Assertion Supported AI assessment confidence: 95% certainty 4/5 debate potential 3/5

Google is the only tech giant with successful custom AI chips

Dwarkesh Patel · AI Scaling, Alignment, and the Path to Superintelligence — With Dwarkesh Patel · May 15, 2024 · at 12:46

Dwarkesh Patel discusses hardware moats in frontier AI development on the Big Technology Podcast.

0:00 / 0:11exact quote · 11.3s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“Google is the company that actually has a successful accelerator program for AI chips already with their TPUs, which other companies are trying, but don't yet have to replace, you know, Nvidia GPUs.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

Google is currently the only major tech company with a successful in-house AI chip accelerator program (TPUs) viable for replacing Nvidia GPUs.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

As of May 2024, Google was the only major tech hyperscaler with a mature, full-scale in-house AI chip accelerator program capable of training and serving frontier foundation models without primary reliance on Nvidia GPUs. Google trained its flagship Gemini 1.0 and Gemini 1.5 models end-to-end on clusters of TPU v4 and TPU v5p accelerators rather than Nvidia hardware. In contrast, competing in-house chip programs had not reached viability for frontier LLM training: AWS relied on Nvidia GPUs for Anthropic's Claude 3 training while Trainium 2 remained under development; Meta deployed MTIA solely for internal recommendation/ranking inference while training Llama 3 on tens of thousands of Nvidia H100 GPUs; and Microsoft Azure's Maia 100 was in early pilot testing while OpenAI and Microsoft trained GPT-4 and GPT-4o on Nvidia GPU clusters.

Google trained Gemini 1.0 (Ultra, Pro, Nano) and Gemini 1.5 Pro end-to-end on TPU v4 and TPU v5p clusters, demonstrating a fully viable alternative to Nvidia GPUs for frontier model training.
AWS's primary AI partner Anthropic trained Claude 3 on Nvidia GPU clusters, with Trainium 2 clusters not yet deployed at frontier scale as of May 2024.
Meta trained its flagship Llama 3 models entirely on two 24,576-GPU clusters of Nvidia H100s, utilizing its in-house MTIA chips only for recommendation workloads.
Microsoft Maia 100 was in internal testing and early pilot deployment as of May 2024, with Microsoft and OpenAI relying exclusively on Nvidia A100/H100 clusters for GPT-4 and GPT-4o.

Checked against

[1] youtube.com [2] uncoveralpha.com [3] ycombinator.com [4] substack.com [5] blog.google [6] bwbx.io [7] substack.com [8] medium.com

AWS Trainium Anthropic Claude 3 training May 2024 "Gemini: A Family of Highly Capable Multimodal Models" "TPU" Google TPU Gemini training custom AI chip May 2024 hyperscaler custom AI chips comparison TPU Trainium Maia MTIA 2024 "Llama 3" "H100" "MTIA" training hardware April 2024 AWS Trainium Inferentia Meta MTIA Microsoft Maia status May 2024 Microsoft Maia 100 status May 2024

assessed as of 2026-09-02

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

Google Nvidia TPU

⚑ Spot an error? Report it

More from Dwarkesh Patel

Prediction Not checkable as stated

Patel: Individuals will eventually train superintelligences in basements

“The cost of training, the systems is declining so fast that Literally you will be able to train a super intelligence in a basement at some point in the future.”

Dwarkesh Patel Jun 18, 2025 ▶ 56:49 Dwarkesh Patel: AI Continuous Improvement, Intelligence Explosion, Memory, Frontier Lab Competition

Opinion

Dwarkesh: AGI requires further algorithmic progress, not just current model scaling

“I don't think we're just right around our corner from AGI and it's just a little additional dash of something. That's all it's going to take. I think, you know, people often ask if all AI progress stopped right now and all you could do is collect more data or …”

Dwarkesh Patel Jun 18, 2025 ▶ 1:56 Dwarkesh Patel: AI Continuous Improvement, Intelligence Explosion, Memory, Frontier Lab Competition

Insight

Dwarkesh: Lack of continual learning prevents LLMs from replacing human labor

“I think a big bottleneck these models have is their inability to learn on the job, to have continual learning. Their entire memory is extinguished at the end of a session. There's a bunch of reasons why I think this actually makes it really hard to get human-l…”

Dwarkesh Patel Jun 18, 2025 ▶ 2:21 Dwarkesh Patel: AI Continuous Improvement, Intelligence Explosion, Memory, Frontier Lab Competition

Prediction Not checkable as stated

Patel: Reinforcement learning may not generalize beyond verifiable domains

“I still think I I'm like, I'm not confident that this will generalize to domains that are not so verifiable or text-based.”

Dwarkesh Patel Jun 18, 2025 ▶ 7:34 Dwarkesh Patel: AI Continuous Improvement, Intelligence Explosion, Memory, Frontier Lab Competition

Prediction Not checkable as stated

Patel: Online continual learning is not imminent for current AI architectures

“And the reason I don't think that's around the corner is just because there's not, there's no obvious way, at least as far as I can tell, to just slot in this online learning into the models as they exist right now.”

Dwarkesh Patel Jun 18, 2025 ▶ 8:11 Dwarkesh Patel: AI Continuous Improvement, Intelligence Explosion, Memory, Frontier Lab Competition

Assertion Not checkable as stated

Patel: Pre-training scaling is seeing diminishing returns

“Pre-training, which is this idea that you just make the model bigger that has had diminishing returns.”

Dwarkesh Patel Jun 18, 2025 ▶ 9:13 Dwarkesh Patel: AI Continuous Improvement, Intelligence Explosion, Memory, Frontier Lab Competition

Google is the only tech giant with successful custom AI chips

More from Dwarkesh Patel

Turn any episode into a week of clips.