Licensing was the obvious first door and Turvey knew the roadmap. Then he stopped.

Judge William Alsup, reviewing the record fourteen months later, wrote that had Turvey kept up those conversations, he might have reached agreements to license copies for AI training from publishers, just as another major technology company soon did with one major publisher. Judge William Alsup had famously presided over the Authors Guild v. Google fallout and early tech IP cases, and recently handled the Bartz v. Anthropic class action. Turvey let those conversations wither.

What replaced them was a purchase order.

His team emailed distributors and retailers about bulk-buying print copies. Anthropic spent many millions of dollars on millions of print books, often used. Vendors stripped the bindings, cut the pages to size, ran them through scanners, and threw the paper away.

The books cost many millions of dollars. What the licenses would have cost, nobody found out.

What the record shows

The internal program was called Project Panama.

A planning document described the goal without decoration: destructively scan all the books in the world. The court records came out in January, and the Washington Post reported them.

Snopes rated the viral version mostly true with undetermined elements. I thought that to be about right, but I would not have considered that the verified part would actually be worse than the exaggeration.

What the court said

Two rulings have been constantly conflated.

First, Alsup found that training a language model on books is fair use. He called the use ‘exceedingly transformative’. Separately, he found that converting a lawfully purchased print book into a digital file was also fair use, for an entirely different reason: The digital copy replaced the print copy. No new copies entered the world. One went in, one came out.

That second ruling runs on first-sale doctrine, the rule that says once you buy a physical object you may do what you like with it. It’s why used bookstores and used music & audio stores are legal and why libraries can lend. It exists to stop publishers from controlling a book after they’ve already sold it.

When I ran supply chain for a tech startup in Cambridge in the ‘90s, I had to do that conversion too. The product called for User Manuals and we printed them: fat little tech guides, written by engineers who didn’t need them rather than those who could speak with a VoC voice. Soft-bounds in lay-flat or spiral binding, page ‘blues’ approved by physical review, print (4-weeks), packed and shipped with one for every two users.

Then I moved the operation onto a Xerox Docutech and put the library on disk. Just-In-Time (JIT) printing, packing and shipping. Ship what you need and save some trees. Goodbye paper. Hello books-on-disc. The customer got documentation, the cycle took care of itself and the manual and its intellectual property lived on.

So that ruling describes something real, and I’d defend it.

Then there’s the money. Anthropic paid $1.5 billion to authors, roughly $3,000 per work stolen. That settlement covered pirated digital copies pulled from LibGen and PiLiMi. It had nothing to do with the shredding.

But the shredding for Project Panama held two pricing models:

Download the book without paying: $3,000 a title.

Buy the same book used, extract the text, pulp the paper: The cost of the book and the postage.

The law put a meter on acquisition and left everything downstream free. An industry looked at that and pivoted accordingly. No moral compass required and no laws to break.

The piracy came from the same instinct, but earlier. The order quotes Dario Amodei saying Anthropic had many places from which it could have bought the books, and preferred to steal them to avoid the legal, practice and business slog. That phrase eventually cost $1.5 billion. It’s also an honest description of what obtaining consent feels like from inside a company in a hurry.

Destructive scanning was a choice

Destroying the books was a choice, and there was a viable alternative on the table.

404 Media, reviewing Turvey’s deposition, reported that one of the firms Anthropic contracted to do the scanning was Datamation, which advertises both high-volume destructive and non-destructive book scanning. Google Books, the project Turvey came from, used a patented camera process that left books intact and returned them to the libraries they were borrowed from.

Why Anthropic chose the destructive service is the one thing the record doesn’t answer. The internal documents don’t say. What’s known is that cutting the spine and feeding loose pages through a scanner runs faster and costs less than photographing a bound book page by page.

So there were two forks. License or buy. Preserve or destroy. Neither decision broke a law, and only one of them is explained anywhere in the file.

The 1909 compulsory license

As vocalist and band leader for 50 years, I perform other people’s writing. Some songwriting early on, then decades of singing material written by people I never met whose music inspired me. But I never did get to ask them if that was ok. I expect ti would have been.

Regardless, this was legal and it was legal by design.

In 1908 the Supreme Court decided White-Smith Music Publishing Co. v. Apollo Co.

Player pianos ran on perforated paper rolls, and the question was whether a roll counted as a copy of the sheet music. The Court said no. A roll was part of a machine, so the composer was owed nothing.

Behind that case sat the Aeolian Company, the largest maker of mechanical instruments in North America. Aeolian had signed contracts with music publishers for the exclusive right to convert their catalogs into rolls, then bankrolled White-Smith’s suit as a test case. A win would have handed one firm the machine and the sole right to everything it could play.

Congress answered inside a year. The Copyright Act of 1909 made mechanical reproduction a right of the copyright holder and subjected that right to a compulsory license in the same breath.

Anyone may reproduce a published musical work, the holder cannot refuse, and the rate was fixed at two cents a copy. Aeolian got nothing. That provision runs today as Section 115, and it is the reason ‘cover bands’ exist.

Historians disagree about how much of it was anti-monopoly and how much publisher lobbying but either way, a court ruled that a machine’s output owed the writer nothing, and the legislature overruled the court within twelve months.

That covers the record. The stage is another matter, and here I have to tell on my own analogy.

Live performance runs on a separate right, and the venue buys it. The club pays ASCAP or BMI a blanket fee that licenses the room for anything in the catalog. Distribution is where it comes apart. ASCAP pays live royalties off the top-grossing tours as reported by Pollstar, which by its own account reaches somewhat over 2,500 acts a year against a membership above 1.1 million.

Those under that line have to file their own setlists through ASCAP OnStage or BMI Live to see a cent. One writer’s statement, published by his administrator, worked out to about a dollar per song performed in a small New York club. So: fifty years in properly licensed rooms, singing songs whose authors mostly collected nothing for that particular night, because nobody wrote down what I played. That’s the honest version.

The machinery for consent at scale has existed for one hundred and seventeen years. Built with real intent, it still pays the top of the market and ignores the bottom. Whatever gets built for books will be designed by people with the same instincts, and the writers who need it most are the ones the survey won’t see.

The part that belongs to writers

The weak version of this argument goes first.

Writers say their books were used without permission and that nobody asked. For the pirated copies that’s exactly right, and it cost the perpetrator.

For the books the companies bought, permission was never part of the deal. Buying a book and reading it has never required anyone’s permission, and neither does buying it and analyzing it. Or buying it and burning it. A lawyer who hits the word ‘permission’ stops reading.

So here is the version that survives.

When you publish, you grant permission to be read. That's the transaction a writer believes they're in. Somebody pays, somebody reads, and the original thought moves from one head to another. Every part of publishing assumes that shape.

What happened here is that someone bought the object, extracted the reasoning, and discarded the paper. There’s no clause to invoke and no party with any standing. The law literally has nowhere to put it.

In 1908 a court told composers that a machine’s output was not a copy. Congress overruled it within a year. In 2025 a court told writers that a machine’s output was transformative. Nothing has followed. What you sold was a reading. What was taken was a training run.

The Haarlem email

Pieter de Vries runs De Vries & De Vries, an antiquarian shop in Haarlem. He got an email from a woman named Natalia at a company called 2077AI.

A fairly large order, it said. Match the titles against the attached spreadsheet, quote prices, estimate shipping to China. The spreadsheet held 3,001 ISBNs. The email never mentioned AI.

De Vries deals in old and rare books. When an order comes in via email which they rarely did, it’s for a single title. He read the first few lines, marked it spam, and moved on. He found out what he’d declined weeks later, when a Dutch journalist called to ask about it. Several other antiquarian dealers in the Netherlands got the same email and made the same call.

Fortune Magazine asked 2077AI what the books were for and got no answer. Two findings come out of that pointing in opposite directions.

The list was mostly 2020 and 2021 academic titles from Emerald, Elsevier, Wiley, Routledge and Oxford University Press. Business, education, engineering, public policy, medicine. Current monographs, ordinary stock. So the heritage panic is in this instance may be overcooked, and I’d rather say so myself than have a reader catch me on it.

The second finding is the one that matters. A list of current academic monographs was mailed to a rare-book dealer in Haarlem, which means it was generated from ISBN records with nobody checking who would receive it. The buying runs on machine-built lists routed through intermediaries.

The seller can’t see the end use, the paperwork doesn’t state it, and the request arrives in email looking enough like spam that a professional deletes it. Consent was never refused here. It was never solicited, and the sourcing was built so it wouldn’t have to be.

The preservation line

The industry’s answer has been a promise to protect the rare material.

Elon Musk said he’d asked his team to preserve rare books in a library and scan them the hard way instead of cutting the spines. But, here’s how to test it.

A commitment to non-destructive scanning shows up as a purchase order. Volume, vendor, rate per page. That document either exists or it doesn’t, and to my knowledge nobody has published one. Until someone does, the preservation framing of Elon Musk remains just that, a sentence which costs him nothing.

Apply the same test to the sourcing.

In July, 404 Media reported that ISBNdb, which maintains the barcode database behind most books in print, had put up a page pitching bulk acquisition to AI labs. The world’s best AI training data is sitting on a shelf, it said. It also said the optics problem was real, and that “AI company destroys two million books” is not a headline that generates sympathy.

The page has since come down, and the company says the service never launched.

A bookseller told 404 Media his volume went from 20 books a week to hundreds. He was straightforward about the split: the money helps, and he doesn’t like that uncommon books are being pulped.

What I’m left with

I use these tools for a lot of things. I write non-fiction so research, validations of sources, and intellectual fracking. That’s the disclosure, and I’m confident enough in my own voice it’s easier just to put it here.

Back to Turvey and the emails. Anthropic asked publishers what it would cost to license the books. The publishers were willing to talk. That November, HarperCollins offered its authors a deal. Microsoft would use their nonfiction backlist to train a model for three years. The publisher emailed authors one book at a time asking permission, and the fee was $5,000 a title, half to the writer and half to the house.

So there is a price for asking. Twenty-five hundred dollars to the author. The same book bought used and pulped costs a few dollars. The writers who got those emails were not impressed.

One of them, Daniel Kibblesmith is a comedy writer. He spent five years on The Late Show with Stephen Colbert, which brought him five Emmy nominations, and he writes comics for Marvel. In 2017 HarperCollins published his children's picture book Santa's Husband.

Seven years later his agent emailed asking whether he'd license that book to the AI deal for $2,500. He posted the email on Bluesky and answered in one word: abominable.

Licensing was there in the spring of 2024, and Anthropic walked past it. The court file is full of people working out whether this was legal.

There is nothing in it about whether it was decent.

Meeting adjourned.

The conference ? Great similarity to the ‘dotconomy’ I had worked before, lots of adjacent-technology vaporware, three-kids in a garage with a Rent-A-Booth and an uber-amped demo machine making promises built on good ideas but with not a lot of models in production.

But when the conversations got honest, when the people you came to hear hit the stages, the Sharpest People in the rooms kept landing on the same hedge: they were about 75 percent sure this would actually work.

The leading minds. Seventy-five percent. Okay… lets see where this goes.

Then the receipts started arriving. In 2025, MIT researchers reported that roughly 95 percent of enterprise generative AI pilots were producing no measurable return. Companies were spending, experimenting, and mostly stalling.

Now hold all of that against one more number. Between that 75-percent conference and the painful-adjustment quote, OpenAI’s annualized revenue went from roughly $2 billion to roughly $25 billion. While the experts hedged and the pilots stalled, the money grew about twelvefold in a little over two years. That last number deserves a harder look, because of where it sits in the ‘ol chain.

Think of the AI bet as a supply chain.

That is the mechanics of every supply chain: revenue flows upstream, risk flows downstream, and risk concentrates wherever it can no longer be passed along. Nothing new here folks, we’ve seen it before inside the dot-com era, and the loudest voices for the ‘new economy’ were the ones monetizing the belief itself.

I was one of them and when the Great Correction came, the believers absorbed it, and many of the preachers had already banked. My options papered the bathroom wall.

None of this makes Altman wrong about AI.

It means his calm and his warning both come from the top of a chain that gets paid either way, while the adjustment he now predicts will be absorbed at the bottom.

Smooth at the top of the chain. Brutal at the top and that my friends is the point.

From thirty thousand feet, technological revolutions look smooth. The revenue curve and the Davos calm are the view from the plane.

The 75 percent hedge and the failed pilots are the view from the ground, where people actually live and work.

I have been writing about ecosystem disruption for a while now, and the same yes ‘patterns’ keeps returning. So let me put the scoreboard on the table. Here is who lost, who won, and what came out the other side.

The Scoreboard — Six Revolutions

The printing press, 1450s. The losers were the monastic scribes who copied books by hand, and the Church that held a monopoly on interpreting them. The winners were a rising literate middle class, the reformers, and eventually everyone. What came out of it was mass literacy, the Reformation, and the scientific revolution. It also produced the first great flood of misinformation in the pamphlet wars, so the problem of messy, democratized information is roughly six hundred years old.

Mechanized textiles, early 1800s. The losers were skilled handloom weavers, the original Luddites, who smashed the machines because the machines really were taking their living. The winners were factory owners first, then urban workers once wages finally climbed. What came out of it was cheap clothing, and because the early conditions were so brutal, the first labor laws, the weekend, and mass public schooling. The catch is that wages stagnated for close to a generation before they rose. Economists call it the Engels pause. The benefits were real. They were also late.

Farm mechanization, early 1900s. In 1900, roughly 40 percent of Americans worked in agriculture. Today it is 1 to 2 percent. The losers were farmhands, sharecroppers, and small farmers, and the displacement helped drive the Great Migration. The winner was everyone who eats. What came out of it is arguably the single largest source of modern prosperity, because it freed the overwhelming majority of people from subsistence farming and collapsed the share of income a household spends on food.

The automobile, 1900s to 1920s. The losers were the entire horse economy: blacksmiths, farriers, stable hands, and the proverbial buggy-whip makers. The winners were autoworkers, the suburbs, and anyone who wanted to move. What came out of it was mobility and freedom, and it quietly solved a genuine public health crisis, because cities were choking on horse manure. It also created new problems: sprawl, pollution, and tens of thousands of deaths a year. New technology tends to trade old problems for new ones rather than end them.

Telephone switchboard operators, 1920s to 1960s. “Operator” was one of the largest categories of women’s employment in the country. Automatic switching erased hundreds of thousands of those jobs.

But the most useful twist for anyone thinking about AI came later, with the ATM. Everyone predicted the cash machine would wipe out bank tellers.

Instead the number of tellers grew for years, because ATMs made branches cheap enough to open more of them, and tellers shifted into sales and relationships. Automating a task is not the same as ending a job. Tehn, sometimes it is.

The internet, 1995 to 2010. The losers were newspapers, whose classified-ad revenue Craigslist gutted, along with video rental, travel agents, record stores, and encyclopedia salesmen. The winners were the platform companies and consumers.

What came out of it was nearly free access to almost all information, and entirely new ways to earn a living that nobody could have named in 1990.

What These Crises Gave Us

Strip the six down and…

The jobs destroyed were disproportionately the dangerous, tedious, backbreaking ones. The price of whatever became abundant collapsed.

Brand new categories of work appeared that would have been gibberish beforehand, since “web developer” means nothing to someone standing in 1985. And the disruptions repeatedly forced protective institutions into being that we now take for granted: labor law, public education, the safety net.

Crisis has a way of building the scaffolding that the next generation inherits.

What These Crises Cost

Here is the part the view from the plane leaves out, and at the risk of sounding like an AI-tell - ‘they don’t want you to know.”

The people displaced almost never captured the gains. It was usually their children, or a different group altogether, who benefited. A handloom weaver in 1810 did not retrain into a prosperous factory foreman. His grandson maybe did.

The gap between the moment of loss and the arrival of the payoff is measured in decades, not quarters, and the question of who absorbs that gap is not an economic footnote. It is the entire political fight, then and now.

The Hinge — America in the 1890s

If you want a single decade that shows the whole shape, take the 1890s.

In 1890 the Census Bureau announced that the frontier line was gone. Three years later a young historian named Frederick Jackson Turner stood up in Chicago and argued that the frontier had made the American character, and that its closing changed everything.

That was the view from above.

The view from the ground was the Panic of 1893, one of the worst depressions in the country’s history, a farmer revolt that became the Populist movement, and strikes at Homestead and Pullman that turned violent.

For most of a century, the answer to displacement had been simple: go west. The 1890s were the first decade where the displaced had nowhere to go but the cities, and the country shook while it absorbed them.

I agree with Turner but hold that decade close for a personal reason.

My grandmother was born on a farm in 1890, the year the frontier closed, and she lived until 1993. One lifetime, one hundred and three years, 5 months, 20 days.

She started in a kerosene-lit, horse-drawn, farm-majority America and lived through nearly every entry on the scoreboard above: electrification, the automobile, the tractor that emptied the farms, radio, television, the jet, the moon landing, the personal computer. She adapted to all of it, and she had more than a century to do it, one revolution at a time.

That is the quiet variable in every story we tell about adaptation. Time.

The Real Question About AI

So the honest question is not whether we will adapt. We always have. The question is narrower, and it comes down to two in this case, exceptionally powerful words:

Speed and Breadth.

Every prior wave hit one sector at a time and played out over decades. That pace left room to reallocate across a career, or across a generation.

The worry now is compression: change arriving faster than a fifty-year-old can retrain. The reach is massive and scale is almost not a factor - technology that lands on cognitive work broadly rather than on a single trade.

Serious people split on whether that makes this genuinely different or simply the same story running faster.

Notice that even Altman, the optimist, now concedes the near-term adjustment will hurt. His two statements were never in conflict.

The machinery always changes. The human is the one who lives through the change, and that is the part no revolution has ever automated away.

Human at the core.

dw

Vision without action is a daydream. Action without vision is a nightmare. - zen saying

I was inside four companies that met the dotcom crash the way the Andrea Gail met the storm in The Perfect Storm, a Gloucester boat in a movie that opened in June 2000, the same summer the market went under.

Unlike the swordboat, the four companies each got a different ending. One got absorbed. One went extinct. One adapted. One ate everything in sight and then got eaten

A quarter century later, the Dotconomy veterans I know are all noticing the same thing. The AI buildout is producing the same rooms and the same option math.

Option math is Schrodinger’s compensation. Your grant is simultaneously a fortune and a punchline, and the state collapses fresh every time you check. In 2000 we checked hourly.

One piece of period machinery sat underneath all four stories and hundreds more.

The market-cap projection did the raising: a forward estimate of what a company would be worth, future stock price times expected shares, real money raised today against a valuation that had not happened yet. The outcome was that a company did not need a product to be valued, bought, and sold.

The Dotconomy ran on vaporware.

The people running the vaporware were young, smart, and new at it: engineering founders more than business builders, some of whom had never faced a customer, and few of whom wrote anything down. The whole apparatus clung to them like lampreys to a shark: the Market, the VCs, the analysts and the influencers before we had a word for them.

Start with the room I remember best.

Absorbed: ATG

Art Technology Group was a real company. Cambridge software, an e-commerce platform that major retailers actually ran, revenue that went from $32.1 million in 1999 to $163.3 million in 2000. A five-x year.

Their first dress code was: Wear shoes in winter.

The Spend flew like the curve would never bend, and the purest proof, as it often was for every doctom at some point, the ATG User Conference Paris, 2000. ATG had sunk millions into its 2000 user conference with swag, dinners, first-class travel, gifts and parties all booked, and somewhere in the run-up the question of calling the whole thing off actually came up. Bolt-on governence or OMG financial realization. Too much had already been spent.

It went ahead. Sales meetings by day, nights like a party at Caligula’s Parisian villa with no budget, and everyone checking the stock price on a BlackBerry in between. The founders were worth about a billion dollars each on paper, a figure that moved day to day with the ticker.

Back home, a recruiting department of about thirty people, internal and full time, hired as fast as the reqs could clear, a large, long blackboard on the backwall keeping score.

I came in after the first stock split, employee #88 with options struck at $48 a share. A stock only splits when the market has already decided it is going up. I arrived on that signal, at the price it produced.

Then one day the desks started emptying.

I should point out there were two screens being watched at that time and through the slide, almost nobody in the building was watching the company. The company was still shipping. But on the Blackberry screen, people were watching their own portfolios go down by the hour.

There was a second screen worth watching. Through the same stretch, the executives were selling. Stock went out by the millions in the windows when insiders were cleared to sell, every sale legal and disclosed on the required schedule, the rats jumping fiscally overboard while the rest of the building sat on options that were underwater and sinking like pre-plastered wallpaper.

And the official record shows why the room felt that way.

By October 2001, as part of the layoff ATG cut 160 jobs, a fifth of the company. It reported a $9 million quarterly loss where there had been a $5 million profit the year before, on revenue down 35 percent. Restructuring charges for 2001 alone came to $75.6 million, more than double a quarter’s revenue.

The irony of layoff day: The stock closed at $1.70, and it rose 20 cents on the news.

If you hold those two numbers next to each other, a $48 strike and a $1.70 print is not options being down, that is options being GONE with no recovery on any horizon that pays them off. Every person at those desks could do the arithmetic, and the arithmetic came out the same no matter what the quarterly results and the C-Suite said.

And we wonder why nobody read the earnings release. (ok, here is what the earnings release said anyway): Thirty million dollars still came in that quarter. The company told the market it would be profitable again within a year and would finish 2002 with more than $50 million in cash, and it was not bluffing.

It cut to survivable size, kept the product, kept the customers, and spent nine unfashionable years rebuilding, most of them under Bob Burke, a Digital Equipment veteran who ran the turnaround through to the sale. In November 2010, Oracle paid $6.00 a share, about a billion dollars, because e-commerce was a gap in Oracle’s product line and ATG’s platform filled it.

The value was real, so the value survived. The independence did not, and neither did the option math that had hired all those recruiters.

Extinct: Opus360

Before ATG, I had already lived the faster version.

In late 1999 I was running call center operations at Opus360, the company behind FreeAgent.com.

The pitch was free agency as the future of work: an exchange matching independent consultants with companies buying project help, more than 32,000 registered consultants by the time we went public.

My team built the member services center from nothing in about six months. New building, new machines, new telephony and infrastructure… GoldMine CRM, computer-telephony integration, a Nortel Meridian PBX.

The phones were live for the IPO.

The IPO raised $77 million, 7.7 million shares at $10.00, and it priced on April 7, 2000. The Nasdaq had peaked on March 10. Twenty-eight days separated the top of the market from our first trade, and nobody inside could feel it yet.

Then, the layoffs came fast, mine included, and by July 31, 2001 the company was gone as an independent entity, reverse merged into Artemis. Fifteen months, IPO to disappearance.

The timing was fatal but tht alone did not sink the ship. Instead, within a year of the IPO came a securities class action. And there was a smaller item that tells you even more: stock options issued to FreeAgent employees that may never have been registered the way securities law requires, cleaned up later for about $100,000.

Small money. The kind of miss that shows the checking function had fallen behind the burn rate. Bolt-on governence.

Intersting, nothing about the thesis was wrong. The gig platforms built exactly this business a decade later and made it enormous. Opus360 had the right idea, a boom-time burn rate, and governance still being assembled while the market turned. Being right about the future turned out not to be a defense.

The idea outlived the company by twenty years and counting.

Adapted: the paper route

Newspapers were supposed to be the prey in this era, and the numbers agreed.

Help-wanted classifieds brought the industry $8.71 billion in 2000. Classifieds overall ran about 40 percent of newspaper ad revenue, and that margin is what kept subscriptions for you and me, cheap. For the newspaper industry, the reader was never the real customer. The advertiser was.

By 2002, help-wanted had fallen to $4.38 billion. By 2011 it was $743 million, less than Monster and CareerBuilder booked between them, and lower in inflation-adjusted terms than 1977.

Publishing took that hit but it did not cave. It went hybrid and absorbed the shocks in order: the crash, then 9/11, then 2008. I had my own eleven months on the bench in there, mid-2001 into 2002, so I take the survival part personally.

I knew the bycycle-to-doorstep delivery interface end of this business before any other part of it from age 12. Decades later I saw the distribution side at industrial scale at PCF, Publishers Circulation Fulfillment, the operation that put the Globe and the Times on doorsteps around Boston, more than a million papers a day at its peak.

What was that boring delivery layer actually worth?

January 2016 answered in public. The Globe moved its home delivery from PCF to a cheaper vendor, ACI Media Group, and flipped the switch overnight on December 28.

Within days, tens of thousands of papers were going undelivered. About 150 routes had no driver at all.

The new routing software drew routes that made no sense to anyone who had ever thrown a paper onto a porch. Two thousand subscribers cancelled in the first stretch.

Globe reporters ended up hand-delivering their own Sunday editions.

The CEO conceded that an overnight cutover was ‘more disruptive’ than anyone had imagined, which is what an executive says when an operating model was assumed instead of tested, and not a peep of governence.

Within two weeks the Globe handed half the routes back to PCF. Within a few months, all of them. The company the Globe had priced as a commodity turned out to be what actually got the paper onto the doorstep every morning.

Without the drivers and the routes, a newspaper is a website with a $10 million printing press.

Predator to Prey: Monster

Monster spent the late nineties and early 2000s eating the classifieds column alive. Then it did the thing evolution predicts but very few business books do. It partnered with the prey.

I ran the first 57 partner engagement and integration meetings of Monster’s newspaper alliance program, all of them inside six months, three of those months on the road, often two or three states in the same week.

Papers that had watched Monster drain their help-wanted revenue sat down and did the deal anyway, because keeping a share of a shrinking market beats keeping your pride. When I checked recently, the Media Alliances group listed more than 200 newspaper partners but cannot speak to it’s current activity.

For part of that stretch I was also working the other side of the story at PCF. Same years, both worlds: the company that had gutted newspaper economics, and the company that physically delivered the newspapers.

Nobody planning a career would design that overlap, but as 20/20 understanding what happened to this ecosystem, I could not have designed it better.

The papers ran the same keep-them-closer play from their own side. Tribune and Knight Ridder built up CareerBuilder as the industry’s counterweight, with Gannett joining later. So by the mid-2000s, the two apex job boards were partly owned by, and partly partnered with, the industry they had been eating.

None of it saved the predators from the next wave however.

In September 2024, Monster and CareerBuilder merged, a defensive combination between two companies the market had already forgotten about. Nine months later, on June 24, 2025, the merged company filed Chapter 11 with $2.2 million in cash against roughly $400 million in funded debt.

The job boards that once out-earned the entire American newspaper help-wanted market sold at auction for $28 million. About $930 million in private equity investment went ‘pop’.

The court filings gave the cause of death. The company had failed to keep pace with AI-driven hiring tools and job aggregators.

Decoded:

Indeed made job listings free to search and effectively infinite, which killed the paid slot Monster sold. LinkedIn turned the resume database, the asset recruiters paid Monster to search, into profiles that update themselves.

And matching moved inside the employer’s own hiring stack, where human-created AI screens rank applicants on arrival, and human programmed ATS systems exclude specified candidates leaving a destination job board with nothing left to sell on either side of the transaction.

The predator’s own obituary names AI as the thing that ate it, and the AI that ate it is the perfected form of Monster’s founding idea, matching people to jobs at scale. I do not have to argue the bridge from that era to this one. It is in the docket.

Eating the incumbents does not exempt you from becoming one.

The same sort, running now

So run today’s AI field through the four slots. Skip the argument about whether it is a bubble. Ask instead which ending each company is built for.

The ATGs have real product, real revenue, and a cost structure priced for a raise that never stops arriving. They will not vanish. They will lose their independence, and their employees’ option math, and resurface inside something larger with a logo change.

The Opus360s have a thesis, a burn rate, and a governance function still being assembled. Some of the theses are correct. It will not matter. The market picks the turn date. The paperwork is the only part they choose.

The adapters are the unfashionable incumbents quietly going hybrid right now, the way publishing did. Count on them to take the shocks in order and still be standing, with almost nobody writing it up.

And somewhere in today’s field, a company is doing to an incumbent what Monster did to the classifieds, too busy counting that revenue to ask who is building the cheaper, automatic version of its own product. Right behind them.

I think we know how that movie ends.

This will not stay inside tech either by the way.

It did not last time. Power companies are building plants for data-center demand that so far lives mostly in forecasts, forecasts big enough to power tens of millions of homes. The landlords carry the same bet a step further: buildings leased to one tenant and nearly impossible to repurpose if that tenant leaves. A few chipmakers live or die on whether one customer keeps spending.

The help-wanted money left the papers, print runs shrank, and delivery routes consolidated out from under drivers who never owned a single share of tech stock.

It reached the barstools too. I lived in Hopkinton MA in those years, EMC’s hometown, and a favorite watering hole was a Mexican restaurant with a parking lot full of Ferraris, Lamborghinis, and Porsches.

One day the cars disappeared. Then the restaurant. South Street became a ghost boulevard of lost fortunes, lined with empty EMC buildings. EMC itself pulled through, rebuilt as a serial acquirer, one of the survivors.

Nobody reopened the restaurant.

What the crash could not kill

One more thing survived 2000, and the doom coverage always skips it but no governance or program manager misses. Capacity.

The fiber overbuild of the late nineties bankrupted a generation of builders, and then that same fiber became the backbone of everything that came after. Kohlberg Kravis Roberts (the ‘Barabarians At The Gates’ folks) has made the identical argument about today's data centers: the capacity endures even when the builders do not.

I watched the small version of that. The delivery routes PCF ran outlived the decade that marked them for zero. ATG’s platform outlived its stock and now runs inside Oracle. The free-agency idea outlived Opus360 and became an industry. Even the Monster brand came out of the auction still operating, under owners who paid $28 million for what a billion dollars could not hold together.

Companies are temporary. Capacity compounds. After two crashes, I no longer ask whether the bubble is real. I ask which of the four fates my desk is attached to, and whether what I build will outlive the ticker.

It is the question I would put to anyone drawing AI-scale option grants right now, in a building where the catering is very good and the burn rate is somebody else's problem.

Another repeatable pattern.

The MIT Media Lab's “The GenAI Divide: State of AI in Business 2025” report reveals that while $30-$40 billion was invested in Generative AI,

95% of organizations failed to see a measurable return, with only 5% successfully scaling projects.

MIT identified three root causes as the primary drivers behind the failure rate in AI pilots, rather than technology shortcomings. (1) Inability to redesign workflows for AI (2) Difficulty moving from pilot to production (3) Misalignment between AI deployment and high-value business processes

Yet, the organizations that need this work most don't need a full-time executive.

They need someone who has done it before, (and before that, and before that) for 90 days. or 120 days. But Right Now. Like now.

That's the pivot. My Consultancy for Orchestration Fractional Leadership is live.

More on what 30 years of technology transitions taught me -- coming this week.


#FractionalLeadership #CCaaS #AITransformation #DigitalTransformation #ContactCenter #JobSearch