DeepSeek secretly builds its own chips to bypass Nvidia
The budget-friendly AI darling is quietly moving from smart software to custom hardware. They are tired of paying rent to the green GPU giant and want to run their models on their own terms.
The Chinese AI disruptor has been quietly poaching top-tier silicon engineers for nearly a year without posting a single public job opening. Instead of shouting on LinkedIn, the company is using surgical headhunting to build an elite hardware team that already includes chip architects, verification experts, and software engineers.
This secret project is laser-focused on creating a custom chip optimized solely for inference—meaning the actual generation of AI answers rather than the brutally expensive training phase. While training gets all the sci-fi headlines, running these models in the real world is where the actual bills pile up, often devouring over half of an AI company's operating budget. It turns out that buying digital shovels from Nvidia is a great way to go broke, so DeepSeek decided to forge its own steel.
The hardware ambition is already baked directly into their code. Insiders point out that the recent DeepSeek-V3.1 model was specifically designed with FP8 mathematical formatting, a move that looks suspiciously like a software bridge built for their future silicon. The project is still in its infancy, with mass production at least a year away, but the algorithms are already waiting for their physical home.
This DIY chip fever is quickly becoming the ultimate status symbol for cash-rich AI startups. OpenAI recently teased its own chip project, codenamed Jalapeño, in partnership with Broadcom, while Anthropic and Sakana AI are also drawing up blueprints. To fund this expensive hobby, DeepSeek secured a monstrous $7 billion funding round in June 2026, backed by tech giant Tencent and battery titan CATL, valuing the startup at up to $59 billion.
This massive war chest is already being converted into power infrastructure plans. The company is actively scouting for engineers to design massive data centers scaling from megawatts to gigawatts, with potential sites being mapped out as far as Ulaanbaatar.
Building a proprietary semiconductor from scratch is a notorious money pit that has broken many ambitious tech giants. Yet, if a company that conquered the AI world through pure algorithmic efficiency can actually pull off its own silicon, the traditional chip monopoly is in for a very rude awakening.
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.