Cerebras IPO makes billions for Benchmark but VC Eric Vishria almost didn’t take the meeting
Cerebras IPO makes billions for Benchmark but VC Eric Vishria almost didn’t take the meeting
Cerebras IPO 为 Benchmark 带来数十亿美元收益，但风投 Eric Vishria 当初差点拒绝了这次会面
The Cerebras Systems IPO was a smash hit on Thursday, generating billions for itself, its founders, and its major investors. Among the big winners is major shareholder Benchmark, which owns 9.5% of the company. One of the firm’s general partners, Eric Vishria, has been a Cerebras board member since 2016, the year the AI chip maker was founded, having co-led its $25 million Series A round. Cerebras Systems 的 IPO 在周四取得了巨大成功，为其自身、创始团队及主要投资者创造了数十亿美元的财富。主要股东 Benchmark 是大赢家之一，持有该公司 9.5% 的股份。Benchmark 的普通合伙人之一 Eric Vishria 自 2016 年（即这家 AI 芯片制造商成立之年）起便担任 Cerebras 董事会成员，并共同领投了其 2500 万美元的 A 轮融资。
But these billions only happened for Benchmark because Vishria met with the startup almost against his will, he told TechCrunch. “It was five founders and a deck, and it was our first hardware investment in 10 years,” Vishria told TechCrunch about that first meeting. “I had been a venture capitalist for like, 18 months.” (Prior to being a VC, Vishria sold the social browser startup he co-founded, RockMelt, to Yahoo for a reported $60-$70 million in 2013.) 但 Vishria 告诉 TechCrunch，Benchmark 能获得这数十亿美元的收益，纯属偶然，因为他当初几乎是“被迫”与这家初创公司会面的。“当时是五位创始人带着一份演示文稿，而这是我们十年来第一次投资硬件，”Vishria 在谈到那次初次会面时说，“我当时做风投才大约 18 个月。”（在成为风投之前，Vishria 曾将他联合创立的社交浏览器初创公司 RockMelt 以据称 6000 万至 7000 万美元的价格卖给了雅虎。）
Benchmark is famously selective in the companies it chooses, and backs hardware companies so rarely that Vishria was kicking himself for giving time to Cerebras. “Why did I take this meeting?” he kept muttering. At one point, he even messaged his assistant, who manages his calendar, and bugged her: “Why did you let me take this meeting?” Vishria recalls. Benchmark 以挑选公司极其挑剔而闻名，且极少投资硬件公司，以至于 Vishria 当时因为给了 Cerebras 会面时间而懊恼不已。“我为什么要参加这个会议？”他不停地嘀咕。Vishria 回忆道，有一次他甚至给负责管理日程的助理发信息抱怨：“你为什么要让我参加这个会议？”
But his grumpy attitude vanished by the third slide, as co-founder and CEO Andrew Feldman laid out Cerebras’ grand plans. “The first deck is the title slide. The second deck is the team. And I was like, ‘Oh, that team is really good.’ And the third slide is something along the lines of ‘GPUs actually suck for deep learning. They just happen to be 100 times better than CPUs.’ And as soon as he said it, a light bulb went off,” Vishria recalled. “I was like, ‘Oh, my God, of course. Like, why would a graphics processor be the right thing for AI?’” 但当联合创始人兼 CEO Andrew Feldman 展示 Cerebras 的宏伟蓝图时，Vishria 的不满情绪在看到第三页幻灯片时烟消云散。“第一页是标题，第二页是团队介绍。我当时想，‘哦，这个团队真的很棒。’而第三页的内容大致是‘GPU 其实并不适合深度学习，它们只是恰好比 CPU 快 100 倍而已。’他话音刚落，我脑子里就像灯泡亮了一样，”Vishria 回忆道，“我当时想，‘天哪，当然了。为什么图形处理器会是 AI 的最佳选择呢？’”
Still, this was years before Google’s famous Transformer paper — the 2017 research that laid the groundwork for modern AI — which eventually led to ChatGPT. Cerebras was pitching a new kind of giant-sized chip, designed for AI training, one the processor world was not prepared to manufacture. 然而，这发生在谷歌著名的 Transformer 论文（2017 年为现代 AI 奠定基础的研究，最终促成了 ChatGPT 的诞生）发表的前几年。当时 Cerebras 正在推销一种专为 AI 训练设计的新型巨型芯片，而当时的处理器制造行业尚未做好生产此类芯片的准备。
Vishria was intrigued enough to discuss it with some Benchmark partners, who quickly told him that they also didn’t know enough hardware. They said if he wanted this deal, he would have to bring in one of the original Benchmark founders from the 1990s, who did understand. Undeterred, Vishria scheduled a meeting to have Feldman pitch to founding partner Bruce Dunlevie, who grilled the founder about chip packaging and cooling and more. “Most of that meeting was like a dog watching TV for me,” Vishria joked, because he understood so little. Vishria 对此产生了浓厚兴趣，并与 Benchmark 的其他合伙人进行了讨论，但他们很快表示自己对硬件了解不足。他们告诉他，如果想做这笔交易，就必须请出 90 年代 Benchmark 的创始合伙人之一，因为他们才懂行。Vishria 没有气馁，他安排了一场会议，让 Feldman 向创始合伙人 Bruce Dunlevie 进行演示。Dunlevie 对芯片封装、散热等问题进行了严厉盘问。“那场会议的大部分时间对我来说就像狗看电视一样，”Vishria 开玩笑说，因为他实在听不懂。
After the pitch, Dunlevie warned that what Cerebras was attempting would be hard. Others have tried and failed. But he thought this team had a shot. He, however, worried there’d be no market for the chip. Although Vishria didn’t fully understand the tech, he was convinced that if Cerebras “could make AI faster” there would be a market for it, and this team had the chops to succeed, he said. They had previously sold a startup, SeaMicro, to AMD. 演示结束后，Dunlevie 警告说 Cerebras 尝试的事情难度极大，此前已有不少人尝试并失败了。但他认为这个团队有成功的机会。不过，他担心这种芯片可能没有市场。尽管 Vishria 对技术并不完全理解，但他坚信如果 Cerebras “能让 AI 运行得更快”，市场一定会存在，而且这个团队有成功的实力。他们此前曾将一家名为 SeaMicro 的初创公司卖给了 AMD。
“The advantage of having had a successful exit previously, is it erases some of the uncertainty in the venture capitalists’ minds,” Cerebras CEO Andrew Feldman tells TechCrunch. “We hadn’t just fallen off the back of a turnip truck. We were an experienced team.” “此前有过成功退出的经历，其优势在于消除了风投心中部分的不确定性，”Cerebras CEO Andrew Feldman 告诉 TechCrunch，“我们并非初出茅庐，我们是一个经验丰富的团队。”
Hardware is hard
硬件之路艰难
What followed was 8.5 years of grind as Cerebras dealt with struggle after struggle to build its product. Feldman and his Cerebras co-founder and CTO Sean Lie had to invent new cooling methods to prevent a chip of that size from burning when drawing power. They had to invent a machine that could drill 40 screws into the wafer simultaneously without cracking it. And so on. The Benchmark investor repeatedly thought to himself, “What are we doing?” 随后的 8 年半时间里，Cerebras 在打造产品的过程中历经磨难。Feldman 和他的联合创始人兼 CTO Sean Lie 不得不发明新的散热方法，以防止这种尺寸的芯片在通电时烧毁。他们还必须发明一种能同时在晶圆上钻 40 个螺丝而不使其破裂的机器，诸如此类。这位 Benchmark 的投资者曾多次自问：“我们到底在做什么？”
Plus, hardware is expensive. At the point where the company raised half a billion dollars from a long list of investors, its chips were still being developed. It had to raise again in the 2022 VC bear market. “You don’t have a lot of traction on the company yet, so yeah, that was where it got really tough,” Vishria recalls. 此外，硬件开发成本高昂。当公司从众多投资者那里筹集到 5 亿美元时，其芯片仍处于开发阶段。他们不得不在 2022 年风投市场的熊市中再次融资。“当时公司还没有太多的市场表现，所以那段时间确实非常艰难，”Vishria 回忆道。
But around 18 months ago, everything changed. Cerebras’ chips, designed for training and successfully being manufactured by TSMC, the world’s largest contract chip manufacturer, turned out to be even better for inference — running AI models to generate responses, rather than teaching them in the first place. Just as that realization hit, the AI world grew insatiably thirsty for that kind of compute. It had a big customer and revenue. 但大约 18 个月前，一切都变了。Cerebras 的芯片原本是为训练而设计，并由全球最大的芯片代工厂台积电（TSMC）成功制造，结果发现它们在推理方面表现更佳——即运行 AI 模型以生成响应，而非最初的训练阶段。就在这一发现出现时，AI 领域对这种算力的需求变得极其渴求。公司因此获得了大客户和营收。
Instead of another private round, Cerebras tried to go public in 2024, only to wind up stuck in U.S. government scrutiny over national security concerns triggered by a large investment by its only major customer, Abu Dhabi-based cloud provider G42. Public investors also weren’t keen on its dependence on G42 coupled with huge losses. The delay was a blessing in disguise. Today, OpenAI and AWS are large customers, too. Cerebras doubled revenues and declared a profit last year. Cerebras 没有进行新一轮私募融资，而是试图在 2024 年上市，却因其唯一大客户——总部位于阿布扎比的云服务商 G42 的巨额投资引发了美国政府对国家安全问题的审查，导致上市受阻。公开市场投资者也不看好其对 G42 的依赖以及巨额亏损。但这次延迟反而因祸得福。如今，OpenAI 和 AWS 也成为了其大客户。Cerebras 去年营收翻了一番，并宣布实现盈利。
Vishria gives all props to the Cerebras team for “persistence, ingenuity, but also adaptiveness,” he says. But this is also a feather in the investor’s cap for finding a winner so far outside the firm’s usual comfort zone. Vishria 将这一切归功于 Cerebras 团队的“坚持、独创性以及适应能力”。但对于这位投资者来说，能在公司惯常的舒适区之外发掘出这样一个赢家，也是他职业生涯中的一大亮点。
Benchmark owned 17,602,983 shares worth $3.3 billion at the IPO’s opening price of $185 price, and over $5.3 billion if the first day of trading’s price of over $300 price holds. It can’t sell shares until after a six-month lockup expires — a standard restriction that prevents insiders from selling immediately after a company goes public. The firm bought about 80% of those shares in early rounds for around $18 million, various disclosures indicate and Vishria confirmed to TechCrunch. It bought the remainder at pricier later rounds which cost it around $250 million, Cerebras disclosed in its S-1. So all told, the venerable VC firm spent maybe $270 million for this stake that is worth multiple billions or more, depending on how the stock price holds. Benchmark 持有 17,602,983 股股票，按 IPO 开盘价 185 美元计算价值 33 亿美元；如果首日交易价格保持在 300 美元以上，其价值将超过 53 亿美元。在六个月的锁定期结束前，他们无法出售股票——这是一项防止内部人士在公司上市后立即抛售的标准限制。多份披露文件显示，且 Vishria 向 TechCrunch 证实，Benchmark 在早期融资轮次中以约 1800 万美元买入了约 80% 的股份。据 Cerebras 的 S-1 文件披露，其余股份是在后期价格较高的轮次中买入的，耗资约 2.5 亿美元。总而言之，这家老牌风投公司总共投入了约 2.7 亿美元，换取了如今价值数十亿美元甚至更多的股份，具体取决于股价的后续表现。