Physix FrontierDeepSeek's In-House Chip Strategy Signals Shift to Vertical Integration of Model + Hardware
Short-term View: An Inevitable Choice Driven by Cost Pressure and Supply Chain Security
DeepSeek's model capabilities have been validated within the industry, but its business model is highly dependent on computing power rental or procurement. Over the past year, GPU prices have remained volatile, and the supply of high-end chips like the H100 is still constrained by geopolitics. For a model company with rapidly rising valuation, annual computing costs in the billions of dollars are a sword hanging over its profit statement.
From a cost structure perspective, self-developed chips can significantly improve gross margins. Assuming DeepSeek's current inference computing cost accounts for 40% of revenue, if self-developed chips can reduce unit costs by 30%, this translates directly to a net margin improvement of more than 10 percentage points. For a company that needs to prove sustainable profitability to investors, this number is quite attractive.
From a supply chain security perspective, self-developed chips mean holding the core lifeline in your own hands. Relying on third-party GPUs means being subject to price hikes, supply cuts, or performance throttling at any time. Microsoft, Google, and Amazon have long laid out their own chip strategies; the underlying logic is the same: Model as a Service, Chips as Factories. As a pure model company, if DeepSeek doesn't act, it will be building skyscrapers on someone else's foundation in the long run.
But short-term challenges cannot be ignored. Chip design is an asset-heavy, long-cycle, high-failure-rate field. The DeepSeek project has only been launched for a year and is still in the early stages. From tape-out to mass production, and then software stack adaptation, it typically takes 18-24 months. This means that within the next year, DeepSeek will still rely on external computing power, while R&D investment in self-developed chips will drag down short-term profits. Investors need to tolerate a window period of "burning cash to build barriers."
Long-term View: Inference Chips are Key to Differentiated Competition and Ecosystem Lock-in
DeepSeek's choice to specialize in inference rather than training is very smart. Training chips are firmly held by NVIDIA, and the CUDA ecosystem plus large-scale cluster hardware-software synergy is almost impossible to surpass in the short term. But inference chips are a different story.
The market space for inference chips is exploding rapidly. As large models move from training to application deployment, inference computing demand will exceed training. Industry estimates suggest that by 2025, inference computing will account for over 60% of the total, with faster growth rates. Inference chips focus more on low latency, high throughput, and low power consumption, differing from the architectural requirements of training chips. This leaves room for latecomers to overtake via a shortcut.
DeepSeek's advantage lies in "Model-Chip" co-design. It understands its own model architecture, operator distribution, and inference load characteristics best. If self-developed chips can be deeply optimized for DeepSeek's models—for example, customized matrix multiplication units, sparsity computation support, efficient KV cache management—then unit inference costs could be more than 50% lower than general-purpose GPUs. Once this vertical integration works, competitors will find it hard to replicate, because they lack both the model data and the chip design capability.
But the long-term risk lies in ecosystem barriers. If DeepSeek's chips can only run its own models, it becomes a closed system. Customers will worry about lock-in, and developer ecosystems will be difficult to establish. In contrast, NVIDIA's CUDA ecosystem covers almost all AI frameworks, and developers are used to it. DeepSeek must either open its chips to third-party models or prove that its own models' performance far exceeds competitors, making customers willing to sacrifice versatility. This choice will determine its ceiling.
From an investment perspective on valuation logic: If DeepSeek can prove that self-developed chips reduce inference costs to the industry's lowest while maintaining model leadership, its valuation logic will shift from "model company" to "model + chip platform company." Referencing Tesla's valuation premium after developing its own chips and NVIDIA's P/S multiple, DeepSeek's reasonable P/E ratio might rise from 50x to over 80x. But the prerequisite is that the chips reach mass production and expected performance, which requires a 2-3 year verification cycle.
Regarding competitive barriers, the core isn't just chip design itself, but three elements: first, the deep coupling capability between model and chip; second, the usability of the software stack; third, mass production yield and cost control. DeepSeek has advantages on the model side, but the chip team needs to quickly make up for engineering capabilities. I am watching the founder's depth of understanding regarding the chip supply chain and whether they have hired a chip VP with mass production experience.
Final Investment Judgment: I would consider investing after the following conditions are met: First, the chip tape-out is successful and performance reaches 80% of design targets; Second, a joint deployment agreement is reached with at least one cloud vendor; Third, model inference latency on self-developed chips is better than same-generation GPUs. Before this, it is just an imaginative story, not a deployable asset.
Summary: DeepSeek's self-developed inference chips are not icing on the cake, but a necessary step to build a moat for its model empire. Success opens a second growth curve; failure burns hundreds of millions of dollars. But given its judgment on tech trends, I bet it leans closer to the former.
Original Link: https://www.tmtpost.com/8060417.html
