@FinanceYF5: 1/ DeepSeek is playing a trillion-dollar game. It doesn't do programming packages, doesn't do multimodal, and insists on open source — looks like self-sabotage. The truth is: it's not aiming for a few hundred million dollars in business, but to support a $10 trillion Chinese AI hardware ecosystem.

1/🧭 DeepSeek Is Playing a Trillion-Dollar Game

It doesn’t do programming bundles, it doesn’t do multimodal, and it insists on staying open-source—on the surface, it looks like it’s crippling itself.

The truth is: it’s not going after a few hundred million dollars in business—it’s aiming to prop up a $10 trillion Chinese AI hardware ecosystem. 👇 https://t.co/0saQ5ymIbj

2/ 1M context, it only needs 5.48 GB of VRAM

Running the same 1 million tokens: DeepSeek V4 takes just 5.48 GB, GLM5 needs 60 GB, Qwen3 needs 89 GB.

It’s a 1.6 trillion parameter model—the largest in parameters, yet the most memory-efficient.

3/ Cheap memory for scarce compute

The KV Cache is so small it can be offloaded to SSD, and weights can be streamed from LPDDR to VRAM—this bypasses HBM, the hardest bottleneck for China to manufacture.

For NAND, look to Yangtze Memory; for LPDDR, look to ChangXin Memory—domestic supply can ramp up immediately.

4/ Cutting demand while breaking moats

Engram moves knowledge lookup tables into memory, replacing full-layer forward computation with cheap memory queries, directly saving compute.

TileLang lets a single operator kernel run across multiple chips—domestic GPUs and ASICs are no longer locked out by CUDA.

5/ How it makes money: copy OpenAI

OpenAI got warrants from AMD and Cerebras, unlocking shares based on usage, binding both parties’ interests together.

DeepSeek will do the same: sign equity agreements with Chinese memory, ASIC, CPU, and networking chip makers, building domestic hardware into a viable solution.

6/ It’s not selling a model—it’s selling the foundation

Western AI hardware stocks already exceed $10 trillion in market cap. China’s slice is just as big.

DeepSeek uses technology to grow the ecosystem, taking a $1 trillion valuation piece for itself—that’s the answer behind all its seemingly contradictory choices.

Full content:

That’s all, original author @bookwormengr

If you like this topic:

1. Follow me (@FinanceYF5)
2. Like + retweet the first post below

Looking back at the truth about ChatGPT Images 2.0

1/ OpenAI didn’t release an image model

Altman said on stage, “From GPT-3 all the way to GPT-5,” and the official blog’s opening line was “Images are a language, not decoration.”

That’s not a poetic sentence—it’s a strategic statement: from now on, images are part of OpenAI’s backbone, not an add-on.