DeepSeek's latest model cuts expose the hidden assumption holding up the entire US AI trade — not that AI will fail, but that it will succeed too cheaply.
Today, Nvidia closed up another 4% at a fresh all-time high. The Nasdaq edged fractionally higher. The narrative held: artificial intelligence is the infrastructure supercycle of a generation, and Nvidia is its irreplaceable toll booth.
Somewhere in the same trading session, a quieter number slipped past without fanfare. DeepSeek — the Chinese laboratory that shook Western markets in January 2025 — adjusted pricing across its V4 model line. The headline was small. The implication was not.
The Pricing Details
V4-Pro now lists at $1.74 per million input tokens and $3.48 per million output, with a limited 75% developer discount running through May 5, 2026. That brings effective Pro pricing to $0.435 in and $0.87 out. V4-Flash, the lighter sibling, sits at $0.14 and $0.28.
The detail worth pausing over is in the cache-hit column. DeepSeek has cut input cache-hit prices to a tenth of launch levels across the entire V4 line. For real enterprise deployments — agents that repeat the same system prompt tens of thousands of times a day, coding assistants holding large codebases in context, search tools recycling the same document corpus — cache hits are not an edge case. They are the majority of spend. At $0.003625 per million cached input tokens on the promotional Pro rate, long-context inference becomes commercially viable at scale in a way it simply was not twelve months ago.
The one-million-token context window, until now impressive but economically marginal, has become deployable.
The Signal That Matters More Than the Price
Buried in DeepSeek's own documentation is a footnote Western financial media has almost entirely missed.
V4-Pro throughput is currently constrained by compute scarcity. Once Huawei's Ascend 950 SuperPod supernodes ship in volume in the second half of 2026, DeepSeek expects to cut V4-Pro pricing significantly again.
That is the signal. Today's price is not the floor; it reflects hardware scarcity, not long-run cost structure. Reuters has confirmed that V4 was architected specifically around Huawei Ascend silicon, and that the Ascend supernode platform will fully support the model. What customers pay today is what DeepSeek charges while supply is constrained. What they pay in late 2026, if Ascend 950 production scales as planned, will be less.
The real bear case is that the US AI trade collapses precisely because AI succeeded — and succeeded too cheaply.
Nvidia is priced as more than a great company. It is priced as a company whose economics are structurally protected: high demand, high margin, high pricing power, limited substitution. The embedded assumption is that AI compute stays scarce and premium. The Chinese counter-thesis attacks that assumption head-on — not by beating Nvidia on benchmarks, but by showing that high-quality inference can be delivered as an integrated system: domestic model architecture, domestic hardware, aggressive software optimization, aggressive pricing.
For the enterprise buyer deploying agents, internal knowledge tools, customer service automation, or coding support, the question was never whether the silicon came from Santa Clara. The question was whether scale was affordable. DeepSeek is answering yes.
China does not need to ship a single Huawei chip into a Western data center for this thesis to work. It only needs to prove the cost curve can be broken. Google moves more inference onto TPUs. Amazon leans harder into Trainium. Western hyperscalers pour capital into distillation, quantization, and model routing. The idea travels even when the hardware does not.
OpenAI and Anthropic are running out of time. Nvidia becomes a risky bet.
If the inference cost floor collapses in the second half of 2026 — as DeepSeek's own roadmap implies — the narrative scaffolding holding up current private valuations starts to corrode. Both companies should be accelerating toward public markets now, before Chinese open-weight models running on cheap domestic silicon anchor the global price of intelligence at a level that makes monopoly-era revenue projections indefensible.
The US AI bubble does not require AI to fail. It requires only that the market concede one uncomfortable truth: the price of intelligence is falling, and the architecture of that fall is being built in China.
That is not a technology story. It is a valuation story. And valuation stories end.
not investment adivce