Susana SchwartzWhy DeepSeek's $7.4B funding matters in custom-ASICs race
Earlier today, I read with interest TechInsights “The Great Chinese AI Decoupling: The AI Tigers,” which talks of emerging hardware-pipeline archetypes forming under the weight of U.S. export controls and allies’ global sanctions and trade restrictions.
As Chinese companies race to become more self-reliant through domestic fabrication and new supply chains, what will the impact be to the giants of the U.S. market? For example, a formidable “David” to Nvidia’s Goliath is DeepSeek, whose recently announced plans for an inference-specific custom chip could fundamentally change the economic landscape of AI infrastructure. Though much smaller than Nvidia, DeepSeek could shrink the addressable market for Nvidia.
With its astounding $7.4 billion (50 billion yuan) infusion of outside capital, it indicates that Chinese AI labs are becoming heavily capitalized and increasingly capable of scaling infrastructure on their own. DeepSeek now has a much better chance of not only protecting its open-source ecosystem with software-side efficiency, but also potentially weaning itself from both Nvidia and Chinese provider Huawei. This move toward custom ASICs could also open the door for new DeepSeek customers in not only mainland China, but also SE Asia, Latin America, Africa, the Middle East.
Success will depends on whether the company’s hardware-software co-design brings the extreme cost-per-token efficiencies some are expecting it will, and whether it rivals what OpenAI’s Jalapeño offers, or what Google TPUs or Amazon Trainium/Inferentia bring to the table.
The DeepSeek pivot and the fact Huawei is scaling up its Ascend chips to compensate for losses of Nvidia Hopper and Blackwell GPUs, plus news of DUV lithography machines are emerging in China, could unleash a movement toward alternatives for governments, enterprises, and startups that don’t want to rely primarily on U.S. semiconductor monopolies.
Of course, Nvidia still retains an impressive 75% to 80% of the global AI accelerator market (by revenue), but if DeepSeek’s highly optimized software and customized ASICs match Western model capabilities – and at a fraction of the cost – then it’s possible more developers may be enticed to bypass top-tier GPUs in favor of Chinese ASICs, specialized accelerators, and customized in-house silicon (especially in instances where high-volume batch processing benefits from low-cost mixture-of-experts architecture).
RCRTech will continue to follow the shift toward high-volume inference scaling and increasingly commoditized intelligence as more and more AI labs move toward proprietary hardware to optimize cost-per-token economics.
Susana Schwartz
Technology Editor
RCRTech
AI Infrastructure Top Stories
Telco AI, ISAC arms race: Data center connectivity, sovereign AI, and defense networks represent huge opportunities for telco, but companies like Verizon and AT&T, and SK Telecom understand it’s a “long roll of the dice” in terms of a pay-day.
Physical AI’s data sovereignty gap: The network is becoming a potential compliance risk vector, according to FloLIVE president Curtis Govan’s Reader’s Forum on cross-border vehicles, robots, and other physical AI systems.
$143.4B spent on enterprise cloud: Synergy Research Group found that enterprise cloud infrastructure services grew 43% YoY, the market’s fastest expansion in eight years, with AWS, Microsoft, and Google in the top-3 spots.
AI Today: What You Need to Know
Executive Order 14318: The Trump administration plans to repurpose large chunks of federal land to host data centers and the power plants needed to run them, including high‑voltage transmission lines and other equipment.
$100B DC repurposes uranium plant: The Trump administration will convert part of a former uranium enrichment facility in western Kentucky to be a massive AI data center campus, which it says will be powered by new energy infrastructure.
Nvidia-OpenAI 10GW DC: Nvidia is in talks to provide a roughly $250 billion financial backstop for OpenAI to lease a 10 GW data center campus being developed in southern Ohio by SoftBank’s energy subsidiary.
EU $€11.5B Gigafactory push: The European Union committed €11.5 billion ($13B USD) to build seven major AI gigafactories – backed by private investments, with hardware supply letters of intent already signed by Nvidia, AMD, and Qualcomm.
Energy Vault purchases Crusoe DCs: Energy Vault is purchasing 10 Crusoe Spark modular data centers, which Crusoe will lease-back from Energy Vault as powered shell infrastructure for training, fine-tuning, and inference customers.
MediaTek’s $5 Billion AI Pivot: MediaTek approved a $5 billion financing budget to shift focus away from smartphones toward custom ASICS for data centers, with AI accelerator ASIC production expected in Q4, and 2nd-gen chips by early 2028.
