英伟达发家史:从游戏芯片到AI军火商,老黄的狂飙之路

Odaily星球日报Published on 2024-06-06Last updated on 2024-06-06

Abstract

「我渴望活下去的意志,超过几乎所有人想要杀死我的意志。」

英伟达发家史:从游戏芯片到AI军火商,老黄的狂飙之路

6 月 6 日,英伟达股价上涨 5.2% ,市值超过 3 万亿美元,超越苹果,成为全球第二大市值的公司。

从 1999 年上市时的 0.41 美元到如今 1224.40 美元, 24 年间英伟达创造了近 3000 倍回报。

英伟达,最让人羡慕的地方莫过于“不受周期限制”,一直作为底层基建,持续“收税”,无论你做什么,都离不开它。

作为 GPU 的缔造者,英伟达抓住了“PC 浪潮”的机遇,伴随着游戏市场的爆发,走进了千家万户;

接着,游戏业务疲软之际,加密牛市到来,英伟达显卡被广泛用于以太坊等加密货币“挖矿”,闷声发财;

随后,智能汽车产业崛起,它的车载芯片业务也迅速发展;最后,ChatGPT 横空出世,英伟达摇身一变,成为 AI 军火商……

但回顾英伟达的成长史,其也曾一次次在失败破产边缘徘徊,老黄(黄仁勋)曾如此呐喊:我渴望活下去的意志,超过几乎所有人想要杀死我的意志。

英伟达,GPU 缔造者

显卡(GPU)的诞生要从 20 世纪 90 年代说起。

当时,硅谷的一些人提出了一个想法:可以通过专门处理声音的声卡和处理网络的网卡等功能特定的芯片,来减轻中央处理器(CPU)的工作负担。同样的道理,制造一块专门负责电脑图像输出的芯片,也就是显卡(Graphic Card),也是顺理成章的事情。例如,索尼在 1994 年底推出的游戏机 PlayStation,就使用了显卡来处理图像。

然而,当时显卡的技术路径有很多种选择。英伟达找到的突破点是通过并行计算来实现 3D 图形加速,特别是在游戏领域中进行应用。所谓并行计算,即将一个复杂的任务拆分成多个小任务,然后同时处理它们,提高计算效率。

1999 年,英伟达推出了一款名为 GeForce 的显卡。这款显卡专为游戏设计,主打“并行计算”,能够显著提升3D图形处理能力,从而提供更流畅、更逼真的游戏体验。

GeForce 的成功,使英伟达迅速崛起,成为显卡领域的领导者。

当时,并不仅仅只有英伟达在研究图形处理单元,但英伟达成功地将自己与“GPU 发明者”这个标签深度绑定起来。

英伟达当时的市场营销负责人 Dan Vivoli 用“graphics processing unit”(GPU)的概念来推广自家的芯片,他认为,英伟达主要反复强调自己是 GPU 发明者就能成为行业领导者。

后来的确如此,英伟达成为了 GPU 的代名词,英伟达依靠营销 GPU ,为自己开辟了一条新路。

英伟达,加密牛市受益者

英伟达市值,从 2016 年的 140 亿美元涨到了 2018 年高点的 1750 亿美元,两年超过 10 倍的涨幅背后或许离不开加密货币挖矿热潮。

2017 年,加密货币迎来大牛市,吸引了大量矿工抢夺 GPU,GPU 变成了印钞机,全球显卡销量急剧增加,价格也水涨船高。

以矿工使用的英伟达 GTX 1060 型号显卡为例, 2017 年 5 月之前的拿货价约为 1650 元一张, 2017 年 6 月后就涨到了 2900 元左右。

英伟达成为了加密货币大牛市背后的大赢家,天降财富。

受益于加密挖矿潮,英伟达的 2018 财年全年收入创下了 97 亿美元的新高。黄仁勋对外表示,“我们的 GPU 支持着世界上最大规模的分布式超级计算,这就是它在加密货币领域大受欢迎的原因”。此外,英伟达还推出了专门面向挖矿定制的 GTX 1060 3 GB 以及 P 106、P 104 专业矿卡。

2020 年,经过此前 2 年熊市,加密市场再次起航,比特币涨了 2 倍多,以太坊涨了 4 倍,英伟达再次成为“加密繁荣”的受益者。

英伟达闻风而动,积极参与矿业市场,推出了 CMP 系列专业矿卡,这些卡去掉了图形处理功能,拥有更低的核心峰值电压和频率,以提高挖矿性能和效率。

英伟达发家史:从游戏芯片到AI军火商,老黄的狂飙之路

2020 年底,英伟达发布了 RTX 30 系显卡,其中入门级的 RTX 3060 显卡的标价是 2499 元,RTX 3090 显卡定价为 11999 元,而随着加密货币的涨势,RTX 3060 售价高达 5499 元,RTX 3090 更是一路飙至 20000 元。
2021 年一季度财报公开后,英伟达首席财务官 Colette Kress 曾透露,英伟达加密芯片销售额高达 1.55 亿美元,用于“挖矿”的显卡占据第一季度总体销量的四分之一。
就在 2021 年,英伟达全年收入创下 269.1 亿美元的纪录,较上一财年增长 61% ,市值一度突破 8000 亿美元。
然而,好景不长, 2022 年 9 月,以太坊执行层、权益证明共识层完成合并,以太坊区块链网络机制由 PoW(工作量证明机制)转变为 PoS(权益证明机制),显卡挖矿时代逐渐终结。
这一定程度上也影响了英伟达的发展, 2022 年三季度,英伟达营收、净利润双双出现下滑,季度收入仅为 59.31 亿美元,同比下降 17 %,净利润仅为 6.8 亿美元,同比下降高达 72 %。2022 年 11 月 23 日,英伟达股价报 165 美元/股,较去年最高点跌去近一半。
当时无论是《金融失败》等海外媒体,还是国内科技媒体,均看衰英伟达。

英伟达发家史:从游戏芯片到AI军火商,老黄的狂飙之路

困顿之极,不料峰回路转,AI 与大模型的风吹了起来,英伟达再次站上风口。

英伟达,AI 军火商

2016 年 3 月,AlphaGO 击败了李世石,震撼人心,也引发了关于 AI 的讨论热潮。
一个月后,黄仁勋在 GTC China 大会上正式宣布,英伟达不再是一家半导体公司,而是一家人工智能计算公司。
2016 年 8 月,一个历史性时刻诞生,英伟达向刚成立的 OpenAI 捐赠了英伟达的第一台 AI 超级计算机 DGX-1 ,黄仁勋本人亲自将这台计算机送到了 OpenAI 的办公室,当时的董事长埃隆·马斯克用开箱刀打开了包裹。
黄仁勋留下了一句话:“为了计算和人类的未来,我捐出世界上第一台 DGX-1 。”

英伟达发家史:从游戏芯片到AI军火商,老黄的狂飙之路

英伟达发家史:从游戏芯片到AI军火商,老黄的狂飙之路

而后,OpenAI 通过英伟达的超级计算机训练出了风靡全球的 ChatGPT,英伟达后续更新的硬件产品 DGX H 100 遭到市场疯抢,供不应求。
罗马并非一日建成,英伟达在 AI 行业的统治地位始于更早时期的积累。
英伟达的前首席科学家戴维·柯克(David Kirk)早就梦想着将 GPU 的 3D 绘图渲染算力通用化,不仅局限于游戏领域。
在戴维·柯克和黄仁勋的领导下,英伟达于 2007 年推出了革命性的 GPU 统一计算平台 CUDA,释放庞大的算力资源。
但在当时,CUDA 完全没有打动投资者,反而因为打造领先于时代的“超级计算”系统投入巨大,英伟达的利润受到大幅削减,华尔街为此嘘声一片。
一档风靡硅谷的热门播客《Acquired》的主持人 Ben Gilbert 对此评价:“他们当时瞄准的不是一个大市场,而是一个学术和科学计算的晦涩角落,但他们为此花了数十亿美元”。
来自外界的声音并没有影响黄仁勋,十几年坚持在 CUDA 上进行投入,让英伟达有了如今的地位。
黄仁勋将算力视为核心。无论是 AI、自动驾驶、元宇宙、机器人还是加密货币,英伟达都是利用庞大的算力寻找新的机会。
算力,英伟达永恒的武器。

三次失败

2023 年,黄仁勋在台湾大学毕业典礼上致辞,他分享了三个失败的故事,向大学生们传授了英伟达成功的秘诀。

第一次失败,在破产边缘活了下来。

1994 年,Nvidia 的第一个客户是日本游戏公司 SEGA,他们为其游戏主机设计显卡。
但是在第二年,微软发布了 Windows 平台的图形接口 Direct 3D,这一变动让 Nvidia 感到非常慌乱,因为与他们的设计存在冲突。
最终,Nvidia 选择中止与 SEGA 的合约,转而为 Windows 平台开发 GPU。这是一个冒险的决定,因为 SEGA 是他们唯一的客户,却被他们放弃了。Nvidia 的资金只能支持 6 个月,如果在这段时间内无法推出新产品,他们将面临倒闭的风险。
幸运的是,在资金即将耗尽时,距离破产仅一个月的时候,Nvidia 设计出了 Riva 128 这款芯片,并取得了成功。到 1997 年底,Riva 128 的出货量超过了 100 万张,Nvidia 因此得以生存下来。

第二次失败,放弃短期利润,成就了未来的伟大。

2007 年,英伟达发布了 CUDA GPU 加速计算计划,愿景是让 CUD A 成为一种编程模型,可以提升从科学计算和物理模拟到图像处理的各种应用。
创造一种新的计算模型非常困难,自 IBM System 360 推出以来,CPU 计算模型作为行业标准已经存在了长达 60 年时间。
CUDA 需要开发者重新编写应用程序,来展示 GPU 的好处;但要开发这样的程序,先要有一个庞大的用户群和一个庞大的需求来推动开发者去开发。
为了解决“先有鸡还是先有蛋”的问题,英伟达利用了他们已经有大量游戏玩家的 GForce 游戏显卡来建立用户群。但 CUDA 的附加成本非常高,导致英伟达的利润在多年间大幅下降,他们的市值一直在 10 亿美元的水平上下波动。
英伟达多年的低迷表现,也让股东对 CUDA 持怀疑态度。股东们更希望公司专注于提高盈利能力,但英伟达坚持了下来,相信加速计算的时机会到来。
黄仁勋创办了一个名为 GTC 的大会,在全球不知疲倦地推广 CUDA。最终,功夫不负有心人,一个个应用程序真的涌现了,包括 CT 重建、分子动力学、粒子物理学、流体动力学和图像处理等。
直到 2012 年,AI 研究人员发现了 CUDA 的潜力。著名 AI 专家 Alex Krizhevsky 在 GForce GTX 580 上训练出了 AlexNet,引发了人工智能的大爆炸。

第三次失败,英伟达退出手机芯片市场。

还记得雷军与黄仁勋的同台么?
英伟达发家史:从游戏芯片到AI军火商,老黄的狂飙之路
2013 年时,在雷军的邀请下黄仁勋出席小米手机 3 发布会。
年少时就去到了美国的黄仁勋被雷军要求说中文,他说得并不利索,但也用中文自信地喊出:“英伟达的 GPU 是世界最棒的。”
当时,小米 3 旗舰版搭载了英伟达推出的 Tegra 4 处理器的移动版,这也是该系列的绝唱。
当时,移动手机市场兴起,英伟达也进入了移动芯片市场,尽管整个手机市场都非常庞大,英伟达本可以为了市场份额而战,但他们做出了一个艰难的决定:放弃了这个市场。
黄仁勋表示,英伟达的使命是构建能普通计算机所不能的计算机,他们应该致力于实现这一愿景,做出独特的贡献。英伟达的战略撤退得到了回报。

人生建议:经历苦难,降低期待

2024 年,黄仁勋重返母校斯坦福大学,并在商学院进行演讲,分享了一些人生经验。
当主持人问黄仁勋,关于成功,有没有给斯坦福学生的建议?他回答道:“我希望各位有机会经历大量的痛苦和磨难。”
他提到自己最大的优点之一,就是“我的期望很低”。
黄仁勋表示,多数的斯坦福毕业生,对自己有很高的期望,但他们绝对值得拥有高期待,因为他们来自地球最好的大学之一,被同样不可思议的同辈环绕,拥有高期待是非常自然的事。
"对自己期望非常高的人,往往韧性(resilience)也低,”黄仁勋说,“不幸的是,韧性对带来成功至关重要。
黄仁勋强调,“成功不是来自于智慧,而是来自于性格,而性格是经历苦难塑造出来的。

原文链接

Trending Cryptos

Related Reads

Super Spiral Mega-Boom, Micron's Earnings Report Rekindles the Semiconductor Bull Run

On June 25, 2026, Micron Technology released its blockbuster Q3 FY2026 results, significantly exceeding market expectations and reigniting confidence in the semiconductor bull market. Revenue soared to $41.456 billion (vs. ~$35.4B expected), up 346% year-over-year, while GAAP net profit surged nearly 15 times to $28.243 billion. Guidance for Q4 was even more striking, with projected revenue of approximately $50 billion, far surpassing prior estimates. The report highlighted that the AI boom is now fueling growth across Micron's entire product stack, not just HBM. Cloud memory, core data center, SSD, mobile, and automotive businesses all saw revenue growth exceeding 250-600%, with margins hovering around 80%. While HBM4 is already in volume shipment and 2026 capacity is sold out, AI-driven demand is also tightening supply for traditional DRAM and NAND, sustaining a strong pricing cycle. A pivotal development is Micron's shift toward a "demand-first" model. The company disclosed 16 long-term strategic customer agreements (SCAs), most spanning 5 years to 2030, covering about 20% of DRAM and one-third of NAND shipments. These are take-or-pay contracts, with 14 agreements already securing roughly $100 billion in guaranteed future revenue and $22 billion in customer performance assurances. To fulfill this locked-in demand, Micron plans substantial capacity expansion, with Q4 capital expenditure projected at ~$10 billion. This investment, backed by concrete long-term orders rather than cyclical speculation, marks a historic change for the memory industry. Following the earnings release, Micron's stock surged 16% after-hours, lifting the broader semiconductor sector globally. The report served as a powerful signal that AI infrastructure build-out is accelerating, with memory positioned as a central protagonist in the ongoing narrative.

Odaily星球日报12m ago

Super Spiral Mega-Boom, Micron's Earnings Report Rekindles the Semiconductor Bull Run

Odaily星球日报12m ago

Deciphering the Ethereum Foundation's New Structure: Reaffirming Self-Sovereignty Amid Institutionalization Trends

Summary: The Ethereum Foundation (EF) has announced a major restructuring, laying off 20% of its staff and introducing a new five-layer operational framework. This move aims to clarify the EF's mission and reaffirm Ethereum's core principle of self-sovereignty amidst growing institutionalization in the crypto space. The five layers are: 1. **Protocol Layer**: Focuses on maintaining Ethereum's foundational "CROPS" values—Censorship-resistant, Robust, Open, Private, and Secure. This involves core technical work like secure hard forks and mitigating toxic MEV. 2. **Access Layer**: Ensures users can practically exercise self-sovereignty through actions like reading the chain and making transactions. A key principle is the "zero option," meaning a trusted, non-intermediated path must always exist as an alternative to any centralized service. 3. **User Layer**: Bridges the protocol and access layers by grounding EF's work in the real needs of users and organizations. This is seen as crucial for moving beyond a purely research-driven approach and ensuring development effectively serves the ecosystem. 4. **Community Layer**: Responsible for building and maintaining consensus around Ethereum's core values both internally and externally. This involves guarding against centralization, upholding technological neutrality, and preventing short-term commercial interests from undermining CROPS principles. 5. **Institutional Layer**: Manages EF's engagement with institutions, but with the precondition of self-sovereignty. The goal is not to make it easier for institutions to control users, but to demonstrate how Ethereum's technology can enable better integrations. The article argues that while institutional adoption brings legitimacy, it also risks diluting crypto's foundational ethos of decentralization. The new structure represents EF's effort to navigate this tension, upholding its core mission while actively engaging with a broader, more complex ecosystem.

marsbit30m ago

Deciphering the Ethereum Foundation's New Structure: Reaffirming Self-Sovereignty Amid Institutionalization Trends

marsbit30m ago

OpenRouter: How Did This 'AI Model Relay Station' Achieve a $10 Billion Valuation?

OpenRouter: The Model Router Building a $10B+ Company This article explores OpenRouter, a platform that aggregates access to over 400 AI models from 70+ providers (like OpenAI, Claude, Gemini) through a single API. It has grown into a unicorn with a $1.3B valuation by 2026, processing massive scale—reaching 100 trillion tokens monthly. Its core value isn't just being a "model supermarket." For developers building real-world AI applications, managing multiple models for different tasks (e.g., cheap models for titles, powerful ones for long articles) is complex. OpenRouter acts as a critical "model scheduling layer," handling routing, failover between providers, cost optimization, and enterprise features like zero-data-retention policies and budget controls. OpenRouter's business model is a "toll fee": it charges a small platform fee (5.5%) on purchased credits while passing model costs directly to users. Its revenue scales with the tokens flowing through its system, which saw explosive growth as AI apps evolved. Key growth drivers include: 1) The explosion of specialized models, increasing choice complexity; 2) AI apps shifting focus from performance to cost optimization; 3) The rise of AI agents that require more reliable, multi-step model calls. However, risks remain. Large enterprises or cloud providers (AWS, Google Cloud) could build similar internal gateways. Its position between model suppliers and developers could also create future tension over pricing and data control. To stay ahead, OpenRouter must deepen its enterprise features and prove it's more than just a request forwarder.

marsbit49m ago

OpenRouter: How Did This 'AI Model Relay Station' Achieve a $10 Billion Valuation?

marsbit49m ago

Bitcoin Falls Below $60,000 Again; After 20 Months, We've Reached a New Low

Bitcoin Drops Below $60,000, Hitting a 20-Month Low Bitcoin fell below the key $60,000 psychological level again, reaching a low of $59,023—its lowest point in approximately 20 months, dating back to October 2024. While it later recovered slightly to around $60,600, this marks its third significant breach of $60,000 this year. The downturn is attributed to two primary factors. First, U.S. spot Bitcoin ETFs are experiencing their longest streak of net outflows since launch, with nearly $5.94 billion withdrawn over 30 days. This creates sustained selling pressure as Authorized Participants sell Bitcoin to meet redemptions. Second, shifting macroeconomic expectations are adding pressure. Strong U.S. job data and hawkish remarks from Fed officials have increased market pricing for potential rate hikes, reversing the earlier liquidity-driven bullish sentiment and prompting a shift away from risk assets like Bitcoin. Analyst views are mixed. 21Shares maintains a bullish long-term outlook, expecting prices to recover towards $100,000, citing historical post-halving cycles and substantial ETF holdings as a base. In contrast, Arthur Hayes predicts a potential bottom around $40,000 within six months due to persistent Fed hawkishness. CryptoQuant suggests, based on on-chain data, that the market may not find a bottom until prices fall below the average investor cost basis around $53,000, potentially extending the bearish phase into late 2026 or early 2027. The immediate focus is on upcoming U.S. inflation data and Fed signals. Lower-than-expected CPI could offer relief, but confirmation of sticky inflation or continued ETF outflows may lead to further downside pressure. Bitcoin's ability to hold above $60,000 remains a critical test for the near-term market direction.

Odaily星球日报1h ago

Bitcoin Falls Below $60,000 Again; After 20 Months, We've Reached a New Low

Odaily星球日报1h ago

Trading

Spot
Futures

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片