On April 23, DeepSeek pushed an announcement to Hugging Face that most people scrolled past. DeepSeek V4 — a 1.6 trillion parameter open-source model — is launching in the next few weeks. Specs: 81% SWE-bench, $0.30 per million tokens, one-million-token context window, hybrid sparse attention.
That part is impressive but not new. What should stop every AI investor cold is a different sentence buried in the release notes.
V4 runs with no Nvidia CUDA dependency. It was built end-to-end on Huawei Ascend chips and the Cambricon stack. Chinese chip, Chinese model, Chinese software glue — no American silicon anywhere in the pipeline.
That is the first fully frontier-class model built on a non-Nvidia stack. And most of your AI-heavy portfolio is not priced for this yet.
The Moat That Everyone Assumed Was Permanent
For the last three years, the entire AI bull thesis has been powered by one idea: Nvidia owns the picks-and-shovels of the most important industry of the decade, and nobody can catch up.
CUDA was the real moat, not the chips. Every researcher learned CUDA in grad school. Every framework optimized for CUDA. Switching costs were enormous. Even when AMD and Intel shipped comparable hardware, nobody wanted to rebuild their entire workflow around unproven tooling.
That’s why Nvidia’s price-to-earnings ratio stayed elevated while the rest of the market cycled. Investors weren’t paying for chips. They were paying for lock-in.
DeepSeek V4 is the first credible signal that the lock-in is breakable. Not by an American competitor. By a Chinese stack that the US government tried to choke off with export controls.
The implication: Nvidia’s terminal margins probably aren’t 70%. They’re whatever margin the market will bear once a real alternative exists — and that number starts with a 4 or a 5, not a 7.
Why This Didn’t Happen Two Years Ago
A lot of smart people predicted Chinese chip independence. Most were wrong on timing by three to five years. The reason the timing finally broke this month:
One, the export controls accelerated what they were supposed to prevent. Huawei and Cambricon got massive state funding specifically because Nvidia became unavailable. Scarcity created a forcing function that market competition never would have.
Two, DeepSeek’s team figured out how to do more with less. Their previous models already beat US labs on cost-per-performance. V4 is the natural end-state — take the efficiency obsession, pair it with domestic silicon, ship it open source.
Three, open source became the weapon. A closed model on Huawei chips is a curiosity. An open model on Huawei chips is a blueprint. Anyone in the world can now build a frontier AI stack without buying a single H100.
The wealth implication is not just about Nvidia. It’s about what happens when the entire cost curve of AI compute gets redrawn.
What to Actually Do With This
If you own Nvidia: don’t panic sell. The data center buildout is still a multi-year story. But trim the position if it’s more than 10% of your portfolio. The asymmetric risk has flipped — upside is capped by competition, downside is real for the first time since 2022.
If you own the “Mag 7” passively through a cap-weighted index: you already have a 30%+ concentration in AI infrastructure plays. That’s a bet, whether you meant to make it or not. Understand the exposure.
If you’re an AI operator, not just an investor: the opportunity is enormous. Cheaper inference means cheaper apps means more products become economically viable. The winners of the next AI cycle will look more like Shopify and less like Nvidia — companies that take commoditized infrastructure and turn it into recurring revenue.
Look at Intel’s Q1 earnings this week. Up 16% on data center CPU demand for AI agent workloads. That’s a company that everyone wrote off in 2024. The narrative around AI hardware winners is quietly reshuffling in real time.
The Bigger Picture — and Why This Is Good News
This is a rare case where the same event is good for humanity and complicated for your portfolio.
Cheaper AI inference at a trillion-parameter scale means tutoring in every language for free. Medical diagnostics running on a phone. Legal services that no longer cost $400 an hour. Autonomous agents handling the paperwork that eats 30% of small business owners’ weeks.
These are wealth-creating technologies in the broadest sense — they expand what a human with limited capital can accomplish. The person who can’t afford an accountant gets one. The kid in a village gets a tutor. The founder gets a 10-person team at the cost of a laptop.
Wealth built by companies that unlock this is wealth worth owning. Wealth extracted from maintaining artificial scarcity around AI compute probably isn’t. The market is going to spend the next three years figuring out which camp each company falls into.
Your job as an investor is to be roughly right before the market gets there.
The Wealtharian Take
Three quick rules for navigating this:
One, never be 100% concentrated in any thesis, even a “sure thing” like Nvidia in 2024. Moats get broken. 2025 was the last year you could hold AI infrastructure as your only bet and sleep well.
Two, rotate toward the application layer. Companies that turn cheap AI into paying customers will compound for a decade. That’s where the next generation of wealth creators will come from — and a lot of them don’t exist yet.
Three, pay attention to Chinese tech the way American investors paid attention to Silicon Valley in 2005. Geopolitics makes direct investment hard, but understanding what Chinese engineers are shipping tells you where the cost curve goes next. That knowledge is free. Most of your peers aren’t doing the work.
The moat cracked this week. The next decade of AI wealth gets created by the people who noticed.
Want to track your own path to financial independence as AI reshuffles the wealth map? The Wealtharian Wealth Tracker lets you monitor your net worth, FU money progress, and investment milestones in one place. Try it free →