TMTB：与微软首席执行官萨蒂亚·纳德拉和开放人工智能公司（OAI）首席执行官山姆·奥特曼的BG2播客要点

TMTB：与微软⾸席执⾏官萨蒂亚·纳德拉和开放⼈⼯智能公司（OAI）⾸席执⾏官⼭姆·奥特曼的BG2播客要点 TMT BREAKOUTTMT突破NOV 01, 20252025年11⽉01⽇∙PAID∙付费 Linkhereto full pod… 链接在此查看完整播客… Revenue scale & the $100B timing 收⼊规模与1000亿美元的时机 Host:“If you were doing in excess of$100B of revenue in 2028 or 2029, you’d bepositioned to IPO around$1T…” 主持⼈：“如果你在2028年或2029年的收⼊超过1000亿美元，你将有望以约1万亿美元的估值进⾏IPO……” Sam Altman: “How about 2027?…We’d like to fund the company with revenue growth. Themost appealing part of an IPO is letting retail investors buy one of the mostimportant companies.” “那2027年呢？……我们希望⽤营收来为公司提供资⾦。IPO最吸引⼈的部分是让散户投资者可以买到这家最重要公司的股份。” Sam Altman (on trajectory): Sam Altman（关于发展轨迹）： “We’re doing well more revenue than that [reported number]. Revenue is growingsteeply. We’re taking a forward bet that it continues—ChatGPT keeps growing; webecome one of the important AI clouds; our consumer device becomes significant;AI that automates science creates huge value. We might screw it up, but without compute, we can’t generate the revenue or build the models at this scale. We don’thave a date; I assume going public someday. I’d prefer to fund growth with revenue.The most appealing thing is retail investors being able to own one of the mostimportant companies.” “我们表现良好，营收⽐那个[报告的数字]要⾼。营收在急速增⻓。我们在前瞻性下注，希望这种增⻓能持续⸺ChatGPT持续增⻓；我们成为重要的AI云之⼀；我们的消费设备变得重要；能⾃动化科学的AI会创造巨⼤的价值。我们可能会搞砸，但没有算⼒，我们⽆法在这个规模上产⽣营收或训练模型。我们没有具体⽇期；我假设有⼀天会上市。我更愿意⽤营收来资助增⻓。最吸引⼈的，是让散户投资者能够拥有这家最重要公司的股份。” Compute constraints & price elasticity 算⼒限制与价格弹性 Sam Altman:萨姆·奥尔特曼： “If the price of compute per unit of intelligence fell 100×tomorrow, usage would goup by much more than 100×. There’s a ton people want to do that makes noeconomic sense at current cost but would at lower cost. As model capability rises,willingness to pay for much higher intelligence also rises.” “如果明天每单位智能的计算成本下降100倍，使⽤量会增⻓远超100倍。现在有⼤量⼈们想做的事情在当前成本下没有经济意义，但在更低成本下就会有意义。随着模型能⼒的提升，为更⾼智能付费的意愿也会提⾼。” Power, shells & where the bottleneck is 能量、外壳与瓶颈所在 Satya Nadella:萨蒂亚·纳德拉： “The biggest issue isn’t chips—it’s power and getting builds done fast enough nearpower. I don’t have warm shells to plug into—that’s my problem today. Think interms of a fungible fleet: an efficient token factory and high utilization—multiple workloads scheduled across pre-training, mid-training, post-training, RL, andacross geographies. “最⼤的问题不是芯⽚⸺⽽是电⼒，以及如何在靠近电⼒的地⽅⾜够快地完成构建。我现在没有可以直接接⼊的热机箱⸺这是我今天⾯临的问题。要以可替代的机群来思考：⼀个⾼效的代币⼯⼚和⾼利⽤率⸺在预训练、中期训练、后训练、强化学习以及跨地域之间调度多个⼯作负载。 “We sometimes say no to demand—even OpenAI’s—if a dedicated training build inone location isn’t optimal for Azure’s long-term fleet. That’s why we gave OpenAIflexibility to procure elsewhere while we maintain a significant book of business andpreserve capacity for our own high-margin Copilots. Azure could’ve grown fasterthis quarter with more compute—absolutely—but we’re optimizing long-termreturns.” “我们有时会对需求说不⸺即便是对OpenAI也是如此⸺如果在某⼀地点的专⽤训练构建对Azure的⻓期机群不利。这就是为什么我们允许OpenAI在别处采购灵活性，同时我们维持⼤量业务并为我们⾃身⾼利润的Copilots保留容量。Azure本季度本可以通过更多算⼒增⻓更快⸺毫⽆疑问⸺但我们是在优化⻓期回报。” Compute glut risk & on-device future算⼒过剩⻛险与设备端的未来 Sam Altman:萨姆·奥特曼： “There will be a glut—whether in 2–3 years or 5–6, I can’t tell you—but somepeople will get burned, like in every infrastructure cycle. If a very cheap formof energy comes online at scale, a lot of people will be burned by existingcontracts. If we keep reducing cost per unit of intelligence at an unbelievablerate, everyone running personal AGI locally could arrive faster than we think.” “会出现供过于求⸺是两到三年，还是五到六年，我说不准⸺但像每⼀个基础设施周期⼀样，会有⼈遭受损失。如果⼀种⾮常廉价的⼤规模能源出现，很多⼈会被现有合约烧掉。如果我们以难以置信的速度持续降低每单位智能的成本，⼈⼈在本地运⾏个⼈级AGI的到来可能⽐我们预想的更快。” “Someday we will make an incredible consumer device that can run a GPT-5/GPT-6-capable model completely locally at low power. That will be incredible. “有⼀天我们会制造出⼀款令⼈难以置信的消费级设备，能够在低功耗下完全本地运⾏具备GPT-5/GPT-6能⼒的模型。那将⾮常了不起。 Addressing critics & compute commitments回应批评者与计算资源承诺 Sam Altman:萨姆·奥特曼： “People talking with breathless concern about our compute spend would be thrilledto buy shares. If anyone wants to sell, I’ll find a buyer. Revenue is growing steeply;we plan for it to continue. A rare time I’d want to be public is when someone sayswe’re about to go out of business—I’d love for them to short it and then see howthat goes.” “那些对我们算⼒开⽀惊呼不已的⼈，若能买到股票会很⾼兴。若有⼈想卖，我会帮他找到买家。收⼊正快速增⻓；我们计划让这种势头持续下去。当有⼈说我们快要倒闭时，那正是我想公开上市的罕⻅时刻⸺我很愿意看到他们去做空，然后看看结果如何。” Satya Nadella:萨蒂亚·纳德拉： “There hasn’t been a single business plan from OpenAI that they put in and didn’tbeat. The business execution has been unbelievable.” “OpenAI没有⼀项商业计划是他们提交后没有取得胜利的。商业执⾏简直令⼈难以置信。” Azure growth, backlog & supply shaping Azure增⻓、未完成订单与供应调控 Satya Nadella:萨蒂亚·纳德拉： “Absolutely Azure could have grown ~41–42%if we’d had more compute.Our$400B RPO has ~2-year duration, which is why we’re spending capex with high certainty—we need to clear this backlog. It’s diversified across first-party and third-party demand. We’re shaping demand because we’re supply-constrained,balancing long-term returns across Azure and ou