Layoffs will continue until we learn to use AI
Original author: Arnav Gupta, AI Engineer
Baoyu, AI Analyst

In the executive office of our company, a layoff list containing up to 8,000 people lies somewhere. I have a 10% chance of being on that list. In a few days, on May 20th, I’ll find out my fate.

Seeing today’s announcement from Coinbase about “AI layoffs,” I decided to write this article. I made a point of starting before May 20, because I wanted to share some of the most authentic perspectives—free from any personal emotion about whether I stay or go. These thoughts aren’t tied to whether I’m laid off, nor are they limited to just my own company. They come from the genuine voices of my friends working at major enterprises across the industry.

There is now a flood of articles debating whether this new wave of layoffs—widely believed to have begun with Jack Dorsey cutting 40% of Square’s workforce—is truly caused by AI, or merely constitutes “AI-washing” (the practice of using the guise of embracing AI to mask other business failures or the real motives behind layoffs).

I don’t want to overwhelm you with links to news articles and papers—you’ve probably already seen them, or can easily find them with a quick Google search or by asking ChatGPT.

Hyped-up "AI productivity" and elusive evidence

Has AI really made us more efficient? This is a truly controversial and weighty question! If we flip it around and claim that “AI hasn’t changed anything,” I doubt even the most skeptical critics of AI’s value would agree.

In tech companies especially, the rocket-like surge in AI usage is an undeniable reality. Even the most conservative companies, which limit AI budgets and do not provide employees with AI tools, cannot deny that some work is effectively completed by AI—whether employees are quietly using Gemini or Copilot within Google or Microsoft Office suites to edit documents.

As for companies with a forward-looking vision that have plunged into the ocean of AI tokens—the basic units of text processed by AI models, for which enterprises are typically charged based on the number of tokens consumed—such as Uber or Shopify (I’m not including companies like Meta or Microsoft that develop their own large language models, nor Vercel or Cloudflare that actively build AI infrastructure; I’m only referring to pure “users”), their AI usage has gone completely wild.

We’ve become accustomed to it: from 90% to 100% of code generated by AI, to a 2- to 5-fold surge in the number of code reviews (PRs/diffs) submitted weekly, to an annual AI budget of hundreds of millions of dollars being exhausted within just a few months.

However, tech critics and investors like Ed Zitron, Will Manidis, Gary Marcus, and Michael Bury will surely ask you a piercing question: If that’s the case, why haven’t these companies’ revenues grown 2 to 5 times as a result? Why do their apps look almost identical to how they did six months ago? If AI is truly that productive, what exactly have they produced with it? And if they’ve written five times as much code, yet end users notice no difference, what’s the point of all that code? This is an extremely sharp and reasonable question.

Input, Output, and Outcome

We need to take a quick detour into some basic business management principles. When a fast-growing, over-funded, spend-thrift mid-sized company finally runs out of cash, you go to a seasoned CEO for advice. He’ll tell you to bring in McKinsey. The consultants will open their presentation with a blank slide, written in the default Arial font, containing just three words: “Input, Output, Outcome.”

They will explain to you a business truth that everyone understands but often forgets:

Code is merely an investment.

Features, not output.

Users willingly pay for your product—that’s the real achievement.

AI (or products like Claude Enterprise) is essentially a business-to-business software service (B2B SaaS). You'll find that SaaS products vary widely in their pricing and marketing approaches. If a product can directly impact outcomes, they often take a cut directly from those outcomes. Imagine this sales pitch: "Our tool helps you close leads 36% faster. Try it today with just a low service fee of 5% of your sales revenue."

This will absolutely blow away your clients. If, under the same conditions, you closed 100 deals in the past 100 days, now you can do it in just 63 days. The 37 days you save (if my math is right) allow you to close an additional 57 deals! That means your potential sales growth is 57%. Anyone would be thrilled to give up just 5% of their sales commission for a 57% increase in income. And if you don’t use this product, you pay nothing at all.

You might have guessed what I’m about to say—Claude’s token pricing model works completely differently. If your software engineers are addicted to programming with Claude (I just realized both share the initials “cc”), generating 100 million tokens per day, you’d be paying $100 per engineer per day.

Even if some of the code they generated was thrown in the trash because it didn’t run,

Even though some code later caused a severe system incident (SEV) (SEV stands for Severity, a term commonly used by tech companies to refer to critical online incidents that cause service outages) and was urgently rolled back;

Even if some code remains, just to give internal tools a new look so that vice presidents find the data dashboards cuter;

You have to pay for everything in full. Code is merely an "input." While it's generally true that, as long as the direction is correct, more "input" often leads to more "output," which in turn results in better "outcomes," this rule doesn't necessarily hold when you suddenly increase your input by five times overnight. The additional "input" you've added might suddenly become like a fly without a head, completely deviating from the expected "output" or "outcomes."

What exactly is holding us back!

In the past, whenever the CEO or product manager (PM) wanted to do 10 things, the development team would always say they could only manage the two most important ones, and didn’t have time for the remaining eight. Why? Because writing code isn’t child’s play—building a complex, functional software system takes a significant amount of time.

Hmm... but code is almost free now. Why haven't we done the remaining 8 things yet?

There are two answers: one that the CEO and product manager don’t want to hear; another that middle management and senior employees don’t want to hear.

Just because the CEO or product manager has had ten ideas doesn't mean they can actually be turned into real business outcomes. Even if you truly build ten new features (output), it doesn't guarantee that users will adopt them or use your app more as a result (outcome).

In fact, precisely because development resources were previously limited, this "friction" forced everyone to engage in more intense debates, thereby eliminating poor ideas before they consumed too many resources and selecting only the two best ones. Now that writing code is fast and cheap, debating the merits of ideas seems pointless. Even if you try to argue against them, do you really think you can stop the CEO or PM from simply turning around and asking Claude for their own requirements? Don’t even bother trying.

We all know how exhausting this can be. First, you need to get all stakeholders aligned on “why” this needs to be done; then, you need another meeting to discuss “what” exactly needs to be done; and finally, everyone has to debate “how” to do it.

The more teams there are, the more projects get stuck in "alignment hell." In the past, this issue was masked by slow coding. Now, as soon as a decision is made to "do something," someone immediately pulls an all-nighter to build a minimum viable product (MVP)—a product developed at the lowest cost that just enough to demonstrate the core idea—for rapid experimentation, and schedules the next meeting for the very next day.

At the meeting, you were surprised to find that another team had also secretly developed an MVP! Even worse, because you operated under different assumptions, the two products functioned on completely opposite principles.

Of course, you can sit down and take your time to discuss whose assumptions are correct.

But let’s be honest—you and your team, holding unlimited Claude tokens, wouldn’t bother doing it this way. Neither would another team. You’d instantly turn to Claude and have it reimplement the other team’s work exactly as you believe it should be done. And Claude would simply reply, “You’re absolutely right!” and immediately start coding.

What problems can layoffs actually solve?

Alright, thanks for patiently listening to me ramble on about these obvious truths. I know you're here for the core insights.

What is the actual goal of layoffs? According to my assumption, if AI hasn't truly replaced 30% of employees on a one-to-one basis (something we should all agree on—while it outperforms junior white-collar workers in many tasks, it falls short in others—it is not a plug-and-play component, and certainly cannot directly replace 10%, 20%, or even 30% of a company’s workforce).

If that's the case, what's the logic behind the layoffs? Because they immediately address two visible short-term issues.

This is essentially the most basic cash flow arithmetic problem. Clearly, if your engineers are addicted to Claude and spend $100 per day on it (that’s $2,500 per month or $30,000 per year), that amount equals the entire annual salary of a software development engineer (SDE) in India, half of an SDE’s salary in Europe, and a quarter of an SDE’s salary in the United States.

If you make the simplest, most straightforward calculation: assuming a flat company where all employees are SDEs, to maintain the current total payroll expenditure (including Token purchases), you would need to lay off 50% (India), 33% (Europe), or 20% (US) of your staff.

In fact, since AI usage is skyrocketing regardless of everything while the company’s revenue hasn’t seen a corresponding increase, layoffs have become inevitable. Otherwise, the company’s balance sheet would collapse entirely. If your input costs have risen by 50% but your final business outcomes remain unchanged or stagnant, your unit economics across the entire software development lifecycle have completely broken down.

If we had truly learned how to use AI—figured out how to turn a 50% increase in input costs into a 50% increase in revenue—we wouldn’t need to take this step. But since you haven’t learned yet, some of you will have to leave so we can afford to pay Anthropic.

Undoubtedly, the scale of any large company far exceeds what is merely needed for its "survival." This is precisely what characterizes large companies—large organizations are inevitably burdened with "organizational fat," an inevitable outcome of their structural design.

In these companies, even if someone leaves, the system continues to function because others always know what they used to do. In many large corporations, you can even take a six-month maternity leave with confidence—your projects will remain perfectly stable. These are all positive signs! But they also serve as clear evidence: if some people are laid off, the company will not immediately collapse. On the contrary, after an initial few weeks of systemic adjustment, operations will likely accelerate over the following months!

Do you remember the two teams that were stuck over the technical solution? It’s simple: just eliminate one team, and let the remaining team pull a few all-nighters to get the job done—they’ll never have to “align” with anyone again.

We cannot predict what will happen in the long term (or, as economist Keynes put it—“In the long run, we are all dead”), but in the short term, laying off 10-20% of employees at large corporations will only make the pace of work faster.

Over time, large enterprises inevitably accumulate redundancy and inefficiency, just as they accumulate technical debt—they build up massive amounts of “organizational debt.” This is a common ailment of big companies. Cutting 10% of the workforce today won’t prevent the same problems from resurfacing two years later. But when you see everyone boasting that they’re submitting five times more code than before, yet still unable to ship because other teams are blocking them, the most direct and brutal remedy is clear: cut some people, so no one is left to block each other.

This is AI-driven layoffs, even if AI hasn’t directly replaced your position.

Has your employee ID been replaced by a new Claude instance running on a virtual machine? We all know that’s not the case.

Nevertheless, isn’t it true that many workflows in your company—once requiring you to type and click in VS Code, Figma, Canva, or Google Docs—are now simply bypassed as others (the very people who previously relied on you for these outputs) shout a prompt at a large language model and no longer bother coming to you for help? This is also an undeniable fact.

Do these layoffs really count as “AI washing”? In other words—are companies actually facing fundamental issues unrelated to AI (such as over-hiring, declining profits, competitive pressure, or poor business decisions), and are they now using AI as a convenient excuse for layoffs? Well, to some extent, that makes sense.

You might also notice that if you collect all the "layoff emails" sent by CEOs during this period, you could almost think they’ve set up a group chat to coordinate and write them together. “AI-native teams,” “managers who code,” “increased management span,” “flattened structures,” “managing AI agent teams”… you’ll find these fresh terms appearing identically in every email. It’s as if they all fed the same prompt to GPT.

But the truth is that even if these layoffs aren’t directly due to AI replacing you, and even if they’re mixed with elements of “AI washing,” these layoffs are ultimately caused by AI. And this wave of layoffs will continue until we truly learn how to use AI.

Until we learn how to turn massive amounts of AI tokens into tangible business outcomes, not just code investments; until we learn to align organizational speeds with the pace of coding in this new generation; until we figure out how to use this additional productivity to pursue ten more promising new ideas beyond the original two good ones and eight bad ones.

Before we truly understand how AI drives global GDP growth, companies can only cover the annual $70 billion token expenditure (the combined enterprise revenue of OpenAI and Anthropic) by cutting employee salaries—robbing Peter to pay Paul.

Until we learn how to more effectively resolve the bottlenecks between teams, the only solution will always be to erase ourselves from the organizational chart.

In 15 days, I’ll find out my fate. But regardless of the outcome, I think I already know why. Even if I had been the one making the decision in that spacious CEO office in the corner, I’m not sure I could have done any better—I might have made the exact same choice as the other CEOs who formed the group.

Original link

Click to learn about the open positions at BlockBeats

Welcome to the official BlockBeats community:

Telegram subscription group: https://t.me/theblockbeats

Telegram group: https://t.me/BlockBeats_App

Official Twitter account: https://twitter.com/BlockBeatsAsia