Has AMD Really Caught Up With Nvidia? OpenAI’s Mega Deal Reveals a Shift in the AI Race

[karin623]

OpenAI’s stunning $100 billion partnership with AMD has sent shockwaves through the semiconductor world — and raised the question everyone’s asking: has AMD finally closed the gap with Nvidia?

The deal gives AMD a historic role in powering ChatGPT’s next phase, marking a bold leap into TSMC’s 2nm frontier and the company’s first-ever rack-scale systems. Yet behind the headlines lies a more complex story — one that speaks to how AI’s power dynamics, and even the rules of the game, are starting to change.

As global AI demand shifts from model training to real-time inference, the hardware race may be entering a new era — one where today’s underdogs get their second act.

Has AMD Really Caught Up With Nvidia? OpenAI’s Mega Deal Reveals a Shift in the AI Race

Liang-rong Chen

cwnewsroom.substack.com

 

[KevinK]

Two things to look at for AMD from my perspective:

* Benchmarking - AMD does about parity in InferenceMAX benchmarking vs NVIDIA in both TCO and power at the slot level (up to 8 processors), but doesn’t seem to have anything to match the much more powerful and cost-effective NVL-72 rack-level integrated hardware and software.

InferenceMAX™: Open Source Inference Benchmarking

NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B

newsletter.semianalysis.com

* Roadmap - both NVIDA and Huawei have highlighted future product that supports disaggregated inference, with customized chips/system for prefill/long context vs decode. Nothing as advanced yet from AMD AFAIK. Here’s info on Huawei’s plans:

Key products in Huawei's AI chips and computing power roadmap

China's Huawei (HWT.UL) has ended years of secrecy to detail its product roadmap for chips and computing power systems, discussing for the first time how it intends to challenge global industry leader Nvidia .

www.reuters.com

 

[hist78]

OpenAI’s stunning $100 billion partnership with AMD has sent shockwaves through the semiconductor world — and raised the question everyone’s asking: has AMD finally closed the gap with Nvidia?

The deal gives AMD a historic role in powering ChatGPT’s next phase, marking a bold leap into TSMC’s 2nm frontier and the company’s first-ever rack-scale systems. Yet behind the headlines lies a more complex story — one that speaks to how AI’s power dynamics, and even the rules of the game, are starting to change.

As global AI demand shifts from model training to real-time inference, the hardware race may be entering a new era — one where today’s underdogs get their second act.

Has AMD Really Caught Up With Nvidia? OpenAI’s Mega Deal Reveals a Shift in the AI Race

Liang-rong Chen

cwnewsroom.substack.com

By looking at OpenAI’s mega deals with AMD, Nvidia, and Broadcom, I believe several guiding principles are shaping OpenAI’s strategy. Among them, the most important are product maturity, performance, available supply and timeline, and the foundry capacity required to manufacture them.

OpenAI is a newcomer to the fabless semiconductor industry, and Sam Altman has likely already realized that even with billions in prepayments to TSMC, OpenAI cannot simply cut ahead of TSMC’s long-term tier one customers such as Apple, Nvidia, AMD, Broadcom, Qualcomm, and MediaTek. Most of TSMC’s leading edge capacity is likely already reserved by these companies two, three, or even more years in advance.

To address this, OpenAI is leveraging the existing relationships that Nvidia, AMD, and Broadcom have with TSMC to secure more capacity for its own needs. Intel could, in theory, capture some of this opportunity and generate more revenue from OpenAI, but there are several challenges to overcome, including:

1. How much Intel 18A capacity will actually be available for external customers like OpenAI over the next 12 to 18 months?

2. How well does Intel’s 18A process align with OpenAI’s performance and feature requirements?

3. While Intel could use its contracted TSMC capacity to support OpenAI, does Intel have a product (manufactured at TSMC) that can realistically compete with those from Nvidia, AMD, or Broadcom/OpenAI? And does OpenAI want to introduce additional complexity to its data centers operations?

​