Buried in the 8000 words I wrote last week was a worrying story — that Microsoft considered drastic measures to free up capacity in its US-based servers for GPUs to power the AI boom.
In an email shared with me by a source from earlier this year, Microsoft's senior leadership team requested (in a plan that was eventually scrapped) reducing power requirements from multiple areas within the company as a means of freeing up power for GPUs, including moving other services' compute to other countries as a means of freeing up capacity for AI.
While it's hard to imagine, I want you to consider for a second that no company is invulnerable, and as I wrote last week, I have serious worries about the current state of Big Tech and the path it’s currently on.
My last newsletter focused heavily on the ramifications of big tech — primarily Google, Amazon and Microsoft — bankrolling the generative AI boom, both through investment and by heavily subsidizing the rates that OpenAI (and likely Anthropic) pay to run their services, and through pouring billions of dollars of venture capital into AI companies.
What I didn't get into as much depth about was the why — why these companies are so desperate to put generative AI in everything, and how doing so might be genuinely dangerous to their future.
Because, let’s face it, every generative AI product we’ve seen from Big Tech has been underwhelming. They, for the most part, boil down to offering summaries of stuff, finding stuff, or putting a chatbot (which they call an "agent") on stuff, which can, in turn, find you stuff. Sometimes, when they're feeling spicy, a service can help you search for stuff in your organization too, then summarize that, or even come to conclusions based on it, which will be generated using generative AI, which hallucinates.However, to really explain, I need to tell you about how a large chunk of the tech industry makes money, and how the growth-at-all-costs Rot Economy thinking makes the software you use in the workplace so much worse.
I need you to understand that while companies like OpenAI and Anthropic are the more public faces of generative AI, and consumer-facing tools like ChatGPT are the PR vehicles that catapulted them into relevance, the actual revenue — the real dollars — that have underpinned almost every hyper-growth market have come from finding a way to sell it at a monthly cost to the enterprise.
Hey Ed, what about all the discord at OpenAI? Mira Murati (their CTO) left, along with two other execs! I will get to it next week.
In fact, maybe it's worth considering the enterprise as the commercial real estate of the tech industry. And yes, it's time to start getting worried.
Software As A Disservice
Forgive me, I am about to explain something that seems quite niche, but actually affects a remarkably large part of your life.
Core to the business models of multiple tech companies is the humble "SaaS" (software as a service) model, where you're charged a monthly amount on a per-user basis for some sort of cloud-based software that you neither own nor control. To be clear, there's nothing inherently wrong with SaaS. For businesses, it reduces costs by removing the need to run their own infrastructure and employ people to maintain it. It can run theoretically from anywhere and costs are measurable, predictable, and adaptable to the organization's size. And, crucially, businesses don’t have to pay upfront for a license — which can cost thousands, or tens of thousands — and spread the cost across the life of the application.
You can see the appeal. SaaS is one of the most dominant business models in tech, because it fits both the customer profile of "not wanting to run a bunch of infrastructure" and the tech industry's love of trapping people in distinct ecosystems that are hard to escape. While SaaS is generally a good deal for small-to-mid-sized companies, the inevitable sprawl of letting SaaS into your organization means that you're stuck with them.
While managing 100 accounts might be something that your organization can do alone, how are you going to manage 1000? Or 10,000? Managing SaaS applications is a time-consuming and tedious process for large businesses, and now there are even — you guessed it — SaaS applications that can do it for you. What happens if your organization is in Europe and needs to be GDPR-compliant? What happens if you need to make sure your data is held on a server entirely separate to the rest of the company's business? While some SaaS companies offer private cloud (where the application exists on its own dedicated AWS or Azure instance), giving companies the flexibility to choose where and how to store their data, many don’t.
This is the devil's deal of the three-trillion-dollar Software-As-A-Service market (and SaaS spend is expected to crest over $230 billion in 2024). While the convenience of not having to build your own distinct software run on its own distinct hardware is great, or having to pay ungodly sums upfront for software licenses, you are also effectively outsourcing your entire organization's functionality to another company. With every new integration, every new seat, every new add-on their sales team makes you pay for and every new product they graciously train your staff to use, your organization becomes more burdened by the beast of SaaS.
The bigger you are — or the longer you stay — the more powerful the parasite becomes, eventually burdening your organization to the point that you are effectively only as innovative as the SaaS provider you're anchored to.
As an aside, "SaaS" and "cloud" kind of get mashed together sometimes because the whole idea of cloud computing is that you run stuff on other companies' computers, meaning that you have to rent it from them. This is tech's rent-seeking arm, and business was, for a while, booming.
Also, the reason I'm so focused on the sale of cloud software is that compute can, at times, be a smokescreen. As organizations build their own AI integrations — through either paying for compute for their own purpose-built solutions or by integrating OpenAI's GPT — there is likely to be some revenue growth in cloud providers. That being said, we're yet to see this actually happen, outside of OpenAI's alleged $3.4 billion of yearly revenue (it still loses $5 billion a year), and generative AI's multi-billion dollar contribution to Amazon Web Services revenue over an indeterminate amount of time.
I'm telling you all of this is because it's the answer to why whatever business you're working for has such shitty, shitty software. Not only is it inherently difficult to switch providers (because there's basically no incentive for them to help, and no legal imperative that forces them to), but SaaS companies have built entire ecosystems — like developers that cost six figures that develop just for Salesforce and entire companies that exist to sell other companies' (like Microsoft)'s software — specifically to make sure that anybody scaling a business inevitably ends up using their software. Almost every major SaaS company offers some kind of developer program, which in turn creates an entire cottage industry of people who exist entirely to help monopolies extend themselves.
The reason that so many "channel partners" (the organizations built entirely to sell software from Microsoft, Google, Salesforce and other cloud companies) exist is because they handle the smaller, less-profitable (yet just as time consuming) markets, allowing big cloud companies to hire hundreds of thousands of sales people (or the organizations they work for) entirely on commission.
They also make money — though not as much — by selling a variety of managed services. Microsoft's Unified Enterprise plan offers round-the-clock support at a percentage of your annual spend, and while there's nothing wrong per se with offering specialized, tailored support, it's also important to realize that this support is not something that exists as a means of Microsoft making money, but a means of increasing your sunk cost with Redmond.
The goal of these programs, as I've suggested, is to anchor you to an ecosystem. The more money you spend — hiring developers to code specifically for one platform, paying for specialized support from one platform, adding users and add-ons and storing data and doing stuff on a system — the more costly it becomes to leave, and said cost becomes only more burdensome the larger your organization becomes. While not all Software-As-A-Service companies are parasites, the incentive for these companies is less about serving your needs and more about making sure they're so deeply embedded in your organization that they're impossible to remove.
In 2011, Marc Andreessen said that “software is eating the world.” And he was right.
How does this manifest in your daily life?
Well, have you ever used a piece of software at a company you work for that sucks? Was it sold by Microsoft, Salesforce, Google, Atlassian or another big SaaS company? Well, it was probably bought by somebody who doesn't use the software, and it'll cost far more to remove than your annoyance matters. The burdensome presence of software like Microsoft Teams or Salesforce Platform in your life is a result of these organizations using brand recognition to sell into your organization, and once they're in there, their sales teams exist to continually find ways to increase the revenue of each user. The people making the decisions about the software you use — usually C-level executives — are doing so based on a sales pitch tailored to them and their preconceptions of what your job is rather than any firm experience, and thus they will sign year(s) long contracts based on a great sales pitch and the financials that "make sense."
As a result, software from companies like Microsoft and Salesforce only has to be good enough to fool the purchaser, and perhaps a few troublesome "users" that might get trotted into negotiation where both parties may or may not have already agreed it's happening. The classic adage "Nobody ever got fired for buying IBM" is how big SaaS companies make their money, and once they've got you, it's so very, very difficult to make them leave, because the cries of agony about their questionable software come from people altogether separate from those making the decisions.
In fact, a great deal of software decision-making appears to be made entirely for show. Take Microsoft PowerBI, which Nik Suresh (who you might remember from the excellent "I Will Fucking Piledrive You If You Mention AI Again") calls "A Human Rights Violation."
PowerBI is a "business intelligence platform" sold as part of a $54.75-a-head-a-month software package (with an annual commit) that helps you create dashboards of data from your organization that you can, I assume, make judgments based on, or as Suresh puts it, something you "connect [to] spreadsheets, then you drag charts onto a page, and it makes graphs that [he] can't even describe as pretty." Suresh found, in the organization he was working in, that dashboards were barely-viewed, and worse-still didn't auto-refresh, meaning that most people viewed the data and didn't actually see if it was recent, or relevant, or useful.
The thing is, PowerBI doesn't exist so that you like it or that it's useful, but so that you pay Microsoft a monthly fee and don't pay Salesforce a monthly fee to use Tableau, which offers "the world's broadest, deepest analytics platform" for the low, low (starting) price of — and this is for the regular version of Tableau — $15 a user a month, at an annual commitment, and that user can only view stuff, you're gonna need to pay $75 a user a month if you want to do everything with Tableau.
As a result, it's just cheaper to use Microsoft's faux-Tableau no matter how hard it Tablows, because it's part of a bunch of other apps that you need like Word, PowerPoint, and Sharepoint, an "enterprise content management platform" that everybody hates that allows you to build internal documents and host files in a way that makes it near-impossible to share anything outside of your organization. But what're you gonna do? Pay for a whole other thing when Microsoft has you by the balls? You stupid asshole. Get out of my office.
The reason I'm drawing such a vast, ungainly picture of the SaaS market is because behind every single stupid workplace application you're subjected to is the parasitic existence of big tech's SaaS parasite. Whatever organization that's burdened you with some sort of half-baked, half-useful piece of shit business app has done so because the people up top don't care if it's good, just that it works, and "works" can be an extremely fuzzy word. It doesn't matter that Microsoft Teams is universally-loathed and regularly threatens to crash every time you load it. A Microsoft salesperson used its monopoly power to cut your boss a deal to either bundle it with a bunch of other mediocre shit or they saw the name "Microsoft" and said "oh boy! I love Microsoft Word!" and pulled out their credit card so fast it left a gouge in their monitor.
If you're wondering why Zoom has remained the same rat king of mediocre software and labyrinthine settings, it's for the same reason — while its business might be serving this software, it's never really sold to the people who have the misfortune of using it regularly or having to administer it.
This problem only compounds in the public sector, where tech integrations are bought by people who don't use them, and the only thing more languid than a sales cycle is an organization's ability to change. The same goes for banks, or any other highly-regulated space, because said services have to be so customized and cost so much both in time and dollars invested, especially with the vast amounts of reporting and data access controls required in anything touching financial or medical data. Not to worry, big tech's software parasites have entire arms dedicated to these expensive specialized solutions, and can provide them across a number of different mediocre apps that nevertheless make it easier to stay compliant, making it borderline-impossible for smaller companies to compete.
On top of that, SaaS pricing is intentionally vague, and gets more so the larger and more demanding your organization's needs are. While it might say something is $50-a-month-per-head, perhaps you need to make sure data stays in America, or that you have extra storage, or special permissions for some users, or you need some users to access some apps and not others. All of these start costing money in a way that is impossible for you, the customer, to calculate, but will magically seem to increase every year.
Your business software will stay the same level of mediocre because it doesn't have to get any better. The SaaS parasite gets fatter and harder to move the larger you get, trading the convenience of having the software "all under one roof" for a product that you'd actually like.
To be clear, I'm not saying any of this precludes you from buying other software, and in fact products like Microsoft 365 (Redmond’s rat king of software options) will somewhat plug into other services...assuming you can work out how. Perhaps you can use something like Zapier, a piece of software that exists to automate actions between your massive stack of software you own for some reason, because said software doesn't actually do what you need it to do, and is rarely built to be interoperable (as in having its data easily operated on by another piece of software), because making software interoperable would allow you to leave.
The reason I've spent so much energy explaining this is because I want you to know why so much business software fucking sucks. If you're reading this, I'd put the likelihood that your company (or a partner) makes you use some sort of janky, horrible business software that sucks, and the reason it sucks is because this industry is built off the back of outpricing the competition, creating an inescapable gravitational pull on your organization and selling to people who don't actually use it.
Just look at Salesforce, a company that made $1.43 billion in the last quarter selling software that I've never been able to get someone to describe to me. Salesforce has a conference that costs thousands of dollars to attend, that attracts 40,000 people a year, that has hundreds of booths and hours of presentations, and yet does so many things that it effectively does nothing. Perhaps it sells Customer Relationship Management software, or some sort of AI-powered chatbot you can plug into your data, or "MuleSoft's unified platform" which "provides the essential building blocks for connected experiences" by "integrating all your data, systems or AI models" to let you "automate any task." Mulesoft, which Salesforce bought for $6.5 billion (or one OpenAI funding round) in 2018, is yet another vague business automation tool that Salesforce can sell, another tool for growth, and another tool that locks you into their greater ecosystem.
Trust me, this affects your day-to-day life too. Every single one of these companies wants to absorb you into their ecosystem, creating the business equivalent of listening to Apple Music on your iPhone while playing a game from Apple Arcade or using an app downloaded from the App Store, with each purchase making it even more unthinkable to leave. Subscription-based businesses feel like a great deal — $9.99 a month seems a much better deal for Microsoft Office than a one-off payment of $149.99 — but what you're really doing is owning nothing and paying more, though Microsoft offers 6 terabytes of cloud storage on top and 60 minutes of free Skype calls, which is the kind of thing you can do when you're running your own massive cloud empire.
Subscriptions like these have been bleeding into consumer technology for years, because they create annual recurring revenue (ARR), which is catnip for investors, and very easy to use to show growth. Yet "enterprise" SaaS companies — those that target businesses of 1000 employees or more — are the prize pigs of the rot economy, theoretically able to grow forever because every customer can perpetually be made "new" by upselling them new features or integrations as their organizations grow.
But really (as with consumer subscriptions, but at a much, much larger scale) each new integration and strategic move further weighs the SaaS company to the organization. You stay on your iPhone (at least I do) because you've got years of data that's specifically baked into Apple’s ecosystem, but imagine if instead of your messages, it was years of internal organizational data, all connected using tools specifically built for one system, specifically integrated to make sure it all worked together. Enterprise-grade SaaS providers make billions from the Hotel California business model. You can certainly give notice on your contract, but you can never truly leave.
And the SaaS business model is valuable. The combined market cap of the top five public SaaS companies is just under a trillion dollars, and every single company in the top 50 is worth over $4.5 billion. As I said before, there's nothing inherently wrong with the idea of cloud-based software or SaaS. It's impractical to, say, build an entire CRM or productivity stack inside your organization, as would it be to maintain it and keep it secure. These companies — and this industry — have grown out of a fairly noble idea that companies at different scales need different solutions that grow and change with them, and cloud infrastructure is a reasonable idea.
There are, however, two major problems.
- As enterprise SaaS companies grow, so too does the pressure to grow faster, and the only way to do that forever is to build tons of extra products you can hock to your customer so that they don't take their business. This means that large SaaS companies don't do just one thing, but continually find new things to do as a means of making it more convenient (and cheaper) to stay with them rather than work with a competitor, even if the product is inferior.
- Take Dropbox, for example. Dropbox Paper (which comes free with Dropbox accounts) is a clunky Google Docs clone, but it kind of makes sense. We store and work on documents, right? As does HelloSign, which allows you to sign stuff on Dropbox. But wait, why is there now a product where you can search all of your company's internal documents using AI? Because Glean, a company that lets you search all of your company's internal documents with AI, was growing rapidly, threatening Dropbox in some sort of indeterminate way (it's now worth $4.6 billion). It doesn't really make sense for a cloud storage product to do this, but whatever, Dropbox need line go up!
- As an aside, Glean has now raised two rounds of over $200 million $4.6 billion in this year alone, but only had $39 million of "annualized" (meaning projected yearly) revenue as of February 2024. If I had to guess, it’s about as deeply unprofitable as a SaaS company has ever been, and there is no real turning that around.
- Take Dropbox, for example. Dropbox Paper (which comes free with Dropbox accounts) is a clunky Google Docs clone, but it kind of makes sense. We store and work on documents, right? As does HelloSign, which allows you to sign stuff on Dropbox. But wait, why is there now a product where you can search all of your company's internal documents using AI? Because Glean, a company that lets you search all of your company's internal documents with AI, was growing rapidly, threatening Dropbox in some sort of indeterminate way (it's now worth $4.6 billion). It doesn't really make sense for a cloud storage product to do this, but whatever, Dropbox need line go up!
- As a 50-person organization, perhaps it's fine and dandy that you've outsourced everything — your cloud storage, your sales software, your accounting software, your communications software, your advertising software, and most crucially your security software - but once you get to 500, or 1000 people, your entire business is distributed amongst anywhere from five to fifty different software vendors. Every time someone leaves the company, you lose the ability to really understand how your system works, and likely will need to retain someone to both understand the system and possibly even buy software specifically to manage your spend. As a result, it's just easier to use a company that does multiple things, even if it sucks.
You may once again be wondering why I'm going over this, and as I've stated above, understanding this bullshit is how you understand why so many workplace systems suck. Almost every single app and service you use is run by a bunch of little interconnected services, all of which are distributed between a bunch of other service providers. Almost every company you call is using some sort of SaaS platform sold to their its that may or may not function well that it will never, ever replace because replacing it requires both ripping out what's already there and actually understanding how it works, which is much harder when so many organizations have adopted the shareholder supremacist mindset of "laying off people to increase margins."
Seriously, there are entire companies that sell "DevOps" software to help you actually manage all of this bullshit. It's ridiculous. There are over 30,000 different SaaS companies now, and so many of them do the exact same thing, existing as different sides of a coin, with one saying "come and be with us and buy all of our stuff so it's easier on you" and the other saying "our stuff is actually good, buy it because your other stuff sucks," with the quiet part being "but we'd love it if you stayed forever."
Microsoft in particular has monetized this chaos by selling you an entire suite of apps. In a world of chaos, partially created by its hand and maintained on its cloud, Microsoft can bring you the simplicity of having all of those apps in one place, run by one company, with a person you can pay six figures to maintain them for you.
This is the underpinning of both the entirety of the tech industry and a large chunk of the valuation of public tech companies — trillions of dollars of market capitalization is held up by the business model of outsourcing your infrastructure and software and being charged for it on a monthly basis, and inventing new ways to keep you "investing in your infrastructure" by agreeing to pay them a little more a month.
And the reason I've so agonizingly explained it is that I believe the current generative AI boom is almost entirely fueled by the hunger of the Software As A Service parasite for growth...and the slow collapse of growth in tech's favorite business model.
You see, much like the rest of the tech industry, the Software As A Service industry hasn't got any hyper-growth ideas left.
A quick note on growth: I'm about to share a lot of numbers that may seem a bit weird, but it's important to understand that the Rot Economy's growth-at-all-costs demands hit the cloud software/SaaS market particularly hard. These companies are expected to grow, and grow fast, and effectively grow forever at a rate that defies logic and, in turn, creates the worst incentives possible.
Feeding The Parasite
According to research from ChartMogul, SaaS ARR (annually recurring revenue) growth has steadily declined since the height of the pandemic, with a "record low" median year-over-year growth rate of 23-24% — numbers that seem high but are more than half the rate that these companies saw in the middle of 2020. The only segment of the SaaS market that saw growth since 2023 was within companies making between $8 million and $15 million of ARR.
The top quartile has done fairly well, sitting at 85% year-over-year growth (but down from an incredible 197% year-over-year growth at the beginning of 2021), but that rate is plateauing. Overall, ChartMogul notes that growth is stabilizing, but on a much lower tier.
It's also important to note that growth does not mean profit, just like revenue doesn’t necessarily mean profit, and research suggests it's becoming more expensive to acquire and retain customers.
In a paper published by Winning By Design, researchers Jacob van der Kooij and Dave Boyce suggested that SaaS companies may be losing their "Go To Market Fit" — meaning whether your software is something that the market will continue to acquire — and that Net Revenue Retention (how much you're making from customers and expanding their spend minus what you're losing from customers leaving/cutting spend) is dropping, which could be truly lethal for SaaS:
“NRR is a key metric that reflects customer satisfaction, loyalty, and the value customers derive from the product. A declining NRR (Ref. 3) shows customers are expanding more slowly, downgrading their subscriptions, or churning at a higher rate. A reduction in NRR is a critical indicator of degrading GTM Fit, implying the product no longer meets its customer base's ongoing needs or expectations.”
“The decline in revenue growth, coupled with the escalating cost of acquiring new revenue and the significant drop in NRR, points to a likely loss of GTM Fit. Due to the persistent degradation of these metrics over two years, this appears to be systemic, meaning there are no quick fixes. Companies must reevaluate and possibly reinvent their GTM strategies”
I realize this is all a bit technical, but the very basic thing to know is that SaaS companies must grow, and their way of growing is growing and upselling ("expanding") customer revenue, and that specific thing is declining. SaaS sales teams are built to land accounts and then grow them using "customer success" teams that find new ways to "get more out of the software" using the customer's credit cards, which is why so many SaaS companies acquire completely different business units (like Salesforce, which just acquired a data protection and management company for $1.9 billion for some reason) as a means of further penetration into the customer's existence.
And if that business model is dying — if customers are no longer as easy to upsell, let alone retain — then things are going to get desperate.
Worse still, the growth in spend on SaaS has been dropping since the end of 2022. A BetterCloud research report recently showed that for the first time ever the number of SaaS applications in the technology stack at companies decreased, dropping from an average of 130 SaaS applications on average per company (which is an incredibly large amount!) to an average of 112 in 2023. And, to be clear, this is the first drop in its kind in over a decade.
Year-over-year revenue growth has slowed in almost every major SaaS company that reports it - it's down at Atlassian, DataDog (where it's effectively flat, with slight signs of life), Okta, Salesforce, Snowflake, Hubspot, ServiceNow, Workday, and Shopify, and in almost every case has been trending steadily downward since 2022.
An aside: The reason I'm not calling out Microsoft, Adobe or Google here is because they deliberately obfuscate where their revenues are coming from by tying products into large units like Microsoft's "Productivity and Business Processes" unit, which includes LinkedIn, Microsoft's consumer Office subscriptions, and Microsoft's enterprise subscriptions, making $19.6 billion last quarter.
So, the main way that over a hundred multi-billion dollar — and parts of multiple multi-trillion dollar — companies make money is potentially collapsing, or at the very least contracting, likely because (as with the rest of the tech industry), there just might not be that much left to sell. The rise of cloud computing has been truly revolutionary, and for better or worse a massive boon for the tech industry, but now they're running out of ways to grow, either through getting new customers or charging their current ones more.
Want another sign that they're desperate? A study of 16,000 SaaS vendors from last year found that prices increased an average of 12% Year over Year, and that 73% of SaaS vendors had increased their prices in 2023. Buried within the study was another worrying point — that renewals were now taking twice as long to complete, leading vendors to delay negotiations until close to the renewal of the contract as a means of pressuring them into closing quickly or losing access to a necessary service.
"But Ed!" you say, "we're thousands of words into this and I'm still not sure what it has to do with AI!"
The answer is simple: the tech industry's desperate attachment to artificial intelligence is largely fueled by the SaaS industry, because AI is the first meaningful new "thing" they've had to flog to customers in quite some time, and because so many of these solutions are sold in bulk on annual contracts signed by people who aren't the end user, artificial intelligence feels like something that they can put on top of another solution and claim it's new.
Theoretically, AI seemed like manna from heaven. It was an entirely new industry that you can either sell services into (like data storage and processing) or build services using. AI has a near-mythical pedigree that makes braindead CIOs and CEOs that don't really know (or care) what their people do all day sit up and say "wow, I need to make sure we're using the bleeding edge of technology" to investors and partners.
The real dream, of course, is that artificial intelligence would help SaaS companies sell employees — a monthly (or annual) subscription to an artificial intelligence that means you don't have to pay a person to do something, which in turn burdens you to the organization (and their model).
This is why Marc Benioff is so excited to hype "Agentforce," Salesforce's alleged "autonomous AI agents" that cost customers an incredible $2 a conversation. It doesn't really matter that these agents are basically the same "chatbot-that-connects-to-large-database-of-information" thing we’ve seen a million times, but this time connected to a Large Language Model. It's a new thing, a new magical tool that Benioff can sling to current customers of Salesforce's current chatbot.
Every single major SaaS player appears to have some sort of AI doodad. Datadog has its own foundation model, ServiceNow offers "a library of enterprise agents you can customize to fit your workflow," Atlassian promises to let you "do the impossible" with Atlassian Intelligence by "transforming teamwork with the power of AI-human collaboration," Snowflake is "delivering a unified platform for secure deployment of LLMs and ML models" while letting you create your own generative AI applications," and Hubspot "Breeze" claims to "make AI easy" by "powering your entire customer platform" with a "complete Ai solution" that "gives customer-facing teams all the AI tools needed."
Workday? Why, its "illuminate" product is "purpose-built to move finance and HR forever forward" and is "fueled by more than 800 billion transactions a year," and of course Shopify has a chatbot that can generate stuff.
You'll notice, of course, the very obvious problem: that it isn't obvious what any of these AI-powered products do, and when you finally work it out, they don't seem to do that much. Every company promising "AI agents" is really just trying to rebrand the concept of chatbots and provide a front-end to let you connect them to data sources and give them guardrails so they don't scream curse words at your customers or write poetry about how bad the company is. Or offer to sell a $60,000 truck for a single dollar.
These "agents" are branded to sound like intelligent lifeforms that can make intelligent decisions, but are really just trumped-up automations that require enterprise customers to invest time programming them.
In fact, you can boil down a lot of the new AI-powered SaaS "revolution" to offering some sort of AI assistant to either a customer or an employee, then calling it an "agent" because it can connect to a few different databases and handle a fairly limited subset of interactions. Almost every one of these companies is claiming that these systems "reduce drudgery" or "enhances productivity," yet there's no real explanation as to how it might do so or what the drudgery is that's being relieved, nor any unassailable metrics on how productivity improves.
And, more worryingly, they're building all of these tools on top of generative AI, which is extremely expensive and, as I've argued, unsustainable at the rates being charged. This leads SaaS companies to go down one of two paths:
- Use AI as a means of luring people into buying or renewing their premium contracts, and thus offering it for free (as Atlassian is doing, and Box is kind-of doing by offering a limited amount of "tokens" to use AI in its products for free)
- Charging an absolute shit-ton for it, as Microsoft is doing with Microsoft 365's AI add-ons for its software.
Scenario 1 is obviously dangerous, offering a tool that costs a great deal of money to provide only a little value to customers for free, meaning that at best you'll have customers that lose you money every single time they interact with the product. This strategy requires companies to sell so much extra software that they somehow make up for the amount of money they're burning running generative AI.
Scenario 2, however, is the most revealing. As I mentioned in my last newsletter, Microsoft has been trying to charge anywhere from $30 to $50 extra a month per head, with an annual commitment, to add "copilots" to Microsoft 365 subscriptions, each one with a nebulous benefit of "efficiency" and "productivity" and a product that boils down to "using an AI assistant to either write emails or provide insights about a person based on a database" — though I'm not mentioning Copilot Studio, Microsoft's $200-for-25,000-messages-a-month Copilot builder where once again you, the customer, are burdened with finding a use case.
As The Information reported a couple of weeks ago, Microsoft has categorically failed to turn Copilot for Microsoft 365 (a product line that makes at least $15 billion of revenue a year) into a meaningful revenue driver, with “between 0.1% and 1% of the 440 million existing users of Microsoft 365...paying for the new AI features" according to an equity analyst's estimation. The Information reports this is because of bugs with the software, as well as "most people [not finding] it valuable right now."
Nevertheless, The Information estimates that Microsoft is making hundreds of millions of dollars from Microsoft 365 Copilot add-ons, though one has to wonder how many of them renew. And I have one other very serious question: is $30-a-month even profitable?
About a year ago, the Wall Street Journal reported that Microsoft's $10-a-month Github Copilot (which generates code and suggests changes to the software you're building) loses the company on average $20-a-user-a-month, and in some cases costs Microsoft as much as $80-a-user-a-month. While it's possible that Microsoft could have found ways to make Github Copilot more efficient, this seriously suggests that Microsoft 365 Copilot loses money in much the same way, though generating code is a little more compute-intensive.
Github Copilot is arguably one of the most popular generative AI products, and one of the few that you can describe as "almost useful" — I say almost because a study from the beginning of the year suggested that GitHub copilot was "putting downward pressure on code quality," and Snyk research showed generative code actively proliferated bugs.
Yet even there, growth is slowing. In February 2024, Microsoft reported that Github Copilot had 1.3 million paying customers, but in an internal presentation given to Microsoft employees last week (and obtained by this publication), Satya Nadella announced that Github Copilot now had 1.8 million paying users. While this sounds like a lot, it’s actually a staggering reversal from the previous 30% quarter-over-quarter growth and suggests that momentum has rapidly slowed.
One has to wonder if Microsoft365's "Copilots" are equally unprofitable. I'd wager they are.
Again, growth does not mean profit, and it's important to understand how that dramatically shifts the incentives for much of the business software you use in your daily life. These companies aren't engineered to make products that are good, either as profitable/sustainable enterprises or functional, reliable pieces of software, but as things that can be sold to somebody to "fulfill a need" at a scale that's almost impossible to calculate, contracted and upsold in a way that makes it inconvenient or actively destructive to stop using.
This is why so many of these products feel awkward, or actively punishing. It's why productivity tools like Slack are an overwhelming sprawl of different things bleeping and clicking at you, or why software like Trendkite (used for tracking how people are talking about you on social media and in the news) require you to do endless training to get even the smallest bit of functionality out of them — because the goal isn't for you to love it, but have so much time and money invested in it that you'd rather spend thousands (or hundreds of thousands) of dollars a year than have to go through setting something like it up again.
As a result, generative AI kind of felt like the perfect SaaS snake oil, a magical new way to indefinitely charge customers for the thing that they already should have. Instead of coming up with a means of intelligently organizing and prioritizing messages, you can pay Slack (owned by Salesforce) $10-a-user-a-month extra for AI-powered summaries of threads and channels. Instead of having to make Sharepoint a better, more usable product, or empower users to make better internal content, Microsoft will charge $200 a month to add the ability to build a chatbot for your Sharepoint pages. Salesforce doesn't need to make it easy to access consumer data at a glance. No, you have to upgrade to Salesforce's $500-a-user-month "Einstein 1 Sales" product, billed annually, to quickly ask a generative chatbot what it is you're meant to be doing and hope that it doesn't hallucinate, or "confidently lie" as Salesforce product executive Patrick Stokes said earlier in the year.
These products almost always include some sort of half-hearted generative addition too — creating presentations or drafting emails or creating boilerplate content, all things that you can do for free with ChatGPT, Claude or whatever other Large Language Model you can find. Or, even without using an AI tool, and instead making a template for your documents, or downloading an app like TextExpander.
It's unclear what Workday's recently-announced generative AI features will do, but it’s hyping the same shit as everybody else: "accelerating common tasks" ("content creation and summarization"), "delivering real-time AI assistance in the flow of work" (which links to the "Workday AI assistant." Hey, another chatbot!), and, of course, "transforming entire business processes with AI orchestration" which will "will provide every user with a "team" of business process experts, or agents," which, you may have guessed, means another chatbot.
What does Box's AI do? Well, it summarizes documents, "creates content in seconds," and, uh...well...that's about it. What does ServiceNow's AI do? Well, it "helps employees, customers, agents, and developers work smarter with Now Assist, our out-of-the-box GenAI experiences," which you'll be shocked to hear involves creating "agents" that can operate like chatbots that connect to databases, as well as a tool for creating your own custom chatbots. What does ADP's AI do? It offers "ADP assist," a chatbot that offers "smart, personalized insights" by connecting to your data on ADP's servers. You know. Like a summary.
There are exceptions — for example, Autodesk alleges that you can "interact with Maya scene data using natural language text prompts" — but it's unclear what that actually means in practice.
While looking for this, I saw that Qualtrics, a survey company with an $11 billion market cap, has an AI product. Qualtrics' AI, you will be stunned to hear, "delivers actionable insights and practical recommendations," and allows you to "leverage XM-specific AI for optimizing customer support across your business," which in practice means "AI-recommended coaching plans" that also "surface location-specific insights for your managers."
It's unclear how Qualtrics makes money from this product.
In fact, that's kind of the problem with the SaaS-AI boom. It's unclear whether any of this generative AI software is useful, and that's largely because it's unclear whether we really need our emails summarized, our emails written for us, or whether users actually want some sort of chatbot to ask questions. While it might theoretically be useful to search for something across the sea of information on an organization (which is why Glean has raised so much), it isn't clear whether this is a sustainable business (as mentioned above, they only make $39 million a year), or what's stopping any number of big companies from copying the entire product.
The other problem, as I've discussed again and again, is that generative AI is expensive, a cost that more-often-than-not means that these questionably-useful services are being sold at a premium at a time when, according to a study by Zylo, SaaS spend declined 10% year-over-year. As a result, SaaS sales people are having to sell extremely-expensive add-ons to budget-conscious decision-makers, and said add-ons are equal parts underwhelming and commoditized.
The plan, it seems, was to tape a chatbot or some sort of generative process to their software and use it to effectively double revenue. Microsoft365 costs anywhere from $30 to $50 a month per-head, and adding Copilot costs another $30 to $50 a month. Slack costs anywhere from $7.25 to $12.50 a month, and adding AI costs another $10 a month on top. It's that simple. p People want AI and they'll pay for it, right?
Nope! Other than Microsoft — which has barely been able to sell it as it is — I can find no public SaaS company that appears to have attributed any meaningful revenue growth to generative AI. In fact, I think they're doing their best to hide the fact that it isn't selling at all, and costs more money than it makes when it does.
But Ed, I heard...Not so fast! While companies like Amazon live on the edge of truth by claiming that AWS is seeing "a multi-billion dollar run rate" from generative AI, that's more than likely referring to companies running generative AI on Amazon Web Services, which it claims is "seeing very significant momentum in people trying to figure out how to run" on AWS. In essence, Amazon is benefitting from AI in much the same way NVIDIA is. As people dick around trying to work out what this stuff is actually useful for, they need cloud compute to do so.
As I hinted earlier, I see the enterprise and the SaaS market as the commercial real estate arm of the Subprime AI Crisis — an industry desperate for growth attaching itself to an unprofitable, unsustainable technology that they invested in in the hopes that it would fuel a decade or more of revenue growth.
Instead, it feels like hundreds of companies have joined a death cult, making AI-powered summaries, generative features and search so commonplace that they're effectively commoditized, making it an unspoken expectation that everyone has to offer mediocre tools that cost too much to run, thereby making it harder to flog them to customers at a premium. After all, why would you pay more for a feature that seemingly everyone has?
At some point, the putrid margins around generative AI will begin to eat into the already-tenuous profitability of these companies, leading to some, I imagine, having to either vastly increase prices or drop the tools altogether, assuming that the competitive landscape doesn't mean that keeping them is a necessity to compete with others.
If that's the case, many SaaS companies may have added the equivalent of an adjustable-rate mortgage to their tech stacks, all to offer features that are at best sort of cool and at worst actively harmful to a company. And based on the fact that most of these companies have to effectively double prices to offer AI, it's hard to imagine that these features aren't already problematically-expensive to maintain.
Another deeply-worrying eventuality could be a race to the bottom, where growth-hungry SaaS companies either deeply discount or don't charge for AI, pushing their competitors to lower the price on their already-unprofitable products. Much like the feared commercial real estate bubble, SaaS may see its own series of dramatic price cuts to its AI tools as a means of competing and showing growth to the markets...despite the costs staying the same, or possibly increasing.
And, again, much like commercial real estate, AI has locked-in costs, and service providers have little incentive to drop them. Especially in the case of Microsoft, which will likely only ever make money from AI by selling compute.
In all eventualities, there's the most obvious problem of them all: that there doesn't appear to be much revenue growth attributable to these tools, which means that they either have to get cheaper — which would make their costs untenable — or better, which would require these companies to find a way to make them more useful, which none of them seem to be able to do, and is likely impossible. Hallucinations are not a bug, so much as a feature of the transformer model that underpins generative AI. It’s a problem that nobody — not Microsoft, or Anthropic, or OpenAI, or Google — is any closer to solving.
For me to be wrong, there will have to be sudden, remarkable changes in either enterprise software spend, or the capabilities and costs of generative AI itself.
While Marc Benioff may be a masterful carnival barker, Salesforce's "pivot" to "AI agents" is a flimsy attempt to rebrand the same kind of chatbot-maker that several other companies are building, and what I believe will be a failed (and expensive) attempt to turn around Salesforce's slowing growth. And if Salesforce — considered the prize growth-pig of the SaaS hogs — isn't growing, then the rest of the market should be terrified.
It's hard to overstate the significance of a collapse of growth in the SaaS market, as is it hard to overstate how dangerous generative AI is to its fortunes. While these companies had costs before, generative AI is multitudes higher than regular cloud compute costs, meaning that any new revenue growth from this software will be burdened by leveraging an increasingly-expensive solution to a problem that most of them have trouble describing.
And if the revenue never arrives, they'll be faced with the same problem as the rest of the tech industry — that they've run out of ideas to generate growth.
At that point, they'll have to reckon with the fact that there are too many software companies incapable of solving any problem other than "how do we find a new way to charge customers for something?"
