Your AI Bill Is Bigger Than The Salary You Cut - We can help!

Added: 2026-05-07

[00:00:00]So, a fourperson startup just posted their $113,000 monthly AI bill. $113,000.

[00:00:07]And they posted it proudly. Uber burned through their entire 2026 AI budget in 4 months. 55% of companies that fired people to save money on AI now regret it because the AI bills are higher than the salaries that they cut. The hyperscalers are spending $715 billion on AI infrastructure this year alone, and analysts are questioning whether they'll see returns before 2028. Now, meanwhile, the average developer spending thousands per month just to try to keep up, and their invoices keep climbing. Something is fundamentally broken about how we're building with AI, and nobody's connecting the dots. But I'm about to today because we've got some great stuff to show you today. So, make sure you stay to the end because we've also got a free giveaway that we're announcing. So, make sure you stay to the end here.

[00:00:48]Let's dive in.

[00:00:57]Welcome to Starter Pack. I'm Spencer and here at Startupac, we love to build custom software solutions for companies.

[00:01:01]With a decade of executive leadership as a fractional CTO and 25 years of software development, I have helped transform a tech teams and products, including building out custom AI solutions. Now, today we're going to do a full autopsy on the AI cost crisis.

[00:01:14]What's causing it? Who's destroying it?

[00:01:16]And then in the second half, I'm going to walk you through why I built something specifically designed to fix this. You ever heard of free tokens?

[00:01:23]Let's talk about it today and make sure you stay here to the end because I'm I'm this isn't just a teaser, guys. Like, you're going to really like this. So, let's dive into this today. So, a fourperson startup just posted their $113,000 monthly AI bill proudly after Uber burned through their entire 2026 budget in four months. 55% of companies that fire people to save money on AI now regret it. the salaries they cut the AI bills are higher. Nobody is saving clearly now. So if you look at this the fire team to save money now the AI bills are higher than the salaries the salaries that they were. Now this is crazy here and we could dive into this article and dig through this but I've got a ton of different stuff to cover today. Startup CEO said he's proud his fourperson team racked up $113,000 monthly AI bill. That is absolutely insanity. And I'm sure you saw about Uber, right? $3.4 billion R&D budget.

[00:02:13]one of the most wellresourced uh well-resourced tech companies on earth.

[00:02:17]They burned through their entire 2026 AI budget in four months because their engineers couldn't stop using claude.

[00:02:23]Now AI evals are becoming the new compute bot bottleneck. AI evaluations have crossed a cost threshold that uh that changes who can do it. The holistic agent leadership recently spent about $40,000 to run 21,000 agent roll out across nine models and nine benchmarks.

[00:02:42]That is just insane, right? The cost problem starts started before the agents. When Stanford's uh C CRFM released these, the the papers own per model uh accounting showed API costs ranging from $85 to $10,000, right? This is just insane. As you keep going, like the news just keeps getting crazier. So, this poster here, I'm really struggling to see how the back of the envelope math on this works out. There are generally there are generously 4 million characterized software workers in America. That's pretty broad and includes a lot of people who aren't really classical engineers and don't produce that much code. That comes out to nearly $1,000 per month of average CLA spend across every dev in America.

[00:03:25]Yes, there's some international use, but it can't be that much. There's some non-software co-work usage, but again, that doesn't burn that many tokens. So, but he says even if we were to assume 50% of all software workers are using Claude, that comes out to $2,000 spend per month per Claude user. That's 10 times more than the highest max tier subscription. So, almost all of the anthropic revenue has to be API billing.

[00:03:48]So, the only other explanation is that something like 20% of software engineers are not only Claude users, but on the API bills and regularly spending thousands per month. I mean, we're seeing that in these other reports. So, that's kind of crazy, right? because that's what they're showing here is Anthropic is now showing $44 billion in annual recurring revenue. That's up $14 billion just since last month. So crazy.

[00:04:10]All right, the pulse token spend breaks budget, but what's next, right? Last week we covered the slightly perverse top trend of top uh token maxing. This is where Facebook was actually had a leaderboard to see who could use the most tokens, which is just kind of crazy. This week I spoke with software engineers of a large company, another seed uh seed stage place. Both shared almost identical stories. At their latest all hands, company leadership expressed concerns about the fast rising cost of tokens. At both places, token spend increased by 10x in the last 6 months with no sign of slowing down.

[00:04:43]Absolute insanity.

[00:04:45]Um let's kind of jump to the end here because u I think he brings a really good point home here at the bottom. So he says uh So he's like how business manage tokens one let it rips means and then start measuring it and number two curb spending right but the discount exists when spending the millions of dollars again this is just crazy to me I can't even fathom costing that much this guy says the light mass uh in light of massive layoffs at Meta Microsoft and others an Nvidia exec suggested that right now AI isn't saving companies money on labor it's actually costing them more than the humans they currently employ insane and they keep doing this thinking the cost is going to go down, but it's actually just going skyrocketing.

[00:05:27]Never made sense to me. If intelligence became cheaper, demand will increase, thus creating upward pricing pressure.

[00:05:32]Energy to run these models isn't free at the end of the day. And so this the original poster was, what happened to intelligence too cheap to meet? Are all these token prices? I I might as well think for myself.

[00:05:44]I mean, it's just again the flat rate co AI coding subscription era is ending.

[00:05:49]what GitHub copilot cla code and cursor changes in April mean. Right? So the big thing here is GitHub reshape copilot individual plans and temporarily pause new signups for some tiers. Anthropic shortened claude code server side prompt cache from 1 hour to 5 minutes. A change many users only discovered through logs and billing behavior. Cursor move frontier models behind max mode for legacy team enterprise. So you can see a lot of this happening. It's what it is is exactly what we've been predicting here for a while, right? the cost is getting going going crazy. The inference is just not scalable. And then on the flip side, you're seeing the VC starting to say, "Hey, these businesses need to become profitable." So that's why we see Anthropic and Openi talking about these IPOs. They're saying if we can't make this profitable, we can't subsidize this forever.

[00:06:36]what we've been using for the last 3 plus years on the inference um you know on these chat bots has been uh subsidized inference by these VC by this venture capital and other private money here. So you're seeing that that's going to come to an end, right? We're it's going to quickly come to an end. Claude Coast price cha code a pricing change delivered as a cache TTL. So this is kind of like getting that bag of chips that has less chips in it and more air, right? Um let's jump down here to the bottom.

[00:07:07]So the thesis flat rate AC AI subscriptions where customer acquisition tactic for autocomplete agentic coding tools are compute products. Compute products revert to metering. April 2026 is when developers start paying that bill directly. So, we're seeing that start to really come to it. We know that Claude 4.7 has been nerfed because it's definitely gotten worse uh because they just are trying to slow down the crazy spend.

[00:07:32]This poster says, "Oh boy, here we go.

[00:07:34]AI is going to get very expensive.

[00:07:35]Workspace agents will be free until May 6th, 2026 with credit based pricing started on that date." Right? So, ChatGBT has workspace agents and they're going to be free until May 6th and then they're going to get to per token. So, the Claude Code lead uh leads users two.5 billion tokens a month, which is like $1,000 or a month less. He'll probably doesn't even pay for it. I genuinely have no idea who's spending six figures in tokens or how that's possible. Well, I know how it's possible, right? Because, you know, this guy works for Anthropic, so he's getting free tokens, right? and he's bragging that he used 7.7 billion tokens in a month. All right, next one here. For clarity, we're running a small test on 2% of new proumer signups. And this is what So, see what Claude did is Claude Code made a uh uh made a trial run of 2% of their signups and actually turned off Claude Code for free pro and you had to pay for the max subscription in order to get this, right? Uh this is like crazy.

[00:08:33]That's $100 a month minimum 5x jump. So, you know, there um Claude backpedalled really quick and said engagement per subscription is way up. We've made small adjustments along the way, weekly caps, tighter limits, but usage has changed a lot and our current plans weren't built for this. So, surprise, they put these $20 plans on and people are blowing them through. Uh, last article here before we jump into some of this other stuff here because I've I've definitely got more stuff to be ch chatting on here and we're pretty excited to to get to this.

[00:09:04]So, this is the most satisfying plot twist in tech history. We just did an AI layoff due to the rising cost. Turns out AI is getting way too expensive. We just canceled five of our AI subscriptions and hired two mid-level devs instead.

[00:09:17]And that is what you call a plot twist, folks. And that's what's happening all over the place. We're seeing the cost of AI is rising quickly. Companies in 2024 and 2025 laid off developers. They laid off writers. They laid off an analyst and support staff with this clear thesis. AI is going to be cheaper. Now, the data from 2026 is in, and 55% of those companies now openly regret that decision. The AI spend didn't replace the salary cost. It's exceeding it often dramatically. You still need people to review outputs, fix hallucinations, maintain prompts, and manage the tools.

[00:09:50]The salaries were predictable and fixed.

[00:09:52]The API bills are unpredictable and elastic. When your cost center can scale to infinity at 3 p.m. because an agent loop went sideways, you have a different kind of problem than headcount. Now, most companies budget for AI like it's a SAS subscription, flat, predictable, and manageable. But in reality, token consumption follows a power law curve. A small number of use cases or users consume a wildly disproportionate share.

[00:10:17]So, one developer working with Claude Code this past April logged 7.48 billion tokens in a single month. That's one person. Now, multiply that by an agent fleet or an enterprise team or an automated workflow running 24/7. the math doesn't scale the way the pitch deck said it would. So startups and enterprises alike are discovering this midyear uh when the CFO opens the invoices. Now there's a popular narrative that AI is getting cheaper over time but that that might be true for older smaller models but frontier capabilities the stuff that actually moves the needle on complex work is getting more expensive. But we've got a catch here for you. The era of subsidized computing is ending quickly.

[00:10:57]Openi's latest top tier models and Anthropics Max plans have both seen significant prices increases and it's only going to get worse because the subsidies are ending. If you've built your cost model on today's frontier pricing, rebuild it because the floor is about to move out from underneath you.

[00:11:14]So, a detailed breakdown from an from an analyst raised serious questions about how anthropic reaches its report of 44 billion ARR number. The math only works if either a huge percent of engineers are paying thousands per month each or if a small group of super users is consuming at almost like a million dollars a month. The answer is probably both. A enterprise agentic development running at scale and individual POW users burning tokens on complex workflow. Now the thing about this here is we've got an answer for you. I know there's a lot of people who are like oh we just have to do this because that's just the new way to do it and to keep up you've got to use this. Here's the thing. You guys know me. I've been an AI realist for a long time. On this channel, we've been talking about the pros and cons of AI. I'm not anti- AI.

[00:11:59]Quite the opposite. I am anti burning tons and tons and tons of money on AI.

[00:12:06]That's one of the reasons why we put our money where our mouth is. It's one thing for me to get on here and tell you guys, oh, don't go spend on the Frontier models. Stay away from open AI. Sam Alman's evil. Instead, I'm putting our money where our mouth is. AI shouldn't have a meter. Unlimited tokens forever.

[00:12:20]You're a machine. your agent, you use it from anywhere. Open Monoagent.AI is a terminal native coding agent powered by local LLMs. Now, if you think those are toy, hang with me here. But this is 100% open- source, free forever, and installed with a single command. It's proudly built on C.NET because AI tooling should be infrastructure, not a subscription. Now, here's my manifesto.

[00:12:42]I wrote this. I'm really proud of this.

[00:12:45]AI shouldn't be a subscription you rent.

[00:12:47]It should be infrastructure you own sitting on your desk like back this here behind me right serving your code answering only to you and you being in control. Now a lot of people are like wait I don't get it. What's the catch?

[00:13:01]Zero catch here folks. This is our repository here. This is uh open monoagent.ai.

[00:13:07]This is the opportunity to dive in and to start to run this yourself. One single install code here. And you're going to run this on uh you can either run this on a single box where you run the inference and uh the agent together and you have a full coding agent that runs completely free absolutely free as free. And everyone's like now what's the catch? We've actually had a lot of people in comments and stuff saying what's the catch? There's no catch. Do me a favor and go start the repository for me because like do a brother a favor right? So if you start the repository we're almost at 300. I'm excited because we're we're going up fast. I mean this is quick. like we've only only been out for a couple days here and we're still in beta. So, we're still improving this.

[00:13:46]You can see that my my devs were just committing here, you know, six hours ago. We had our latest commit. We're actively still working on the project and still making it better. What kind of things we making it better with? Well, I'm going to tell you some of the stuff we're making this better with. But you install one command and let me jump over here and show you kind of this diagram here and get my big head out of the way here. So, one command and you can set up the inference on your uh GPU or CPU.

[00:14:12]Now, again, one of the things that I want to catch you on here is everyone's like, "Oh, but I bet you've got to have super expensive hardware." We actually have this running for one developer, one agent on a box very similar to this. The a box this size running on CPU with 32 gigs of RAM. We have it running fast enough that this agent runs for one developer. Now, if you want to push it a little bit further, the sweet spot on this right now is running with a 3090, which is what's in this box back here behind me. This other box right here is actually running a little smaller card, and we're doing a lot of testing on that. This box over here, that one over there, is actually running a 5090. It can run up to five to six uh developers at one time, all working against the inference here. Now, let's go back to our grid here. The code then actually can your coding agent will run local on your box. So, you install the agent on your box. We are offering a free secure relay where it's encrypted with your own encryption and we walk you through. It's super easy to set up. You can set this up in about 5 to 10 minutes and you actually have your own inference stack all the way from a local agent that runs in Visual Studio in your code uh and generates code for you and then runs against this. It's works absolutely incredible. And today we're actually just announcing that we're now releasing a mobile app which is a companion app.

[00:15:28]We're going to continue to build on this. For right now, it's a chat app that you can use also against your own infrastructure. This is a great opportunity for you to be able to then own all of your own intelligence. No more uh renting out this intelligence to other people. Now, again, even more still, but wait, there's more. We're going to give even we're doing a giveaway here so that on May 15th, no catch here, we're going to give away one of these boxes, right? We're give away one of these guys that'll give you up to 20 tokens per second. We've optimized this like crazy. And this is a free uh pre-flash, pre-built, ready to run open monoagent.ai out of the box. Plug it in, set yourself up, and you're off and running. So, hit hit the giveaway here and u make sure you go enter for the giveaway. Now, what we're asking is just that you give us a little bit of love on the GitHub repository. Now, this is really exciting and we're really, really excited because I wanted to be able to keep this open for you, right? See, what we want to do is we want to help everybody to be able to run their own stack of AI. We want to be able to help teach people and to give the power back to democratize AI. And this is what we really mean by this. I don't mean democratize AI by you pumping your data up into the cloud for OpenAI or Enthropic. I mean, you own the whole stack from your inference server that can run at home to your own local agent or your chat on your device where you can uh use this agent absolutely anywhere. Now, you know, how does this work? Uh that's where you can dig into the covers. There's no secret to it because we actually give you all of it.

[00:17:03]Now, there's a lot of other really cool features on there that um that I'm not going to highlight today because, you know, our video is already a really long video here, but this is an opportunity for you to be able to dig in and to be able to own your own stack. Open Monoagent.ai isn't designed for the enterprise procurement deck. It's designed for the developer who opens a terminal and wants to build the hardware giveaway that we're running, the open source codebase, the local first inference, all of it's designed to lower the barrier to serious AI assisted development. The developers who are winning right now aren't the ones with the biggest AI budgets. They're the ones with the most efficient workflow. One person using 7 billion tokens in a month can be impressive or catastrophic. It's impressive if you're doing it on your own hardware. It's catastrophic if your company is going to have to pay for that bill. We're building a tool that will maximize the what you shipped side of the equation while taking control of the tokens. That's productivity story that actually holds up. So a 715 bill $715 billion hyp hyperscaling bet is a top-down wager. Open monoagent.aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa is the bottomup counter thesis that lean localaware open- source tooling will outperform bloated cloud dependent stacks for the majority of real development work. Energy costs, compute availability and pricing volatility all favor the team that minimize external dependencies and build on their own stack. The companies that survived the AI cost reckoning won't be the ones who spent the most. They'll be the ones who extracted the most value per token and kept their infrastructure under their own control. Open Mono Agent.ai is being built with this thesis at the foundation, not bolted on after the fact. Now, if you want to be here, dive in, go pull down the repo, test it, run it. It works great. We're really excited about this. I have an R&D team who's been working on this for the last couple of months. We just are unveiling it and you know it's still in beta so we're still rubbing some of the edges off of it but it works really well. It has agents, has sub agents, has a lot of different features. One of the big features that I'm not going to go into today, it's called playbooks. Go check it out on the website. It's awesome.

[00:19:11]It's a feature that absolutely will blow you away. A lot more of them out there.

[00:19:15]So here at Starter Pack, we love to build custom software solutions for companies. So, if we can help you build out your own AI stack, if you want a local AI stack for your company, you can check out openmonoagent.ai. And if you need some help, reach out because we have great teams that can come in and help your company. So, check out startup.com if you want to reach out to us and make sure you check out openmonoagent.ai AI and we'll catch you guys