内容来源：https://www.technologyreview.com/2026/04/08/1135398/mustafa-suleyman-ai-future/

内容总结：

微软AI首席执行官穆斯塔法·苏莱曼近日指出，人工智能发展在可预见的未来不会遭遇瓶颈，其核心驱动力在于计算能力的“爆炸式增长”。这一趋势正以超越传统认知的指数级速度推进，彻底改变了基于线性思维的传统预测模式。

苏莱曼以自身经历说明：自2010年至今，前沿AI模型的训练数据量已增长约1万亿倍。这一飞跃主要得益于三大技术汇流：

1. 硬件性能飞跃：以英伟达芯片为例，其原始性能在六年内提升超过七倍。微软自研的Maia 200芯片在性价比上亦领先同类产品30%。
2. 存储传输革命：高带宽内存（HBM）技术通过立体堆叠，使最新一代HBM3的带宽达到前代的三倍，确保数据高速输送至处理器。
3. 集群规模扩张：借助NVLink和InfiniBand等技术，数十万GPU可连接成仓库规模的超级计算机，作为一个整体认知实体运行。

这些进步共同导致计算效能急剧提升。例如，训练同一语言模型所需时间从2020年的167分钟缩短至如今的不足4分钟，提升幅度达50倍，远高于摩尔定律预测的5倍。软件算法的革新同样迅猛，达到固定性能水平所需的计算量每八个月即可减半。

未来增长同样惊人。领先实验室的年算力增长接近4倍，全球AI算力预计到2027年将增长十倍。综合来看，到2028年底有效算力可能再提升1000倍。到2030年，每年新增的算力能耗可能相当于英法德意四国的峰值用电总和。

苏莱曼认为，这将推动AI从聊天机器人向近乎人类水平的智能体过渡——这些半自主系统能够连续数天编写代码、执行长达数周或数月的项目、进行电话沟通、谈判合同和管理物流。其影响将远超科技行业，所有依赖认知劳动的领域都将被重塑。

面对算力扩张带来的能源挑战，苏莱曼指出清洁能源的发展提供了解决路径：过去50年太阳能成本下降近100倍，三十年电池价格降低97%，为可持续的算力扩展展现了清晰前景。

中文翻译：

穆斯塔法·苏莱曼：人工智能发展短期内不会撞墙，原因如下
算力爆发是我们这个时代的技术主旋律，而这一切才刚刚拉开序幕。

人类进化于线性世界。步行一小时能走完特定距离，两小时则翻倍。这种直觉在草原时代行之有效，但面对人工智能及其核心的指数级增长趋势时，却会引发灾难性误判。

从2010年我开始从事人工智能研究至今，尖端AI模型的训练数据量已增长约1万亿倍——早期系统约需10¹⁴次浮点运算（计算核心单位），而如今最大模型已超过10²⁶次浮点运算。这是真正的爆发式增长，AI领域的一切进展皆源于此。

三大技术突破正共同推动这一进程：
首先，基础算力持续加速。英伟达芯片原始性能在六年内提升超七倍，从2020年的312万亿次/秒增至如今的2250万亿次/秒。我们今年初发布的Maia 200芯片，每美元性能比现有硬件高出30%。
其次，高带宽内存技术让数据输送更快。这项技术像建造微型摩天大楼般垂直堆叠芯片；最新一代HBM3的带宽达到前代的三倍，能以实时喂饱处理器的速度传输数据。
第三，计算规模从“房间”扩展到“城市”。借助NVLink和无限带宽技术，数十万GPU可连接成仓库规模级的超级计算机，形成统一认知实体——这在数年前还无法想象。

这些进步共同催生了算力的跨越式增长。2020年用8块GPU训练语言模型需167分钟，如今同等现代硬件仅需不到4分钟。作为对比：摩尔定律预测同期性能仅提升约5倍，而我们实现了50倍增长。从2012年开启深度学习浪潮的图像识别模型AlexNet仅用2块GPU，到如今最大集群配备超10万块GPU（每块性能远超早期版本），软件革命同样惊人。Epoch AI研究显示，达到固定性能水平所需的算力每8个月减半，远快于摩尔定律传统的18-24个月翻倍周期。部分最新模型的年化部署成本骤降900倍，AI正变得极其廉价。

未来数据同样震撼：领先实验室的算力年增速近4倍；自2020年来，训练尖端模型的算力每年增长5倍；全球AI相关算力预计2027年达1亿H100等效算力，三年增长十倍。综合来看，到2028年底有效算力或将再增1000倍。到2030年，每年新增算力可能需200吉瓦电力——相当于英法德意四国峰值用电量总和。

这一切将带来什么？我认为将推动从聊天机器人到近人类水平智能体的转变：半自主系统能持续数天编写代码、执行长达数周甚至数月的项目、拨打电话、谈判合同、管理物流。这不再是回答问题的初级助手，而是能思考、协作、执行的AI工作团队。当前我们仅处于这场变革的起点，其影响将远超科技领域，所有依赖认知劳动的行业都将被重塑。

当然，能源是显性约束。一台冰箱大小的AI机柜耗电120千瓦，相当于100户家庭用电。但算力需求激增恰逢另一场指数级变革：太阳能成本50年下降近100倍，电池价格30年降低97%。清洁能源规模化之路已清晰可见。

——穆斯塔法·苏莱曼，微软人工智能首席执行官

深度聚焦
人工智能
“卸载ChatGPT”运动呼吁用户取消订阅
对ICE的抵制正演变为更广泛的反AI公司与特朗普政府结盟运动。

OpenAI全力打造全自动研究助手
专访OpenAI首席科学家雅库布·帕乔斯基，探讨公司新宏伟目标与AI未来。

《宝可梦GO》如何为配送机器人提供厘米级世界视图
独家报道：Niantic的AI子公司正利用玩家众包的300亿张城市地标图像训练全新世界模型。

这家初创公司欲变革数学家工作方式
Axiom Math正在免费提供强大AI工具，但其能否如公司所愿加速科研仍有待观察。

保持联系
获取《麻省理工科技评论》最新动态
探索专属优惠、头条新闻、即将举办的活动等精彩内容。

英文来源：

Mustafa Suleyman: AI development won’t hit a wall anytime soon—here’s why
The compute explosion is the technological story of our time. And it is still only just beginning.
We evolved for a linear world. If you walk for an hour, you cover a certain distance. Walk for two hours and you cover double that distance. This intuition served us well on the savannah. But it catastrophically fails when confronting AI and the core exponential trends at its heart.
From the time I began work on AI in 2010 to now, the amount of training data that goes into frontier AI models has grown by a staggering 1 trillion times—from roughly 10¹⁴ flops (floating-point operations‚ the core unit of computation) for early systems to over 10²⁶ flops for today’s largest models. This is an explosion. Everything else in AI follows from this fact.Three advances are now converging to enable this. First, the basic calculators got faster. Nvidia’s chips have delivered an over sevenfold increase in raw performance in just six years, from 312 teraflops in 2020 to 2,250 teraflops today. Our own Maia 200 chip, launched this January, delivers 30% better performance per dollar than any other hardware in our fleet. Second, the numbers arrive faster thanks to a technology called HBM, or high bandwidth memory, which stacks chips vertically like tiny skyscrapers; the latest generation, HBM3, triples the bandwidth of its predecessor, feeding data to processors fast enough to keep them busy all the time. Third, the room of people with calculators became an office and then a whole campus or city. Technologies like NVLink and InfiniBand connect hundreds of thousands of GPUs into warehouse-size supercomputers that function as single cognitive entities. A few years ago this was impossible.
These gains all come together to deliver dramatically more compute. Where training a language model took 167 minutes on eight GPUs in 2020, it now takes under four minutes on equivalent modern hardware. To put this in perspective: Moore’s Law would predict only about a 5x improvement over this period. We saw 50x. We’ve gone from two GPUs training AlexNet, the image recognition model that kicked off the modern boom in deep learning in 2012, to over 100,000 GPUs in today’s largest clusters, each one individually far more powerful than its predecessors. Then there’s the revolution in software. Research from Epoch AI suggests that the compute required to reach a fixed performance level halves approximately every eight months, much faster than the traditional 18-to-24-month doubling of Moore’s Law. The costs of serving some recent models have collapsed by a factor of up to 900 on an annualized basis. AI is becoming radically cheaper to deploy. The numbers for the near future are just as staggering. Consider that leading labs are growing capacity at nearly 4x annually. Since 2020, the compute used to train frontier models has grown 5x every year. Global AI-relevant compute is forecast to hit 100 million H100-equivalents by 2027, a tenfold increase in three years. Put all this together and we’re looking at something like another 1,000x in effective compute by the end of 2028. It’s plausible that by 2030 we’ll bring an additional 200 gigawatts of compute online every year—akin to the peak energy use of the UK, France, Germany, and Italy put together. What does all this get us? I believe it will drive the transition from chatbots to nearly human-level agents—semiautonomous systems capable of writing code for days, carrying out weeks- and months-long projects, making calls, negotiating contracts, managing logistics. Forget basic assistants that answer questions. Think teams of AI workers that deliberate, collaborate, and execute. Right now we’re only in the foothills of this transition, and the implications stretch far beyond tech. Every industry built on cognitive work will be transformed. The obvious constraint here is energy. A single refrigerator-size AI rack consumes 120 kilowatts, equivalent to 100 homes. But this hunger collides with another exponential: Solar costs have fallen by a factor of nearly 100 over 50 years; battery prices have dropped 97% over three decades. There is a pathway to clean scaling coming into view.Mustafa Suleyman is CEO of Microsoft AI.
Deep Dive
Artificial intelligence
A “QuitGPT” campaign is urging people to cancel their ChatGPT subscriptions
Backlash against ICE is fueling a broader movement against AI companies’ ties to President Trump.
OpenAI is throwing everything into building a fully automated researcher
An exclusive conversation with OpenAI’s chief scientist, Jakub Pachocki, about his firm's new grand challenge and the future of AI.
How Pokémon Go is giving delivery robots an inch-perfect view of the world
Exclusive: Niantic's AI spinout is training a new world model using 30 billion images of urban landmarks crowdsourced from players.
This startup wants to change how mathematicians do math
Axiom Math is giving away a powerful new AI tool. But it remains to be seen if it speeds up research as much as the company hopes.
Stay connected
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.