The Hermit Opens the Door: Liang Wenfeng's Two Bets

In April 2026, DeepSeek did two things at once. On April 24, the Hangzhou lab released V4, a 1.6-trillion-parameter base model, with its official hardware verification checklist naming Huawei's Ascend 950PR alongside Nvidia GPUs. Days earlier, per TMTPost, the company had quietly opened its first external fundraise — targeting more than US$300 million at a valuation the Chinese press put at US$20 billion to US$30 billion, roughly double the US$10 billion mark the lab was last associated with. Tencent and Alibaba were reported to be in talks to participate.

Read separately, these are a model release and a funding round. Read together, they are the same bet. Liang Wenfeng is raising outside capital for the first time in DeepSeek's history to fund the exact compute transition V4 already began.

The chip math

The pivot is legible only if the numbers work. Huawei's Ascend 950PR inference chip entered mass production in March 2026, with industry reporting putting per-card performance at roughly 2.87x the Nvidia H20 on inference workloads, and a purchase price of about one-third to one-quarter of an Nvidia H200. Absolute compute lands at roughly half an H200's, which is the trade-off the framing turns on: for training-heavy frontier work, Nvidia still wins. For inference at scale, the price-performance math now favors domestic silicon.

V4 is the proof of concept. The model's Hugging Face release notes and the DeepSeek technical report cited inside the TMTPost coverage describe a CUDA-to-CANN migration — Nvidia's software stack to Huawei's — and report per-token compute cost at 27% of the prior V3.2 generation, with KV cache usage at 10%. Both figures are DeepSeek-self-reported and have not been independently benchmarked in our sources, but they are also why V4 is officially the first frontier model to ship with the Ascend 950PR in its supported-hardware list rather than as an experimental footnote.

The strategic logic is what the Crossing the River analysis of Liang's move frames as a paired, deliberate bet. The lab that refused outside capital for years is taking it now — not because it ran out of money, but because the compute bill required to run a 1.6T-parameter model on domestic silicon at frontier scale is a different category of cost than anything DeepSeek has absorbed before. Alibaba, ByteDance, and Tencent have reportedly ordered hundreds of thousands of Ascend chips between them, per the same TMTPost reporting. DeepSeek is now a buyer in that same market.

What's preserved, what's new

Liang Wenfeng still controls the company. He is reported to hold 84.29% of DeepSeek directly and indirectly, with near-100% voting rights. Outside investors are buying into a structure, not a seat at the table — and the structure preserves the things that made the lab unusual: open technical reports, model-weight releases, a research posture that treated frontier work as something done in the open.

What is genuinely new is a domestic-silicon default. Per TMTPost, V4 is text-only — no native multimodal — at a moment when Google's Gemini 3.1 Pro and OpenAI's GPT-5.4 are both full-modality. The Ascend 950PR's current generation is optimized for inference; training a multimodal frontier model on it is a different problem. Liang is not just buying chips. He is choosing a research path that the chips can carry.

The fundraise also reframes the talent story. Per TMTPost, four senior researchers have left in recent months: Luo Fuli to Xiaomi's MiMo (November 2025), Wang Bingxuan to Tencent's Hunyuan (late 2025 / early 2026), Ruan Chong to Deeproute.ai (January 2026), and Guo Daya reportedly to ByteDance Seed in April 2026 — a move our sources do not independently confirm. ByteDance Seed is reported to have lost around 70 technical staff in the past year and has responded with mass special option grants. The framing in the Chinese press is that Liang's fundraise is partly about creating option pricing for the people who stay — giving DeepSeek the ability to issue equity at a known valuation rather than lose researchers to competitors who can.

The broader market

DeepSeek is not raising into a vacuum. The same TMTPost piece places the round inside a competitive funding cycle: Zhipu and MiniMax have listed in Hong Kong; Moonshot is reported to be raising at US$18 billion and above; OpenAI closed a US$122 billion round in March 2026 at an US$852 billion post-money valuation; Anthropic closed a US$30 billion Series G in February 2026 at US$380 billion post, with annualized revenue reportedly on track from US$9 billion toward US$30 billion. In a market where US frontier labs are repricing on a monthly basis, a Chinese lab going from US$10 billion to US$20–30 billion is a doubling, not a transformation.

OpenRouter's token-volume data, cited in the same coverage, put Chinese AI models at 12.96 trillion tokens processed in the week of March 30 to April 5, 2026 — roughly 4.3x the US volume over the same period. The market Liang is raising into is already the larger one by throughput, even if it remains the smaller one by dollar valuation.

Jensen Huang, at GTC in March 2026, declared the inference era and pointed to a 10,000x surge in AI compute demand over two years. In an April 15, 2026 Bloomberg podcast — not independently verified against a primary transcript in this reporting — he pushed back on the framing of chip export controls as a national-security "enriched uranium" question.

The idealist grows up

The read that Liang is selling out does not survive the technical record. The read that he is being forced off Nvidia by export controls misses the agency in the choice. What he is doing, on the evidence, is solving the single point of failure in his own research mission: the assumption that a frontier lab can stay independent, open, and Chinese-built on someone else's compute.

V4 is the first time that assumption is no longer load-bearing. The fundraise is the capital to scale it. The two moves, on the same week, are the answer Liang reached — and they are deliberately paired.