DeepSeek Open-Sources Its Core AI Infrastructure: The Distribution Logic Behind the Code

Open-Sourcing Code Is Also Open-Sourcing Value

On September 30th, DeepSeek^Ad released another wave of open-source contributions — not models this time, but operators, a programming language, compute libraries, and communication libraries: the foundational infrastructure that AI runs on. The official description states these components correspond one-to-one with the NVIDIA ecosystem.

In plain terms: AI models and applications that previously only ran on NVIDIA chips can now run on domestic Chinese chips without rewriting any code. What looks like a purely technical move actually touches on one of the most fundamental questions of the AI era: when technology is no longer scarce, where does the core value actually flow?

By releasing these low-level tools to the public, DeepSeek effectively transferred that value to the broader ecosystem. From this angle, open-sourcing is itself a distribution mechanism — it disperses value from a handful of companies to an entire ecosystem.

---

The components DeepSeek open-sourced involve several key technical concepts. Operators (Kernels) are the smallest execution units in AI model computation — matrix multiplication, convolution, attention mechanisms, and so on — and their low-level implementations differ completely across chip vendors. Compute libraries package commonly used operators; NVIDIA's CUDA ecosystem includes highly optimized libraries like cuDNN and cuBLAS, which form a core part of its competitive moat. Communication libraries (such as NVIDIA's NCCL) handle data synchronization across multi-GPU and multi-node training, directly affecting large-scale training efficiency. What DeepSeek released is essentially an interface layer that is "isomorphic" to NVIDIA's CUDA for domestic chips — developers don't need to modify upper-level code; the underlying drivers adapt automatically. This breaks the lock-in effect NVIDIA has long built through its software ecosystem (rather than raw hardware performance alone): even when domestic chips met performance benchmarks, the absence of a software ecosystem made migration prohibitively expensive.

The Shift in Expectations Hidden in Consumer Data

To understand the deeper significance of this open-source release, we need to read the current economic signals correctly. According to public data, retail sales of communication devices grew 27.3% year-on-year in August, while automotive sales fell 18.5% — a gap of roughly 46 percentage points. One is exploding; the other is in sustained decline.

Another set of figures is equally telling: total retail sales of consumer goods for the first eight months grew only 1.1%, yet household deposit balances grew 9.7% year-on-year. This suggests the issue isn't overall spending power — it's a structural shift in consumption patterns. Essential goods like food, grain, and tobacco continued growing, while big-ticket items such as cars, furniture, and home renovation — especially anything downstream of real estate — contracted broadly.

Income data reinforces this anxiety. The median growth of per capita disposable income lagged slightly behind the mean; net property income (interest, rent, dividends) grew only 1.6%; and migrant workers' monthly income rose just 2.3%. Layered on top of demographic trends — 23% of the population is over 60, 15.9% over 65 — all these figures point to the same conclusion: people feel uncertain about the future, and their first instinct is to stop making large purchases.

The Biggest Structural Risk of the AI Era: Unsold Goods

The anxiety behind these numbers isn't just about money — it's about whether there will still be jobs in the future. And that concern is deeply intertwined with expectations around AI development.

AI enhances supply-side capacity. It can produce more things at lower cost. But demand depends on distribution — people need income before they can afford what's being produced. If technology races ahead while distribution fails to keep up, you end up in an awkward situation: goods keep getting better and cheaper, but fewer and fewer people can afford to buy them.

This is the biggest structural risk of the AI era: not the inability to produce, but the inability to sell. The imbalance between supply and demand is the real trap that technological transformation should be wary of.

Two Paths: Value Concentration vs. Value Diffusion

On September 29th, the United States rebranded "artificial intelligence" as "superintelligence" at a White House luncheon. This change means far more than a definitional dispute — it raises the bar for "whether AI should be regulated": after all, how do you regulate something smarter than yourself?

Both countries are developing AI, but they're taking completely different approaches — and this determines how the resulting value will ultimately be distributed.

The American model is private-sector-led: technology is closed-source, value is concentrated, and the returns primarily flow to capital. China's model is state-capital-anchored: the National Integrated Circuit Industry Investment Fund (Phase III) alone registered with approximately 344 billion yuan in capital and has taken stakes in multiple private AI companies; state-owned enterprises have opened up domestic computing resources, creating a channel for technology to spill over into public benefit.

Closed-source technology concentrates value in a few companies; open-source technology distributes it across an entire ecosystem. DeepSeek's decision to release its foundational tools is a microcosm of China's distribution logic — using open-sourcing to achieve a social redistribution of technological value.

Jobs That Won't Be Replaced, and the "Employer of Last Resort"

Following this logic to its conclusion, the future employment landscape begins to take shape. As AI and robotics become capable of handling nearly all material goods production, some jobs remain difficult to automate in the near term: childcare, education, elderly care, and skilled installation and repair work. How many people these roles require and what types they involve can actually be estimated.

One way to absorb this labor is through a state-level labor coordination system — the state becoming the "employer of last resort": those willing to work more earn more; those who don't work receive a basic safety net; and social insurance is all tied to this platform.

This isn't fantasy. There's a classic concept in economics called the "Employer of Last Resort" (ELR), and some countries have already implemented related policies. China may have an easier path to making this work, since state-owned platforms and social security systems already exist as ready-made infrastructure.

---

The "Employer of Last Resort" (ELR) is one of the core policy proposals of Modern Monetary Theory (MMT), systematically developed by economists Hyman Minsky and Pavlina Tcherneva, among others. The core logic is that a sovereign currency-issuing nation can never run out of money for employment purposes — the government can serve as a guaranteed employer of last resort, offering public jobs at a base wage to eliminate involuntary unemployment. Unlike Universal Basic Income (UBI), ELR emphasizes "work for pay" rather than direct transfers, and is considered more effective at maintaining labor skills and social participation. Argentina's Plan Jefes and India's MGNREGA are real-world examples of varying degrees of implementation. With widespread AI displacement of routine labor on the horizon, the ELR discussion is experiencing a revival in economics circles — its feasibility depends heavily on a government's administrative capacity and fiscal willingness.

Two Concrete Indicators Worth Watching

What will ultimately determine people's wellbeing isn't what AI produces, but how effectively this coordination system gets built.

The article offers two publicly verifiable indicators — more reliable than any forecast:

First, watch for the large-scale emergence of state-backed human resources coordination platforms. If and when these are established, there will be public filings and official registrations.

Second, track the scale of "service procurement" and "employment subsidy" line items in fiscal expenditures. This is the most direct accounting of the state paying for labor. If this spending grows significantly, it signals the path is actually being pursued.

Back to the Original Question

If technology is no longer scarce, what actually becomes valuable? Perhaps not computing power, not data, and not even money — but a healthy body.

The author's conclusion carries a somewhat idealistic tone: let those who understand technology do the technical work; let ordinary people focus on living well and staying healthy — without anxiety, without chasing trends. As medical technology potentially breaks through, healthy life expectancy may meaningfully extend. At that point, we may find that what's truly scarce is no longer wealth, but health itself.

A note of transparency: everything above is observation and inference based on public data, and does not constitute any investment advice. DeepSeek open-sourcing its foundational operator library is an established technical fact, but the conclusions about distribution logic and labor coordination reflect the author's personal views — readers are encouraged to engage critically.