DeepSeek R2 leak points to 1.2T params and Huawei chips

Silicon Valley hasn't even fully recovered from the DeepSeek R1 shockwave, and the next storm is already brewing.

Venture capitalist Deedy shared leaked technical specs on X about DeepSeek's upcoming flagship model, DeepSeek R2. While DeepSeek hasn't officially confirmed any of these figures, reports originating from Chinese sources are quickly making the rounds across the global AI ecosystem.

1.2T parameters at a fraction of the cost

According to the leak, DeepSeek R2 will adopt an advanced hybrid Mixture of Experts architecture. The total parameter count allegedly doubles R1 to reach 1.2 trillion, with 78 billion active parameters per forward pass.

Leaked benchmarks claim an 89.7% score on C-Eval 2.0 and a 92.4% score on COCO vision benchmarks, reportedly aiming to go toe-to-toe with OpenAI's GPT-4 Turbo and Google's Gemini 2.0 Pro.

Then there is the pricing bomb: R2 is rumored to be 97.3% cheaper than GPT-4o. The leak points to just $0.07 per million input tokens and $0.27 per million output tokens. If accurate, DeepSeek isn't just undercutting the competition, it is practically giving high-end inference away for free.

Ditching US chips for Huawei Ascend

The most geopolitical headline in the leak sits under the hood. DeepSeek R2 was reportedly trained on 5.2PB of data using a compute cluster heavily powered by Huawei's Ascend 910B chips.

The cluster reportedly delivered 512 PetaFLOPS of FP16 compute and achieved an eyebrow-raising 82% hardware utilization rate.

If true, this marks a massive strategic pivot away from the US chip supply chain. Instead of scrambling for sanctioned or slowed-down Nvidia hardware, DeepSeek appears to be mastering localized silicon at a massive scale.

Why it matters

First, take every leaked metric with a massive grain of salt until official weights or technical reports drop. But if even a portion of this is true, the implications are staggering.

For enterprise developers and startups, pricing this low would completely rewrite the unit economics of building AI agents. Workloads that once burned thousands of dollars monthly on proprietary frontier APIs could suddenly drop to a few dozen bucks. Trying R2 the moment it drops is a complete no-brainer for anyone managing serious inference bills.

For the broader industry, we think the Huawei hardware angle is the real story. An 82% utilization rate on an Ascend 910B cluster would prove that frontier models no longer require full reliance on Nvidia's CUDA ecosystem. It seems poised to validate China's domestic semiconductor push, while applying immense pricing pressure on Western AI labs that are still justifying massive training clusters and high API costs.