The pendulum got tired of swinging I verified the monolith vs. microservices debate of 2026 against 65+ primary sources -- from Shopify and Amazon to the European Union, China, and India. Conversation about the investigation (7 min). Kilowatto and Leia are characters; both voices are synthetic. This conversation summarises the piece, it does not replace it. The full text, with sources and charts, is on kilowatto.com. --- Kilowatto: Leia, quick one. Big tech companies like Google and Amazon are built on microservices, right? Leia: Yes, that is how they achieve scalability and reliability. Kilowatto: That is what almost everybody answers. Shopify operates a single Ruby on Rails application, handling tens of terabytes of data per minute during Black Friday traffic peaks, without fragmenting the application code into separate services. Leia: Hang on. What? Kilowatto: Hey! Let's go straight to the bottom of it, because today I have a Deep Dive: The pendulum got tired of swinging. I verified the monolith vs. microservices debate of 2026 against 65+ primary sources -- from Shopify and Amazon to the European Union, China, and India. Leia: Alright Kilowatto, where do we start? Kilowatto: I found that 42% of organizations that adopted microservices are consolidating services back into larger units is widely cited, but I couldn't confirm it against a primary document. Leia: Where did you see that 42% figure? Kilowatto: It's attributed to a 2025 CNCF survey, but I reviewed their reports and couldn't find it, even in the 2024 annual survey report or the November 2025 announcement. Leia: Is 42% the only figure mentioned? Kilowatto: No, I did find that 46% of backend developers work with microservices, which is consistently reported, but the 42% figure is the most-cited in the 2026 architecture debate. Leia: You said it's cited in at least eight independent sources, right? Kilowatto: Yes, with consistent detail, including an 18% to 8% drop in service mesh adoption, but I still couldn't verify the 42% figure against a primary document. Leia: So it's not necessarily false, but... Kilowatto: Exactly, it's possible it was reported in a webinar or side panel, but I couldn't find it, so I'm treating it as widely repeated, not independently verified. Leia: What about the other figures, like the 46%? Kilowatto: I confirmed that one directly against the foundation's original statement, so that one is solid. Leia: And what about the 18% to 8% drop in service mesh adoption? Kilowatto: That's also consistently reported, but again, I couldn't verify the 42% figure, which is the most-cited one. Leia: So we should be careful with that 42% figure, right? Kilowatto: I found Shopify's majestic monolith handles tens of terabytes of data per minute, disproving the idea that monoliths can't scale. Leia: Tens of terabytes, how many exactly? Kilowatto: Well, it's tens of terabytes per minute, during Black Friday traffic peaks, with the database layer partitioned by store. Leia: Per minute, that's a lot, right? Kilowatto: Look, the platform operates a single Ruby on Rails application, with components separated by code boundaries, not network calls. Leia: What's the tool they use for that purpose? Kilowatto: I found they developed Packwerk, which imposes strict boundaries internally. Leia: So it's not just one big mess, it's organized? Kilowatto: Exactly, the components run in the same process, without network calls between them, that's the definition of a modular monolith. Leia: Is that a new concept? Kilowatto: Well, the literature from 2025-2026 identifies it as a third way between classic monolith and microservices, with only 15 peer-reviewed primary studies between 2020 and May 2025. Leia: Only 15 studies, is that a lot? Kilowatto: Hang on, it says something important, industry practice is years ahead of the formal research that supports it, according to the Future Internet journal. Kilowatto: I found that AI orchestrates AI, choosing the modular monolith, suggesting a shift towards consolidation over distribution. Leia: What's the modular monolith? Kilowatto: Well, it's a central process with total authority, modules invoked on demand, clear boundaries, and communication that never leaves the orchestrator's control. Leia: Is that like a microservices mesh? Kilowatto: No, it's not - there's no point-to-point channel between them, which is key to this finding. Leia: How did you determine that? Kilowatto: I looked at Anthropic, OpenAI, and Cognition, and they all adopted this pattern independently by 2026. Leia: What about the A2A protocol? Kilowatto: It joined MCP under the Linux Foundation in December 2025, standardizing a hierarchical pattern, not a flat mesh. Leia: So, what does this mean for distribution vs consolidation? Kilowatto: The answer is towards a modular monolith with disposable subagents, not a distributed mesh of autonomous microservices. Leia: Is this what the AI builders are saying, or just what they're doing? Kilowatto: It's what they're doing, not their marketing talk - that's what's interesting about this finding. Leia: Where can I see the details of how you arrived at this? Kilowatto: I found academia is behind the industry in understanding modular monoliths in the cloud, with only 15 peer-reviewed primary studies between 2020 and May 2025. Leia: 15 studies, is that a lot? Kilowatto: Well, it's not enough to establish a consensus on a clear definition of the term, according to a systematic literature review published in October 2025. Leia: What year was that review published? Kilowatto: It was published in 2025, and another study from 2026 found similar conclusions about the lack of solid empirical bases for decisions. Leia: How many sources did the 2026 study review? Kilowatto: The 2026 study reviewed 67 sources between 2022 and 2026, and still found a lack of rigorous research on the topic. Leia: Is 67 a big number for a study? Kilowatto: Look, what I did not expect was that even with 67 sources, the study couldn't find a consensus on the definition of modular monoliths. Leia: You mentioned 2020, what was the end year for the 15 studies? Kilowatto: The 15 studies were published between 2020 and May 2025, and it's surprising that there's still no clear definition of the term. Leia: Is May 2025 the latest data you found? Kilowatto: Well, I also found a 2026 paper on arXiv that proposes abandoning the microservice unit in favor of universal module interfaces, which shows the vocabulary itself is starting to feel insufficient. Leia: What's the title of that 2026 paper? Kilowatto: I found a consistent pattern in serious sources: start with a modular monolith, not a flat monolith or microservices. Leia: What's a modular monolith? Kilowatto: Well, it means defining clear domain boundaries from the first commit, even if everything runs in a single process. Leia: How many sources said that? Kilowatto: It was in every serious source I reviewed, from the Amazon Prime Video case to the academic systematic literature review. Leia: Is that for all projects? Kilowatto: Look, it's a consistent pattern, but I didn't find any one-size-fits-all answer. Leia: What's the benefit of a modular monolith? Kilowatto: Extracting a well-defined module to an independent service later is cheap; undoing a poorly defined microservices mesh is extremely expensive. Leia: How expensive? Kilowatto: I mean, it's extremely expensive, but I didn't find a specific figure. Leia: Is this pattern new? Kilowatto: It was in the literature I reviewed, including the 2025-2026 literature. Leia: What year did the academic systematic literature review come out? Kilowatto: That is the short version. The piece has the sources, the charts and the parts we skipped. Leia: The full investigation is on kilowatto.com, under Deep Dives. Look for "el-pendulo-se-canso-de-columpiarse".