商业 · AI营收会计

OpenAI营收相差200亿美元：这是会计口径差异，并非销售额消失

合作伙伴销售额的会计处理解释了OpenAI据报9月营收运行率为500亿美元与此前估算之间的差额，但并未解决AI基础设施扩张背后的经济效益问题。

播出 文字 阅读需 7 分钟209 个来源HeyDay News Desk · About usWatch report
与 HeyDex 一起探索

通过 RSS 关注

原语言视频

文字实录 · 正在加载

译自英文原文。

新闻追踪第 5 天

1 篇报道130 家媒体209 个来源5 个视频

  • 暂无标注日期的事件
  1. 第 1 次更新OpenAI营收相差200亿美元：这是会计口径差异，并非销售额消失当前位置

10月8日，CoreWeave股价下跌7.77%，投资者面对的是OpenAI营收报道中约200亿美元的差额。市值损失是真实的。销售额似乎消失的说法则尚未得到证实：较大的营收数字对云合作伙伴交易采用了不同的计算方式。59 7

据报道，OpenAI告诉投资者，截至9月底，其年化营收增速接近500亿美元，而不是9月底流传的680亿至700亿美元。《金融时报》率先报道了较低数字，CNBC称已予以证实，路透社也作了报道。这些是关于向投资者披露信息的报道，并非公司公布的经审计业绩。1 16

这一差异影响重大，但并不能证明需求已经崩塌。投资者曾调整OpenAI的数字，以近似采用Anthropic对云合作伙伴销售额的处理方式。OpenAI自身的数字只计入其从某些合作伙伴销售中保留的份额；较大的估算则计入客户支付的全额。同一项底层业务因呈现方式不同，产生了不同的标题数字。7

关键事实

  1. ~$50bn据报道的9月底年化营收运行率 116
  2. $68bn–$70bn包含合作伙伴销售总额的此前估算 17
  3. $70bn+据报道的年底目标，并非已实现营收 172

一笔付款，两种营收数字

按照报道中Anthropic对云合作伙伴销售额采用的总额法，客户全额付款计入营收，合作伙伴所得份额则计为费用。按照OpenAI对某些合作伙伴销售额采用的净额法，只有OpenAI所得份额计入营收。OpenAI此前的估算是向投资者提供的总额化数字，旨在使比较更接近同类对比。7

据称，根据对客户关系控制权及交付情况的判断，两种处理方式都符合公认会计原则。这一说法并不意味着对任何一家公司的数据进行了独立审计，也不意味着两家公司的营收数字都可以直接比较。7

这里尤其需要谨慎理解净额一词。10月9日，彭博电视的评论员提出，约500亿美元的数字可能是扣除微软和亚马逊等供应商付款后的营收。更准确的会计解释是，只计入合作伙伴销售中留给OpenAI的份额，而不是减去所有云基础设施账单。若将500亿美元称为利润，或称为扣除所有计算成本后的营收，就会把一个有用的区别变成另一个误导性标题。4 5 7

年化处理还带来了另一个局限：它将较短时期的销售速度推算到全年。这并不意味着OpenAI已经在2026年收取了500亿美元、赚得了这笔利润，或已通过合同确保这笔收入全部具有经常性。具体测算窗口和公式尚未公开，总额与净额数字之间逐笔调节的明细也未公开。16 7

这也让与Anthropic进行整齐的排名比较变得困难。Anthropic据报在7月底的年化营收达到约650亿美元，这一数字于8月公布。将采用合作伙伴销售总额口径的数字，与OpenAI采用合作伙伴销售净额口径的9月数字相比较，混用了不同会计口径和相隔两个月的观测值。这些数字无法证明哪家公司拥有更大的可比营收业务。178 179 7

增长考验仍在前方

投资者演示材料并未描述一家业务萎缩的公司。CNBC报道称，第三季度总营收运行率增长77%，企业业务的运行率增长107%。这是7月至9月期间的环比变化，并非同比增幅。企业业务的百分比衡量的是增长，而不是企业业务占总销售额的比重。具体的年化计算公式仍未公开。1 120

路透社的报道显示，2026年初，OpenAI的年化营收约为200亿美元。这提供了有用的历史背景，但并非一套经过独立审计、口径完全一致的数据序列，不能据此稳妥地计算之后的每一项增长率。16

另一个700亿美元数字确实存在，但指的是未来。彭博社10月9日报道，OpenAI预计到2026年12月31日，年化营收将达到或超过700亿美元，增长主要来自企业业务。这是年底预期，不是此前经总额调整的9月估算，也不是对2026日历年总收入将达到700亿美元的预测。172 6

如果9月约500亿美元的运行率与12月目标采用一致口径，OpenAI需要在最后一个季度将年化营收运行率提高约40%。这是有条件的算术推算，并非已经实现的增长。企业业务预计贡献的具体金额，以及9月指标与预测之间的完整调节说明，均未公开。172

不过，市场反应剧烈。英伟达10月8日收盘下跌2.94%，报230.48美元；甲骨文下跌5.48%，报135.69美元；CoreWeave下跌7.77%，报81.58美元。纳斯达克综合指数下跌1.25%，收于27,193.34点。这些精确数据细化了CNBC对当日跌幅约3%、6%、8%和1%的概述，并不代表另一轮独立抛售。60 100 59 87 1

AI供应商跌幅更大，且下跌恰逢营收澄清，这支持了这样的解读：市场对基础设施支出背后需求的疑虑，促成了股价下跌。但这无法将该报道的影响与美国国债收益率上升或其他市场压力区分开来。OpenAI是一家非上市公司：下跌的是供应商的交易股价，并非OpenAI股价暴跌。1 167

万亿美元雄心需要更清晰的证据

会计口径问题出现之际，另一场规模庞大的融资行动也在推进。彭博社和路透社9月29日报道，OpenAI正在讨论至少300亿美元的新融资，投前估值约为1.4万亿美元——即尚未计入这笔新资金。谈判处于早期阶段，条款可能变化，截至10月9日尚未确认交易完成。据称，这笔资金将作为过桥融资，而首次公开募股（IPO）将推迟至2026年之后。36 41

3月的融资轮次可提供规模参照，但其估值并不能直接比较。该轮融资承诺资金1220亿美元，投后估值为8520亿美元，其中已包含所筹资金。附带条件的承诺和分期付款安排也意味着，总额不能证明每一美元都已以现金到账。此前的融资和最近的讨论，均不能证明新提议的资金已经到账。187 190

更大的问题在于，付费需求能否支撑计算基础设施的扩张。彭博情报的曼迪普·辛格认为，前沿AI营收需要与资本支出同步增长，才能缓解市场对回报的担忧。更多的token消耗——即更多模型处理量——本身并不能证明营收按比例增长，也不能证明利润率有吸引力。这些披露并未同时提供经核实的OpenAI利润或自由现金流业绩。5 6

彭博社9月25日报道的另一项高盛预测称，美国五大超大规模云服务商的AI基础设施支出在2027年将约为1.2万亿美元；2026年预计约为8000亿美元，而当时市场对2027年的普遍预期约为1.1万亿美元。这些是行业预测，并非OpenAI的预算，也不是已经花出去的钱。54

彭博电视讨论的明年支出范围更广，为1.2万亿至1.3万亿美元，但其统计范围尚未厘清。不应以此取代高盛口径较窄的预测。将几家云公司的基础设施支出与一家AI实验室的营收相比较，也无法得出衡量整个行业回报的有意义指标。6 54

已知

  • 9月数字、此前经总额调整的估算和12月目标，描述的是不同指标或不同时间点。 17172

未知

  • 公开记录没有提供详细的营收调节说明、经核实的盈利情况，也没有确认拟议融资已经完成。

后续

  • OpenAI能否在12月31日前，以明确且一致的口径达到至少700亿美元的年化营收运行率，以及付费企业需求将作出多大贡献。

这次澄清排除了一种错误结论：不同会计呈现之间180亿至200亿美元的差额，并不意味着销售额相应减少。更难回答的投资问题依然存在。OpenAI必须将快速增长转化为足以支撑其雄心的商业模式。仅仅报出一个更大的数字，并不能回答这个问题。

播出原貌 28 行
  1. CoreWeave shares fell 7.77% in a sell-off that hit artificial-intelligence suppliers. OpenAI’s September revenue pace was reportedly about fifty billion dollars a year, below the sixty-eight to seventy billion headline investors had seen. Different ways of counting partner sales explain the gap. The next test is whether OpenAI can reach its year-end target on a consistent accounting basis.
  2. OpenAI reportedly gave investors the fifty-billion-dollar figure for September’s end. An annualized run rate projects a recent sales pace over a full year. That pace can change, and the company still has operating costs to pay.
  3. The figure doesn't tell us how much cash OpenAI collected during 2026, or how much profit it made. It also doesn't guarantee those sales are contractually recurring. The correction concerns what investors thought the revenue headline measured.
  4. The earlier sixty-eight to seventy billion dollars came from investors adjusting OpenAI’s numbers to match Anthropic’s partner-sales method. OpenAI’s own figure counted its retained share of certain partner sales. That produces an eighteen-to-twenty-billion-dollar measurement gap without showing an equivalent loss of sales.
  5. A customer payment can support two different revenue headlines. The underlying demand can stay exactly the same. The sixty-eight-billion and near-seventy-billion estimates were variants of the same adjusted figure, rather than separate streams of revenue.
  6. Anthropic counts the full customer payment through cloud partners as revenue, then treats the partner’s cut as an expense. OpenAI records only its own share of certain partner sales. Each approach depends on who controls the customer relationship and delivery.
  7. Both treatments have been described as complying with generally accepted accounting principles. That description doesn't amount to an audit of either company. Earlier explanations suggested OpenAI’s figure came after payments to Microsoft and Amazon. The more precise explanation concerns its share of partner sales. Computing costs still have to be paid separately.
  8. Anthropic’s reported pace reached about sixty-five billion dollars at July’s end. OpenAI’s fifty-billion-dollar observation came two months later. Those different dates and accounting bases prevent these figures from telling us which company had the larger comparable revenue business.
  9. OpenAI’s investor presentation also described rapid expansion during the July-to-September quarter. Its total revenue run rate grew seventy-seven percent, while its enterprise run rate grew one hundred and seven percent. Those percentages describe sequential growth within the third quarter.
  10. Enterprise customers are businesses paying for AI services. The one-hundred-and-seven-percent figure measures how fast that business grew. It doesn't mean enterprise supplied that percentage of total sales. OpenAI expects these customers to drive much of its next increase.
  11. OpenAI began 2026 at a reported annualized pace of about twenty billion dollars. A figure of about six billion circulated in 2024. Those older numbers give a sense of expansion, but changing measurement details make them unsuitable for a precise, like-for-like growth calculation.
  12. OpenAI reportedly still expects its annualized pace to reach at least seventy billion dollars by December thirty-first. That is a separate seventy-billion-dollar figure, with a future deadline. Enterprise growth is expected to supply much of the increase.
  13. If September and December use the same measurement basis, reaching that target requires about twenty billion dollars more in annualized pace. That's roughly forty percent growth during the final quarter. OpenAI competes for business customers with Google Cloud and Meta, but their effect on its pricing and growth hasn't been quantified.
  14. Nvidia closed down 2.94% on October eighth. Oracle lost 5.48%, and CoreWeave fell 7.77%. The Nasdaq Composite declined 1.25%. The heavier losses in AI suppliers show where concerns about demand behind the infrastructure buildout were concentrated.
  15. These closing percentages refine the earlier rounded descriptions of three, six, eight and one percent. The revenue clarification coincided with the sell-off, and concern about AI demand may have contributed. Rising Treasury yields also pressured markets, so the revenue news alone can't explain every percentage point lost.
  16. Nvidia finished at $230.48, Oracle at $135.69, and CoreWeave at $81.58. The Nasdaq closed at 27,193.34, losing 345.35 points. OpenAI is privately held, so the immediate losses reduced the value of investors’ holdings in other companies.
  17. OpenAI was also asking prospective investors to consider another enormous financing round. Discussions disclosed in late September sought at least thirty billion dollars at a valuation of about one point four trillion dollars before adding the new funding.
  18. March’s round brought one hundred and twenty-two billion dollars in commitments at an eight-hundred-and-fifty-two-billion-dollar valuation that included the new money. September’s proposed valuation excludes new funding. Comparing the two headlines directly would mix different valuation bases.
  19. The new discussions were early-stage, and terms could change. No closing had been announced through October ninth. The proposed round would provide bridge financing while a public offering was deferred beyond 2026. Investors were weighing that trillion-dollar valuation as the meaning of the revenue headline shifted.
  20. Amazon committed up to fifty billion dollars in the earlier round. Nvidia and SoftBank committed thirty billion each, alongside other investors. Some funding depended on conditions or scheduled payments, so the commitment total doesn't show that all the cash had arrived.
  21. The larger investment question is whether paying demand can support the computing buildout. More AI processing doesn't automatically produce proportionally more revenue. Token consumption measures model processing, while prices and margins determine how much that activity can earn.
  22. Revenue growth has to cover computing, staffing and other costs before it can produce profit. OpenAI’s disclosures don't give investors a complete picture of profitability or cash generation. A faster sales pace alone can't settle the question of returns.
  23. Goldman Sachs forecasts an even larger industry buildout. Its estimate covers the five biggest American hyperscalers, the companies operating enormous cloud-computing systems. It puts their AI infrastructure spending at about eight hundred billion dollars in 2026.
  24. Goldman expects about one point two trillion dollars in 2027, against a Wall Street consensus of about one point one trillion. These figures cover several cloud companies’ infrastructure plans. They describe expected spending across that group, rather than OpenAI’s own budget.
  25. The underlying analysis projects about fifty-four percent growth in 2027, after nearly one hundred percent in 2026. The rounded eight-hundred-billion and one-point-two-trillion endpoints imply fifty percent. Rounding explains why those displayed totals don't reproduce the underlying growth rate exactly.
  26. A broader estimate put next-year spending by hyperscalers and newer cloud providers at one point two to one point three trillion dollars. The boundaries of that group weren't clear. Comparing several companies’ spending with one lab’s revenue can't measure the sector’s overall return.
  27. The accounting explanation accounts for the smaller OpenAI headline without demonstrating a sudden collapse in demand. But OpenAI hasn't publicly supplied the detailed reconciliation or calculation behind its September pace. Investors still need a consistent basis for judging whether its growth can support the valuation.
  28. OpenAI’s next measurable test arrives on December thirty-first. It expects at least a seventy-billion-dollar annualized pace, with enterprise customers driving much of the increase. Reaching that level on a clearly stated, consistent basis would show how much paying demand has actually grown.

来源

209 个来源129 家媒体5 个视频3 篇通稿

来源地图

追溯本篇报道的依据。图谱展示其关联；下方链接可打开本文引用的来源。

来源图谱将在您阅读时加载。

探索完整的新闻编辑室图谱 ↗

译文更新于

修订记录

  1. r1首次发布。