On 16 June 2026, the audited 2025 financial statements of OpenAI leaked via independent journalist Ed Zitron, were independently reviewed by the Financial Times, and made their way into an Ars Technica write-up that hit the front page of Hacker News within hours. The headline number — a $39 billion "net loss" — is misleading, and almost every angle in the post is downstream of one line item that the casual coverage has underweighted. The story is not that OpenAI is losing money. The story is the shape of the loss: where it goes, who it goes to, and what the trajectory implies about the IPO that the company is now filing for.
The 2025 numbers, as reported in the audited statements (revenue, R&D, cost of revenue, sales & marketing, loss from operations, headline net loss), tell a coherent story when you stack them. Revenue: $3.7B in 2024, $13.07B in 2025. Loss from operations: $8.78B in 2024, $20.92B in 2025. R&D: $7.81B in 2024, $19.18B in 2025. Of that 2025 R&D, $10.59B was paid to Microsoft as part of the cloud and compute partnership. Cost of revenue (inference-time compute, primarily): $2.65B in 2024, $7.5B in 2025. Sales and marketing: $1.11B in 2024, $5.73B in 2025. The headline net loss of $39B includes a roughly $30B one-time accounting charge tied to the company's 2025 conversion to a for-profit structure. Strip that out, per the FT's reporting, and the 2025 net loss is closer to $8B — which is still enormous, but the order of magnitude is different.
Angle 1: The headline $39B is a one-time charge, not a run-rate
This is the most important framing correction. The $39B "net loss" number that hit the front page is not what OpenAI is burning through 2026. It is a paper charge related to the conversion from a non-profit capped-profit structure to a fully for-profit one. The mechanism: when investor valuations shift during a structural reorganization, the accounting books revalue prior commitments, and the difference lands on the income statement as a one-time hit. The FT cited "a person familiar with the matter" putting the 2025 net loss at roughly $8B without that charge. $8B is still a 64% revenue multiple in losses. It is not the apocalyptic $39B figure that the Reddit threads are running with, and that distinction matters for how serious readers read the rest of the line items.
The $20.92B "loss from operations" number, by contrast, is a run-rate. That is the number that reflects what OpenAI spent, day-to-day, to operate in 2025 — and it grew 138% year-over-year, against revenue that grew 253%. As a percentage of revenue, operating losses improved from 237% in 2024 to 160% in 2025. The unit economics are getting less bad. They are not yet close to zero. The company has guided to profitability by 2030, and the loss-from-operations trajectory is consistent with that guidance if the cost-growth curve bends and the revenue-growth curve does not.
Angle 2: Microsoft is the single largest line item that is not a line item
The $10.59B of $19.18B R&D paid to Microsoft in 2025 is the story, and the Ars Technica write-up flags it but does not foreground it. That is more than half of OpenAI's entire R&D spend, going to one supplier, on a compute contract that is — per public reporting on the 2023 partnership extension — capacity-constrained and price-fixed through at least 2030. This is not a vendor relationship. It is a structural dependency.
The implication: OpenAI's "loss from operations" is, in a real sense, a Microsoft rent bill. The company can grow revenue as fast as it wants, but if its marginal inference cost is set by Azure compute pricing and the partnership cap is what it is, the operating-loss trajectory is bounded by the unit economics of the Azure deal. The 2025 numbers make this concrete. Cost of revenue went from $2.65B to $7.5B — a 183% jump — which tracks with the inference volume growth ChatGPT saw in the same window (900M weekly active users reported, of which roughly 50M are paid subscribers). Inference is now the second-largest cost line, behind R&D, and it is the one that scales with usage. R&D, by contrast, is mostly fixed (training runs) plus the Microsoft commitment.
Angle 3: The paid-subscriber math is the actual unit-economics story
OpenAI reports 900M weekly active ChatGPT users, of which roughly 50M are paid subscribers. At a blended subscription price point somewhere in the $20-$25/month range (the Plus tier, weighted by the smaller Pro and Team populations), the annual subscription revenue run-rate is plausibly in the $12-15B neighborhood. The remainder of the $13.07B 2025 revenue is API access (ChatGPT Enterprise, the OpenAI API for third parties) plus a smaller Microsoft Azure resale line. Of those three streams, the subscription one is the only one with positive gross margin at any reasonable scale; the API is inference-cost-heavy; the Microsoft resale is mostly a pass-through.
Per-paid-subscriber unit economics: $20.92B operating loss / 50M paid subs = roughly $418 of operating loss per paid subscriber per year. If you assume the average paid subscriber is generating around $240/year of subscription revenue (Plus tier at $20/month × 12), OpenAI is losing $1.74 for every $1 of subscription revenue. The unit economics are still deeply negative. The improvement from 2024 (where the multiple was worse, on a smaller subscriber base) is real. The gap to break-even is still large.
The strategic question this raises: what happens to the paid-subscriber base when local models cross the threshold for the "good enough" workflows? This blog covered the Vicki Boykis "running local models is good now" inflection two days ago; the implication there is that 25-50% of the workflows that currently route to ChatGPT Plus are now viable on a local Gemma 4 26B. If even 10% of paid subscribers migrate to local, the unit-economics curve bends the wrong way. The 2025 financials are the high-water mark for "people pay $20/month for a frontier chat." The 2026 and 2027 numbers will show whether that base holds.
Angle 4: The Microsoft $30B charge is a tax on the IPO structure, not a tax on the business
The single largest accounting event of 2025 was the conversion from capped-profit to for-profit, which is the structural prerequisite for the IPO paperwork OpenAI is now filing. The roughly $30B charge is the fair-value re-measurement of the prior investor commitments against the new equity structure. This is the kind of line item that shows up once, in the year of conversion, and never recurs. Auditors (and the SEC) will flag it. Analysts will adjust for it. The press will, eventually, stop quoting it.
The more durable read is the operating-loss line, the R&D-to-Microsoft line, and the cost-of-revenue growth rate. Those three are the things that compound. A company can absorb a one-time $30B accounting charge and survive. A company whose cost of revenue grows 183% year-over-year cannot, at this rate, sustain 160% operating losses indefinitely. The 2030 profitability guidance requires cost-of-revenue growth to slow, R&D-to-Microsoft to stay flat or decline (i.e., the Azure partnership terms to renegotiate), and the revenue line to keep compounding at 50%+ CAGR. Two of those three are within OpenAI's control. The middle one is not.
Angle 5: What the S&M jump tells you about the ChatGPT business
Sales and marketing went from $1.11B in 2024 to $5.73B in 2025 — a 5.16× increase, far outpacing the 3.53× revenue growth. As a percentage of revenue, S&M went from 30% to 44%. This is the line item that says the most about the underlying business. Frontier AI labs that are growing primarily by word-of-mouth and developer adoption (Anthropic, the open-weights tier) spend single-digit percentages of revenue on S&M. OpenAI is now spending nearly half of revenue on customer acquisition.
The HN thread had two comments that triangulated this from different angles. "iaaan" reported physical billboards for ChatGPT in the Portland, OR area, and asked what return those have. "themafia" replied at the top level: "I don't understand the 'sales and marketing' cost…It's so polarizing I can't imagine how that $5.7B is being spent." A follow-up reply by "dylan604" suggested the line item is paying for influencers to set up "kool-aid stands." Neither framed it in S&M-as-percentage terms, but both are pointing at the same phenomenon: OpenAI is now in the customer-acquisition-cost regime that consumer software companies enter when organic growth plateaus. The 900M weekly active number is large. The 50M paid conversion — 5.5% — is not. The reason the conversion rate is not improving is that the $20/month price point is now competing with a local tier that crossed the "good enough" threshold.
Angle 6: The IPO is the strategic context for the leak
OpenAI is filing SEC paperwork for an expected IPO. The leaked statements are from 2025; the IPO will price on 2026 numbers plus a forward projection. The question the prospectus has to answer is: at what 2027-2028 revenue and cost-of-revenue trajectory does the operating loss line bend to zero? The 2025 audited statements are the historical baseline; the S-1 will project forward. Every dollar of Microsoft R&D, every dollar of inference cost, every dollar of S&M is now a number that an underwriter has to defend at a roadshow.
This is the part of the story that is genuinely novel and that the front-page coverage has not emphasized. The leak is not a leak for its own sake; it is a leak into the middle of an SEC review. The numbers, the trends, and the trajectory are now public record in a way that constrains what the S-1 can claim. Operating losses improving from 237% to 160% of revenue is a real story and a defensible narrative. A $39B "net loss" that the average reader will not parse as a one-time charge is a story that hurts the IPO, and the company's communications team will spend the next 90 days working to reframe it.
The original take: the per-subscriber line is what the 2026 numbers will be judged on
The most common read of the 2025 financials in the press and the HN thread is that OpenAI is "losing billions." That is true and it is not useful. The more useful framing is: OpenAI is a $13B-revenue business that is losing $20.9B from operations, of which $10.6B is a single Microsoft contract. The 2026 numbers — when they leak, or when they appear in the S-1 — will be read against three questions, not one.
- Did paid-subscriber growth keep pace with 2025's pace, or did the 900M-weekly-active / 50M-paid gap close at all?
- Did cost of revenue grow slower than revenue, or faster? (The 2025 numbers had cost of revenue growing 183% against revenue at 253% — a favorable ratio, barely.)
- Did the Microsoft R&D line stay flat, or did the 2026 number push above $11B? If it pushed above $11B, the IPO narrative is "we are growing into a structural cost we cannot control." If it stayed flat or dropped, the narrative is "we are scaling past the fixed compute commitment."
The 2025 financials, read this way, are not a "losing billions" story. They are a story about a $13B business whose next 18 months will be read at the per-subscriber and per-inference-call level. The pre-2025 AI-lab financials (Anthropic, Mistral, Cohere) are private and not directly comparable. The closest public comp is Google's "Other Bets" line, which includes DeepMind and runs an operating loss on a much larger revenue base (the specific 2025 figure should be checked against Alphabet's most recent 10-K before quoting; the directional read is "comparable-scale operating loss, vastly larger revenue"). OpenAI is making the same bet — that the AI line will eventually be large enough to absorb its own R&D cost — on a tighter runway, with a single-supplier compute dependency that is not Google's.
What this means for you
If you are a developer or a small team paying for ChatGPT Plus, the 2025 financials do not change your short-term calculus. The price is not going up in 2026; if anything, the S&M line item is evidence the company has pricing room. The thing worth tracking is the paid-subscriber base: if 2026 shows a flattening or decline, the price-stability assumption breaks.
If you are a startup building on the OpenAI API, the cost-of-revenue trajectory is the line that matters. Inference pricing has reportedly been declining sharply year-over-year on the public benchmarks (the rule-of-thumb figure is in the 70-80% range, though the exact rate depends on which benchmark and which model family you anchor to); the question is whether OpenAI can keep pricing flat or pushing lower while its own cost-of-revenue grows. If cost-of-revenue growth in 2026 outpaces 2025's 183% rate, the unit economics on the API tighten, and either pricing has to rise (unlikely during an IPO year) or the company has to renegotiate the Microsoft deal.
If you are a founder or an enterprise buyer, the Microsoft dependency is the strategic line item. Every API call routed through OpenAI is, indirectly, routing through Azure. The diversification argument — "we are not locked into one cloud" — does not hold for OpenAI-routed workloads. The 2025 financials are the first time this dependency has been quantified in audited statements; it was speculated about for years, and the $10.59B number makes it concrete.
What to do this week
# Step 1. Pull the full Ars Technica article (primary source) so the
# numbers above are not the only version of the story you are
# anchoring on:
curl -sL --compressed --max-time 20 -A "Mozilla/5.0" \
"https://arstechnica.com/ai/2026/06/leaked-financial-docs-show-openai-is-losing-billions-of-dollars-a-year/" \
-o /tmp/openai-2025.html
# The audited numbers ($3.7B/$13.07B revenue, $7.81B/$19.18B R&D,
# $10.59B Microsoft, $2.65B/$7.5B cost of revenue, $1.11B/$5.73B
# S&M, $8.78B/$20.92B loss from operations) are all in the body
# of that article; FT's $8B-adjusted-net-loss framing is in the
# same write-up.
# Step 2. If your stack runs on the OpenAI API, run a one-week
# shadow of token usage and pricing against a local-tier model
# (Gemma 4 26B or Qwen 3 30B-A3B). The point is not to migrate.
# The point is to know what fraction of your API bill is on
# workflows the local tier now covers — that fraction is the
# negotiating room you have if 2026 cost-of-revenue growth
# forces OpenAI to push API pricing.
# Step 3. If you are an enterprise buyer, file the question with
# procurement: "What fraction of our AI spend routes through
# Azure, via OpenAI, and is that the diversification posture we
# think we have?" The 2025 financials are the first public
# evidence that the answer is "more than you assumed."
# Step 4. Read both HN threads (48577208, the post-Ars write-up
# thread, and 48550465, the prior thread where Ed Zitron first
# surfaced the leak). The simonw comment in the 48577208
# thread is the explicit pointer between the two. The 48550465
# thread is where the "what the people who were paying
# attention already knew" framing originates — read both
# before you form a position on the 2025 numbers.
Related reads from this blog
- Your Local Model Is a Faster Google (And Now It Loops, Too) — the Vicki Boykis "local models are good now" post. The 75% threshold for agentic work on a local model is what the "did paid-subscriber growth keep pace in 2026" question is, at the user level. The 2025 financials are the macro read; the local-model post is the per-workflow read.
- An AI Agent Submitted Code to Fedora. Maintainers Merged It. — the cost-control angle at the per-task level. The Fedora incident is what happens when a frontier-tier agent runs unattended; the $5.73B 2025 S&M line is the company-level version of the same dynamic.
- An AI Agent Burned $6,531 on AWS to Scan a Hobby Network Nobody Asked It to Scan — the dn42 incident as a per-task cost-control case study. The relevant line in the 2025 financials is cost of revenue: $7.5B, growing 183% YoY. The dn42 incident is the per-task version of what that line aggregates to at scale.
Disclosure
Drafted with AI assistance. Primary source: Kyle Orland, "Leaked financial docs show OpenAI is losing billions of dollars a year," Ars Technica, 16 June 2026 —
curl -L --compressed, 18 June 2026. Audited figures (revenue $3.7B/$13.07B; R&D $7.81B/$19.18B incl. $10.59B to Microsoft; cost of revenue $2.65B/$7.5B; S&M $1.11B/$5.73B; loss from operations $8.78B/$20.92B; net loss $5B/$39B with ~$8B adjusted 2025 net loss net of a ~$30B for-profit-conversion charge) are from the Ars article, which sourced them from Ed Zitron's leak and the FT's review. 900M weekly active / 50M paid subscribers, $122B round, $852B valuation: same source. HN item 48577208 (197 points, 116 comments at API snapshot) via Algolia HN Search, 18 June 2026. The 237%→160% of revenue, $418/sub/year, and >50% of R&D to Microsoft figures are this blog's arithmetic on the source line items, not direct claims. The "iaaan" / "themafia" / "dylan604" HN comment references are direct quotes from the Algolia API response. The $20-$25/month subscription range is a read of public Plus/Pro/Team pricing, not a verified blended average. The "10% migrate to local" scenario and the 25-50% / 75% local-workflow thresholds are thought experiments that reference the Vicki Boykis piece linked in Related reads, not direct claims. The "Other Bets / DeepMind in the same neighborhood" framing is this blog's directional read of Alphabet's 10-K; the specific 2025 figure should be checked against the filing before quoting. The per-subscriber/per-inference framing in the original-take section is this blog's editorial position.
Sources
- Kyle Orland, "Leaked financial docs show OpenAI is losing billions of dollars a year," Ars Technica, 16 June 2026 —
https://arstechnica.com/ai/2026/06/leaked-financial-docs-show-openai-is-losing-billions-of-dollars-a-year/ - Ed Zitron, "OpenAI Losses Increased Nearly 8X in 2025, with Spending Hitting $34B," Where's Your Ed At (Zitron's newsletter), 16 June 2026 — the original source of the leak; the HN item 48550465 links to this piece at
https://www.wheresyoured.at/exclusive-openai-financials/ - Hacker News thread (197 points, 116 comments at time of writing; numbers move as the thread ages) on the Ars Technica article, item 48577208 —
https://news.ycombinator.com/item?id=48577208 - Algolia HN Search API metadata for item 48577208 (the source for the point/comment counts and commenter references) —
https://hn.algolia.com/api/v1/items/48577208 - Vicki Boykis, "Running local models is good now," 15 June 2026 (referenced as the "75% threshold" framing for the per-subscriber migration scenario) —
https://vickiboykis.com/2026/06/15/running-local-models-is-good-now/
