IBM just had its worst day since 1968, and the takes were… predictably fucking stupid.
Yesterday (Tuesday, July 14th, 2026 for those of you who are time travelers from the future), IBM lost roughly a quarter of its entire market value in a single session. That is the worst single-day drop in the company’s history going back to at least 1968, a stretch of time that includes the emergence of the personal computer, the dot-com bust, the iPhone revolution, Justin Timberlake bringing sexy back, the 2008 financial crisis, and now the AI-revolution.
A 113-year-old company just had its worst ever single day. That’s something to think about. But the question is, should we care, or did the market have a collective freakout, and use IBM as a proxy for their collective freakout? To examine this, we need to understand what IBM did that was so catastrophic. So, let’s see here, this 25% single day drop was because…
(Author checks notes for CEO eating a live baby on television. Hmmm… Nope. Did a data center get bombed? Hmmm. Nothing about that here. Well then, they must have started a beef with Taylor Swift fans? Also, no.)
IBM did this by preannouncing that operating earnings landed at $2.93 a share instead of the $3.01 the street wanted. So for eight cents, (mind you, they’re still making money per share) — the internet promptly decided this was a referendum on the soul of artificial intelligence. We’re talking about a slight dip in earnings per share, while many other companies developing frontier models and selling to consumers at a loss are literally burning tens of billions per quarter.
Okay then. I guess I’ll just go fuck myself then, because clearly I’m running my business the wrong way. I’m supposed to be burning cash and building hype, right?
The world of AI-bros-with-hot-takes arrived on schedule, because these are deeply unserious people who have staked their professional careers on being right about this kind of thing. Mind you, these are the same people who were also talking about “big data” in 2014, and then “the blockchain revolution” in 2020, and at some point pivoted to AI.
So what really happened here? IBM tanked because their CEO won’t stop being a wet blanket about AI. IBM tanked because mainframes are dinosaurs. IBM tanked because they didn’t pivot hard enough into agents, or pivoted too hard, or pivoted in a direction that offended someone’s feelings on LinkedIn. Just keep in mind that every one of these people has a newsletter, and several of them have a course for the low price of only $199 per month.
Want to know what actually happened? It’s boring, but I’ll tell you. IBM’s customers did not leave. They redirected their budgets into things that are experimental. They took the money they used to spend on software and mainframe hardware and they moved it toward AI servers and memory. Since there is currently a global memory shortage, the bill went up, while the new, shiny thing they bought did less than expected. CEO Arvind Krishna said IBM “faltered.” I’m not even sure we can call it that. I think we’re seeing exactly what happens when people feel obligated to buy into an AI-hype cycle, lest they put their job in jeopardy. Everyone must be AI-enabled, even if it doesn’t really make business sense.
Business customers are making some pretty fucking weird decisions right now, in my humble opinion. The money is flowing toward the promise of agentic productivity. It is not flowing toward the delivery of agentic productivity. Companies are buying the theory that eventually the money they spend will have a positive outcome. Most of the agentic products I have seen in real production environments are really, REALLY rough around the edges. (Or so buggy, and miserable to use, that everyone uses workarounds.)
You know what else is a thing? Actual earnings.
I do this thing in my head that we all should do. I look at outcomes, and in business, those outcomes are either red or black. In IBM world, earnings are what matter. They’re making money, alright? This isn’t some weird hand-waving thing where there is an AI-narrative they’re supposed to follow. IBM didn’t become a global brand for over 100 years because of vibes. They aren’t known for big keynotes where they hype things up. They are known for being a business-facing company that makes reliable products that scale when your enterprise needs things to simply do the job.
The agentic boom is a wonderful narrative. I fucking love how I can share insane click-bait bullshit on LinkedIn. It’s sick, man. The AI-takeover has been a wonderful narrative for about three years now, and I’m going to milk it for all the content I can on TikTok and LinkedIn, because I’m a content creator, and narratives sell advertisements. The funny thing about AI is that the date it’s going to take over just keeps sliding to the right.
This is not because the models are bad. The models are, at this point, great. They’ve been capable of doing real coding work for like a year now. When Opus 4.6 hit, I was like, “Damn, good enough.” (Mind you, I still write all my own tests and review every signle thing it produces, but the fact that I can speed-read the process means senrior software engineers are living the good life right now. You still need to learn to code if you want to use these models, but if you know how to code, it’s so damn good to be an engineer.)
The models are the least of our problems. The problem is everything the models have to touch to become useful. That’s where we reach the SaaS cliff of death. Everything the models need to touch is owned by somebody whose entire business depends on you not touching it.
I know this is true because I tried to do it. Not as a thought experiment. As a favor to a homey.
The barbecue joint that should have been an easy weekend project that was not, in fact, an easy weekend project.
I have a friend who runs a small barbecue business. I have built custom software for 20 years now. I can self-host a model. I can build an agent. I can stand up an MCP server. In theory, I have all the skills I need to make a small business effficient. I looked at my friend’s operation and thought, correctly, that I could save him real time and real money by wiring his systems together and letting an agent handle the tedious parts. Inventory. Ordering. Scheduling. Delivery orders. Customer emails and SMS messages for specials. Catering order management. You know, the stuff that takes hours of human review time.
I was wrong, but I was not wrong about my skills. I was wrong about the world all these products live inside.
His food comes from Sysco. Sysco was not receptive to letting me connect to their API, and there was no clean option to connect to whatever database Sysco keeps his order history in. This is not an oversight from Sysco’s IT team. Sysco does not want an agent optimizing my friend’s orders down to exactly what he needs. Sysco wants to sell my friend more food. They don’t give a fuck about his food waste problem. They call food waste “profit” in their business. The friction is not a bug in their system. The friction is their system. A perfectly efficient inventory agent is, from Sysco’s perspective, a machine designed to reduce their revenue, and they are not going to hand me the keys to build it. So the alternative is to stand up a database from scratch and inventory every pound of chicken wings, and every bag of flour by hand, which is to say, adding a layer of technical debt, and a database that my friend will not be able to effectively manage… forever. Which is a stupid fucking idea.
His point of sale is Toast. Toast has his sales data, his menu data, and increasingly, his customer data. They want the customer data most of all. They want to own the customer journey. They want to own the SMS interactions. They want to own the email blast about the weekend brisket special. That is not a courtesy they are providing because they like my friend. That is a subscription they are charging for, and the way you keep charging for a subscription is by making sure the thing it does cannot be done anywhere else — which means making sure my agent cannot reach in and just take my friend’s own customer list and do something smarter with it. As strange as it sounds to say out loud, the point of sale company does not want the restaurant to have easy programmatic access to the restaurant’s own customers. They want to have access to every single restuarant they work with, and to provide other sales products, because Toast is in the business of making money for Toast — not selling more BLT’s during happy hour.
The beer distributor, sensing that it was 2026 and everyone was contractually obligated to ship something with “AI” in it, built a chat portal. The chat portal helps my friend figure out what beer he might want to stock. This is a feature that no human being has ever requested, because the way you find out what beer to stock is that a beer rep. This is an actual person, who shows up and tells you what is coming in seasonally, and that interaction is pleasant and takes ten minutes and involves free samples. Nobody wanted that replaced. It was not a job crying out for automation. It was a guy named Dave with a cooler in his trunk, and Dave is great. We like Dave.
The scheduling software, the payroll company, the clock-in and clock-out system: All of them slapped an AI layer on top of the exact same product they sold last year. Each layer is one more thing my friend has to log into, one more dashboard, one more “insight” he did not ask for, generated on top of data he cannot export.
Then there is the stuff that I thought would be super easy — and it was actually the most annoying. I wanted an agent to various sports schedules so my friend would know which games were on which nights, so he could staff the bar accordingly. You would think this is a solved problem in the year 2026. It is not. Most of the web is now defended by captchas specifically designed to stop exactly the kind of automated retrieval I was attempting. ESPN, which has every schedule for every league in a database somewhere, does not offer an accessible way in, because ESPN needs you to load the page so you can see the ads so they can sell the ads. ESPN does not want you to automate away their page views. They want you to visit ESPN and look at advertisements for a Ford F-150. The schedule is advertising bait. The schedule converts you, a person looking for information about when the Dallas Cowboys play, into a person who is now configuring a 2026 Ford F-150 on the Ford website, which Ford calls a conversion.
Nobody is rebuilding a goddamn thing. They are painting the barn a new color.
This is the thing that is actually happening. Every company already selling software in this space is “adding AI” to a product that was architected years ago. This is because most SaaS companies need to make their product (loudly clears throat)… better? No one is starting from a clean slate. There is no clean slate. What the fuck are we even talking about here? Am I the only motherfucker on this planet who has worked in this environment for the past two decades? There is always a legacy system, a decade of accumulated decisions, a database schema that probably came from 2016, when it was moved from some managed Oracle database, and there’s still some weird shit going on with it, and now there is a thin coat of retrieval-augmented generation or a bolted-on MCP server sitting on top of all.
Cool. Now you have a spoiler on a 2012 Toyota Sienna minivan.
This does not make the minivan fast. It makes the minivan… a minivan with a spoiler and 160,000 miles on it. The graham cracker crumbs are still in the cupholders, and the mystery stain on the carpet? It’s still there. (Ketchup? Pomegranate juice?) The AI layer is not integrated into the business logic. It is a translator standing outside a building it is not allowed to enter, describing the building to you based on what it can see through the windows. The owners of the building have no incentive to let you inside.
You are just adding little stuff.
It’s just paint. There’s nothing wrong with paint. I repainted the interior of my whole house, by myself. Did I hate myself for doing it? Yes. Did I drastically underestimate how long it would take? Also, yes. Were my shoulders and lower back on fire for a month after the whole process? Yes, once again. Did I save money? I mean, sure.
The reason it is only a coat of paint, is because all the other shit would mean tearing down walls, brick, filling your house with dust, consulting with architects and engineers, and… you know, spending a shit ton of money. You should not start implementing your own “open floor plan” in your home without making damn sure removing those walls will not result in the structural integrity of your home being like grandpa when he doesn’t have his little blue pills for extra help.
Even the things that were “major” improvements to my home? The new floors in the basement? The old carpet we ripped out and new carpet we put in? The little decorative copper tiles in the bar? Those are just adding stuff to the existing structure. They look a little different, and sometimes it even feels substantial, but it’s really not.
Side note: I’m taking on a crazy project this weekend. I’m adding built-in shelves to my office. I ordered some solid hardwood shelves that were on sale, and I have a wood-shop in my basement that will allow me to slightly modify these shelves in a way that make them look like they’re “custom” built ins. It’s pretty cool what you can do with some pre-made stuff, when you know a thing or two.
See the trend here? All of these AI startups, mine included? We’re additive. We have to work with what you have. Often, what you have is pretty dogshit. This lack of interoperability is corporate suicide if you start building on top of things that can’t support their weight.
The moment Toast makes it trivial to move your customer data out, Toast loses the thing that lets it charge you every month. The moment Sysco lets an agent order precisely and only what you need, Sysco sells less food, which reduces their profits. The friction between these systems is not a temporary state of engineering immaturity that a better protocol will fix. The friction is built into their product, and they’re betting you won’t take a chainsaw to your own arms because you want to remove your nail polish.
The friction is the thing keeping these businesses alive right now. The friction is the only reason half these companies still have pricing power.
The bottleneck is not me, or any other well-intentioned engineer. It is late-stage capitalism and fuckery.
Put this in your shitty newsletter and smoke it. The bottleneck for agentic AI is not intelligence, models, engineers, or our capability to build some genuinely dope shit. We have plenty of thoughts on how ideal systems will work. The bottleneck is that every single integration required to build a truly agentic system is a negotiation with a company whose financial incentive is to say “Fuck off and don’t email me again.”
To build the thing everyone keeps promising is on the horizon, you would have to get through every other company’s business model. You would have to convince Sysco and Toast and the payroll vendor and the scheduling app and ESPN to each voluntarily open a door that, when opened, makes their numbers less black, and more red. You are asking billion dollar enterprises to stop making money so that a restaurant in a strip mall can be more efficient. They understand exactly what I’m asking for, and they’re not going to oblige. Their answer is an API with limited access, stupid rate limits, and an agentic system they built. They want to charge you for agents they built, and they want to charg you per seat.
Every incumbent right now is saying the same two things at once, and they do not notice the contradiction because the contradiction would drop their stock price.
They say they are all-in on AI.
And they say they are building it on top of their existing product, inside their existing walls, on their existing terms.
Which is another way of saying: “We want the valuation bump for claiming AI enablement, and we want none of the IT disruption that real infrastructure change would cause.” They want to sell you the future as a feature while making absolutely sure the future never arrives in a form that gives customers more ability to manage their data, outcomes, etc. Because that’s how they make money.
This isn’t an engineering problem at all. There is no patch that fixes it. This is about incentives, and incentives do not get patched. If you start “patching incentives” then your C-Suite is jettisoned and replaced by people who understand the difference between red and black.
So why is IBM going to be fine?
Because IBM already won the fight, you dipshit. The fight that actually matters is not “who has the best agent.” It is “who already owns the enterprise relationship.” Their customers are not consumers. Their customers are banks and governments and airlines and the kind of institutions that measure relationships in decades. They have IBM code running processes so important that nobody at the company knows how they work, and that’s why no one is touching that DB2 database from 2002. The capex on that shit is damn near zero, and it’s working, so they’re not going to let “AI enablment” anywhere near it.
Those long-time IBM customers did not leave on Tuesday when the EPS missed by eight cents. Their customers are dabbling right now in AI. A few of their customers took some of the usual IBM money, and they played with local inference, or tried figuring how how to ingest documents into a vector database. IBM happened to be sitting on the wrong side of that particular quarter’s spending mix. That’s barely a problem. It is not the death of IBM. The market just has a finite amount of dollars to spend on IT, and right now the market is adding some built-in shelves in the office they bought on Wayfair. Don’t worry. IBM still builds the houses. The houses last 100 years. Those built ins will be gone in 15 years when you sell the house, and the next owner decides your old office is going to be the home gym, and tears it all out. The structure stays, while the decorative shit is hauled away.
IBM is not about to blow up its own business model to chase the agentic dream. That’s risky. They are doing exactly what Sysco and Toast are doing, just with a bigger balance sheet and enterprises that are larger than my buddy’s BBQ joint. They are protecting the moat. The difference is that IBM’s moat is deep as fuck. Ain’t nobody trying to traverse that fucking moat, dog. A quarter of their market cap can evaporate like someone pointed a death ray at it, and the enterprise contracts do not care, because the enterprise contracts were never priced on quarterly EPS.
So what are we doing here? We’re dragging the IBM CEO for not being sufficiently horny about AI? Really? He is not out of touch with the future. He understands, better than the people mocking him, how much friction sits between business-value and production systems. This isn’t a thing you do on a PowerPoint slide. Building massive production enterprises that are reliable is hard, man. If you can happen to do this in a way that actually saves time, money and resources, while also meeting a cybersecurity threshold that doesn’t put everything at risk, you’re going to make it a century. That’s why he’s the CEO of IBM. He runs the unit. He has seen the integration tickets. He knows that “just integrating apps” is not a hand-waving thing you do. That’s a thing that consultants put on PowerPoint slides with arrows that connect. Those arrows are, it turns out, actually a shit ton of work.
There are no more expensive icons in the world than arrows on a PowerPoint slide that connect the “customer data” box to the “application” box.
The last part, then I swear I’m done ranting.
We have built an economy that sells the fantasy of “frictionless AI” to the exact same people who are, in every other meeting, engineering the friction on purpose. The boom keeps getting pushed to the right. The hard part of our “agent powered future” is that every useful thing an agent could do requires reaching into a system owned by someone who makes their money by keeping your agents out. Efficiency for you is churn for them. Your agent is their attrition.
So the money pours into hardware. That’s why RAM costs more than cocaine. (Or so I have heard. I don’t know the price of cocaine, for the record.) That’s why gamers are literally spending “nice used car” levels of money on their GPU’s. We’re throwing cash into servers and memory, data centers and personal desktop AI you own. These are the picks and shovels, the Levi’s denim, and the pans. You can sell a pick to a man who has been told “thar’s gold in them hills.” You just cannot promise him he will be allowed to dig.
Also… one last thing. (He shouts from afar, as the miner has already walked into the distance.) We forgot to tell you something really important about those hills where the gold is buried.
“Oh, by the way, the people selling you those picks and shovels? Yeah, they also put landmines in the hills where all the gold is. Good luck out there!”