China’s national semiconductor fund put money into an LLM company for the first time ever. That’s the headline from DeepSeek’s May 6, 2026 funding announcement. $3-4 billion raised at $45-50 billion valuation, Big Fund leading, Tencent and Alibaba in the round. The valuation was $20 billion a few weeks earlier.
V4 dropped on April 24 and everything moved.
If you’ve been watching DeepSeek from a distance, stop.
This is the part where it becomes unavoidable.
State Money Changes the Equation
Liang Wenfeng owned 90% of DeepSeek before this round.
Let that sink in. A quant hedge fund called High-Flyer funded a frontier AI lab from trading profits, no VC, no strategic checks, no board decks. That’s not supposed to work. It worked.
Now they’re taking China’s Big Fund money. 344 billion yuan. First LLM investment ever for that fund. Why? Retention. ByteDance, MiniMax, Alibaba have been running salary campaigns against DeepSeek’s research team. Capital goes to keeping people, not building demos.
The open-source AI world just got state-backed. That’s new.
The Huawei Thing Is Bigger Than It Looks
DeepSeek V4 runs on Huawei Ascend chips.
Natively. Not a port, not a compatibility shim, native on Ascend 910C silicon.
US export controls were supposed to starve Chinese AI labs of competitive compute. Nvidia can’t ship the best chips to China. Huawei fills the gap with something good enough.
And now DeepSeek’s models run on it, optimize for it. And the Big Fund pays to keep the whole thing moving.
Jensen Huang said on a podcast: “If DeepSeek comes out first on Huawei, that would be a terrible outcome for our country.” He wasn’t being dramatic.
He was being accurate.
The export control strategy assumed Chinese labs would stay behind. They didn’t.
Inference is cheap when you control your own stack.
The Discount Is a Trap (Maybe)
DeepSeek V4 Pro is 75% off through May 31.
Their API has been 5-10x cheaper than OpenAI or Anthropic for over a year. I don’t love using the word “trap” but here’s the pattern: low entry price, get developers locked into workflows, raise later.
OpenAI started free, got people dependent, then introduced token caps. Copilot killed flat-rate billing June 1. Every US vendor is moving toward pay-per-use because that’s where the margin is.
DeepSeek’s discount might be the same play.
Or it might just be genuine price competition. I genuinely don’t know which. What I know is that the moment your workflow requires their API, you’ve got a vendor dependency. That’s true for DeepSeek and it’s true for OpenAI. At least DeepSeek’s weights are downloadable.
Run the model locally if you can. If you can’t, use the API but keep your prompts portable. The bill matters less than the lock-in.
What Actually Changes for You
Nothing today. That’s the honest answer. DeepSeek V4 was available before the funding. The API discount was available before the funding. The open weights were available before the funding.
What changes is the timeline. $4 billion in state-linked capital means DeepSeek isn’t going away. They’re not going to run out of GPU time. They’re not going to have a funding crunch in 18 months. The competitor that’s been keeping US API prices honest is now the most stable open-source lab on the planet.
Test your current workload against DeepSeek V4 Pro.
One task. Pick something you’re paying $10+ a month for through OpenAI or Anthropic. Run it through DeepSeek’s API at the current discount. Compare the output. If it’s within 90% of quality, you’re spending too much.
Download R1 or V4 weights if you’ve got GPU available. A 4090 handles 7B models fine. 70B needs more but inference is getting cheaper by the quarter.
The AI bill you’re paying exists partly since you haven’t found the alternative yet. The alternative just got better funded.
Sources
– TechCrunch — DeepSeek funding details
– Economic Times. Big Fund lead, Tencent participation
– Open Source For You. Open-source angle, Huawei chip optimization
– Seoul Economic Daily — Jensen Huang quote, Big Fund background
– Hacker News. Developer discussion on DeepSeek pricing, open-source competition
