Nvidia Has No Clear No. 2: Who Runs the 60 Direct Reports When Jensen Huang Retires
Nvidia's strategy wasn't planned — it was locked in after tactics like three teams, two seasons and shipping a whole cow proved themselves. But the company still has no obvious No. 2.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
NV1 lost because it led on tech but wasn't compatible with the mainstream
Nvidia's first product, NV1, used 3D graphics acceleration, while the market at the time was broadly on 2D standards. They judged the market would move quickly from 2D to 3D, but the transition was slower than expected. The hottest game at the time, Doom, was still 2D, so NV1's newest graphics chip couldn't run the best-selling game. Priem also added audio to the chip, but used his own standard, incompatible with the then-popular Sound Blaster. The result was a product that won praise but not sales — customers fed back that your graphics display is pretty good, but take out the audio and I'll buy it.
— Zhou KanglinOne contract clause saved a dying Nvidia
During the NV2 era, Nvidia signed an agreement with the Japanese game company Sega. Huang added a clause to the contract: as long as I can produce a chip prototype and it works on your motherboard, you have to pay me — rather than requiring me to mass-produce. That million dollars of revenue gave the company financial breathing room. Yajun says the contract originated from a passage in the contract at the time; without that passage, NV3 might never have come out, and the company might have died then.
— Wang YajunRiva 128 won the bet on redundancy and a simulator
Riva 128 was Nvidia's comeback product from the brink. The plan no longer chased its own leading-edge technology elements, but followed the technology route the market recognized, packing in the most transistors. The company had only that little bit of money left and could afford only one tape-out. To guarantee success, they did something few had done before: simulating and debugging the chip's operation on a simulator. They also left some blank space and redundancy in the chip design, and that redundancy ultimately saved it — and that decision wasn't made by Huang, it was made by architect Priem.
— Wang YajunThree teams, two seasons was fought out, not planned
The chip development cycle at the time was about 18 months, which would miss two PC selling seasons. Nvidia split its teams into three: while one team designed the chip, the other two supported it, compressing the development cycle to under six months. This strategy wasn't thought up from the start — it was forced out of the situation during Riva 128, when the books only had enough cash for nine months, and after buying the emulator the estimate dropped to six. After doing it, they found the company could work this way, and only then locked it in. 3DFX later tried to copy it, and found that what Nvidia did was just execution and a strict spec sheet.
— Dong ShimingShipping a whole cow: holding the low end for just over a dollar
Chips need quality testing after manufacturing, and poorer-quality products used to be scrapped. Nvidia realized it could lower the clock speed on these products and push them to market as low-end products, spending just a little to fix them — a bit over a dollar — but holding a market segment so others simply couldn't get in. This is ‘shipping a whole cow’ — one cow eaten many ways, nothing wasted. This strategy also wasn't dreamed up in the office; it came out of Huang talking with employees about cost control and discovering a special packaging cost of about a dollar or so.
— Wang YajunWall Street didn't buy CUDA, and Huang didn't waver
While developing CUDA, the whole Wall Street market was very disapproving. At offline meetings people kept asking: are you going to spend a lot on new product development, will it hit gross margin hard — which left Huang quite depressed. Then the financial crisis hit, the stock fell terribly, but Huang never wavered, and the company kept investing in CUDA, looking for more use cases and a better architecture language. By now everyone sees Nvidia isn't just a hardware company, and the CUDA ecosystem makes it very hard for users to migrate to other chips.
— Dong ShimingPriem thought the architecture was his; Huang said it was Nvidia's
Priem was an excellent architect; Riva 128 and three teams, two seasons were both built on his clever chip architecture design. But after the company reached a certain stage, he believed this is my architecture, while Huang believed it isn't your architecture, it's Nvidia's architecture — one person can't do everything. Priem was very possessive, and would delete or change the architecture without notifying colleagues. Later Huang demanded that what he deleted be restored and that next time he follow company process. Priem found that hard to accept, and combined with problems in his marriage, he ultimately chose to stop managing chip architecture and gradually left.
— Dong ShimingNvidia has no obvious No. 2
At the end the show discusses whether Nvidia has a No. 2: if Huang retires, how does the company run, who manages those 60 people. There's currently no very obvious news report or information indicating one person can take on that role, pick up Huang's baton, and keep leading the company in executing its existing systems and culture. Because the way he's chosen makes Nvidia an extension of his own personality — but if he retires, that's actually a problem.
— Zhou KanglinIn their own words · checked verbatim
What he particularly disliked was customers ordering a milkshake, because he felt the milkshake had a long production cycle and a long cleanup time afterward, so whenever he saw a customer about to order a milkshake, he would try as hard as he could to persuade them, and then help them switch to a cola.
他特别不喜欢的是客户去点奶昔 因为他觉得奶昔这个产品 制造周期又很长 吃完以后的清理时间也很长 所以呢 当他看到客户要去点这个奶昔的时候呢 他都会尽可能的说服他们 然后呢 帮他们去换成可乐
Dong Shiming9:54
He realized that for this kind of product with quality problems, I can actually lower its clock speed and push it to market as a low-end product. That way I only need to spend a little money to fix it, a bit over a dollar, but I can hold a market segment that others simply cannot get into.
他意识到这种质量有问题的产品 其实我是可以把它的频率调低 把它作为一个低端产品 推到市场上 这样我其实只需要花一点点的钱 来修复它 一块多美元 但是我可以站住一块市场上 其他人根本就不可能进来
Wang Yajun1:00:43
This strategy isn't that from the start I thought it up and decided we want to do it this way, and then we set it up. Rather, because they were in that environment, they had done it this way and succeeded, so they just said, let's verify whether the company can really operate the way it did at the limit back then.
这个战略不是说我一开始我就想好拍脑袋 说我们想这么去干 然后我们再弄好 而是他们因为是在这种环境下 他们这么干过 而且成功了 所以他们只是说 要来验证一下公司 到底是能不能按照当时的 一个极限的方式去做
Dong Shiming1:04:45
People asked him, if your stock falls that much, what do you think? He said, nothing much to think — do what needs to be done. Do it the way it was done before, still do it that way, right? Have our customers changed? Customers haven't changed. Our products changed? Products haven't changed either, they're right there. So we just do what we need to do.
别人问他 你股价跌那么多的话 你会怎么想 他说 没什么怎么想 该怎么干就怎么干 以前怎么干 还怎么干 对吧 那我们客户变了吗 客户也没变 我们产品变了 产品也没变 就在那边 所以我们该干啥干啥
Wang Yajun1:16:56
Second really is nothing. Before, if we didn't discuss this number thing, we'd think first and second is just a slogan. But when we think about who still makes money after Nvidia, and how much they can make, then in that situation, you can only be first.
第二真的就没什么 我们以前可能 如果不讨论这个数字的东西 会觉得第一第二 这个只是一句口号 但是我们想想 英伟达之后 谁还赚钱 能赚多少钱 所以在这种情况下 你只能做第一
Wang Yajun1:42:06
Figures
| Revenue from the NV2 contract | 1 million USD | 29:24 |
| Runway left during the Riva 128 period | about 6 months | 1:04:45 |
| Chip development cycle after compression | under 6 months | 54:40 |
| Market cap decline during the NV30 period | 80% | 45:33 |
| Number of Top5 emails Huang reviews daily | about a hundred | 51:38 |
| Number of people Huang directly manages | 60 | 54:40 |
| Apple's share of industry profit | 80% | 1:18:57 |
Glossary
- CUDA
- Nvidia's software platform that lets GPUs be used for general-purpose computing rather than graphics alone.
- Riva 128
- Nvidia's comeback graphics card, which succeeded through transistor count and simulator debugging.
- pilot in charge
- In Nvidia's project system, the person fully responsible for a given project.
- Top5
- Employees email their boss their five most important items; Huang reviews them every day.
How to listen
Founders and executives at tech companies watching Nvidia's strategy and organization and wanting to learn its management methods, plus investors researching the compute and chip industry.
There's a stretch of garbled transcription around 1:19:57 that can be skipped.