内容来源：https://www.wired.com/story/ai-could-democratize-one-of-techs-most-valuable-resources/

内容总结：

【AI芯片霸主英伟达面临新挑战：AI技术正削弱其软件护城河】

在人工智能芯片领域，英伟达长期占据主导地位，其芯片设计与配套软件生态已成为全球AI基础设施的核心支柱，推动公司市值突破4万亿美元。然而，随着AI技术本身的飞速发展，这一格局可能正在悄然改变。

当前，多家科技巨头已投入自研芯片赛道：苹果通过定制芯片提升终端设备性能，谷歌、亚马逊、Meta等企业则为其云计算平台开发专用处理器。但即便硬件性能接近，企业仍需克服编程优化难题——让软件在特定芯片上高效运行需要高昂的工程成本，而英伟达凭借成熟的软件生态形成了显著优势。

如今，初创公司正利用AI技术试图打破这一壁垒。例如，Wafer公司通过强化学习训练AI模型，使其能自动优化底层硬件代码，并与AMD、亚马逊等企业合作提升其芯片的软件适配效率。该公司创始人埃米利奥·安德雷指出，当前高端芯片在理论算力上已与英伟达产品持平，“真正的护城河在于芯片的可编程性”，而AI可能很快将削弱这一优势。

更深远的变化可能发生在芯片设计环节。由前谷歌工程师创立的Ricursive公司正开发AI驱动的芯片设计工具，旨在通过自然语言交互简化芯片布局与验证流程。该公司已获得34亿美元估值，其技术方向预示着未来AI可能同时参与芯片与算法的协同设计，形成“算力提升→芯片优化→算力再提升”的递归式创新循环。

行业观察指出，若AI能持续降低芯片设计与软件优化的门槛，更多企业将有望打造垂直整合的算力体系，从而动摇现有市场格局。英伟达的软件生态优势虽仍坚固，但AI对自身基础设施的反向赋能，或许正在打开新一轮竞争的大门。

中文翻译：

英伟达是AI芯片领域无可争议的王者。但得益于它助力构建的AI技术，这位冠军可能很快将面临日益激烈的竞争。

现代AI的运行依赖于英伟达的设计架构，这一优势推动其市值突破4万亿美元大关。英伟达每一代新芯片都能让企业通过连接数百乃至数千个处理器，在庞大的数据中心内训练更强大的AI模型。其成功秘诀之一在于提供配套软件来协助编写新一代芯片程序，但这项优势可能即将不再独特。

初创公司Wafer正在训练AI模型执行一项极其困难却至关重要的任务——优化代码，使其在特定芯片上实现最高效运行。该公司联合创始人兼CEO埃米利奥·安德烈表示，他们通过对开源模型进行强化学习，训练其编写内核代码（即直接与操作系统硬件交互的软件）。Wafer还为Anthropic的Claude和OpenAI的GPT等现有编码模型添加"智能控制框架"，增强其编写直接运行于芯片的代码能力。

如今众多科技巨头已拥有自研芯片。苹果等公司多年来通过定制芯片提升笔记本电脑、平板和手机的软件性能与能效；而谷歌、亚马逊等企业则通过自研芯片优化云计算平台性能。Meta近期宣布将与博通合作部署搭载新型芯片的十亿瓦级算力设施。部署定制芯片还需编写大量适配代码以确保在新处理器上流畅高效运行。

Wafer正与AMD、亚马逊等企业合作，帮助优化其硬件上的软件运行效率。这家初创公司已从谷歌的杰夫·迪恩、OpenAI的沃伊切赫·扎伦巴等人处获得400万美元种子轮融资。

安德烈认为其公司的AI驱动方案有望挑战英伟达的统治地位。目前多款高端芯片在原始浮点性能（衡量芯片基础计算能力的关键行业指标）上已可比肩英伟达顶级产品。"顶尖的AMD硬件、亚马逊Trainium芯片和谷歌TPU都能提供与英伟达GPU相当的理论浮点运算能力，"安德烈表示，"我们的目标是实现每瓦特算力的智能化最大化。"

他指出，具备芯片代码优化能力的性能工程师身价高昂且供不应求，而英伟达的软件生态使其芯片编程维护更为便捷，这令即使大型科技公司也难以独立突破。例如Anthropic与亚马逊合作在Trainium芯片上构建AI模型时，就不得不从头重写模型代码以实现硬件效率最大化。

当然，如今包括Claude在内的众多AI模型已具备超人类的代码编写能力。安德烈推测，AI吞噬英伟达软件优势的时代可能为期不远。"真正的护城河在于芯片的可编程性，"他针对英伟达的代码优化软件库表示，"现在或许是时候重新思考这道护城河是否依然坚固了。"

除了简化不同芯片的代码优化，AI可能很快将降低芯片设计门槛。由前谷歌工程师阿扎莉娅·米尔霍塞尼与安娜·戈尔迪创立的初创公司Ricursive Intelligence，正在开发用AI设计计算机芯片的新方法。若其技术取得突破，更多企业或将涉足芯片设计领域，开发能更高效运行自身软件的定制芯片。

"我们正攻克芯片设计中最耗时的环节——物理设计与设计验证，"兼任斯坦福大学助理教授的米尔霍塞尼指出。芯片设计堪称全球最复杂精密的工作之一，工程师需要在硅片上布局海量元件以优化各项功能，设计完成后还需经过反复测试验证才能交付晶圆厂生产。

米尔霍塞尼与戈尔迪在谷歌期间曾开发出用AI优化芯片关键组件布局的技术，该方法不仅革新了谷歌的自研处理器设计流程，如今更在业界广泛应用于各类芯片的特征排列。Ricursive志在更进一步，通过自动化更多芯片设计环节并整合大语言模型，最终让工程师能用自然语言描述芯片修改需求或提出疑问——正如现在用自然语言生成应用程序，未来或许也能"自然语言设计芯片"。

虽然技术仍在开发中，但米尔霍塞尼透露公司已证明能优化芯片设计的更多维度。这种自动化设计前景令投资者兴奋不已：Ricursive在短短数月内以40亿美元估值融资3.35亿美元。

戈尔迪认为，未来AI可能实现芯片与算法的协同设计以提升整体效能。她指出AI通过自我调整芯片与代码，或将形成递归式改进循环："我们正进入一个新时代——通过投入更多算力来设计更快更好的芯片，这将在芯片设计领域创造新的规模效应法则。"

您如何看待AI设计自研芯片的前景？欢迎通过[email protected]邮箱分享见解。

本文节选自威尔·奈特《AI实验室》通讯专栏，往期内容可通过此处查阅。

英文来源：

Nvidia is the undisputed king of AI chips. But thanks to the AI it helped build, the champ could soon face growing competition.
Modern AI runs on Nvidia designs, a dynamic that has propelled the company to a market cap of well over $4 trillion. Each new generation of Nvidia chip allows companies to train more powerful AI models using hundreds or thousands of processors networked together inside vast data centers. One reason for Nvidia’s success is that it provides software to help program each new generation of chip. That may soon not be such a differentiated skill.
A startup called Wafer is training AI models to do one of the most difficult and important jobs in AI—optimizing code so that it runs as efficiently as possible on a particular silicon chip.
Emilio Andere, cofounder and CEO of Wafer, says the company performs reinforcement learning on open source models to teach them to write kernel code, or software that interacts directly with hardware in an operating system. Andere says Wafer also adds “agentic harnesses” to existing coding models like Anthropic’s Claude and OpenAI’s GPT to soup up their ability to write code that runs directly on chips.
Many prominent tech companies now have their own chips. Apple and others have for years used custom silicon to improve the performance and the efficiency of software running on laptops, tablets, and smartphones. At the other end of the scale, companies like Google and Amazon mint their own silicon to improve the performance of their cloud-computing platforms. Meta recently said it would deploy 1 gigawatt of compute capacity with a new chip developed with Broadcom. Deploying custom silicon also involves writing a lot of code so that it runs smoothly and efficiently on the new processor.
Wafer is working with companies including AMD and Amazon to help optimize software to run efficiently on their hardware. The startup has so far raised $4 million in seed funding from Google’s Jeff Dean, Wojciech Zaremba of OpenAI, and others.
Andere believes that his company’s AI-led approach has the potential to challenge Nvidia’s dominance. A number of high-end chips now offer similar raw floating point performance—a key industry benchmark of a chip’s ability to perform simple calculations—to Nvidia’s best silicon.
“The best AMD hardware, the best [Amazon] Trainium hardware, the best [Google] TPUs, give you the same theoretical flops to Nvidia GPUs,” Andere told me recently. “We want to maximize intelligence per watt.”
Performance engineers with the skill needed to optimize code to run reliably and efficiently on these chips are expensive and in high demand, Andere says, while Nvidia’s software ecosystem makes it easier to write and maintain code for its chips. That makes it hard for even the biggest tech companies to go it alone.
When Anthropic partnered with Amazon to build its AI models on Trainium, for instance, it had to rewrite its model’s code from scratch to make it run as efficiently as possible on the hardware, Andere says.
Of course, Anthropic’s Claude is now one of many AI models that are now superhuman at writing code. So Andere reckons it may not be long before AI starts consuming Nvidia software advantage.
“The moat lives in the programmability of the chip,” Andere says in reference to the libraries and software tools that make it easier to optimize code for Nvidia hardware. “I think it's time to start rethinking whether that's actually a strong moat.”
Besides making it easier to optimize code for different silicon, AI may soon make it easier to design chips themselves. Ricursive Intelligence, a startup founded by two ex-Google engineers, Azalia Mirhoseini and Anna Goldie, is developing new ways to design computer chips with artificial intelligence. If its technology takes off, a lot more companies could branch into chip design, creating custom silicon that runs their software more efficiently.
“We are going after the long poles of chip design—physical design and design verification,” says Mirhoseini, who is also an assistant professor at Stanford University, in reference to two of the main challenges involved with chip design.
Designing computer chips is one of the most consequential—and tricky—jobs on the planet. Chip engineers need to figure out how to arrange a vast number of components across a piece of silicon to optimize different functionality. After a chip is first designed, its performance has to be carefully tested and verified in an iterative process before the designs can be sent off to a foundry.
Nvidia’s designs are crucial for modern AI, with each new generation of chip allowing companies to train more powerful AI models using hundreds or thousands of processors networked together inside vast data centers.
Mirhoseini and Goldie developed a way for AI to optimize the layout of key components of computer chips while at Google. The approach transformed how Google designs its own processors, and it is now widely used in the industry to help arrange features on different chips.
Ricursive aims to go further, however, by automating more elements of chip design and integrating large language models into the process. The goal is to enable engineers to use natural language to describe changes or ask questions about a chip. Just as one can vibe code an app, perhaps eventually it will be possible to vibe design a chip.
Ricursive is still developing its technology, but Mirhoseini says the company has already shown that it can optimize more aspects of chip design.
The prospect of automating chip design in this manner has some investors salivating: Ricursive has raised $335 million at a $4 billion valuation in just a few months.
Goldie says it may ultimately be possible to have AI codesign both chips and algorithms to make them more powerful. She says that having AI tweak its own silicon and code could form a recursive kind of AI improvement. “We are moving into this new regime where we can just spend more compute to design faster and better chips—creating a kind of scaling law for chip design.”
What do you think of AI designing its own silicon? Leave a comment or send an email to [email protected] to let me know.
This is an edition of Will Knight’s AI Lab newsletter. Read previous newsletters here.