Back in 1991, Steve Jobs stood in front of his NeXT team with a chalkboard and laid out a market prediction that still feels eerily fresh. He had just realized the workstation world wasn’t one market, it was two.
The old “science-and-engineering half” as he described it was crowded with Sun, HP, and DEC. But a brand-new “professional half” was emerging, full of business users who weren’t scientists or engineers yet still craved workstation power for custom apps, networking, and real collaboration.
Pro Workstations Became a Thing
Jobs called it the professional workstation segment, and he put hard numbers on it. In 1990 the whole market was roughly 50,000 units (Sun had about 40,000 of them and an 80 percent share). Jobs forecast it would double to 100,000 units in 1991 and then triple to 300,000 by 1992.
Growth would come from two groups which are PC and Mac owners stepping up for better development environments, and mainframe terminal users (remember 3270 emulators?) pulling database apps down to the desktop for speed, better interfaces, and lower costs.
The three things actually driving purchases? First is custom mission-critical applications. The second driver is great productivity tools that spread beyond the core developers. And the third leg of the market stool was what he called “interpersonal computing”, described as real group collaboration over the network.
He even said interpersonal computing would start as an afterthought but become the biggest reason to buy within 24 months.
Jobs was so tuned in to the market sentiment that he said interpersonal computing would start as an afterthought. Keep that in mind as you read on. It means a lot as we move to the AI part of the story.
The whole presentation was a crisp, practitioner-level read on why people upgrade. And it worked for a while. NeXT did land early wins in finance, publishing, and higher ed. The broader commercial workstation slice did expand exactly as he described. But the exact 300,000-unit explosion never quite materialized for the segment as a whole, and NeXT’s lifetime hardware sales topped out around 50,000 units before the company pivoted to software.
Jobs Predicted Enterprise AI Adoption Patterns
Fast-forward 35 years and the story feels like it is repeating, except the “desktop” now is AI-assisted code. Tools like Anthropic’s Claude Code, OpenAI’s Codex-powered agents (the ones powering GitHub Copilot and beyond), and Google’s brand-new Antigravity (the agent-first IDE with multi-agent orchestration) are bumping into the identical three drivers Jobs named. Enterprises see the same 3× velocity on custom apps.
The savvy enterprise buyers feel the productivity pull on the promise of collaboration and efficiency as the upside. Yet the resistance is almost word-for-word the same for AI tools as it was for NeXT workstations.
The key friction points were (and still are) security concerns, governance headaches, integration with legacy systems, and the very real fear of losing control once the mission-critical stuff moves off the mainframe (or the tightly locked CI/CD pipeline).
Let’s walk through where 1991 meets AI adoption patterns in 2026, because I keep seeing it in the field and on whiteboards with platform teams today.
The Custom-App Driver Is Back (and Louder Than Ever)
Jobs said every single customer in that 1991 professional segment had one mission-critical custom app they absolutely had to write. The development environment, networking, and database connectivity became non-negotiable.
Sound familiar? (side note is that this is also a hint of the growth for sovereign AI)
Today the mission-critical app is still the killer. Only now it is the internal platform, the compliance workflow, the customer-facing microservice that has to ship in days instead of quarters. Claude Code, Codex agents, and Antigravity let a single developer (or a small pod) prototype, iterate, and productionize that app at speeds that feel like the 3× boost Jobs’ customers raved about after NeXT’s developer camp.
Early enterprise pilots show that same lift. Internal benchmarks I have seen from a couple of large financial services orgs put Claude Code at roughly 2.8× faster end-to-end for CRUD-plus-business-logic work compared with a senior engineer working solo. Antigravity’s multi-agent setup (one agent plans, another codes, a third writes tests and security gates) is pushing some teams even higher. The whiteboard version looks beautiful as you draw the old loop of “ticket → spec → code → PR → review → merge,” then erase half the steps.
If you think of the phrase “resistance is futile”, this is the where the opposite is true. Hoping there is no resistance is what is what’s futile. The moment that custom enterprise app touches real data or real money, the security and compliance teams show up with the same questions the mainframe folks asked in 1991.
Who owns the IP? Where did the code come from? Can we audit every prompt and every generated line? Enterprises are not saying “no” to the productivity, instead they are saying “not yet, not without guardrails that make the 3× feel like 1.2× again.”
We’ve been given the 10x promise before. That is about as popular (and laughed at) as the “single pane of glass“. Yes, I wrote about that in 2017 and it is still true today.
Productivity Apps Spread the Same Way (and Hit the Same Wall)
Jobs also noticed that after the first custom-app sale, a second wave would hit. That was where less technical folks (e.g. administrative and marketing teams) wanted on the network too, which meant the productivity suite (Lotus tools, WYSIWYG WordPerfect, Adobe Illustrator) suddenly mattered. The machine moved from the developer’s desk to the wider organization.
AI coding tools are running the same playbook whether planned or not. Codex agents started life as pair-programming helpers for engineers. Now product managers are asking Claude Code to generate Jira tickets that actually match the spec, while ops folks use Antigravity to spin up monitoring dashboards without opening a PR. The productivity layer is expanding exactly like Jobs predicted.
The boost is real. McKinsey’s latest (early 2026) developer survey puts AI-assisted coding at 35–45 percent time savings across the average enterprise dev org. But again, we can’t get away from the very human heuristic that leads to resistance.
Once those tools touch shared repositories or customer data, the governance conversation flips from “how fast can we go” to “how do we keep the audit trail and prevent prompt-injection backdoors?” Antigravity’s early versions, for example, had a few very public persistent memory leaks that let agents keep context across sessions in ways that made security teams lose sleep. Same story, different decade.
Interpersonal (Now Multi-Agent) Computing Is the Next Big Reason
Jobs saved the biggest long-term bet for last: interpersonal computing. Not just email, but real collaboration that improves group productivity. He said it would take education, but within 24 months it would lead the buying criteria.
We are watching that exact shift with agentic AI. Claude Code and Codex agents already hand off tasks inside a team. Antigravity was built from the ground up as a multi-agent workspace where one agent can critique another’s code, a third can run security scans, and a fourth can summarize the whole thread for the architect. It’s designed to be interpersonal computing on steroids.
And just like we predicted demand, the enterprise pushback is identical to what played out in the market in 1991.
It’s typical for Enterprise buyers to say “We love the demo, but can we run this behind our firewall without leaking proprietary prompts? Does every agent conversation get logged for compliance? What happens when two agents start a side conversation the human never sees?”
The education cycle Jobs described is happening again, only this time the “customer” is the CISO and the legal team.
What the 1991 Predict for 2026 Enterprise AI Adoption Curves
If you map the old table to today’s AI tooling, the shape is the same:
- 1990 baseline: small but real beachhead (Sun at 80 percent).
- 1991 forecast: double, then triple.
- Reality: meaningful growth, but the exact hockey stick was tempered by integration friction and economic headwinds.
AI coding tools are on the same curve. 2025 was the “Sun sells 40k into the new segment” year. 2026 is the doubling year, but only if the resistance gets solved. The vendors know it. Anthropic, OpenAI, and Google are all racing to ship enterprise-grade governance layers (private VPC instances, prompt auditing, deterministic output modes). The winners will be the ones that turn the 3× dev velocity into something the compliance team can actually sign off on.
The practical takeaway for platform leaders and architects reading this is simple. Watch for the teams that already have one painful mission-critical custom app. That is where the first beachhead forms. Then look for the second wave (productivity tools spreading to adjacent roles). Finally, keep an eye on the multi-agent collaboration layer; it will be the reason the budget gets approved organization-wide.
Jobs was right in 1991 when he predicted the upgrade path being driven by custom apps, productivity, and group computing. The resistance is real, but so is the upside. The organizations that solve the governance piece without killing the 3× speed will own the next professional segment, whatever we end up calling it in 2026.
Here is the original video for folks who want to watch it. It’s a great presentation and I refer back to it regularly which is what got me thinking about mapping to today’s market. Enjoy!
