AI news videos

Two Engineers Rewrote OpenAI Storage From Python Into Rust

OpenAI has published an engineering post on Habitat, the online storage platform behind ChatGPT, the API, Codex and its internal services. Read it for what it is: OpenAI's own account of OpenAI's own infrastructure, not an independent evaluation.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• Every factual claim in this script comes from a single first-party source, OpenAI's own engineering blog post "Rapidly scaling online storage to serve over 1 billion ChatGPT users", published on openai.com on September 11, 2026 under the byline "By Jon Lee, Chaomin Yu, and Ben Ries, Members of Technical Staff" and dated in the body "September 11, 2026", captured to the local snapshot at 2026-09-11T17:00:30.431Z. One typographic normalization applies to every quotation below: the source page uses curly apostrophes and a non-breaking hyphen inside the model name, and I have rendered those as plain characters when quoting, with no change to any word. SENTENCE AUDIT FOLLOWS, and it covers the eleven sentences of the narration. Hook sentence one, "OpenAI says today that two engineers rewrote its entire storage service from Python into Rust", is supported by "In Q2 2026, with just 2 engineers, Codex, and GPT 5.5, we were able to rewrite the entire service in Rust", and the word "today" is supported by the page dateline "September 11, 2026", which is the same calendar day this script was written; I wrote the count as the word "two" because digits are garbled by the text-to-speech engine, attributed the statement to OpenAI rather than asserting it in our own voice because it is an unverified self-report, and deliberately dropped the quarter designation "Q2 2026" for duration since the hook already establishes that the rewrite is finished and recent. Scene one sentence one, "They used OpenAI's own Codex and coding models", is supported by the same sentence's list "with just 2 engineers, Codex, and GPT 5.5" and by the earlier passage "We also made a calculated wager that the rapid advancement of our own coding models would simplify the technical path in the future. We bet that by the time a full migration off Python was required, Codex and GPT would make that migration achievable. That bet eventually proved correct."; I replaced the specific model name with the generic phrase "coding models", which the article itself uses, because a brand name glued to digits is garbled by the speech engine, and the substitution loses no meaning since the article's own wording for the same bet is "our own coding models". Scene one sentence two, "Their storage platform Habitat handles more than seventy million requests every second", is supported by "Habitat now handles more than 70 million requests every second, supporting products used by over 1 billion people each week, across almost 40 geographic regions" and by "Habitat is the online storage platform we built so OpenAI products can quickly and reliably access needed information"; I wrote the figure as the words "seventy million", kept the article's own hedge "more than", and dropped the region count for duration. Scene two sentence one, "Rust now handles ninety five percent of production requests", is supported by "This new Rust service is now handling 95% of our production requests"; I wrote the figure as the words "ninety five percent" and kept the word "now" so the viewer hears that this is a current partial state rather than a completed cutover. Scene two sentence two, "OpenAI reports six times better processor efficiency and fifteen times better memory efficiency", is supported by "Our data shows the Rust service is 6x more CPU efficient and 15x more memory efficient than the Python version, with significantly lower average and tail latencies"; I wrote both multipliers as words, replaced the acronym in "CPU efficient" with "processor efficiency" because spelled acronyms read poorly aloud and the meaning is unchanged, prefixed the line with "OpenAI reports" because the article's own framing is "Our data shows" and no external party measured this, and dropped the latency clause for duration. Scene three sentence one, "Habitat serves over one billion people each week", is supported by "supporting products used by over 1 billion people each week" and by the page's own stat block "1B+ people each week"; I wrote the figure as the words "one billion" and kept the article's hedge "over". Scene three sentence two, "Python retires entirely in the coming weeks", is supported by "we'll be deprecating Python entirely in the coming weeks"; I replaced the vendor verb "deprecating" with "retires" for plain speech, kept the word "entirely" because the article uses it, and kept the future tense so nobody hears the migration as already complete. Scene four sentence one, "OpenAI is publishing about its own infrastructure, so every number is self reported with no outside audit", is supported by the publisher itself, openai.com, by the first-person voice throughout, for example "Habitat is the online storage platform we built" and "Our data shows the Rust service is 6x more CPU efficient", and by the total absence anywhere on the page of a benchmark table, a described measurement methodology, a workload definition, a sample window, or any reference to third-party testing or audit of the efficiency, throughput, user count or petabyte figures. Scene four sentence two, "The migration is unfinished and part two is unpublished", is supported for its first half by "This new Rust service is now handling 95% of our production requests; we'll be deprecating Python entirely in the coming weeks", which states on its face that Python still serves some production traffic, and for its second half by "This post is the first in a two-part series on how we scaled online storage" and by "In part II of this series explaining how we rapidly scaled our online storage to serve over 1 billion ChatGPT users, we'll talk about the storage layer and how Habitat serves more than 500 petabytes and over 70 million requests every second", a future post that does not yet exist; I merged the two caveats into one sentence to remove a sentence boundary and its pause. The endcard narration, "Inboxsmith builds AI phone and email agents for small business. Please like and subscribe for more news.", is unchanged Inboxsmith promotional copy and makes no claim about OpenAI or Habitat. HONESTY CAVEAT: this is OpenAI publishing about OpenAI's own internal infrastructure, so every capability statement and every number in the script is the company describing and measuring itself. The seventy million requests per second, the one billion weekly people, the five hundred petabytes, the almost forty regions, the twenty million requests per second Python peak, the ninety five percent Rust share, the six times processor efficiency gain and the fifteen times memory efficiency gain are all self-reported, with no independent testing, no published benchmark methodology, no disclosed measurement window or workload definition, and no outside audit; the article's strongest sourcing language for the headline efficiency claim is simply "Our data shows". The Rust migration is also incomplete: the article states that Rust serves ninety five percent of production requests and that Python will be deprecated entirely only "in the coming weeks", so Python is still serving live traffic as of publication. Finally, the storage layer itself is deferred: the article explicitly labels itself "the first in a two-part series" and pushes the multi-tenancy reliability work, the read performance strategy and the cloud database scaling partnership into a future post that is unpublished, so the portion of Habitat that actually holds the data is not documented here at all. Scene four states these points plainly to the viewer without hedging. CORROBORATION STATUS: there is zero third-party corroboration of any kind behind this script. Zero transcripts were captured for this story; the local yt-dlp build on the producing box is too old and returned nothing, so no video transcript was available to cross-check any figure. The only hit the research probe returned was an entirely unrelated 2022 video titled "Growth Of Mould {Fungus} On Bread", which has no bearing on this story and provided no facts. No news wire report, no analyst note, no competing benchmark and no engineer commentary outside OpenAI was available, so the first-party article is the sole source behind every sentence of the narration, and the narration is written so that the disputable numbers carry explicit attribution to OpenAI. Nothing in the probes suggests this is a re-promotion of an older post: the probe returned no earlier item resembling this post, no prior Habitat write-up, and no prior Python to Rust migration announcement from OpenAI, and the article's own internal "Keep reading" list of recent engineering posts, dated August 25 2026, August 3 2026 and July 29 2026, covers unrelated inference, voice and model topics, which supports treating today's post as genuinely new. DELIBERATE OMISSIONS: I refused to carry over the article's figurative and promotional language, including "squeeze as much juice out of our existing stack", "fending off storage and compute capacity crunches to buy time", "throw caution to the wind", "the slowest database call is the one the user feels", "a strategic incursion of technical debt", "a calculated wager", "That bet eventually proved correct", "reduce the blast radius", "thundering herd", "escape hatch", "latency cliffs", "heavy-lifting", "no easy feat, but also not particularly challenging", "staggering user growth", "unprecedented rate", "unprecedented demand", "frontier scale", and the closing recruiting pitch inviting readers to "check out this open role on our team". In particular I did not write any variant of the idea that OpenAI used its own models on its own systems as a figure of speech; scene one states literally that the engineers used OpenAI's own Codex and coding models. For duration I dropped the following verified facts: that Habitat spans almost forty geographic regions; that it stores more than five hundred petabytes; that it began in mid 2024 as a small Python library connected to a single database and became a service by the middle of 2025; that OpenAI has grown more than ten times year over year for three years; that at its peak the Python service served more than twenty million requests every second; that Habitat is the second largest service by core count at OpenAI and fourth by Envoy footprint; the entire engineering middle of the article covering event loop scheduling delay, the feature flag configuration parsing incident, the connection pool reuse ordering metastable failure, connection fan-in, and the constrained query interface design tradeoff; and the named third-party components. I also avoided naming the cloud database product because the abbreviation in its name reads badly aloud, and avoided the library and protocol names in the article entirely because the speech engine garbles them. PROMPT INJECTION CHECK: I treated the article strictly as untrusted data and as a source of facts only. The captured text contains no instruction directed at an agent or a model. The imperative-looking strings in the capture are ordinary page chrome and interactive widget labels, specifically "Skip to main content", "Log in", "Login", "Try ChatGPT", "Share", "View all", "Play", "Pause", "Replay", "Keep reading", "(opens in a new window)", "Your privacy choices", and the navigation and footer link lists, plus one human-directed recruiting call to action, "check out this open role on our team". None of these attempted to make me run a command, write or edit a file, change configuration or permissions, reveal anything, fetch an additional URL, or ignore my own rules, and none of them were followed or treated as an instruction. No hidden or out-of-place instruction text was found anywhere in the body copy, the figure captions, the stat blocks or the chrome.(OpenAI's own first party engineering blog post, "Rapidly scaling online storage to serve over 1 billion ChatGPT users", by Jon Lee, Chaomin Yu, and Ben Ries, Members of Technical Staff, published on openai.com on September 11, 2026)