The U.S.–China AI competition has reached a consequential inflection: China is beginning to export not only competitive AI models, but also the software infrastructure that makes those models economical to operate.
Meanwhile, new American financial disclosures expose the scale of capital commitments required to sustain frontier leadership. Three developments on October 20–21 materially strengthen The Velocity Edge's central thesis.
The Three Developments
October 20–21
-
On October 20, DeepSeek released DeepInfer — the inference framework behind its production services — under the Apache 2.0 license, providing optimized computing kernels, distributed execution, speculative decoding and scheduling techniques. Crucially, it supports both Nvidia Blackwell SM100 GPUs and Huawei Ascend 950 accelerators.
Limitations matter: the release supports only DeepSeek-V4.1-Flash, offers offline inference rather than a full serving service, excludes certain caching and load-balancing components, and its published benchmarks use Nvidia GPUs — comparable independent Huawei results are not yet available. Still, DeepSeek is attempting to commoditize a second layer of the AI value chain: the software required to operate intelligence at scale. signal: After model weights, inference optimization is the next layer China is making accessible — strengthening diffusion without establishing frontier parity.
-
On October 20, Reuters reported details of Anthropic's draft IPO prospectus. The $80.7 billion headline loss includes roughly $72 billion in noncash charges and must not be confused with operating cash consumption. The more consequential signal is the enormous contractual exposure to future computing capacity — $518 billion in minimum commitments over about ten years.
The disclosures sharpen the central investment question: can rapidly expanding revenues support the growing commitments required to maintain model leadership? The American frontier is increasingly a long-duration financial obligation as well as a technological advantage. signal: Compute utilization, contractual flexibility, pricing power and revenue durability now define the frontier investment case.
-
On October 21, Reuters reported that Samsung and SK Hynix are upgrading their Chinese NAND memory plants as domestic manufacturer YMTC intensifies competition. Per Omdia figures cited by Reuters, YMTC captured 38.5% of China's NAND market in Q2, up from 28% in 2025; Samsung's share fell to 33.5%. Upgrades remain subject to U.S. equipment-export restrictions.
This is a different market from the advanced accelerators that determine frontier training capacity. But it is tangible evidence that China's substitution strategy is generating competitive pressure — and a second-order effect: export controls may constrain access while simultaneously reinforcing incentives for domestic substitution. signal: Whether those incentives ultimately produce superior manufacturing economics remains an open question.
The Hidden Pattern: Two Economies, Two Regimes
Strategic emphasis, not exclusive capability
United States — Frontier Capital Regime
Technological leadership → infrastructure commitments → premium capability → monetization requirements.
China — Diffusion Economics Regime
Domestic hardware alternatives → open inference infrastructure → lower deployment barriers → wider industrial adoption.
The Second-Order Effect: The Next Market Rotation
Consumption vs. profit
These models are not mutually exclusive — American firms build efficient inference, Chinese firms invest heavily in infrastructure. The emerging hypothesis is that frontier technology could remain concentrated while the software economics of deploying intelligence become progressively more competitive, creating substantial differences in earnings growth across the value chain.
If independent operators can reproduce DeepSeek's optimizations, competition among inference providers may intensify. Lower prices could stimulate more workloads — increasing aggregate demand for computing while compressing margins for less efficient operators.
That creates a potentially divergent outcome: advanced accelerator manufacturers may continue benefiting from infrastructure demand; cloud operators may face greater pressure to demonstrate utilization; model providers could see further pricing competition; and industrial enterprises may capture an increasing share of AI's economic benefits through productivity. The distinction is between growth in intelligence consumption and growth in the profits earned by producing it — the two should no longer be assumed to move together.
What Investors Should Watch Next
The decisive tests
The near-term test is independent benchmarking of DeepInfer on Nvidia and Huawei hardware — measured through cost per completed workload, throughput, latency and energy consumption.
For American frontier companies, attention should turn to the relationship between contractual obligations, recurring revenues and free-cash-flow generation. For Chinese semiconductor suppliers, market-share gains must be evaluated alongside manufacturing economics and margin sustainability.
The next material inflection would occur if Chinese inference optimization demonstrated competitive performance across domestic accelerators at a materially lower total deployment cost — moving the contest beyond cheaper models toward a substantial challenge to American infrastructure economics.
America still leads the AI frontier. China is challenging the cost of accessing and deploying it. The next market leaders will be determined not by who produces the most intelligence, but by who captures the economic value of its diffusion.