HN Debrief

Leaked OpenAI financials show $38.5B loss and compute burn

  • AI
  • Startups
  • Economics
  • Infrastructure
  • Public Markets

The article presented leaked OpenAI numbers as evidence of catastrophic losses, centered on a reported $38.5 billion 2025 net loss and massive spending on compute, R&D, and sales. The useful framing that emerged is that the headline number is not the business. Multiple commenters pointed to Financial Times reporting that about $30 billion of the 2025 loss came from a non-cash accounting charge tied to OpenAI’s restructuring from its earlier ownership setup into a public benefit corporation. On that view, the operating loss is closer to $8 billion, still enormous but much less sensational than the headline suggests.

Treat AI lab financial headlines like startup metrics with extra accounting noise. The practical test is simpler: separate one-time restructuring charges from recurring model, compute, and go-to-market costs, then ask whether the company still has pricing power once model quality converges and switching stays cheap.

Discussion mood

Mostly skeptical and corrective. Readers thought the headline was sensational because the $38.5 billion figure appears dominated by a one-time accounting charge, but they were still doubtful about long-term AI lab economics because training, pricing pressure, and weak moats could keep recurring profits elusive.

Key insights

  1. 01

    The $38.5B loss is mostly accounting

    The giant loss figure is inflated by a non-cash charge from OpenAI’s restructuring, not by a sudden collapse in underlying operations. The accounting treatment reflects investor rights that had to be marked as liabilities as the company’s valuation rose, so the number says more about ownership mechanics than about cash burn in the business.

    Do not use net loss alone to judge AI lab viability. Strip out capital structure effects first, then compare recurring revenue, compute, and R&D needs.

      Attribution:
    • tptacek #1
    • ThrustVectoring #1 #2
    • datadrivenangel #1
    • ltononro #1
  2. 02

    Inference may already be gross profitable

    The leaked split between revenue and cost of revenue suggests serving existing models is not obviously underwater. That does not settle the whole business case, but it does kill the simpler claim that every token is being sold at a loss and that usage growth automatically deepens the hole.

    If you buy or build on model APIs, watch cost of revenue separately from total operating loss. A lab can have a usable serving business even while the full frontier race remains cash destructive.

      Attribution:
    • nl #1
    • maxnevermind #1
    • HlessClaudesman #1
    • barrkel #1
    • danielmarkbruce #1 #2
    • spongebobstoes #1
  3. 03

    R&D classification hides the hard question

    The real fight is not over accounting vocabulary but over economic necessity. If training new models, refreshing knowledge, and securing compute are mandatory to stay relevant, then pushing those costs into R&D does not make them strategically optional. It just moves them below the gross margin line while the business still depends on them.

    When evaluating AI vendors, treat ongoing training and model refresh as recurring competitive spend unless there is clear evidence the model can stay differentiated without it. Reported gross margin alone is not enough.

      Attribution:
    • surgical_fire #1 #2
    • ElProlactin #1
    • simoncion #1
    • lrae #1
  4. 04

    The leak undercut the loudest doom framing

    Several commenters noted that the leaked figures weaken the most dramatic anti-OpenAI takes, especially claims built on misunderstanding consolidated accounting or on the headline loss alone. The documents still support concern about cash needs, but not the idea that the numbers reveal hidden fraud or that the service business is fake on its face.

    Be careful about using polemical industry critics as primary sources on financial structure. For decisions, rely on the underlying line items and what they imply operationally.

      Attribution:
    • N_Lens #1
    • sho #1
    • ElProlactin #1
  5. 05

    Enterprises can mix premium and cheap models

    A practical threat to revenue growth is architectural, not ideological. Enterprises can use a top model for planning and routing, then hand execution to cheaper hosted or local models. If that workflow becomes standard, frontier labs keep the prestige role but lose a lot of the token volume that supports their revenue projections.

    If you sell AI products, design for multi-model routing now. If you invest in model providers, discount forecasts that assume most downstream work stays on the most expensive model.

      Attribution:
    • JSR_FDED #1 #2
    • ted_dunning #1
    • ZaoLahma #1
  6. 06

    AI economics look more like media infrastructure

    A recurring comparison was to utilities, video streaming, and live media rather than to classic software. Those businesses have real delivery costs, massive infrastructure bills, and weaker operating leverage than SaaS. That framing makes it easier to see why huge revenue can coexist with stubbornly thin profits.

    Use infrastructure and media comps, not pure software comps, when pressure-testing AI model businesses. Expect lower long-term margins unless distribution or lock-in becomes much stronger.

      Attribution:
    • bxk76 #1
    • dofm #1
    • hackcasual #1
    • Ekaros #1
    • echelon #1
  7. 07

    IPO viability depends on whether training is mortgage or optional growth

    The sharpest valuation analogy compared model development to the mortgage on an apartment building, not to discretionary expansion. If frontier performance must keep improving to defend revenue, then training spend behaves like core capital cost. That makes an IPO less about 'when do they stop losing money' and more about whether public investors will fund a permanently expensive arms race.

    Before backing an AI IPO, decide whether new training cycles are optional growth bets or unavoidable maintenance capex. That single assumption changes the valuation case more than any headline revenue number.

      Attribution:
    • bayarearefugee #1
    • grey-area #1
    • cmiles8 #1
    • etempleton #1

Against the grain

  1. 01

    $13B revenue this fast is extraordinary

    Even critics of the valuation conceded that going from a research lab to roughly Fortune 500 scale revenue in just a few years is unusual by any tech standard. That speed suggests genuine market pull, not just hype, even if the cost structure remains ugly.

    Do not confuse a questionable valuation with lack of product-market fit. A company can be strategically important and still overpriced.

      Attribution:
    • simonw #1 #2
    • jacobgold #1
  2. 02

    OpenAI still has product stickiness

    Against the 'no moat' claims, some users said ChatGPT and Codex remain materially better in real workflows, especially coding and review, and that brand plus tooling matter more than benchmark deltas. If many users feel that way, switching will be slower than raw model comparisons imply.

    Benchmark parity is not enough to win accounts. If you compete with OpenAI, invest in workflow integration and user experience, not just model quality.

      Attribution:
    • arjie #1
    • jacobgold #1
    • vunderba #1

In plain english

cost of revenue
The direct costs tied to delivering a product or service, such as compute or hosting, before overhead like research or marketing.
gross margin
Revenue minus direct cost of revenue, usually expressed as a percentage of revenue.
inference
Running a trained AI model to generate outputs for users, as opposed to training the model.
IPO
Initial Public Offering, when a private company first sells shares to public market investors.
moat
A durable competitive advantage that makes it hard for rivals to take customers or profits.
non-cash accounting charge
An expense recorded on financial statements that does not involve money leaving the business at that moment.
open-weight models
AI models whose trained parameters, called weights, are released so others can run or adapt them.
public benefit corporation
A corporate structure that is allowed to pursue stated public goals alongside profit.
Qwen
A family of AI models developed by Alibaba.
SaaS
Software as a Service, software delivered over the internet by subscription.

Reference links

Financial reporting and source material

Accounting and company disclosures

  • Google Q1 2026 10-Q
    Cited to show standard public-company separation between cost of revenues, R&D, sales and marketing, and G&A.

Model pricing and market competition

Company and product references

Space and valuation analogies

Culture and political references