One year after DeepSeek’s R1 reasoning model triggered the largest single-day stock wipeout in U.S. history, Wall Street appears to have dismissed the Chinese AI upstart. On Jan. 27, 2025, the S&P 500 shed over $750 billion, with Nvidia alone losing $590 billion in market value, according to Dow Jones Market Data. Yet U.S. hyperscalers are projected to spend more than $600 billion on AI infrastructure this year, seemingly unperturbed by the Hangzhou-based lab’s cost efficiencies.
DeepSeek, owned by quantitative hedge fund High-Flyer Quant, challenged the narrative that frontier AI demands exorbitant compute resources. Its R1 model matched elite U.S. large language models at a fraction of the cost—reportedly $294,000 to train, per Reuters—using clusters like Fire-Flyer built from 10,000 Nvidia A100 GPUs acquired before U.S. export curbs. Founder Liang Wenfeng, who holds a 55% stake in High-Flyer with 99% voting rights, funded the lab through the firm’s R&D budget, as detailed in Reuters.
High-Flyer’s pivot from AI-driven trading—managing over 70 billion yuan ($10 billion) with 56.6% returns in 2025, ranking second among large Chinese quants per PaiPaiWang—to foundational models underscores China’s strategy of leveraging financial muscle for tech disruption. Liang, a Zhejiang University engineering graduate, co-founded the fund in 2015 and spun off DeepSeek in 2023 after venture capitalists balked, according to corporate records cited in Fortune.
High-Flyer’s Hidden Fuel
High-Flyer’s 2025 performance, with strategies by co-founder Xu Jin averaging 58.6% gains and CEO Simon Lu’s products at 56%, bolstered DeepSeek’s war chest amid U.S. chip sanctions, as reported by Bloomberg. This financial resilience enabled innovations like the V3 model, which topped open-source rankings on Hugging Face and powered 40% of advanced Chinese AI tasks in programming and design, per TrendForce.
DeepSeek’s open-source push propelled Chinese models to 15% global share by November 2025, surpassing Google and OpenAI in some benchmarks. Japanese firms built six of their top 10 models on DeepSeek and Alibaba’s Qwen, noted Nikkei via TrendForce. Microsoft president Brad Smith highlighted its rapid adoption in Africa, where free access and local data sovereignty outpaced Western platforms, according to Financial Times.
Yet bans in Italy, Denmark, Czech Republic, Australia, and scrutiny from Germany, India, and the U.S. reflect security fears. Italy’s antitrust closed a probe after DeepSeek committed to hallucination warnings, per Reuters. U.S. lawmakers urged adding DeepSeek to the Pentagon’s Chinese military aid list.
Efficiency Breakthroughs Reshape Compute Race
DeepSeek kicked off 2026 with a paper on Manifold-Constrained Hyper-Connections (mHC), co-authored by Liang, proposing stable scaling of massive models without instability. Omdia analyst Lian Jye Su called it a ‘breakthrough’ rippling across rivals, while Counterpoint’s Wei Sun deemed it ‘striking,’ as covered by Business Insider. The arXiv upload signals R2 preparations, delayed from mid-2025 due to chip shortages and Liang’s performance demands, per The Information.
Internal tests suggest upcoming V4, due mid-February, outperforms Anthropic’s Claude and OpenAI’s GPT in coding, handling long prompts for complex software, according to Reuters citing The Information. V3.2 and V3.2-Speciale achieved gold in IMO 2025 math, 96% AIME, rivaling GPT-5 via Mixture-of-Experts and sparse attention.
Bloomberg noted mHC preserves efficiency at scale, challenging U.S. hyperscalers’ capex frenzy. Constellation Research views it as ending the ‘muscle head era’ of brute-force compute, shifting to engineering prowess.
Wall Street’s Cost Blind Spot
Post-R1 panic, Nvidia rebounded; Jim Cramer noted AI stocks breaking out as DeepSeek fears receded, per CNBC. Morningstar’s Brian Colello foresaw accelerated 2026 spending. Yet MarketWatch‘s Christine Ji argues hyperscalers overlook persistent Chinese efficiencies amid fragile economies.
Gartner analyst Haritha Khandabattu called R1 a ‘repricing’ of cost curves and China’s edge; later V3/R1 updates didn’t jolt markets like new architectures might. Hugging Face dubbed it the ‘DeepSeek Moment,’ catalyzing U.S. open-source like ATOM project.
CNBC questioned why no repeat frenzy: incremental releases versus R1’s shock. South China Morning Post sees mHC signaling bigger models cheaper, with Liang’s hands-on research foreshadowing flagships.
Global Ripples and Investor Reckoning
DeepSeek’s traction in developing nations narrows AI gaps, per Microsoft’s report via Euronews. Premier Li Qiang met DeepSeek’s Liang a year ago, now MiniMax CEO, signaling Beijing’s global AI push, per SCMP.
High-Flyer-inspired quants like Baiont and Wizard ramp AI; mutual funds deploy DeepSeek, per Reuters. Shenzhen subsidizes compute at 4.5 billion yuan.
As V4 looms, Wall Street risks underestimating DeepSeek’s trajectory—from hedge fund side project to open-source leader. Liang’s vision: China leading, not following, per Economic Times.
WebProNews is an iEntry Publication