Chapter 1: What Financialization Cuts Is Not R&D, but the Word "No"

Part One: What Shape Is the Structure? ｜ 📖 Who's Collecting at the Door ｜ 📖 Full Edition

---

Chapter 1: What Financialization Cuts Is Not R&D, but the Word "No"

Prologue: An Overlooked Sequence

Most people believe companies cut R&D because they run short of money. That explanation is too clean, and too safe.

The real sequence has one more layer. First, the R&D department shifts from developing new things to maintaining old ones. The industry calls this PE (Product Engineering): patching existing systems, troubleshooting customer problems, fine-tuning specifications. The shift happens quietly. It needs no announcement and never shows up in the financial statements.

By definition, the US accounting standard ASC 730 actually excludes routine improvements, troubleshooting and adapting existing products for customers from R&D. But the company draws that line itself, and external disclosure requires only a single total, with no split between "developing new products" and "maintaining old systems."¹ At many companies the line on the income statement is already called "research, development and engineering." The result is that a company can show R&D spending rising steadily for ten years without developing a single new thing, and no outsider can tell.

The hollowing-out comes before the layoffs.

By the time the layoffs begin, the people being cut are no longer doing research. They are people doing maintenance work under an R&D title. "Slashing R&D" sounds alarming, but the department stopped doing R&D long ago.

Even that is not the decisive step. This chapter is about what financialization actually wants to remove: not R&D, but the word "no."

1. The Veto Is the Target

There is a capability inside companies that is scarcer than the ability to develop things, and more dangerous: the ability to say "no."

The most fundamental function of a technical team is not writing code or drawing circuit diagrams. It is having the standing, the data and the confidence to say "this can't be done" when management commits to a timeline.

Once that sentence is spoken, the numbers stop looking good.

The quarterly outlook looks worse, the story for investors looks worse, and so does the CEO's roadmap for earnings per share (EPS). The whole cycle of "announce, watch the share price react, deliver, announce again" breaks at the first step.

So the problem is not that R&D staff are expensive, though they do sit near the top of the payroll. The problem is that their very presence is an institutional veto on unrealistic promises. As long as these people are there, every grand roadmap hits a wall internally: someone can answer the question "how long will this take, technically?"

Worse, when this veto succeeds, it leaves no evidence. A bad promise that was stopped never happened, the accident never happened, so there is no way to count what was avoided; yet every delay and every instance of "not cooperating" it caused is recorded in full. The success of prevention cannot be verified, while failure can be counted. Chapter 6 returns to this asymmetry of measurement (see Chapter 6).

Pure cost-cutting layoffs keep the cheapest engineers. Layoffs aimed at the veto clear out the whole group that can speak up.

2. Replacement, Not Deletion

In 2013, I watched one complete version of this operation at a company.

First the original R&D team was laid off in full. Then a group of project managers was brought in to manage R&D, and people from lower-cost regions were hired to do the actual development work.

Once those three steps were done, the R&D department still existed on the org chart. Headcount was about the same, and spending may even have risen: project management costs money too, and the communication overhead of working across several locations does not show up under R&D in the accounts. Every external indicator looked normal.

Only one thing was gone.

Project managers are trained to judge not "can this be done" but "when can it be delivered." A schedule can be revised, extended, recommitted, but it never says "this cannot be done at all." Putting project managers in charge of R&D swaps the veto for the power to schedule.

And the execution layer hired from low-cost regions was cut off from the decision-makers by language, culture and even time zones. This is not a flaw in execution but a structural effect: distance itself acts as a firewall, keeping the signal "it can't be done" from reaching the people who need to hear it. The top can promise; the bottom just does the work; the feedback loop is broken.

That was 2013, nearly ten years before ChatGPT appeared.

3. The Numbers You Can See

If this were a one-off case, it would not be worth writing about. It is worth writing about because the whole trend shows up in numbers.

The economist William Lazonick has spent years tracking how S&P 500 companies distribute their profits. Between 2003 and 2012, the 449 companies that stayed listed spent 54% of their earnings on share buybacks and 37% on dividends, leaving less than a tenth for every other use, productive investment included.²

In the twelve months to September 2025, S&P 500 buybacks hit a record of about US$1.02 trillion; dividends over the same period also set a record at about US$664.9 billion, for total shareholder return of about US$1.685 trillion.³

A more precise indicator is the share of companies whose buybacks shrank the share count used to calculate EPS by more than 4% in a year. That means buybacks large enough to lift earnings per share by themselves, even when actual profit doesn't grow. In the third quarter of 2025, that share was 17.1%; a year earlier, it was 13.6%.³

The numbers are rising.

In a 2000 paper, Lazonick and Mary O'Sullivan named this shift: from "retain and reinvest" to "downsize and distribute." Lay off experienced, and usually more expensive, employees, and distribute the cash to shareholders.⁴ This is not a side effect they observed; it is their definition of the shift itself.

4. Promises Settle Later Than Promotions

Once you understand the logic of the veto, a deeper structure comes into view.

Say that in companies like this, a promise takes eighteen months to blow up, while promotions come round every twelve. The person who made the promise is always standing one floor above their own ruins.

No one needs to plan this. All it takes is two mismatched cycles. Announce a grand technology roadmap, watch the share price react, get promoted; by the time the roadmap fails to deliver, the person who promised it is no longer in that seat. The successor inherits the mess, and the fix is to announce a new roadmap.

IBM's "Roadmap 2015" is the textbook case of this cycle. In 2010, then-CEO Sam Palmisano promised operating earnings per share of "at least $20" by 2015, and said about 70% of free cash flow would go back to shareholders, roughly $50 billion in buybacks and $20 billion in dividends.⁵ Note that "operating earnings" was itself a company-defined non-GAAP measure that excluded acquisition-related charges and some pension costs.⁶

In October 2011, IBM announced the succession; in January 2012, Ginni Rometty became CEO. She refused to abandon the target and kept it alive through financial engineering: buybacks, a lower tax rate, layoffs and business sales. In the second quarter of 2013 alone, IBM booked about $1 billion in charges for "workforce rebalancing," that is, layoffs.⁷ On October 20, 2014, in a disappointing quarterly report, IBM formally admitted the $20 target could not be met.⁵

Look at the timing. The person who made the promise left the CEO's chair in early 2012, just as the person who would swallow its consequences took over; four years passed between the promise and its formal abandonment.

This is not a metaphor. It is a case with public dates.

5. The Highly Credentialed Over-Promiser

Once the veto is removed and promises lose their calibration, another kind of person comes in.

The environment in which PhDs learn to estimate has one feature: you control the problem's definition, the scope can be negotiated all the way to the deadline, and the cost of failure is deferred and diffuse. Research genuinely cannot be estimated accurately. That is not a character flaw but the nature of the work.

The break comes with migration. They carry the estimating habits of research into delivery, and delivery requires accurate estimates.

In an organization where the veto still exists, over-promising gets stopped; the technical team says "that timeline is impossible." Once the veto has been replaced, over-promising is not only allowed but rewarded, because a promise is booked the moment it is announced, while failure settles eighteen months later.

One layer deeper: financialized companies need credibility. Once the technical team has been hollowed out, the company loses the ability to judge internally whether technical promises are real, and needs an external, verifiable, third-party-certified credibility to stand in for it.

What is the cheapest credibility you can buy? An academic credential.

It is issued by a third party, it is on the public record, and laypeople are in no position to question it. Putting a doctorate in the CTO or chief scientist's seat costs far less than actually sustaining an engineering team, and the market may react just as well.

The organizational sociologists James Westphal and Edward Zajac studied a phenomenon they called "symbolic management": a company formally adopts a program, while actual commitment is decoupled from it, more symbol than substance. They found that the stock market reacts positively to governance measures that signal alignment between the CEO and shareholders, even when those measures are never actually implemented.⁸

If the market rewards announcing rather than executing, companies have a systematic incentive to buy credibility itself. That is how academic credentials get bought: not as research capacity, but as a signal.

As far as I can tell, the literature has studied "symbolically adopting a policy" but not "symbolically adopting a person." That is a gap.

6. A Loop

Put all these parts together and you get a loop.

Falling returns on real investment bring on a focus on shareholder value. Companies shift from retaining earnings for reinvestment to downsizing and distributing. The most expensive technical staff, who are also the most experienced, leave first. The remaining R&D starts vetoing unrealistic numbers, and the vetoes make the numbers look bad. So R&D is replaced wholesale: not cut, but swapped for a group without the ability to veto. Promises lose their calibration; outside credibility is needed to replace internal judgment, and the highly credentialed are hired as a signal. Their training makes them habitually underestimate what delivery requires, and the organization's incentives reward over-promising. Delivery fails; more layoffs make up the numbers; back to the start.

The loop also has a self-protecting mechanism. People who see through it usually don't stay; those who remain either can't see it or benefit from it. So the fact that "no one with enough weight ever objected" becomes evidence in its own right, and that evidence was manufactured by the departure of those who objected. Critics who do stay meet another response: "you're just jealous." How that move works, and why it is so hard to refute, is left for Chapter 7 (see Chapter 7).

The defining feature of this loop is that every step is a rational response to the step before, and no step requires anyone to be a villain.

That is the part most worth recognizing. No one designed this system with bad intent; the system completes the loop on its own.

7. Every Round of Technology Comes With a Story

Same Name, Different Contents

The R&D department is still there. Headcount is about the same; spending may even have gone up. R&D's share of revenue in the annual report is stable, and analyst reports still praise the company's "R&D strength."

But what goes on inside is no longer R&D.

I have seen the same technique applied to language: the name stays, the contents are swapped, and outsiders can't tell. Passing off fake R&D as real and passing off a replaced language as the original are structurally the same thing.

AI's role here is not that it drove this shift, but that it gave the shift a respectable name. In 2013 you could not say out loud, "I laid off the entire R&D team and replaced them with people who don't know the work." By 2024, "we are using AI to replace R&D labor and improve efficiency" could not only be said aloud but put into an investor presentation. In "The Efficiency Trap: Sequel" I wrote one line about this: AI is not the cause; it is the license.⁹

The Technology Is Real, and So Is the Story

Let me be clear first: AI really is useful. Just as computers were really useful, the internet was really useful, and the railway, 180 years ago, was really useful. Saying that AI is only packaging with nothing inside is not true.

This chapter is saying something else. Every round of genuine technological progress arrives together with a financial story. The technology is real and the story is real; they happen at the same time, but they are not the same thing. Only by seeing this can you tell which changes come from the technology itself and which come from distribution carried out in the technology's name.

Start with the railway. Andrew Odlyzko, a mathematician and financial historian, considers Britain's railway mania of the 1840s one of the largest technology manias in history. The railway transformed Britain; there is no doubt about that. But reliable figures already existed at the time showing that the projected traffic on many new lines could not support the promised revenues, and investors bought anyway, among them Charles Darwin and John Stuart Mill.¹⁰ The bubble burst in the late 1840s. The railways stayed; many of the shareholders did not.

Then the computer. In 1987 the economist Robert Solow left a famous line: you can see the computer age everywhere but in the productivity statistics.¹¹ The productivity computers brought did not show up in the numbers until the late 1990s; before that, they had already become a story. In 1990 Michael Hammer published "Reengineering Work: Don't Automate, Obliterate" in the Harvard Business Review, and "reengineering" quickly became one of the most common banners for the layoff waves of the 1990s.¹² The shift Lazonick described, from "retain and reinvest" to "downsize and distribute," rolled out in full during exactly that period.

Then the internet. In 1998 and 1999, a company had only to announce that it was changing its name to include ".com" for its shares to earn an average abnormal return of about 74% over the ten days around the announcement, regardless of how much the company actually had to do with the internet.¹³ After the bubble burst in 2000, the same researchers found the reverse also held: dropping ".com" from a name lifted the share price too, with abnormal returns of about 64% over the sixty days around the announcement.¹⁴ The internet itself did not disappear; over the next twenty years it reshaped the entire economy. But in 1999, what the market paid for was often not the technology but the name.

This is the mirror image of the case at the start of this section. The R&D department keeps its name while its contents are swapped out; the .com company keeps its contents while a new name is swapped in. Both say the same thing: the market looks at the label, and the reckoning comes many years later.

Why This Round Is So Fast

Each round has moved faster. In Britain, about ten years separated the first railway boom of the mid-1830s from the great mania of the mid-1840s. From the spread of computers to Solow's remark to the productivity finally showing up took more than ten years. From Netscape's IPO in 1995 to the bubble bursting in 2000 took about five.

AI took six months. ChatGPT launched in November 2022; in May 2023, IBM CEO Arvind Krishna told Bloomberg that about 30% of the company's roughly 26,000 back-office roles that don't face customers, around 7,800 jobs, could be replaced by AI and automation within five years, and that hiring for those roles would be paused or slowed.¹⁵ A company spokesman later added that there was no blanket hiring freeze. This is the same IBM known for "Roadmap 2015" in Section 4.

Speed is itself a mechanism. The technology's benefits take years to verify, but the story is booked on the day it is announced. This is the same asymmetry as Section 4's "promises booked at once, failures settled later." The faster a technology arrives, the further verification falls behind; before the numbers are in, the layoffs, restructurings and revaluations are already done. By the time anything can actually be settled, the people who made the announcements have usually moved on.

So speed is not just a feature of this round; it is also a ruler. When the next technology arrives, measure how long it takes from launch to the first layoff announcement that cites it. The shorter the gap, the further the story is running ahead of the technology.

A Pessimism Misread Twice

The mainstream view says AI, like earlier automation, will eliminate some jobs and then create more new ones, and the people displaced will eventually be absorbed. Every so often a report announces how many tens of thousands of jobs have vanished, how many more tens of thousands have appeared, and that the net gain is positive.

I am pessimistic about this. But this kind of pessimism is often filed under Luddism: fear of technology, resistance to progress. That is a misreading, and it is a misreading twice over.

The first misreading is of the Luddites themselves. Around 1811, textile workers in the English Midlands and North smashed machines, but they were not against machines; many of them already operated machines. What they opposed was how factory owners used the machines: to bypass the guilds' quality standards, push down wages, and replace skilled workers with cheap labor that had served no apprenticeship. They were attacking the way the machines were deployed and the way the gains were distributed. "Luddite" was later simplified to "anti-technology," and that is the version written by the winners.¹⁶

The second misreading is of my argument. I am not saying "technology puts people out of work," but "financialization puts people out of work through technology." The difference lies in one key mechanism.

Economics has what is called the compensation mechanism: the costs saved by automation turn into lower prices, new investment and new wages, which create new demand and so absorb the people displaced.¹⁷ That is roughly how the industrial era worked. But this channel rests on one premise: the money saved flows back into the real economy.

Financialization cuts the channel. The savings no longer go into reinvestment; they flow into buybacks, dividends and propping up share prices, which is exactly Lazonick's "downsize and distribute." In other words, AI is not the starting point, only an accelerator: a system that long ago stopped distributing has been handed a more powerful tool.

As for the figures on "how many lost, how many created," they count the number of jobs, not their quality, and certainly not who fills them. The economist David Autor has studied job polarization: middle-income work gets hollowed out, and most of the people squeezed out move down into low-paid service work, not up.¹⁸ A large share of the so-called new jobs is only an illusion that keeps people competing ever harder while they slide downward.

Daron Acemoglu has another concept, "so-so automation": technology that replaces labor without raising productivity much.¹⁹ It doesn't make the pie bigger; it only moves income from one group's hands into another's.

To be fair, the mainstream's strongest rebuttal is that every pessimistic prediction in history has been wrong, and that most of today's occupations didn't exist a few decades ago. So the focus of this debate should not be "are there enough jobs" but "where does the money go." The first is the mainstream's question; the second is mine.

AI did not invent this playbook. The railway, the computer and the internet have all played the same part, and the next round will have a new name.

8. How to Check

If all of the above is only an argument, another argument can refute it. So it needs methods of verification that can actually be carried out. Here are four, none of which require inside information.

First, the gap between product generations. How many years separate two generations of the same product line. A lengthening gap is direct evidence that development capability is draining away. The data is public, checkable and comparable across companies.

Second, patent structure. The ratio of first filings to continuations and improvements. A rising share of continuations is PE's fingerprint: you are maintaining old inventions rather than creating new ones.

Third, the wording of job ads. No one dresses up a job description for investor relations; it is a direct projection of real internal need, it carries a timestamp, and it can be collected over long periods. When a company's senior engineer postings shift from "architecture design, building from scratch" to "maintaining existing systems, troubleshooting customer issues," that is first-hand evidence of hollowing-out.

Fourth, the wording of senior appointment announcements. If a press release spends most of its length on degrees, alma maters and publication records rather than on the specific deliverables the person will be responsible for, the appointment itself is a signal to the outside, not an arrangement for the inside.

9. Two Objections to Address

First, PE is sometimes normal. A mature product line scaling back development and moving into maintenance is a reasonable allocation of resources. The line between pathology and normality lies not in any single product line but in scope: one product line going PE is normal; every product line going PE at once, while the company still tells the world it is a technology company, is the symptom.

Second, every company's failure can be given a technical explanation. IBM was too late moving to the cloud, Intel misjudged its process technology, Boeing's 787 outsourcing strategy failed. None of these explanations is wrong. But they can explain "why it failed," not "why, while failing, it was still buying back shares and paying dividends on a large scale." These are two different questions: technical error answers the first, financialization answers the second.

Separate the two questions and the argument holds.

Summary

Looking back over this chapter, PE, the veto, replacement, the roadmap and academic credentials all come down to one thing: in an organization built around numerical promises, the word "no" has no place. Clearing out the people who say "no" requires no villain, only a structure that rewards announcement and defers settlement; and every new round of technology gives that structure a name it can say out loud. This round, the name is AI. The technology is real, and so is the story; only by separating the two can you see what is actually happening in the world.

This chapter looked inside companies: how the money saved drains away from R&D and from employees. The next chapter looks outside companies: how the same savings leave their mark on the food you eat every day.

As for how the people pushed out of such a structure should see themselves, that question is left for the conclusion.

---

← Preface: Structures Don't Need Villains	📖 Back to Contents	Chapter 2: Maxim's →

Footnotes

1. FASB, ASC 730-10-50-1 (disclosure of total R&D costs for each period); ASC 730-10-55-2 (activities that are not R&D, including routine improvements, troubleshooting, and adapting existing capability for a customer). Text quoted in Internal Revenue Service, "FAQs – IRC 41 QREs and ASC 730 LB&I Directive". ↩

2. William Lazonick, "Profits Without Prosperity," Harvard Business Review, September 2014. ↩

3. S&P Dow Jones Indices, "S&P 500 Q3 2025 Buybacks Post Modest 6.2% Gain to $249.0 Billion…", press release, December 18, 2025. ↩ ↩²

4. William Lazonick & Mary O'Sullivan, "Maximizing Shareholder Value: A New Ideology for Corporate Governance," Economy and Society, 29(1), 2000, pp. 13–35. ↩

5. "IBM: When a Company Veers Off the Roadmap", Advisor Perspectives, January 21, 2015. ↩ ↩²

6. IBM, 2013 Annual Report (Form 10-K, Exhibit 13), Management Discussion, "Operating (non-GAAP) Earnings." ↩

7. IBM, "IBM Reports 2013 Second-Quarter Results", press release (Form 8-K exhibit), July 17, 2013. ↩

8. James D. Westphal & Edward J. Zajac, "The Symbolic Management of Stockholders: Corporate Governance Reforms and Shareholder Reactions," Administrative Science Quarterly, 43(1), 1998, pp. 127–153. ↩

9. Nebula Walker (星忘塵), "The Efficiency Trap: Sequel — You Think You Won" (〈效率陷阱・續章：你以為你贏了〉), Mythogen Engine [in Chinese], July 28, 2026. ↩

10. Andrew Odlyzko, "Collective Hallucinations and Inefficient Markets: The British Railway Mania of the 1840s", University of Minnesota, 2010. ↩

11. Robert M. Solow, "We'd Better Watch Out," New York Times Book Review, July 12, 1987, p. 36. ↩

12. Michael Hammer, "Reengineering Work: Don't Automate, Obliterate," Harvard Business Review, July–August 1990. ↩

13. Michael J. Cooper, Orlin Dimitrov & P. Raghavendra Rau, "A Rose.com by Any Other Name", Journal of Finance, 56(6), 2001, pp. 2371–2388. ↩

14. Michael J. Cooper, Ajay Khorana, Igor Osobov, Ajay Patel & P. Raghavendra Rau, "Managerial Actions in Response to a Market Downturn: Valuation Effects of Name Changes in the dot.com Decline," Journal of Corporate Finance, 11(1–2), 2005, pp. 319–335. ↩

15. Brody Ford, "IBM to Pause Hiring for Jobs That AI Could Do", Bloomberg, May 1, 2023. ↩

16. E. P. Thompson, The Making of the English Working Class, Victor Gollancz, 1963; Brian Merchant, Blood in the Machine: The Origins of the Rebellion Against Big Tech, Little, Brown, 2023. ↩

17. Marco Vivarelli, "Innovation, Employment and Skills in Advanced and Developing Countries: A Survey of Economic Literature," Journal of Economic Issues, 48(1), 2014, pp. 123–154. ↩

18. David H. Autor & David Dorn, "The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market," American Economic Review, 103(5), 2013, pp. 1553–1597. ↩

19. Daron Acemoglu & Pascual Restrepo, "Automation and New Tasks: How Technology Displaces and Reinstates Labor," Journal of Economic Perspectives, 33(2), 2019, pp. 3–30. ↩