Over the past seven days, two of China’s leading large-language-model builders—DeepSeek and Zhipu—have quietly let slip that they are running the numbers on building their own AI chips. The industry whispered it first as a rumor; now it has crystallised into a headline that reads less like a breakthrough and more like a cost-benefit spreadsheet.
This is not a hardware announcement. It is a confession.
Both companies have admitted that the “arithmetic problem” of outsourcing inference to NVIDIA GPUs has finally become too expensive to ignore. The margin compression on their API services, combined with the geopolitical uncertainty of advanced chip access, has forced them to ask: is it cheaper to mine our own silicon than to keep renting someone else’s?
The answer, depending on the algorithm, could reshape not just their own futures, but the entire frontier between AI software and the hardware it runs on.
Context: The Riddle of Vertical Integration
To understand why DeepSeek and Zhipu are even entertaining this, we need to step back into the mechanics of AI infrastructure. Both companies operate massive clusters of NVIDIA GPUs—H800s and A100s—either leased from cloud providers or housed in their own data centres. The cost of inference per token for their flagship models (DeepSeek-V3, Zhipu GLM-4) is largely dominated by GPU depreciation and electricity. At scale, those costs become a direct drag on gross margins.
For reference, a single NVIDIA H800 GPU used for inference can cost anywhere between $10,000 and $25,000 in total cost of ownership over its lifetime. If a company processes hundreds of billions of tokens per month—as DeepSeek and Zhipu likely do—the math quickly shows that switching to custom ASICs (Application-Specific Integrated Circuits) optimised specifically for their transformer architectures could slash inference costs by 60–80%.
But here’s the rub: a custom ASIC project requires $50 million to $200 million in upfront engineering and tape-out costs, plus another 18–24 months of development before the first chip is validated. That is a high-risk, long-cycle investment in an industry where model architectures change every few months.
Yet the arithmetic problem is not purely financial. It is also about supply chain sovereignty. With US export controls tightening access to advanced semiconductor fabrication, Chinese AI companies face an existential risk: if NVIDIA can no longer ship its newest GPUs to China, where does the next generation of inference compute come from? Building a proprietary chip—even if it underperforms NVIDIA’s latest—insulates the company from that supply shock.
Core: The Dual-Narrative Analysis – Cost Engineering Meets Narrative Engineering
Based on my experience auditing tokenomics and infrastructure stacks for Web3 protocols, I’ve learned that every “savings” story has a hidden cost. DeepSeek and Zhipu are currently telling two narratives simultaneously, and they don’t perfectly align.
Narrative One: The ROI Spreadsheet. Let’s run the back-of-the-envelope numbers. Suppose DeepSeek currently spends $10 million per month on GPU inference rental. A custom chip that cuts that cost by 70% saves $7 million per month, or $84 million per year. If the chip development program costs $120 million over two years, the payback period is roughly 17 months after deployment. That’s attractive—if the chip actually works at the quoted performance and if model architecture doesn’t change so radically that the chip becomes obsolete.
But here’s the catch: China’s domestic foundries (SMIC, for example) can at best achieve 7nm-equivalent processes, while NVIDIA is already shipping 4nm and 3nm chips. That means a custom chip built domestically will lag in raw transistor density. To compensate, DeepSeek or Zhipu would need to implement sophisticated chiplet designs or use more advanced packaging—both of which drive up cost and risk.
Narrative Two: The Strategic Hedge. Beyond the spreadsheet, there is a softer but equally powerful rationale. By publicly exploring self-designed chips, DeepSeek and Zhipu signal to regulators, investors, and customers that they are not entirely dependent on foreign technology. This narrative can unlock government subsidies, preferential loan terms, and even diplomatic goodwill. In a geopolitical environment where “tech sovereignty” is a buzzword, that intangible value can be more valuable than the hardware itself.
However, the two narratives create a tension. If the chip is purely a strategic hedge, the company can tolerate a lower ROI. But if it is purely a cost-saving measure, the payback period must be tight. The market is currently pricing in both possibilities, making it difficult for analysts to assess true risk.
I have seen this pattern before in the blockchain space. Projects often announce “Layer-1 scaling solutions” that are really just repackaged Ethereum clients—the announcement itself provides more value than the actual technical improvement. The same could be true here: the announcement of chip exploration may boost valuation more than the chip itself ever will.
Contrarian: The Pragmatism Test – Why Self-Made Silicon Might Be a Trap
Every evangelist wants to believe that vertical integration is the purest form of decentralisation. But in hardware, “do it yourself” can quickly become “did it all wrong.”
Consider the historical data. In the AI accelerator market, only two non-NVIDIA players have succeeded at meaningful scale: AMD (with its MI-series) and Google (with its TPU). Both had existing hardware expertise and decades of semiconductor design experience. DeepSeek and Zhipu are software-first companies. Their core competency is in training algorithms, not in digital logic design or chip verification.
I once worked with a DeFi protocol that decided to build its own custom blockchain to “escape Ethereum gas fees.” They spent 18 months and $30 million before realising that the developer ecosystem they’d abandoned was worth far more than the cost savings. The same trap awaits any AI company that underestimates the value of NVIDIA’s software stack—CUDA, cuDNN, TensorRT, and the entire PyTorch ecosystem. Those are not easily replicated, even with a competent chip team.
Moreover, the time horizon works against them. Even if DeepSeek and Zhipu tape out a competitive chip in 2026, by then NVIDIA will have released its Vera Rubin or Rubin Next architecture, potentially widening the performance gap again. The bear case is that self-designed chips are a multi-billion-dollar treadmill that never ends.
A more pragmatic path would be hybrid: partner with an existing Chinese chip company (like Cambricon or Huawei Ascend) to co-design inference accelerators, sharing the risk and the software stack while retaining some customisation. But that would dilute the narrative of “self-developed.” And narrative, as every community founder knows, is often the most important asset.
Takeaway: The Vision Forward – Arithmetic Is Only Half the Equation
The arithmetic problem of chip development is solvable on paper. But the real test lies in execution, culture, and timing.
DeepSeek and Zhipu are making a bet that the future of AI belongs to those who control their own compute. That bet mirrors the ethos of early Bitcoin miners who built their own ASICs to secure the network. But Bitcoin’s ASICs were simple—single-purpose SHA-256 engines. An AI inference chip must be general enough to handle rapidly evolving transformer variants, multimodal models, and potentially completely new architectures.
“Trust is the only protocol that matters.” If these companies can execute the chip program without diluting their core mission—building safe, powerful, accessible AI—then the trust they earn from the market will compound far beyond the cost savings. But if the chip distracts them from their model quality, the arithmetic will turn negative.
“Code is law, but people are the context.” The teams behind these chips will determine success or failure. Neither company has publicly disclosed its chip leadership. That is a signal worth watching.
“Community over coin, always.” In the end, the most valuable resource DeepSeek and Zhipu have is their user community and developer ecosystem. A chip that fragments that community—by making it harder to run models on existing infrastructure—would be a losing bet no matter how cheap the tokens.
The arithmetic is still in progress. But one thing is clear: the era of passive dependence on NVIDIA is ending. The next chapter will be written by those who dare to build their own foundations—and by those who know when to stop building and start buying.
What will the spreadsheet ultimately say? And more importantly, what will the community say? That is the equation I am watching.