Full Circle

Two racks, sixty-two years apart.

On April 7, 1964, IBM announced the System/360 — a family of compatible mainframes priced from about $133,000 to $5.5 million, in 1964 dollars. It was the safest purchase in enterprise computing: proprietary big iron from a single dominant vendor, with a proprietary software layer underneath everything you built. The phrase everyone remembers — nobody ever got fired for buying IBM — didn't come from an IBM ad. It came from the customers. It was the sound of an entire profession de-risking itself.

In 2026, Nvidia sells the GB200 NVL72: seventy-two GPUs fused into a single liquid-cooled rack, roughly $3 million apiece, with a proprietary software layer — CUDA — underneath everything you build. Jensen Huang introduces its successors by invoking the System/360 by name. Enterprises are buying them on-prem, behind their own firewalls, to run models on data they will not send to anyone's cloud. Nobody is getting fired for buying Nvidia.

Between those two racks sits the entire history of business software: custom, then packaged, then custom again, then rented, and now — with the marginal cost of writing code collapsing toward zero — bespoke once more, running against big iron you own. The circle looks complete. It isn't, quite. It's a spiral, and the difference between a circle and a spiral is the whole story.

The first era of business software was custom by necessity. There was no market of applications to buy. You leased the machine, and you wrote — or IBM's systems engineers helped you write — the payroll system, the inventory system, the reservation system, against your exact business, encoding workflows nobody had ever mapped. Organizations weren't adapting themselves to software. Software was being poured, hot, into the shape of the organization.

Here's the detail that gets forgotten: that software was free. IBM bundled programs and systems-engineering services with the hardware at no separate charge — until the Justice Department came knocking, and on June 23, 1969, IBM unbundled, creating priced “Program Products” and, with them, the independent software industry. The custom era existed partly because software had no market price. The industry that would eventually sell you a locked-down workflow was born the day software got an invoice line.

And the labor writing all that custom code didn't yet know what it was worth — by design. Programming was deliberately framed as clerical work. The six women who programmed ENIAC weren't even introduced at its 1946 public demonstration; hardware was the prestigious work, and coding was treated as transcription. The profession spent the next two decades discovering that the transcription was the hard part — a discovery the 1968 “software crisis” made official, and salaries spent the next fifty years pricing in.

One honest caveat before the wheel turns: the eras overlap. Timesharing bureaus — Tymshare, GE's information services division — were selling metered, centrally hosted, pay-per-use computing by 1966. Rented software is as old as custom software. The stages of this story are tides, not walls.

The swing away from the glass house started with minicomputers in the seventies and broke fully open with client-server in the early nineties. PowerBuilder shipped in 1991, Visual Basic the same year, and corporate IT spent the decade feverishly replacing green screens with custom Windows applications on cheap boxes. This — not SOA — is where “highly customized on commodity hardware” actually begins. SOA arrives in the mid-2000s as the cleanup crew: web services, ESBs, and middleware trying to tame the integration mess a decade of two-tier client-server had created. (B2B integration is older still — EDI standards were moving purchase orders between companies by the late seventies.)

The labor, meanwhile, became a culture. The clerical workers of 1964 professionalized into the highest-paid mass profession in history, and organizations that now understood exactly what business value software carried found themselves competing for the people who could build it.

But the era's most important development is the one the “custom era” framing skips: the first cookie-cutter wave happened here, not in the SaaS era. SAP shipped R/2 in 1979 and R/3 in 1992, and packaged ERP conquered the Fortune 500 through the nineties on eight-figure implementation budgets. Fortune called SAP “the ten-ton messiah of enterprise-wide computing” with the Big Six consultancies as “its true disciples.” Locked-in workflows, “vanilla implementations,” adopt-vendor-best-practice ideology, and the high-cost implementation-partner industry — every feature we blame on SaaS was invented by the ERP wave, running in your own data center.

The era closed with a dress rehearsal for our own: offshoring. Y2K remediation handed India's services industry the world's legacy codebases, and for a decade the promise was that the cost of coding would fall toward zero. It fell — and total software costs didn't, because specification, coordination, and rework turned out to be where the money lived. Remember that result. It's about to be relevant again.

The ASPs of the late nineties tried hosting everyone's client-server apps one instance at a time and died of the economics. Salesforce, founded in March 1999, won with multi-tenancy and theater: the “End of Software” campaign, complete with hired protesters picketing a Siebel conference in 2000. Then AWS launched S3 in March 2006, the infrastructure layer commoditized, and for fifteen years the application layer captured the rent. At the peak, the average enterprise ran hundreds of SaaS subscriptions.

What SaaS actually did to the customization question is subtler than “cookie cutter.” It took the ERP era's locked workflow and moved it out of your data center, so that even the optionof deep customization disappeared. Configuration, yes — an ecosystem of implementation partners and consultants (the Big Six disciples, re-platformed) will happily bill you for it. But the workflow itself ships with the subscription. And here is the strategic cost, the one this blog keeps circling: if your workflow is the vendor's best practice, it is also your competitor's workflow. Process innovation stopped being a place you could build alpha, because everyone was renting the same process.

That bargain made sense as long as building was expensive and buying was cheap. Both halves of that sentence just changed.

In December 2024, Satya Nadella said the quiet part on a podcast: business applications are “essentially CRUD databases with a bunch of business logic,” and “the notion that business applications exist — that's probably where they'll all collapse in the agent era.” Bill McDermott — who ran SAP, the original ten-ton messiah, and now runs ServiceNow — joined the chorus. When we first wrote that SaaS was dying it was a forecast; by February 2026 the market was repricing the entire application layer. Median public SaaS revenue multiples fell from about 6x at the end of 2024 to the low 3s by spring 2026, and Salesforce's own Agentforce has cycled through three pricing models in fourteen months — per conversation, per action, per “digital worker” — which is what it looks like when a business model is searching for the floor.

The famous case study cuts both ways, so tell it honestly. Klarna announced in 2024 it was shutting down Salesforce and Workday in favor of internal, AI-assisted systems. It did drop them — but its CEO later admitted they didn't “replace SaaS with an LLM”; they consolidated, built some things, and bought other SaaS. And Klarna's AI customer-service push overreached badly enough that by 2025 it was rehiring humans. The bespoke turn is real. It is not a free lunch.

Because here is the concession that keeps this thesis honest: the cost of coding is collapsing toward zero; the cost of software is not. Architecture, security, governance, maintenance — the things that make code an asset instead of a liability — did not get cheap. Veracode's testing finds AI-generated code carries security flaws at rates that should terrify anyone vibe-coding their ERP replacement, and the nineties already ran this experiment: the custom client-server estates of 1995 became the unmaintainable legacy that the SaaS wave was hired to euthanize. Offshoring taught the same lesson from the other direction — cheap coding never made software cheap, because knowing what to build was always the expensive part. The winners of the fourth turn won't be whoever generates the most code. They'll be whoever directs, reviews, and governs it — and whoever architects the cost structure around it.

And the hardware side of the circle: the long tail of bespoke software needs somewhere private to think. Gartner forecasts $80 billion in sovereign-cloud spending for 2026; Nvidia's DGX line explicitly recreates the old range, from “personal AI supercomputers” on desks to SuperPOD “AI factories” in enterprise data centers; 37signals' very public cloud exit put eight-figure savings on the board for owning your own iron again. A top-of-the-line System/360 cost $5.5 million in 1964 dollars — tens of millions today. A $3 million NVL72 is big iron at a steep inflation-adjusted discount, which is precisely why the circle can close this time: the mainframe came back cheap enough for the upper-middle market. One caveat carried forward from the means-of-production argument: owning the rack still isn't independence. You've swapped a SaaS dependency for a silicon one, and Nvidia collects the toll no matter who wins.

Now the uncomfortable part, because a thesis this tidy deserves a stress test. The instinct is to call AI the biggest efficiency shock in economic history. The measured record, so far, says otherwise. Robert Gordon's data on the American “special century” puts total-factor-productivity growth at 1.89 percent per year from 1920 to 1970 — the age of oil, electricity, and the internal combustion engine. Daron Acemoglu's 2024 estimate for AI is about 0.6 percent — total, over a decade. US productivity growth actually decelerated in 2025 while the AI boom raged; only about a fifth of American firms report using AI at all; and the METR trial found experienced developers were 19 percent slower with AI tools while believing they were 20 percent faster. Energy spending peaked near 14 percent of US GDP in 1981. Oil and gas still books $4–6 trillion a year. The entire AI sector's revenue is perhaps a tenth of that. What rivals oil today is only the spending on AI — roughly $660 billion of hyperscaler capex committed for 2026, more than the world spends drilling for oil — and an economy of capex is not yet an economy.

So the skeptics are winning on points. Here is why the fight isn't over: electrification looked exactly like this. Paul David's famous study showed factory electrification produced no measurable productivity gain for its first thirty years — the gains arrived only after factories reorganized around the new input. And the early sectoral data is doing something: since 2024 the three most AI-exposed US sectors have grown productivity at more than twice the rate of everything else. Judging AI's efficiency realization in 2026 may be like judging the dynamo in 1900. Both readings are honest; anyone selling you certainty in either direction is selling.

What history is unambiguous about is what technology does to knowledge workers, because we've run the experiment a dozen times. The Bell System employed about 342,000 switchboard operators at the 1950 peak — one in thirteen working American women — and automatic switching reduced the occupation to a rounding error; the cleanest research finding is that the labor market adjusted and the displaced individuals often didn't. The typewriter — get the direction right — created the secretarial profession; it was the word processor that killed the typing pool, after office-support work peaked at nearly 13 percent of all US employment in 1980. “Computer” was a human job title until the machine took the name. VisiCalc and its heirs erased some 400,000 bookkeeping clerks and added 600,000 accountants and analysts, because cheap analysis meant everyone bought more of it. ATMs famously grew teller employment for two decades by making branches cheap. The New York Times printed its last hot-metal page in 1978 and a craft that took years to master vanished in fifteen. Travel agents fell by half while travel boomed. Elevator operators — ninety thousand of them — held on for forty-five years after the automatic elevator was invented, until a 1945 strike convinced New York to trust the button.

The pattern across all of them: tasks get automated, occupations get redefined, the parent field usually grows — and the incumbents pay the personal price while the aggregate statistics smile. That is the worker's-eye view this blog has taken before, and nothing about this turn repeals it. But notice the loop specific to our story: programming began as underpaid clerical work, professionalized into the highest-paid mass profession in history, and the fourth turn's explicit pitch is to make coding clerical again — most users of AI coding tools already aren't developers. The value never lived in the typing. It lived in knowing what to build. That was true of the ENIAC six in 1946, true of the offshoring wave in 2002, and it is true of the agents in 2026.

Computing has a name for this. In 1968, Myer and Sutherland described the “wheel of reincarnation”: function migrates out of the center into a specialized peripheral, the peripheral grows until it becomes a computer of its own, and the function migrates back. Centralize, decentralize, repeat. So no — cycles are not news. What makes a turn worth writing about is what's different when you come back around, and this time three things are.

The scarce input rotated again. In 1964 it was compute. In 1990 it was developers. In 2006 it was distribution. In 2026 it is alpha — proprietary data, process, and judgment — which is why the hardware is moving back on-prem: not because the cloud stopped working, but because the one thing AI can't commoditize is the thing enterprises now refuse to ship to someone else's API. The invariants held. Lock-in never dies; it relocates — vendor, package, estate, subscription, silicon. And the consultants survive every turn: the Big Six became the SaaS integrators became the AI-transformation practices, billing at every revolution of the wheel. And the bargain inverted. For forty years, renting your workflow was the rational trade because building was expensive. Now the build cost is collapsing, and the strategic question inverts with it: which of your workflows are commodity — rent those, nobody ever built alpha in expense reporting — and which ones are your alpha, the idiosyncratic processes you flattened to fit the subscription? Those, for the first time since 1992, you can afford to own again. Not because coding agents make software free — they don't — but because they collapsed the cost floor that made bespoke irrational.

The skeptic's line is that nobody is building a homegrown CRM to replace Salesforce, and the skeptic is right — and aiming at the wrong target. The fourth turn isn't about rebuilding commodity software. It's about the million idiosyncratic tools that were never built at all because they cost $80,000 and six months, and about the locked workflows that get to be load-bearing decisions again instead of subscription defaults. Software isn't dead — it stopped being a product and became a capability, and the deep systems of record will outlive every prediction in this post; disruption calls are directionally right and 3–10x slow. But the direction is set. The wheel is turning toward owned workflows on owned iron, sixty-two years after the last time that sentence was true.

Two racks. Same bet: that the organization which encodes its own way of working — instead of renting someone else's — earns more than the market average. In 1964 that bet cost $5.5 million and a programming staff nobody valued properly. In 2026 it costs $3 million and a director who knows what to build. The circle is complete. The spiral continues.

This post is the wide shot of an argument this blog has been making in close-up: Own the Means of Production on where value goes when supply shocks hit, The SaaSpocalypse on the repricing, Software Is Dead on why the timing will be slower than the direction, and Once in a Lifetime on who pays for the turn.