Who Actually Runs Elixir in Production, Fact-Checked
A viral infographic ranked 15 companies using Elixir by valuation. Fact-checked: nine rows hold up, three are stale or unsupported, four name the wrong system.
In this post
TL;DR: A graphic called “The 15 Most Valuable Companies Using Elixir” made the rounds in August 2026, ranking companies by valuation with a one-line “evidence” tag next to each logo. Nobody checked it before it started getting reposted as fact, so I did — row by row, against primary sources: engineering blogs, live job postings, GitHub commit logs, funding announcements. The result: 9 of 15 rows verify as CURRENT, 3 are AGING (real but old), 1 is STALE (Pinterest), and 2 are UNSUPPORTED (Spotify, Together AI). Four rows have the system mislabeled or attributed to the wrong entity — Whatnot’s describes the wrong product, SumUp’s the wrong language era, and RBC’s and Adobe’s describe a division and an acquired subsidiary rather than the parent company. One of the two unsupported rows corrects my own reporting.
Last verified: September 2, 2026. Every verdict below is a snapshot, not a permanent rating — job postings get pulled, engineering blogs go dark, GitHub orgs get renamed. Treat the source table as the durable part; treat the verdicts as dated.
Method: what counts as evidence here
Four tiers, in order of trust, and only the top three count as anything:
- First-party engineering content — an official case study (elixir-lang.org/cases, a company’s own eng blog), a conference talk by a named employee, or live source code in the company’s GitHub org. This is what “CURRENT” or “AGING” is built on.
- First-party job postings — a live “here’s our current stack” posting from the company itself. Real signal that Elixir exists somewhere internally, but it names an intent to hire, not a shipped system. On its own this caps a row at “hiring signal only,” never “CURRENT” in the strong sense.
- Third-party reporting with named sources — a trade-press piece quoting a named engineer, e.g. Steve Cohen at Pinterest or Thanos Vassilakis at RBC. Weaker than first-party, but it’s a person on the record, not an aggregator.
- Agency listicles and SEO blog posts — “10 Companies Using Elixir in Production” pieces from dev shops selling Elixir contracting. These are not evidence. Every one I found in this research (Netguru, Curiosum, Selleo, Monterail, Prograils, Freshcode, among others) traces back to the same handful of primary sources above and adds nothing independently. If a listicle is the only thing repeating a claim, that claim doesn’t survive contact with a source check — which is exactly what happened to the original infographic’s Spotify row.
The other axis is currency. I used a straightforward scale: CURRENT (evidence dated within the last 24 months, or live source code with recent commits), AGING (2–5 years old, nothing newer, nothing contradicting it either), STALE (evidence over 5 years old, or a newer account of the same system that no longer mentions Elixir), UNSUPPORTED (the only “evidence” is a listicle, a self-contradicting source, or — in Together AI’s case — a posting that’s since been pulled with no archive of it left).
If you evaluate vendor or acquisition-target technology claims for a living, this is the same exercise as technical diligence on a smaller scale: don’t take the deck’s word for a stack claim, go find the primary source and check the date on it. It’s the kind of gap-finding I do in fractional engagements before a term sheet gets signed, not after.
The 15 rows
Notes on the rows that need explaining
Apple: real, but it’s one internal team, not “production systems”
The only evidence is a live July 2026 job posting from Apple’s Environment and Supply Chain Innovation team — internal carbon-accounting and supply-chain tooling inside Operations, not iOS, iCloud, or any consumer product. The posting is unambiguous that Phoenix/LiveView run there today, but Apple has published zero engineering content about it. “Production systems” on the infographic implies something company-wide; what’s actually documented is one internal-tools org, and two of the three Apple Elixir job reqs I found are already expired — this reads like a small, persistent team, not a platform bet.
RBC and Adobe: real systems, wrong entity
Both of these rows attribute a division or an acquired subsidiary’s stack to the parent brand. RBC’s Elixir story is RBC Capital Markets, the investment-banking arm, described by its Head of R&D on a 2022 podcast episode — greenfield trading and risk tooling with LiveView outperforming React for hundreds of live-updating charts. That’s a real, credible story about a division, not the CAD$2T-asset universal bank.
Adobe’s is the more interesting case: the actual Elixir shop is
Frame.io, the video-collaboration platform Adobe
acquired for $1.275B in October 2021. Frame.io
rewrote its entire backend in Elixir in 2018 — three years before
Adobe bought it. Adobe inherited an Elixir platform; it didn’t choose
one. The freshest evidence — Adobe’s own elixir-styler
linter, maintained by engineers who are, by their own public profiles
(reported), on the Frame.io team, for a “single codebase (140+
contributors),” per its own README — was still getting merged commits a
week before this research. So Elixir is genuinely alive under the Adobe
umbrella; it’s just not Photoshop or Creative Cloud.
Pinterest: is it still using Elixir?
No — not verifiably, and the evidence points the other way. Pinterest
really did build Guardian, a
real-time spam and rules engine (reported — the Pinterest
Engineering post is behind a Medium block that returns the text only as
a search snippet), in Elixir starting in late 2017, and an earlier
notifications/rate-limiting system before that. The famous “$2M/year
saved” figure traces
to a single 2018 book interview with the engineer who wrote it,
repeated by every listicle since — no Pinterest publication states that
number itself. The newest public description of Guardian I could find,
dated
2025 — a third-party write-up of a 2025 Pinterest conference talk —
describes it as a Kafka + Flink + Iceberg streaming
architecture with StarRocks for analytics. Elixir isn’t mentioned
anywhere in it. Nobody at Pinterest has said “we left Elixir” on the
record — the case for STALE is silence plus a replacement architecture,
not a confession. Pinterest’s own Elixir open-source repos are dormant
too: elixometer
hasn’t been pushed since 2023, riffed since
2020, and elixir_thrift
no longer exists.
Spotify: the infographic’s own source contradicts it
This is the cleanest example of the chart laundering an agency blog post into a claim its source doesn’t support. What’s documented is one internal debugging tool for the artist advertising platform, built by Joel Kemp around 2021, described in his own article title (reported — Medium blocks direct fetch) as “Internal Web Applications.” The agency write-up everyone cites for “Spotify runs on Elixir” says outright, in the same piece: “the listener-facing backend services are done with Java.” A Hacker News thread in 2024 called this out directly, quoting that exact line back at someone repeating the claim. Nothing newer than 2021 exists. “Production systems” is not a defensible label here.
Whatnot: a great story, mislabeled
This is the row that annoyed me most, because the real story is better than the infographic’s version. Elixir at Whatnot is the Live Service — the system behind live auctions, live chat, and stream state — not the “marketplace backend,” which is Python; it’s built on Phoenix Channels, PubSub, and Horde per a 2022 Whatnot engineering post (reported — Medium blocks direct fetch). Whatnot’s own February 2026 engineering post about scaling for the MrBeast livestream is explicit about the split, in the same sentence: “our Elixir-based Live Service and Python-based main backend.” That post documents 583,000 concurrent viewers on a single stream and even names an Elixir-specific failure mode under peak load (distributed Erlang proxy calls timing out) — the kind of detail nobody fakes. It’s the best-documented row in the set — first-party, recent, and specific enough that nobody could have invented it. It’s also mislabeled.
Ramp, SumUp: real, scoped, and older than the label suggests
Both rows are accurate about what runs on Elixir but understate how narrow it is. Ramp’s Elixir footprint is two specific services — the real-time card authorizer and the risk analysis platform — inside a company whose “largest service is in Python,” per Ramp’s own engineering post. A Ramp engineer confirmed the authorizer was still live in a 2023 Hacker News thread, and a job posting mirrored within the last year still describes the Elixir stack in the present tense (the mirror carries no hard date), but the load-bearing engineering post is from 2021. SumUp’s row conflates two different eras: the company’s core transaction-processing engine that scaled it to millions of payments is mostly written in Erlang, per its own engineering manager; Elixir arrived in 2019, specifically for SumUp Bank in Brazil. “Banking backend” is accurate for the Brazil bank and misleading for SumUp’s global payments core.
Correcting my own post: Together AI
I published a post in July 2026 claiming Together AI runs Elixir/Phoenix on the BEAM for its identity, auth, and RBAC layer, sourced from one Greenhouse job posting. Re-checking that URL for this piece: it now resolves to Together AI’s job-board index, not a posting, and no Wayback snapshot of it exists — the citation is permanently unverifiable from the source I used. A separate Together AI Elixir posting, mirrored on a job-board aggregator, was removed on March 3, 2026 — four months before my post ran. Whether the exact req I cited was still live on publication day can’t be established either way. Together AI’s current 58-role board has zero Elixir listings, and its public GitHub org surfaces Python, Go, Rust and TypeScript across its most-visible repos, and no Elixir.
The Elixir at Together AI looks inherited, not chosen — the only Elixir postings named the CodeSandbox team, acquired in December 2024, and even then the scope was auth and org management on a subsidiary product team, not the AI platform.
My July post did hedge — it said outright that “a list of logos proves nothing on its own.” What didn’t hold up was a more specific, present-tense claim in that same post: that Together AI “put that on Elixir and Phoenix… and is hiring engineers into it by name.” That sentence is no longer supportable. The post also never disclosed the CodeSandbox provenance — the only thing tying CodeSandbox itself to Elixir is a 2018 community forum thread by a non-employee — which reframes “Together AI chose Elixir” into “Together AI acquired a company that already ran Elixir.” The funding numbers in that post — the $800M Series C, the $8.3B valuation, both from July 2026 — still hold; the technical claim doesn’t.
What the chart gets right
Strip out the mislabeling and the stale rows, and the underlying pattern the infographic is gesturing at is real. The best-supported rows in this set — PepsiCo, Discord, Supabase, Remote, Whatnot, Adobe/Frame.io — cluster around three patterns: real-time fanout at scale (Discord’s per-guild processes, Whatnot’s live auctions, Supabase’s WebSocket realtime), correctness-under-concurrency paths (Ramp’s authorizer, PepsiCo’s sales-intelligence pipeline), and whole-backend commitments (Remote’s monolith, Frame.io’s full rewrite). That’s not a coincidence — it’s what the BEAM is actually good at, and it’s the same reason I wrote about why I reach for Elixir on AI startup backends: the concurrency model matches the shape of the problem, not the hype cycle. If you want the toolbox rather than the company logos, the libraries I actually reach for in 2026 is the more useful read.
It’s also not just Elixir wearing a public-facing name. Erlang/BEAM systems keep turning up inside infrastructure that nobody markets as “Erlang shops” — I wrote about one recently in Cisco’s network config engine, which turned out to be secretly Erlang under the hood. The pattern in that post and this one is the same: the BEAM shows up wherever concurrent, fault-tolerant state management matters, whether or not the company advertises it.
Discord and Remote are the two rows I’d point at if someone asked
“does this actually work at scale, in 2026, for real production
traffic”: Discord has a 2026
first-party post naming live services (discord-guilds,
discord-sessions) and walking through a real incident;
Remote has both a 2025
elixir-lang.org case study and a 2026
conference keynote describing a monolith that’s grown from one team
to fifty. Those are the rows I’d trust without a second source.
The takeaway
Half the value of an infographic like this is that it’s directionally right — Elixir genuinely runs production infrastructure at companies with real valuations, in exactly the workloads you’d predict. The other half of the value, the part that actually matters if you’re deciding whether to bet a hiring plan or a stack choice on this pattern, is knowing which four rows are describing the wrong system or the wrong company, which one has been superseded by a different architecture, which three are running on evidence two to five years old, and which two shouldn’t be on the list at all. Screenshots don’t carry dates. Sources do.
If you’re weighing an Elixir bet for your own stack, or trying to validate a vendor’s or acquisition target’s technical claims before you commit to them, that’s exactly the kind of gap-finding I do for founders and technical teams — let’s talk about your engagement.