Cloud repatriation already showed us where AI goes next

By Rob Linton ·2026-08-12

Everyone renting frontier models today is making the same bet everyone made on cloud in 2010, and the same second act is coming. I say that as someone who was on both sides of the first round at once, which needs some explaining.

Amazon named Rob Linton an AWS Community Hero in 2015 (that’s me). Through the early 2010s I was evangelising cloud as hard as anyone in Australia: I wrote one of the first practitioner guides to AWS migration (Packt, 2011) and founded the country’s first AWS user group. At the same time, and to the visible confusion of almost everyone I pitched it to, I was building Podzy: an encrypted, fully on-premise alternative to Dropbox. The received wisdom said on-premise was backwards. Organisations were spending serious budgets getting into the cloud, and the topical security question of the day was whether the cloud could be trusted at all.

Both positions were right. That is the whole point of this essay.

Was going all-in on cloud wrong?

No. Renting the new capability while it is young is the correct first move, and I would give the same advice again today. You trade money for speed, you skip the capital outlay, and you learn what your workload actually is before you commit to owning anything.

Sarah Wang and Martin Casado at a16z compressed the whole pattern into one line in 2021: “You’re crazy if you don’t start in the cloud; you’re crazy if you stay on it.” Their estimate put roughly $100B of suppressed market value across the 50 top public software companies, driven by cloud’s drag on margins at scale. The initial play and the long play point in opposite directions, and both are smart in their season.

The mistake is not renting. The mistake is never noticing the season change.

Which companies actually left the cloud?

The best-documented case is Dropbox, and I concede the irony freely, having spent three years building an on-premise alternative to it. From 2013 Dropbox built Magic Pocket, its own custom storage infrastructure, and by early 2016 it reported storing over 90% of user data on its own kit. The S-1 shows what that did to the economics: infrastructure costs fell by $39.5M in 2016 and a further $35.1M in 2017, and gross margin climbed from 33% to 67% in two years. Dropbox stayed hybrid, keeping AWS for the remainder and for European data localisation. The company Podzy was built to compete with became the proof of Podzy’s infrastructure thesis.

Netflix is the example everyone gets wrong, in both directions. It never repatriated compute; it closed its last data centre in January 2016 and its current annual report still says the vast majority of its computing runs on AWS. But Netflix owns its entire delivery layer: Open Connect, its own hardware, built from 2012 because, in its own words at the time, “it now makes economic sense for Netflix to have one as well”. By March 2016 it carried 100% of Netflix’s video traffic. As Ken Florance put it, essentially everything before you hit play happens in AWS. Everything after happens on hardware Netflix owns. That is not a company that rejected the cloud. That is a company that worked out precisely which layer was core, and bought that layer.

The pattern kept running. 37signals left the cloud loudly from 2022: a $3.2M annual cloud bill cut to $1.3M by 2024, roughly $700K of Dell hardware paid back inside a year, the same team running it all, and projected savings the company now puts well past ten million dollars over five years. GEICO, ten years into a cloud migration, told it plainly through its infrastructure VP: bills up 2.5x with reliability worse, on a hyperscaler spend reported around $300M, and a repatriation program underway. In a 2024 Citrix-commissioned survey of 1,200 IT leaders at large companies, 94% had been involved in a cloud repatriation project within the previous three years. Steady-state, well-understood, margin-critical workloads came home. Experiments stayed rented.

Why is AI replaying the same pattern?

Because the economics rhyme almost embarrassingly. Frontier-model APIs are the new 2010 cloud: the fastest possible way to get capability you do not yet understand, at a price that eventually forces the question. The bills are now doing the forcing. TechCrunch reported in June 2026 that Uber blew through its entire 2026 AI coding budget by April, that a Priceline contract renewal came back four to five times more expensive, and that one company hit a $500M model bill with no usage limits in place. Gartner had already predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, with escalating costs among the named causes.

And the quiet builders are doing what the quiet builders did last time. Bloomberg trained its own 50-billion-parameter model on 363 billion tokens of its proprietary financial data. BNP Paribas signed a multi-year agreement with Mistral expressly covering on-premises deployment across the bank. Airbnb’s CEO says its customer-service agent leans heavily on open-source Qwen models because they are fast and cheap. A 2025 a16z survey of 100 enterprise CIOs found open-model adoption highest at the largest enterprises, for exactly two stated reasons: on-premises deployment and customisation. HPE’s AI order backlog passed $5B this year with 64% of cumulative orders coming from enterprises and governments rather than hyperscalers. The workloads that are steady, understood and line-of-business-specific are being brought onto models and infrastructure the organisation controls. Call it sovereignty if you like, but cash the word out: your keys, your custody, your bill.

My read: the companies custom-training open models on their own workflows today are the Dropboxes of 2013. They are spending effort now that looks unnecessary to everyone paying API invoices, and in three years the margin gap will look obvious in hindsight.

The honest caveat

Owning is not automatically right, and the aggregate numbers say the crowd has not turned. Zynga spent over $100M building its own data centres to escape AWS, then shut them and moved back in 2015; it owned infrastructure for a workload it could no longer predict. Menlo Ventures’ 2025 enterprise survey found open-source models’ share of enterprise LLM usage actually fell to 11%, from 19% a year earlier, even as spend roughly tripled to $37B. Measured by volume, the market is still sprinting the other way.

I do not read that as evidence against the turn. I read it as 2013 again: the year the crowd was maximally in the cloud was the same year Dropbox started building Magic Pocket. But the caveat stands on its own terms. Repatriate a workload you cannot predict and you get Zynga. The second play is only smart when the workload is steady-state, core to the business, and understood well enough to size. Rent the experiment. Own the capability.

What is the second smart play?

The upshot: treat frontier APIs the way a disciplined company treated cloud in 2010. Use them to learn fast, and meter them. Identify the two or three line-of-business workflows where AI is core rather than experimental, and start building the owned version now: an open model, custom-trained on your own data, on infrastructure you control, sized to the workload you have measured. The extra work is the moat. It was thirty years of the same question that taught me this: the organisations that end up in control of their data, their models and their bills are the ones that started building before the invoices made the argument for them.

Net: the initial play is renting intelligence. The long play is owning it. The companies doing the extra work quietly, again, are the ones who will look smart in three years, again.