug pull shrunk usage quota by 96% (8 hours vs. previously 168 hours of continuous usage in a week).
While the stock market may show weakness in the AI trade, token consumption continues to grow, and cloud providers like Amazon have already warned that compute availability may remain tight until the end of 2028, or beyond as new supply coming online has already been reserved.
So, where does that leave you?
You’ve built your entire business assuming a certain price and volume supply of tokens.
Ensuring you have continued access going forward without sacrificing perpetually increasing margins to the frontier labs is a critical question of business continuity at this point.
When I compared the raw API cost of my token spend in early 2026, I was getting $12k worth of tokens on a $200/m OpenAI Codex plan. As of last week, I’m down to $2k in effective usage.
If I do nothing, I inevitably am going to see my cost for tokens rise from a current negligible $200/m, up to thousands or tens of thousands of dollars per month.
Silent nerfing on token quality
Beyond simply rugging users on shrinking usage quota and rising prices, cloud AI providers have now become notorious at silently rugging users on the actual quality of the tokens themselves.
How often have we now heard complaints like this?
The latest model releases. The benchmarks are incredible. Influencers are hyping it up. Everyone’s going to lose their job, this model can do anything except impregnate your wife.
And then a week later, “why is it dumb now?” What happened?
Frontier labs have a variety of knobs and controls at their disposal to adjust the cost of serving each marginal token to users. Terms vary at companies but effectively the thinking difficulty and accuracy can be adjusted significantly to make models cheaper to serve, but at the cost of quality.
Models are simply a huge set of linear algebra matrices, rows and rows of numbers crystallized as “weights”. Each model takes millions or billions of dollars of training to pass terrabytes of content through a linear algebra matrix, and computing the next version. The final set of numbers are the ones which are frozen as the “weights”, and are used on the faster and cheaper inference path to generate and respond with tokens to your requests.
In full precision, the full set of “weights” can be hundreds of gigabytes, even terrabytes. The latest leaked Anthropic Mythos-tier model was clocked at over 7.2TB.
To run a model in less RAM, the numbers comprising the weights can be shrunk down with some precision loss to a smaller “quantization”. For example, uncompressed weights which use floating point 16-bit numbers can be shrunk down to 4-bit or even 1-bit integers with sometimes minimal quality loss (when done well).
Models also are measured in the size of the linear algebra matrix, captured in the metric of “parameters”. Frontier models now stretch into the trillions of parameters, but extremely capable models which can run on your laptop can be accomplished with tens of billions of parameters. Model compression can also be done by stripping out parameters which are less critical to most tasks. So instead of serving an expensive 2 trillion parameter model, a 450 billion (or even 27 billion) parameter model could be swapped to and for a lot of tasks, you’d never know. But for some tasks, you absolutely would notice.
Frontier labs leverage these and other techniques to drastically cut inference costs after the initial new model launch. Launch week is their “Costco chicken” loss leader. Millions in subsidized compute for the highest quality and accuracy tokens.
Then, the focus on profit margins rears its head, and customers get silently degraded to smaller, cheaper, and less accurate quantization model versions, never notified that the product they are paying for has materially changed.
Is there any way out from the multi-dimensional consumer fraud from cloud frontier AI labs?
Government crackdowns loom
While Dario every time he needs to fundraise endlessly FUDs that AI will take your jobs, develop bioweapons, and probably kill your wife and family, the government unfortunately has started to listen to him seriously.
Remember when the government blocked Fable for a few weeks for all users, and then restricted use to only American citizens within the United States?
Government restrictions on frontier model access will continue. And frontier labs have even started self-restricting access, keeping the best models in-house for themselves, and limiting full access to approved partners (ie. Anthropic’s Project Glasswing, OpenAI’s Daybreak Red/Blue), nerfing capabilities for general users especially in cybersecurity, biology, and other perceived risky topics.
Discussions have arisen around even requiring full KYC (Know/Kill your Customer) identity verification as a requirement for model access. Given conversations with AI models still do not yet benefit from attorney-client privilege style data protection laws, tying your chat records to full KYC identification sets us on trajectory to live in a full surveillance state dystopian nightmare.
Furthermore, regulations have already been passed to restrict government agencies and government contractors from using certain Chinese opensource models. How long will it take for frontier labs to convince the government to ban all opensource models (their primary competition) and force the public to forever only use their duopoly for access to tokens?
If only there was a way to self-host models privately, locally, and without anyone’s permission...
Use Local AI to build your private STR (Strategic Token Reserve)
While most are used to AI being only served from expensive and controversial data centers, predominantly Chinese open source models have for the past year been closing the gap with the cloud frontier labs.
Almost any Macbook Pro from the past three years can already run extremely capably local models like Qwen3.8-27b at 40tok/s (tokens per second). This speed is entirely usable in chat or programming applications. And you can adjust the quantization or parameter model size to tradeoff accuracy against speed or memory requirements, and it always is served the same way – no silent degraded output like with cloud plans.
Mac Studio M5 Ultras, Nvidia DGX Sparks, or other custom Nvidia/AMD GPU builds now offer the ability to host near-frontier sized models from your basement. And they don’t cost $ millions for a full rack, like the data center GPUs.
Starting around $20,000 today, you can lock-in usable token supply that no cloud provider can ever again rug pull or jack up the prices or silently degrade quality. The process at a high level looks like this:
Buy or use existing hardware
Pull the latest DeepSeek, Qwen, Z GLM, or Kimi model weights from Hugging Face
Run on a local inference engine which exposes an OpenAI compatible API
Point your Hermes, Claude Code, or other applications at that URL
Profit
But, why bother? Even with cloud plans at $300/m, the break even point for $20k in hardware is 5.5 years. Would the hardware even last that long?
The point, as mentioned above, is to skate to where the puck is going. Cloud plans are going away. Quota is shrinking. Model quality degrades randomly. Prices are rising.
What happens to the break even if cloud plans are $1,000/m? Under 2 years.
How about $5,000/m? 4 months.
Even $10,000/m? 2 months.
Beyond the shifting prices and quota of cloud plans, local hardware is in tight supply, with prices being pushed up monthly as RAM shortages have rippled across the industry from data center build outs.
Nvidia DGX Sparks were selling for $4k per unit earlier this year. They now are now close to $10k.
Mac Studios have limited production runs, and once the last generation M3 Ultras sold out, the only place to find them was the used market, and prices subsequently doubled.
Beyond the price increases, at various times hardware has simply been unavailable for purchase at all. All retailers showing “Out of stock”, and nearly no listings on the used market.
Given rising cloud plan prices, the demand for alternatives like local compute will also rise. I expect local hardware shortages and rising prices to continue for the next 24 months. Micron just cancelled another factory expansion. RAM supply is not keeping pace.
AI stocks may fall in the next year, but AI hardware prices probably will not.
For example, the Mac Studio M5 Ultra 512GB opened for pre-orders on October 28, and every time I refreshed the page, the shipping dates kept slipping further and further out. Orders submitted immediately ship in 3 weeks.
Now you’ll be lucky to get any order fulfilled within 5 months. Once they sell out, you’ll be stuck waiting until the M7 Ultras expected in late 2027 or early 2028.
The time to secure your compute is now. But how?
Get your feet wet: download LM Studio (or mtplx, DwarfStar, TensorFold, or any of the much more performant new LLM engines), pull some models, and try them yourself locally. Point your Hermes, Claude Code, or OpenCode at it even. Turn your wifi off. The tokens will keep flowing. The first time you get this working, your mind will unlock and you will see that the future is self-hosted.
Build your personal eval: AI benchmarks are done with a repeatable set of tasks that every model is tested upon, this is called an eval. Collect a couple common tasks that you currently use cloud AI for, then use OpenRouter to test available opensource models against your common tasks. You can even get your favorite cloud agent to automate this and run your eval against all the top opensource models by size in an afternoon for likely under $35 in token spend.
Choose your target model: based on your personal eval results, you can confidently know what sized model you need given your actual workloads, and then how much RAM you’ll need to run that model. For many, Qwen3.8-27b is sufficient and only requires 64GB of RAM to run. Qwen3.8-Flash-Next works with 128GB. For GLM-5.3-Flash, you’ll want 256GB. GLM-5.3 or DeepSeek v4 and you’ll be getting closer to 512GB.
Determine your concurrent token volume needs: if your hardware needs to support multiple users, each with multiple parallel agents running all day, you’ll need a much beefier setup than someone that only uses Hermes for low priority programming tasks overnight. Keep a daily log of how you use AI in an average day, check your token usage if your cloud plan reveals it, and try to determine your daily token usage, and minimum required aggregate tok/s during peak usage. If you can sacrifice some peak tok/s by spreading out your work overnight or longer stretches, you can get by with a much cheaper setup than buying hardware to support huge performance spikes but otherwise largely sit idle most of the time. Only you can decide whether the jump from a $15k Mac Studio to a $70k Nvidia RTX Pro 6000 rig is necessary.
Buy your hardware, pull your models, choose an engine, profit.
Migrate from Grok Bot to Hermes, OpenMausBot, or any number of open source agent platforms where you can bring any model (local or cloud), with no lock-in.
Local AI is not yet Plug-And-Play
Local AI today is akin to getting into computers in the 1970s or 1980s (and with the hardware prices to match). There will be some tinkering, and some research & security hardening is required, but the upsides in self-sovereignty, privacy, performance consistency, and hedging future supply/price risk will make it worth the upfront investment for many readers.
While hardware prices may eventually come down, for the foreseeable future any hardware investments you make will have an effective floor when considering the risk since reselling for near MSRP (or even above purchase price) is always an option if the setup doesn’t work for your needs.
At least this is the set of arguments that I made to my CEO when arguing to purchase $60k in hardware for AI inference earlier this year. While initially aghast at the price tag, as we discussed the tradeoffs and looming risk of cheap AI cloud plans disappearing, he soon was as local-AI pilled as I was, and approved the purchase plan.
Inflation adjusted, top-tier AI hardware is not dissimilar to what we faced in the 1970s and 1980s.
When the Macintosh launched in 1984, it cost more than a new car. Today’s top of the line Mac Studios, let alone full Nvidia RTX Pro 6000 builds, return to that paradigm. Early adopters pay the piper, but the opportunity cost of not adopting local AI could be far steeper for those that wait. I’m still kicking myself for not snagging a few Nvidia DGX Spark’s when they were at $4k instead of $10k.
For more on local AI, I post on X, write on Substack, and am CTO and solo engineer for a local AI startup.
I’ll be answering questions in the comments on the Wednesday paid article so everyone here has a chance to read everything.
You can also DM if you want additional help to adopt local AI for your business. (BowTiedStack)
The future will be self-hosted, or slavery. Secure your compute.
Disclaimer: None of this is to be deemed legal or financial advice of any kind. These are *opinions* written by an anonymous group of Ex-Wall Street Tech Bankers and software engineers who moved into affiliate marketing and e-commerce.
Old Books: Are available by clicking here for paid subs. Don’t support scammers selling our old stuff
How ETH is Staked: Covered (here)
Crypto: The DeFi Team built a full course on crypto that will get you up to speed (Click Here)
Crypto Taxes: We have a suggested Tax Partner and 25% discount code, for information see this post.
Security: Our official views on how to store Crypto correctly (Click Here)
Social Media: Check out our Instagram in case we get banned for lifestyle type stuff. Twitter will be for money