Analysis of 50 public software companies finds that repatriating workloads from public cloud infrastructure could help them reduce their cloud spend by 50%
Andreessen Horowitz : Tweets: @zackkanter , @jasnonaz , @lizrice , @jaycuthrell , @zackkanter , @zackkanter , @stratelogical , @sandykory , @divinetechygirl , @teej_m , @matt_levine_1 , @martin_casado , @knieriemen , @anshublog , @jaimenovoa , @mjasay , @a16z , @levie , @carnage4life , and @carnage4life Tweets: Zack Kanter / @zackkanter : Excellent *financial* analysis of using *commoditized* cloud infrastructure (vanilla servers). It misses: i) the (long-term devastating) cultural cost of recruiting world-class engineers to do undifferentiated heavy lifting; ii) it's unfeasible to recreate noncommodity infra. 1/n https://twitter.com/... @jasnonaz : This is a great point and one of the main reasons I'm so enthusiastic about tools like @getdbt, @_hex_tech and @LookerData - they are all extremely _fun_ to use. https://twitter.com/... Liz Rice / @lizrice : “The popularity of Kubernetes and the containerization of software, which makes workloads more portable, was in part a reaction to companies not wanting to be locked into a specific cloud” Do you plan to run your own data center when your company grows? https://twitter.com/... @jaycuthrell : It's still fun to look back at how @netflix was described in the tech press as being 100% public cloud even as Open Connect Appliances (OCAs) seeded the planet in major facilities based broadband providers and internet exchange points just as @Akamai did 20 years ago.🤓☁️🤔 💸📈📉 https://twitter.com/... Zack Kanter / @zackkanter : @martin_casado @sarahdingwang This certainly is a cautionary tale of what can happen if you don't think carefully about your cloud spend - just not in the way the piece contends. https://twitter.com/... https://twitter.com/... Zack Kanter / @zackkanter : Every sufficiently-funded software CEO on earth will tell you that their constraining factor is hiring great engineering talent - repatriating commodity servers to save on COGS means increasing engineering headcount requirements, definitionally making the constraint worse. 3/n @stratelogical : Outstanding thread! The a16z piece makes some fair points, you *might* save money. Buuuut like Bezos says in this video, breweries shouldn't be making their own electricity, even if it saves them money, they should be focused on making the best beer. https://twitter.com/... https://twitter.com/... Sandy Kory / @sandykory : Spot on analysis. “AWS is so expensive for unicorns” is rife with selection bias. Similar to “SaaS startups always underprice” which is easy to say when ignoring that the overpricers are dead. https://twitter.com/... @divinetechygirl : I've been thinking about this paradox a ton lately, especially as it relates to modernizing applications, and what they're calling repatriation. Definitely appreciate the level of research and detail here. Worth the read. https://twitter.com/... TJ Murphy / @teej_m : “The boringness of the work” is vastly underappreciated idea. My pet theory - Slack (a boring idea) was only good because it was born out of Glitch (a rad idea). You'd never assemble the early Slack team from scratch because workplace communication is BORING. https://twitter.com/... Matt Levine / @matt_levine_1 : .@zackkanter nailed this... Cloud service provider margins will inevitably compress as companies grow more thoughtful around the steps they can take to better manage their cloud infra spend. But for 99.999% of companies, “re-build AWS ourselves” is not one of those steps. https://twitter.com/... @martin_casado : Ironically I'll bet it was a similar margin analysis which pushed the public clouds to explore alternate supply chains (and DIY) for servers, NICs, network firmware, etc. Clearly not a core competency. But the leverage via COGs reduction is huge https://twitter.com/... Greg Knieriemen / @knieriemen : Reading lots of hot takes of @martin_casado and @sarahdingwang's research “The Cost of Cloud, a Trillion Dollar Paradox” and most seem triggered by the word “repatriation” and entirely miss the actual point: “Cloud Optimization”. https://a16z.com/... /1 Anshu Sharma / @anshublog : iPhones have 40% gross margins. Literally hundreds of billion. You should consider making your phone at home. https://twitter.com/... Jaime Novoa / @jaimenovoa : The Cost of Cloud, a Trillion Dollar Paradox If you're operating at scale, the cost of cloud can at least double your infrastructure bill https://a16z.com/... Matt Asay / @mjasay : But, I mean, besides all the *realities* of rolling your own cloud... https://twitter.com/... @a16z : “You're crazy if you don't start in the cloud; you're crazy if you stay on it.” by @martin_casado & @sarahdingwang https://a16z.com/... Aaron Levie / @levie : @Carnage4Life We went the opposite direction, because what this analysis misses is the opportunity cost on R&D spend that could be better spent on software innovation, as well as the rapid rate of new services from the public clouds that you'd never be able to build on prem. Dare Obasanjo / @carnage4life : This analysis provoked by an A16Z funded company, Optimyze, hand waves about the expertise needed to build in-house versions of what AWS & Azure provide as a service. In addition to product stasis caused by rewriting your infrastructure. Dropbox isn't a poster child for growth. Dare Obasanjo / @carnage4life : A16z argues that SaaS companies can save 50% on their service costs by moving from AWS to on-premise systems which directly translates to increased market cap from increased profitability. Dropbox saving $75M over 2 years is the canonical example. https://a16z.com/...
Context & Ripple Effects
The hyperscalers' cost advantage has been framed almost entirely on their side of the ledger: Platformonomics' CAPEX gap analysis showed IBM and Oracle tens of billions behind Amazon, Google, and Microsoft, and Amazon's capex grew another 69% in 2020 to $54B (per Platformonomics' 2020 breakdown) — scale that reads as an unassailable moat. A Logic Magazine interview with an AWS engineer separately argued AWS is deliberately easy to migrate onto and hard to leave, compounding the lock-in story.
Andreessen Horowitz's new analysis of 50 public software companies attacks that framing from the buyer's side, claiming repatriating workloads could cut cloud spend by roughly half, with Dropbox's reported $75M two-year savings as the canonical proof point. The rebuttal came fast and public: figures including Zack Kanter and Aaron Levie argue the model ignores the cultural cost of recruiting engineers to do undifferentiated infrastructure work and is largely unfeasible to recreate — turning a spreadsheet argument into a fight over what the cloud premium actually buys.
First-order effects
- SaaS companies' COGS lines are suddenly under board-level scrutiny: the a16z math gives any large AWS/Azure tenant a concrete template for questioning its committed-spend contracts, using Dropbox's $75M saving as the benchmark.
- AWS and Azure now face a public narrative that their largest customers are overpaying by design, forcing account teams to defend value beyond raw compute price just as the lock-in mechanics surfaced in the AWS engineer interview get wider attention.
Second-order effects
- Hyperscalers' rational counter-move is deeper discounting and bespoke commitments for at-scale tenants rather than list-price concessions, which pressures margins exactly where Platformonomics' capex tracking shows the providers are spending most aggressively to build capacity.
- Tooling and services vendors gain a new sales motion around hybrid and repatriation readiness — cost-modeling, colocation partnerships, and migration-out services — monetizing doubt about cloud-first orthodoxy without anyone actually leaving.
Third-order effects
- If even a minority of large workloads repatriate successfully, the infrastructure market bifurcates into commodity compute bought like a utility and premium managed services paid for explicitly — eroding the blended-margin model that funded the hyperscalers' capex arms race.
- The durability of the answer depends on execution risk: Dropbox's outcome is one data point, and the critics' core claim — that replicating internal infrastructure competence is unfeasible for most companies — means the realistic endpoint is selective repatriation by the largest tenants, not an industry-wide exit. The pattern echoes Amazon's own history of rebuilding servers in-house (the Sun-to-HP/Linux transition) when scale made it rational, suggesting repatriation decisions track tenant size more than ideology.
The trend: Cloud infrastructure purchasing is splitting from a default into an explicit financial decision, with hyperscaler lock-in increasingly priced and contested by the largest tenants rather than accepted.