TL;DR: For those who know me, you’ll hear me say “IT is a fashion industry.” I thought I was being humorous, maybe even novel. I wasn’t. Eric Abrahamson wrote the actual theory of “management fashion” in 1996, and it maps onto the data platform market with unsettling precision. The Oracle-and-Teradata season gave way to Cloudera’s, Cloudera-and-Teradata’s gave way to Snowflake’s, and Snowflake’s is giving way to Databricks’. Each season change got sold as rational and progressive, each vendor built a bigger show, a keynote, to sell it, and the gap between seasons has compressed from eight years to four. Other industries run technology cycles too, but they treat the technology as a durable asset with a return to prove, not next season’s dress shoes. At Zetaris, we’re working to shift data teams toward that same idea: provable returns, not next season’s look.
Izzy Lamberti is mid-fill, sticks spinning like helicopter blades, when the beat locks into the opening of “Everybody Wants to Rule the World.” Behind her a film washes across the backdrop; models hit the runway on the downbeat. The crowd is packed against the barricade, phones up, and by the time the hi-hat (the pair of foot-pedal cymbals that opens back up to drive a chorus) kicks back in, half of them are singing along to a forty-year-old Tears for Fears song like it was written just for this moment. That description is imagined, but the performance wasn’t. Izzy Lamberti performed live over the track for a Chick.nyc x Runway7Fashion show during New York Fashion Week in September 2025, and the clip cleared 36,000 likes before most of the industry even saw what the models were wearing (Lamberti, 2025).
Now imagine that same energy, but swap the building and locale. Same lighting, same projection wall, same packed floor with phones up, a beat the crowd already trusts doing the work of making whatever comes next feel inevitable. This isn’t Fashion Week, it’s a data trade show. I don’t know if anyone drummed live over Tears for Fears at the Databricks Data + AI Summit in June 2026. Yet the keynote stage ran the identical play NYFW ran with Izzy Lamberti: borrow a hook the crowd trusts, then use its momentum to launch the new thing before anyone’s asked questions. Here, the hook was the crowd itself. The room already believed the Databricks story, and that belief did the work a drum solo does. What launched into it was LTAP, pitched as the architecture that finally unifies transactions, analytics, and streaming on a single copy of storage (Databricks, 2026a).
I scrutinized LTAP in a companion piece, “Databricks LTAP ‘Solves’ the Decades-Old Data Pipeline Problem!?” The work is real and well-executed, but it is not a new idea. I don’t want to dig into whether the look on the runway was new. I want to look at why IT needs a “drummer and a beat” to sell what should be an engineering and financial decision.
There’s a name for what happens on IT fashion show stages, and it isn’t a made-up marketing term. Eric Abrahamson, a Columbia Business School professor, called it “management fashion” in a 1996 Academy of Management Review paper, and thirty years on his framework describes the data platform market with scary precision. He defines management fashion as “a relatively transitory collective belief, disseminated by management fashion setters, that a management technique leads rational management progress” (Abrahamson, 1996). Two things have to be true at once for a technique to become fashionable: it has to look rational (an efficient means to an important end) and it has to look progressive (new, and an improvement on what came before). A technique that’s merely old and effective doesn’t qualify; neither does something that’s new but can’t be dressed up as smarter than its predecessor.
“Fashion setters” is Abrahamson’s term for consulting firms, industry analysts, gurus, and trade media. Setters don’t force fashions onto buyers. They race to sense what buyers want, build an argument that makes it sound inevitable, and disseminate that argument before their competition. Abrahamson’s proof case was quality circles: between 1980 and 1982, nearly a third of large U.S. organizations adopted them, and one estimate put Fortune 500 adoption at 90 percent; by 1987, more than 80 percent of those adopters had abandoned the technique (Abrahamson, 1996). A textbook bell curve, from rise to fall, in less than a decade. Quality circles didn’t get worse between 1982 and 1987; they just stopped being fashionable.
Here’s the sequence I’ve watched play out inside the data platform IT sub-sector:
I’ve left out Google BigQuery, AWS Redshift/Athena, and Microsoft’s Fabric/Synapse on purpose, because they’re variations on similar runways rather than distinct seasons. To be clear, I’m not saying any of these platforms is bad. Each one solved something real for somebody. The point is the fashion cycle rhythm.
The opening scene wasn’t much of an exaggeration. Snowflake Summit drew roughly 2,000 people to its inaugural 2019 event; by 2023 that was nearly 12,000 in-person attendees, and 2025 attendance topped 20,000 (Snowflake Inc., 2018, 2023; Vendelux, 2025). Databricks’ Data + AI Summit (born as Spark Summit) climbed the same curve from the other direction: roughly 3,000 attendees a decade ago, 16,000 in-person plus 40,000-plus virtual in 2024, and 22,000 in-person plus 65,000 virtual worldwide in 2025 (Sector 6, n.d.; Databricks Blog, 2024; Janea Systems, 2025). These aren’t user groups anymore. They’re fashion weeks, complete with a keynote runway, press coverage designed to land the same week, and a purchasing rhythm that resets on schedule with the announcement calendar.
It’s tempting to reach for Oracle and Teradata here and argue the real stalwarts are the vendors who skip the fashion show entirely. But nope. Oracle runs its own keynote circus called Oracle AI World, projected to draw 20,000-plus attendees in 2025 (Oracle, n.d.). Teradata isn’t standing still so much as shrinking on schedule: revenue slid in four of the last six years, from $1.92 billion in 2021 to $1.66 billion in 2025 (WallStreetZen, n.d.; Teradata Corporation, 2026). Both companies are fine. Neither is dead. But “they skipped the fashion game” isn’t why.
What actually survives a fashion cycle looks like one of two things, and neither is an incumbent’s logo. The first is survival as concept: the brand dies, but the idea gets absorbed into infrastructure nobody keynotes about. Hadoop is dead as a label, but AWS built Elastic MapReduce on its architecture in 2009 and Google built Dataproc on it in 2015, and both still quietly run the world’s Spark and Hadoop workloads (Amazon Web Services, 2009; MSV, 2015). Presto survived something harder, an actual schism: when Facebook trademarked the name in 2019, the creators renamed their fork Trino in 2020, and both lineages are still alive, with the shared concept now running underneath Amazon Athena (Trino, 2020, 2022; Ksolves, n.d.; nOps, n.d.). Kafka outgrew its own codebase: Redpanda and WarpStream reimplement its wire protocol with none of the original code, and Confluent itself bought WarpStream in 2024, buying back an independent copy of its own protocol (Li, 2020; Estuary, n.d.; AutoMQ, n.d.). And Spark is doing this in real time, underneath the very keynote this piece opened with: it’s the engine inside LTAP, Databricks’ commercial vehicle since 2013, and simultaneously boring commodity infrastructure on EMR, Dataproc, Synapse, and Fabric, with no Databricks license in sight (Bigeye, n.d.-b; Reintech, n.d.).
The second way to outlive the runway is survival as brand, and it’s rarer: a product nobody’s financially incentivized to hype. PostgreSQL started as a Berkeley project in 1986, has no IPO and no 20,000-seat keynote hall behind it, and has still been one of DB-Engines’ biggest ranking climbers, moving up quarter after quarter without a single fashion setter pushing it (PostgreSQL Global Development Group, n.d.; Red Gate, 2025).
None of these four survived by staying loyal to the company that built them. Cloudera cratered, Presto split in two, Kafka’s creators had to buy back their own protocol, and Spark is commoditizing on clouds with no stake in Databricks’ success. What survived is the idea, precisely because it became separable from whoever originated it. The thing has to be useful enough, for long enough, that hype stops deciding whether it survives.
My intuition going in was that the gap between “it’s cool to use Snowflake” and “it’s cooler to use Databricks” was shorter than the gap between Teradata and Snowflake, and the dates back that up. Call Cloudera and Hadoop’s peak cultural moment roughly 2012–2014 and Snowflake’s roughly 2020–2021: an eight-year gap between fashion peaks. Call Databricks’ peak roughly 2023–2024, when its valuation first pulled decisively ahead of Snowflake’s, a three-to-four-year gap from the last one. The cycle didn’t just continue, it roughly halved.
Compare that to industries that also run technology cycles but aren’t built on quarterly hype. Hydraulic fracturing was first field-tested in 1947 and in commercial use by 1949. It took roughly fifty years, until the late 1990s and early 2000s, before shale extraction became economically mainstream and touched off the shale revolution (Engineering and Technology History Wiki, n.d.; NES Fircroft, n.d.). The Boeing 737 has gone through four generations (Original, first flight 1967; Classic, 1984; Next Generation, 1997; MAX, certified 2017), averaging roughly seventeen years between generations, and the CFM56 engine family powered every 737 built from the Classic’s 1984 debut until the MAX’s LEAP-1B arrived in 2017, a single engine architecture in production for over three decades. CFM’s next clean-sheet engine design isn’t expected before the mid-to-late 2030s (Simple Flying, n.d.; Wikipedia, n.d.; GE Aerospace, n.d.).
Here’s the part I think gets missed when people cite oil and gas or aviation as “the slow industries.” The long cycle isn’t evidence that those industries are slower-witted about technology. It’s evidence that the technology is treated as a durable asset that has to earn a return, and someone’s job depends on proving it did. A refinery doesn’t swap a catalytic cracker because a rival refinery announced a new one at a conference. It runs a cost-per-barrel and yield analysis, weighs that against the capital and downtime cost of the swap, and only moves when the numbers clear a hurdle rate someone will be held accountable for. An airline doesn’t reengine a fleet because the new engine looked good on a brochure. Both industries employ people with the actual job title asset manager, whose performance review is a utilization curve and a depreciation schedule, not a season’s silhouette. The long cycle is the return showing up. It’s the opposite of stagnation, it’s discipline.
That’s the gap I think our sub-sector actually needs to close, and it’s the reason Zetaris exists in its current form. Right now, the conversation in most data organizations sounds like the crowd at that keynote: “We moved to the new widget because the new widget is what cool engineers use now.” That’s a sociopsychological answer wearing a technical costume. The conversation an oil and gas asset manager has instead sounds nothing like that: “Here is the utilization we’re getting out of the asset we already paid for, here is what an upgrade would actually buy us in yield, and here is the payback period.” One conversation is about looking current. The other is about return.
Our bet: you must hack the data platform fashion cycle by changing the conversation, not by refusing to participate. An overlay architecture queries Oracle, Kafka, Iceberg, RESTful APIs, and Databricks data where it already sits, instead of re-platforming it every time a new keynote lands. That turns the existing estate into exactly the kind of asset an O&G portfolio manager already knows about: something with a utilization rate, a cost-per-query, and a return that compounds over time. It’s not a wardrobe that depreciates to zero the moment a rival vendor’s model walks the runway in a newer color or style. That’s not a rejection of new engines or new formats (e.g., Iceberg genuinely earned its adoption). It’s a rejection of the idea that adoption has to mean replacement, and that engineers should be talking about which widget they’re wearing this year instead of what the estate is actually returning.
Put the pieces together and the shape is unmistakable. Abrahamson’s theory predicts fashions require a technique that can be sold as both rational and progressive with fashion setters racing to define what that means, and fashion shows to disseminate the rhetoric. The data platform sub-sector has all three: a rhetoric engine (analyst reports, vendor blogs, conference keynotes) that reliably recasts “new” as “necessary,” fashion setters (the hyperscalers, the systems integrators, the platform vendors themselves) competing to be first to name the next architecture, and eye poppingly large fashion shows (20,000 seats and climbing) where the new season is unveiled to standing ovations. The cycle time between seasons has compressed from roughly eight years to roughly four. Oil and gas and aviation prove the long-cycle alternative isn’t industry-specific caution; it’s an asset-management discipline that happens to produce a much better return than chasing the runway. And durability inside our own sub-sector, once you look past any single vendor’s keynote calendar, still comes down to one of two things: getting absorbed so completely into infrastructure that the brand becomes irrelevant (i.e., Presto, Kafka, and Spark), or never needing a fashion campaign in the first place (i.e., Postgres).
I’m not arguing for never migrating. Elastic compute on any type of cloud, separated storage and compute, and open table formats are real advances. The argument is narrower and I think more useful: before you fund a migration, ask whether you’re gaining real technical or business benefits or buying a season’s fashion. Capital One’s much-publicized exit of its on-premises data center footprints by 2020 is held up as the industry’s proof point that full migration works (Amazon Web Services, n.d.; SiliconANGLE, 2020). It’s a genuinely impressive achievement, and it’s also worth noting that two years later, Capital One’s own team candidly told the press that cloud migration introduced new problems into its data estate that hadn’t existed before, like fragmentation, added complexity, and new governance gaps (BigDATAwire, 2022). The celebrated case study and post-implementation honesty are both true. Capital One’s achievement is an example of the migration trap: each generation of platform promises to retire the complexity of the last one, and each one later becomes the complexity the next fashion promises to retire.
Fashion isn’t a slur, and I’m not arguing tech keynotes are pointless theater, or that your last migration was a mistake. Abrahamson’s real point, the one people miss when waving off “IT Fashion” as trivial, is that fashions aren’t purely sociopsychological. They compete with genuine technoeconomic performance gaps for the right to shape your thinking and your roadmap. Sometimes that IT fashion fad is actually the trend worth paying attention to because there are honest technical gains, but sometimes it’s just this season’s color from an industry that measures success in conference attendance instead of cost-per-query. Other industries like oil and gas or aviation don’t have that problem, because nobody gets promoted for looking fashion-forward at the wellhead or on the tarmac. Ask the question an asset manager asks before the question a fashion follower asks: what does this new feature or technology return, and would it still be worth adopting if it was chartreuse? If you can’t answer that without mentioning who else was in the room, you’re not migrating, you’re shopping.