grimshot

[grimshot @grimshotlll]

@benitoz Ok so TPU v7 > GB300 and TPU v8i > Vera Rubin?

[grimshot @grimshotlll]

@dnystedt @jukan05 Expected, since TPU v8i > Vera Rubin.

[grimshot @grimshotlll]

@The_AI_Investor Not a problem, Jalapeño > GB300 Ultra and TPU > VeaRubin.

[grimshot @grimshotlll]

@firstadopter Google is dumping Broadcom. Anthropic and OpenAI will soon follow. Broadcom guiding 2x increase in 2027 AI revenue and another 2x in 2028 is disappointing. 2029 and 2030 revenue will be back to <100B once contracts end.

[grimshot @grimshotlll]

@Alex_Intel_ Why is Intel dropping from $140 -> $90?

[grimshot @grimshotlll]

@SemiAnalysis_ Jalapeno > GB-Ultra. TPU v8i > Vea Rubin. Let that sink in.

[JT @lokoyacap]

Broadcom over the last 90 days has been busy it seems (h/t to WFC)👀

2/27/2026 10-Q:

Purchase Commitments: $54 million

"Certain multi-year customer contracts in our semiconductor solutions segment and infrastructure software segment...the remaining performance obligations under these contracts as of February 1, 2026 were approximately $45.0 billion"

5/29/2026 10-Q:

Purchase Commitments: $128.1 billion (yes, not a typo)

"Certain multi-year customer contracts in our semiconductor solutions segment and infrastructure software segment...the remaining performance obligations under these contracts as of May 3, 2026 were approximately $164.6 billion. These commitments include obligations under a long-term contract for custom AI accelerators entered in the fiscal quarter ended May 3, 2026."

[grimshot @grimshotlll]

@The_AI_Investor You trying to suggest that in 2023 almost 100% of AI computes are Nvidia GPU? Wow

[grimshot @grimshotlll]

@benitoz Bro that’s ASIC 🤣 so is Nvidia building ASIC now?

[grimshot @grimshotlll]

@jukan05 Why exclude networking in 26 27 AI revnue?

[grimshot @grimshotlll]

@Ric_RTP Everyone is saying Nvidia GPU is the best, suddenly there is a faster chip from a startup, funny

[grimshot @grimshotlll]

@vikramskr People been saying Nvidia can do better ASIC internally if they choose to, and they don’t because they get higher margin with GPU, why pay the 20B?

[grimshot @grimshotlll]

@Midnight_Captl Why is Nvidia buying Groq if their chips are generations ahead? Am I missing something? Did Jensen just agree ASIC competitiveness?

[grimshot @grimshotlll]

@benitoz @sundeep Why is Nvidia buying groq if their chips are already few generations ahead?

[FUNDA @FundaAI]

Over the past two days, there’s been a lot of discussion on X about TPU vs. GPU TCO, especially after Semianalysis released its report on the topic last week. But many of these debates are misleading. From the very beginning, TPUs were never designed to be compared on single-card performance. The core design philosophy of TPU is to optimize TCO at the superpod level. Therefore, if you want to compare TCO, you must evaluate a TPU superpod together with its optical interconnect stack (OCS + BiDi transceivers). And since Google has never disclosed the MFU of its superpod, most single-card TCO comparisons are not particularly meaningful.

Based on what we’ve learned:

1.Some large customers’ initial benchmarking shows that, at the 10,000-card scale, TPU v7 outperforms GB200. However, the JAX ecosystem still lags behind PyTorch/CUDA in ease of use.

https://t.co/g6RzVhbb7y Google’s internal use cases, TPU v5p has a better TCO than H100, even though its BF16 compute performance per chip is less than 40% of H100.

The most fundamental reason Google can outperform Nvidia at the 10,000-card scale is that Google uses OCS + BiDi transceivers to scale up to that size. For v7 Ironwood, the scale-up bandwidth reaches 4800 Gb. Nvidia GPUs, on the other hand, scale out to 10,000 cards, with bandwidth at 400 Gb (GB200) and 800 Gb (GB300). We expect that in the next generation of v7, Google will further increase scale-up bandwidth to 9600 Gb.

Therefore, comparing single-card performance between V8 and Rubin alone cannot fully capture the reality of TPU vs. Nvidia. This is also why we’ve consistently emphasized that optical interconnects are the biggest alpha in the TPU supply chain. $NVDA $GOOG $LITE $COHR

[grimshot @grimshotlll]

@Midnight_Captl All of you look super frightened with a thing that you claimed not competitive with NV, chill