Item — HN Reader

Google is still one of the Nvidia's largest customers. Google Ironwood (TPU v7) is roughly 12 months (generation) behind Nvidia in performance, but they get incredible savings in cost by having their own chips.

Where Google has advantage is massive Superpods that scale better than Nvidia racks.

1 comments

Google doesn't and won't have enough TPUs regardless. Nvidia owns the supply chain (confirmed with them trying to or has taken over Apple as TSMC's biggest customer).

That's because Google is once removed from TSMC, unlike Nvidia and Apple

Google and Broadcom are in co-design partnership (and now also MediaTek). Google defines the architecture. Broadcom provides the essential infrastructure (IP blocks) that makes the chip work on silicon and works tight with TSMC to make it happen in new nodes. Spending billions to help Broadcom get into par with Nvidia is not what they want.

Apple, for example, designs nearly the entire SoC in-house (CPU, GPU, NPU) and work with TSMC hand in hand. They buy only specific chips and block designs like Wi-Fi, Bluetooth, and RF.

> Apple, for example, designs nearly the entire SoC in-house

I already mentioned Apple lost their top spot. They're also are paying a 100% price hike for Samsung RAM.

> Spending billions to help Broadcom get into par with Nvidia is not what they want

Point being if Apple doesn't have enough whether Google is once removed or not still won't have anywhere near enough to compete. This is about supply. Nvidia reserved capacity. You can't get it.

Even if Nvidia GPUs were worse - they have 10-100x more of them. What would you as a large customer want? Tesla for example has many buildings racking GPUs and more in the works. Specialized hardware like what OpenAI is doing with the new "spark" will just be "experimental" and side projects because better or not there won't be enough for real demand.

Apple gets enough. Their need for chips is not growing as high as Nvidia's.

I think you are trying to say that Nvidia has lock-in due its massive reservations and it starves others out. That's partially true real bottleneck currently is CoWoS/CoWoS-L packaging, not the chips. Nvidia is reaping the benefits of being pioneer who helped to make it possible. I'm not sure of the exact numbers, but I think Nvidia has over 60% capacity reservation of TSMC CoWoS.

That kind of lock in is only temporary. Anyone can reserve ahead of time and take the risk of failing to get competitive node and suffer the losses.

ps. Nvidia is paying 50% to 100% premiums over standard prices just to secure "urgent" additional capacity beyond their original reservations. If others want that capacity they can bid.