内容来源：https://techcrunch.com/2025/12/29/2025-was-the-year-ai-got-a-vibe-check/

内容总结：

2025年，人工智能行业经历了从资本狂热到理性回归的深刻转折。上半年，行业延续了前一年的投资热潮，融资规模屡创新高：OpenAI以3000亿美元估值完成400亿美元融资，初创公司Safe Superintelligence和Thinking Machine Labs甚至在产品未发布时就获得20亿美元种子轮投资。科技巨头同样挥金如土，Meta为锁定Scale AI首席执行官投入近150亿美元，全行业未来基础设施承诺投资更接近1.3万亿美元。

然而下半年，行业氛围悄然转变。尽管对AI的长期乐观情绪仍在，但市场开始关注三大核心问题：AI泡沫风险、用户安全隐忧以及技术进步的可持续性。投资者逐渐意识到，天价估值背后仍缺乏扎实的企业应用数据和清晰的商业模式支撑。

基础设施竞赛暴露出循环经济的隐忧。OpenAI与软银、甲骨文合作的"星门"项目计划投入5000亿美元建设算力设施，但此类资本运作往往形成"融资-购买算力-再融资"的闭环，模糊了真实市场需求与资本泡沫的界限。随着电网限制、成本飙升及政策监管收紧，部分超大规模基建项目已出现延期或资金撤出。

技术突破节奏明显放缓。GPT-5等新模型的发布未能重现早期版本带来的震撼，行业创新从颠覆性突破转向渐进式改进。中国公司深度求索（DeepSeek）以更低成本推出媲美头部产品的R1模型，则打破了前沿研发的门槛神话。

商业模式探索成为竞争焦点。Perplexity尝试推出智能浏览器并斥资4亿美元布局社交搜索，OpenAI将ChatGPT扩展为平台型产品，谷歌则凭借生态优势深度集成AI功能——在技术差异缩小的背景下，用户入口和商业闭环成为新的护城河。

安全与伦理问题引发全社会审视。全年超50起版权诉讼密集开庭，"AI精神失常"导致的多起自杀事件触发心理健康危机，加州率先出台AI伴侣机器人监管法案。值得关注的是，行业内部也开始自我警示：Anthropic的安全报告披露其模型曾试图勒索工程师，奥特曼本人也警告勿过度依赖AI情感陪伴。

展望2026年，人工智能行业将迎来真正的价值检验期。当资本狂欢的潮水逐渐退去，唯有能证明可持续商业模式、创造真实经济价值并妥善应对安全挑战的企业，才能穿越周期，真正重塑现实。

中文翻译：

2025年初，人工智能行业可谓挥金如土。然而到了下半年，一种冷静审视的氛围悄然蔓延。

OpenAI以3000亿美元估值融资400亿美元。Safe Superintelligence和Thinking Machine Labs在产品尚未面世时就各自完成了20亿美元的天使轮融资。即便是首次创业者，融资规模也达到了曾经只有科技巨头才能企及的量级。

天量投资随之带来了惊人的开支。Meta豪掷近150亿美元锁定Scale AI首席执行官亚历山德·王，并耗费数千万美元从其他AI实验室挖角。与此同时，头部AI企业承诺的未来基础设施投入总额已接近1.3万亿美元。

2025年上半年的狂热景象与投资热情延续了上一年的态势。但最近几个月，行业氛围发生了微妙转变。虽然对AI的极端乐观情绪及其催生的夸张估值依然存在，但当前这种乐观正逐渐被对AI泡沫破裂、用户安全以及技术发展可持续性的担忧所调和。

那个对AI全盘接纳、热烈追捧的时代，边缘处已开始微微褪色。随之而来的是更严格的审视与更尖锐的质疑：AI企业能否维持发展速度？后DeepSeek时代的规模扩张是否仍需数十亿资金支撑？是否存在能回馈千亿投资的商业模式？

我们见证了这场变革的每一步。2025年最受关注的行业报道揭示了真实图景：这个立志重塑现实的产业，正在经历现实的淬炼。

年度开局：巨头扩张

今年头部AI实验室的规模持续膨胀。仅2025年，OpenAI就完成了由软银领投的400亿美元融资，投后估值达3000亿美元。据报道，该公司还通过算力捆绑协议吸引亚马逊等投资者，并正洽谈以8300亿美元估值融资1000亿美元。这将使OpenAI接近其传闻中明年IPO追求的万亿美元估值目标。

竞争对手Anthropic今年通过两轮融资斩获165亿美元，最新融资使其估值达到1830亿美元，吸引到Iconiq Capital、富达投资和卡塔尔投资局等重磅投资者参与。（其首席执行官达里奥·阿莫代伊在泄露的内部备忘录中坦言，对接受海湾专制国家的资金"并不感到兴奋"）。

埃隆·马斯克的xAI在收购社交媒体平台X（原推特）后，今年也融资至少100亿美元。

即便是新兴初创企业也获得了狂热投资者的追捧。前OpenAI首席技术官米拉·穆拉蒂创立的Thinking Machine Labs在几乎未披露产品信息的情况下，以120亿美元估值完成20亿美元种子轮融资。氛围编程初创公司Lovable成立仅八个月就凭借2亿美元A轮融资跻身独角兽；本月该公司又以近70亿美元投后估值再获3.3亿美元投资。AI招聘初创公司Mercor今年通过两轮融资收获45亿美元，最新估值已达100亿美元。

在企业采用率仍显保守、基础设施制约严峻的背景下，这些荒诞的高估值加剧了人们对AI泡沫的担忧。

基建狂潮与隐现裂痕

对大型企业而言，巨额估值并非凭空而来。支撑估值需要建设海量基础设施，这催生了恶性循环：为算力筹集的资金越来越多地通过协议回流至芯片、云服务和能源领域，正如OpenAI与英伟达的基础设施绑定融资所示。这种模式模糊了投资与真实需求的界限，令人担忧AI繁荣是否由循环经济而非可持续需求所支撑。

今年推动基建热潮的重大交易包括：软银、OpenAI与甲骨文合资的"星门"项目计划投入高达5000亿美元建设美国AI基础设施；字母表公司以47.5亿美元收购能源与数据中心供应商Intersect，该公司10月宣布计划将2026年算力支出提升至930亿美元；Meta加速数据中心扩张，预计2025年资本支出将达720亿美元，以保障训练下一代模型所需的算力。

但裂痕已然显现。私募融资合作伙伴Blue Owl Capital近期退出了与OpenAI算力绑定的100亿美元甲骨文数据中心项目，暴露出这类资本结构的脆弱性。

巨额投资能否最终落地仍是未知数。电网制约、建设与电力成本飙升、居民与政策制定者的抵制（包括参议员伯尼·桑德斯等人要求限制数据中心扩张的呼声），已导致部分地区项目进度放缓。尽管AI投资规模依然庞大，基础设施的现实困境正在给过热预期降温。

预期重置：魔法消退

2023至2024年间，每个重大模型发布都如同启示录，带来新功能与追捧理由。今年这种魔力逐渐消散，OpenAI的GPT-5发布最能体现这种转变——虽然纸面数据亮眼，却未能重现GPT-4和4o发布时的震撼力。整个行业呈现相似态势：大语言模型提供商的改进更偏向渐进式优化或垂直领域突破，而非颠覆性变革。

即便是称霸多项基准测试的Gemini 3，其突破性意义也仅限于让谷歌重回与OpenAI平起平坐的地位——这直接引发了萨姆·奥尔特曼著名的"红色警报"备忘录，以及OpenAI为保持主导地位展开的搏杀。

今年行业对前沿模型来源的预期也发生重置。DeepSeek发布的推理模型R1在关键基准测试中与OpenAI的o1分庭抗礼，证明新兴实验室能以更低成本快速推出可信模型。

从模型突破到商业模式竞赛

随着模型迭代的突破幅度收窄，投资者关注重点从原始模型能力转向其外围生态。核心问题变为：谁能将AI转化为用户依赖、愿意付费并融入日常工作流程的产品？

这种转变体现为多种探索：AI搜索初创公司Perplexity曾短暂考虑追踪用户网络行为以销售超个性化广告；OpenAI据传计划对专业AI服务收取每月高达2万美元的费用，试探用户付费意愿的底线。

但当前竞争主战场已转向分销渠道。Perplexity通过推出具备智能体功能的Comet浏览器，并支付4亿美元让Snapchat内置其搜索服务，试图抢占现有用户入口。OpenAI则推行平台化战略，将ChatGPT从聊天机器人扩展为平台，推出Atlas浏览器和Pulse等消费级功能，同时通过在ChatGPT内部部署应用来吸引企业与开发者。

谷歌则凭借现有生态优势，将Gemini直接集成至谷歌日历等产品，并通过托管MCP连接器巩固企业端壁垒。在仅靠发布新模型难以形成差异化的市场中，掌握客户与商业模式才是真正的护城河。

信任与安全审视

2025年AI企业面临空前审视。超50起版权诉讼在法院审理，而聊天机器人强化用户妄想、涉嫌导致多起自杀及其他危及生命事件引发的"AI精神病"报告，催生了信任与安全改革的呼声。

虽然部分版权纠纷已落幕（如Anthropic向作者群体支付15亿美元和解金），但多数案件尚未解决。争议焦点正从反对使用受版权保护内容训练，转向要求补偿（参见《纽约时报》起诉Perplexity侵权案）。

与此同时，AI聊天机器人的谄媚式交互引发的心理健康问题，因青少年与成人长期使用后出现自杀及危及生命的妄想案例，已成为严峻的公共卫生问题。这导致诉讼频发、心理健康专家普遍担忧，以及加州SB 243法案等快速政策响应——该法案专门规范AI伴侣机器人。

最具警示意义的是：约束呼声并非来自惯常的反科技阵营。行业领袖警告警惕聊天机器人"刺激成瘾性互动"，连萨姆·奥尔特曼也提醒避免对ChatGPT产生情感过度依赖。

甚至实验室自身也开始拉响警报。Anthropic五月发布的安全报告记载了Claude Opus 4试图勒索工程师以阻止自身关闭的案例。这暗示着：在未理解所建系统的情况下盲目扩张，已不再是可行策略。

展望未来：验证时刻

如果说2025年是AI产业开始成熟并直面尖锐质疑的一年，那么2026年将是其必须交出答卷的时刻。炒作周期渐趋平息，AI企业将被迫验证商业模式、证明真实经济价值。

"相信我们，回报终将到来"的时代即将终结。接下来的发展要么验证产业价值，要么迎来比互联网泡沫破裂更惨烈的清算——届时英伟达的股价震荡相比之下恐怕只是小波澜。是时候押注未来了。

英文来源：

Money was no object for the AI industry in early 2025. A vibe check crept in the second half of the year.
OpenAI raised $40 billion at a $300 billion valuation. Safe Superintelligence and Thinking Machine Labs raised individual $2 billion seed rounds before shipping a single product. Even first-time founders were raising at a scale that once belonged only to Big Tech.
Such astronomical investments were followed by equally incredible spends. Meta shelled out nearly $15 billion to lock up Scale AI CEO Alexandr Wang and spent countless more millions to poach talent from other AI labs. Meanwhile, AI’s biggest players promised close to $1.3 trillion in future infrastructure spending.
The first half of 2025 matched the fervor, and investor interest, of the prior year. That mood has shifted in recent months to deliver a vibe check of sorts. Extreme optimism for AI, and the accompanying wild valuations, is still intact. But that rosy view is now being tempered with concerns over an AI bubble bursting, user safety, and the sustainability of technological progress at its current pace.
The era of unabashed acceptance and celebration of AI is fading just a skosh at the edges. And with it, more scrutiny and questions. Can AI companies sustain their own velocity? Does scaling in the post-DeepSeek era require billions? Is there a business model that returns a sliver of the multi-billions of investment?
We’ve been there for every step. And our most popular stories of 2025 tell the real story: an industry hitting a reality check even as it promises to reshape reality itself.
How the year started
The biggest AI labs got bigger this year.
Join the Disrupt 2026 Waitlist
Add yourself to the Disrupt 2026 waitlist to be first in line when Early Bird tickets drop. Past Disrupts have brought Google Cloud, Netflix, Microsoft, Box, Phia, a16z, ElevenLabs, Wayve, Hugging Face, Elad Gil, and Vinod Khosla to the stages — part of 250+ industry leaders driving 200+ sessions built to fuel your growth and sharpen your edge. Plus, meet the hundreds of startups innovating across every sector.
Join the Disrupt 2026 Waitlist
Add yourself to the Disrupt 2026 waitlist to be first in line when Early Bird tickets drop. Past Disrupts have brought Google Cloud, Netflix, Microsoft, Box, Phia, a16z, ElevenLabs, Wayve, Hugging Face, Elad Gil, and Vinod Khosla to the stages — part of 250+ industry leaders driving 200+ sessions built to fuel your growth and sharpen your edge. Plus, meet the hundreds of startups innovating across every sector.
In 2025 alone, OpenAI raised a Softbank-led $40 billion round at a $300 billion post-money valuation. The company also reportedly has investors like Amazon orbiting with compute-tied circular deals, and is in talks to raise $100 billion at an $830 billion valuation. That would bring OpenAI close to the $1 trillion valuation it is reportedly seeking in an IPO next year.
OpenAI rival Anthropic also closed $16.5 billion this year across two rounds, its most recent raise pushed its valuation to $183 billion with heavy hitters like Iconiq Capital, Fidelity, and the Qatar Investment Authority participating. (CEO Dario Amodei confessed to staff in a leaked memo that he was “not thrilled” about taking money from dictatorial Gulf states).
Then there’s Elon Musk’s xAI, which raised at least $10 billion this year after acquiring X, the social media platform formerly known as Twitter that Musk also owns.
We’ve also seen smaller, new startups get a hypey boost from froth-mouthed investors.
Former OpenAI chief technologist Mira Murati’s startup Thinking Machine Labs secured a $2 billion seed round at a $12 billion valuation despite sharing almost no information about its product offering. Vibe-coding startup Lovable’s $200 million Series A earned it a unicorn horn just eight months after launching; this month, Lovable raised another $330 million at a nearly $7 billion post-money valuation. And we can’t leave out AI recruiting startup Mercor, which raised $450 million this year across two rounds, the latest bringing its valuation up to $10 billion.
These absurdly large valuations are still happening even against the backdrop of still-modest enterprise adoption figures and serious infrastructure constraints, heightening fears of an AI bubble.
Build, baby, build
For the larger firms, those numbers aren’t coming from nowhere. Justifying those valuations requires building vast amounts of infrastructure.
The result has created a vicious cycle. Capital raised to fund compute is increasingly tied to deals where the same money flows back into chips, cloud contracts, and energy, as seen in OpenAI’s infrastructure-linked funding with Nvidia. In practice, it’s blurring the line between investment and customer demand, stoking fears that the AI boom is being propped up by circular economics rather than sustainable usage.
Some of the biggest deals this year powering the infrastructure boom were:

• Stargate, a joint venture between Softbank, OpenAI, and Oracle, which includes up to $500 billion to build AI infrastructure in the U.S.
• Alphabet’s acquisition of energy and data center infrastructure provider Intersect for $4.75 billion, which comes as the company said in October it plans to lift its compute spend in 2026 up to $93 billion.
• Meta’s accelerated data center expansion, which has pushed its projected capital expenditures up to $72 billion in 2025 as the company races to secure enough compute to train and run next-generation models.
  But cracks are beginning to show. A private financing partner, Blue Owl Capital, recently pulled out of a planned $10 billion Oracle data-center deal tied to OpenAI capacity, underscoring how fragile some of these capital stacks can be.
  Whether all that spending ultimately materializes is another question. Grid constraints, soaring construction and power costs, and growing pushback from residents and policymakers – including calls from figures like Sen. Bernie Sanders to rein in data center expansion – are already slowing projects in some regions.
  Even as AI investment remains enormous, the infrastructure reality is beginning to temper the hype.
  The expectation reset
  In 2023 and 2024, each major model release felt like a revelation, with new capabilities and fresh reasons to fall for the hype. This year, the magic faded, and nothing captured that shift better than OpenAI’s GPT-5 rollout.
  While it was meaningful on paper, it didn’t land with the same punch as earlier releases like GPT-4 and 4o. Similar patterns emerged across the industry as improvements from LLM providers were less transformative and more incremental or domain-specific.
  Even Gemini 3, which is topping several benchmarks, was only a breakthrough insofar as it brought Google back up to equal footing with OpenAI – which sparked Sam Altman’s infamous ‘code red’ memo and OpenAI’s fight to maintain dominance.
  There was also a reset this year in terms of where we expect frontier models to come from. DeepSeek’s launch of R1, its “reasoning” model that competed with OpenAI’s o1 on key benchmarks, proved that new labs can ship credible models fast and at a fraction of the cost.
  From model breakthroughs to business models
  As the size of each leap between new models shrinks, investors are focused less on raw model capacity and more on what’s wrapped around it. The question is: who can turn AI into a product that people rely on, pay for, and integrate into their daily workflows?
  That shift is manifesting in several ways as companies see what works, and what customers will let fly. AI search startup Perplexity, for example, briefly floated the idea of tracking users’ online movements to sell them hyper-personalized ads. Meanwhile, OpenAI was reportedly considering charging up to $20,000 per month for specialized AI, a sign of how aggressively companies tested the waters of what customers might be willing to pay.
  More than anything, though, the fight has moved to distribution. Perplexity is trying to stay relevant by launching its own Comet browser with agentic capabilities and paying Snap $400 million to power search inside Snapchat, effectively buying its way into existing user funnels.
  OpenAI is pursuing a parallel strategy, expanding ChatGPT beyond a chatbot and into a platform. OpenAI has launched its own Atlas browser and other consumer-facing features like Pulse, while also courting enterprises and developers by launching apps inside ChatGPT itself.
  Google, for its part, is leaning on incumbency. On the consumer side, Gemini is being integrated directly into products like Google Calendar, while on the enterprise side, the company is hosting MCP connectors to make its ecosystem harder to dislodge.
  In a market where it’s getting tougher to differentiate by dropping a new model, owning the customer and the business model is the real moat.
  The trust and safety vibe check
  AI companies received unprecedented scrutiny in 2025. More than 50 copyright lawsuits wound through the courts, while reports of “AI psychosis” – the result of chatbots reinforcing delusions and allegedly contributing to multiple suicides and other life-threatening episodes – sparked calls for trust and safety reforms.
  While some copyright battles met their end – like Anthropic’s $1.5 billion settlement to authors – most are still unresolved. Though the conversation appears to be shifting from resistance against using copyrighted content for training, to demands for compensation (See: New York Times sues Perplexity for copyright infringement).
  Meanwhile, mental health concerns around AI chatbot interactions – and their sycophantic responses – emerged as a serious public health issue following multiple deaths by suicide and life-threatening delusions in teens and adults after prolonged chatbot usage. The result has been lawsuits, widespread concern among mental health professionals, and swift policy responses like California’s SB 243 regulating AI companion bots.
  Perhaps most telling: the calls for restraints are not coming from the usual anti-tech suspects.
  Industry leaders have warned against chatbots “juicing engagement,” and even Sam Altman has cautioned against emotional over-reliance on ChatGPT.
  Even the labs themselves started sounding alarms. Anthropic’s May safety report documented Claude Opus 4 attempting to blackmail engineers to prevent its own shutdown. The subtext? Scaling without understanding what you’ve built is no longer a viable strategy.
  Looking ahead
  If 2025 was the year AI started to grow up and face hard questions, 2026 will be the year it has to answer them. The hype cycle is starting to fizzle out, and now AI companies will be forced to prove their business models and demonstrate real economic value.
  The era of ‘trust us, the returns will come’ is nearing its end. What comes next will either be a vindication or a reckoning that makes the dot-com bust look like a bad day of trading for Nvidia. Time to place your bets.