DeepSeek in 2026: Inside the $74B Pivot from Efficiency Lab to Hyperscaler Infrastructure Giant

1-Gigawatt AI Data Center in Inner Mongolia

DeepSeek is planning a massive 1 GW AI data center campus in Ulanqab, Inner Mongolia, about 350 km northwest of Beijing . The facility will combine company-owned infrastructure with leased capacity from neighboring operators . At least part of the capacity is expected to go online by late 2027 or early 2028 . Ulanqab is an established hub under China's "East Data, West Compute" initiative, attracted by cheap electricity and cold weather that reduces cooling costs . This project represents DeepSeek's first hyperscaler-scale campus and a major departure from its previous reliance on borrowed or rented compute clusters .

In-House Inference Chip Development Amid U.S. Export Controls

In July 2026, Reuters reported that DeepSeek has begun developing its own AI inference chip — focused on data-center inference workloads, not training . The goal is to reduce reliance on Nvidia and Huawei as U.S. export controls tighten access to advanced semiconductors . DeepSeek has been in talks for about a year with chip design, foundry, and memory companies to advance the effort . This effort comes after the U.S. held off placing DeepSeek on the Entity List as of June 2026, but the threat of future blacklisting remains active . Notably, a U.S. official claimed in February 2026 that DeepSeek may have trained its latest model on Nvidia's banned Blackwell chips, reportedly located in an Inner Mongolia data center .

"Substantial" API Price Hike After V4-Flash Surge

On August 6, 2026, DeepSeek officially announced a "comprehensive price increase for DeepSeek API services" described as "significant" in magnitude, telling developers to adjust usage accordingly . The price hike is driven by explosive demand: the V4 Flash model alone consumed up to 8 trillion tokens in a single day on the OpenCode platform . Across OpenRouter, V4 Flash processed 7.22 trillion tokens in the week ending July 27 . Current (pre-hike) pricing is $0.14/$0.28 per million input/output tokens for V4 Flash and $0.435/$0.87 for V4 Pro . The company had already introduced 2× surge pricing during two daily Beijing peak windows, a structural shift from flat-rate pricing .

IPO Preparations on Shanghai's STAR Market

DeepSeek is planning a potential mainland initial public offering on Shanghai's STAR Market, and the $74 billion fundraising round is structured as a pre-IPO capital raise . The company is stepping up efforts to finance infrastructure, chips, and research ahead of a possible listing . No specific IPO timeline has been publicly set, but the entire fundraising strategy is being positioned as preparatory for an onshore Chinese IPO .

Strategic Context: The Capital-Intensive Pivot

Taken together, these moves signal a fundamental pivot: DeepSeek is transitioning from a lean, efficiency-obsessed AI lab into a capital-heavy, vertically integrated AI infrastructure company — building its own data centers, designing its own chips, raising tens of billions in equity, and preparing for a public listing, all while grappling with demand that far outstrips its current compute capacity. The API price hike is the clearest signal that its earlier ultra-low-price strategy was unsustainable at current usage volumes. As one analysis put it, DeepSeek "won on price cuts" and is now "crushed by its own demand" .