Huawei's Tao's Law is going viral. What are the signals that programmers should really understand?

This article is from the "Silicon-based Time" series of columns by Jiang Tao, the founder of CSDN. Starting from the "Tao (τ) Law" that He Tingbo of Huawei made popular, it leads to a question that programmers should be more concerned about - when the 80-year foundation of the chip industry can collapse, do the two iron laws of "code + agility" that the software industry relies on still hold? At the end of the article, the real data from CSDN's internal R & D experiment is attached.

In recent days, the news that He Tingbo of Huawei announced the "Tao (τ) Law" at ISCAS 2026 in Shanghai has gone viral. People's Daily reposted it, and the A-share market has speculated on a round of "logic folding concept stocks". The whole network is asking a question: Can Huawei catch up with TSMC by 2031?

Leave this question to industry analysts.

As someone who has been dealing with code for decades, after reading that paper, I have a profound feeling -

Its truly dangerous part doesn't lie in "chips".

It uses the most hardcore semiconductor engineering to give us code writers a wake-up call:

The "premise" that your industry relies on,

may be like the von Neumann architecture,

and has started to fail.

01 What exactly does this paper say?

In one sentence: For sixty years of Moore's Law, what has been scaled is not "space", but "time".

Why are transistors made smaller and smaller? It's not for the sake of smallness itself. It's because - smaller → shorter distance → lower latency → less time.

The exact words in the paper: Spatial scaling is only a tool for compressing time.

He Tingbo's team used six years to mass-produce 381 types of chips, and made this "time-first" logic work. This autumn's new Kirin chips will be the first to use the "logic folding" technology among them, with a 53.5% increase in density and a 41% improvement in energy efficiency.

But for programmers, what they should really chew on is not these numbers. It's the subtext of "the premise fails" in the paper -

It points out one thing: The 80-year iron law of von Neumann's "computation is for computation, storage is for storage, and data is moved back and forth" is loosening.

Why? Because in 1945, the most expensive thing was computation, and moving data didn't matter. Today, the numbers given in the paper are: more than 80% of the energy consumption and 70% of the cost are spent on "moving data". The most expensive thing has changed, and the smartest design back then has become the biggest burden.

It's not that von Neumann was wrong.

It's that the premise he had back then is gone.

All things that "still follow the premise that is no longer valid" will be like the old x86 architecture - still running, but starting to age slowly.

02 What programmers should really pay close attention to is this inference

If even the most hardcore and physics-based semiconductor industry can see its 80-year foundation collapse -

What about the software industry?

In the past twenty years, the software industry has firmly stood on two iron laws:

The first one: Code is the core asset. Whoever has more engineers and a thicker code library is stronger.

The second one: Agility is the development method. Take small steps quickly, have a sprint every two weeks, and align frequently.

These two are cracking at the same time now.

Let's talk about the code part first.

Code itself is becoming the output of Agents - you no longer "write" code, you describe your intention, and the Agent generates it.

Let me tell you a real thing that happened inside CSDN: We have a full Rust project called AtomCode. None of the three core architects can write Rust. But they can write the architecture specifications clearly and define "what kind of output is considered qualified", and leave the rest to AI. Language is no longer a barrier, and architectural judgment is.

The work of programmers is changing from "writing code",

to "defining what kind of code is correct".

The former is a craft, and the latter is a judgment.

Now let's talk about the agility part.

The whole meaning of agility is to let a group of people take small steps quickly and align frequently in an uncertain situation.

But when "a group of people" becomes "one person + 100 Agents", the sprint review every two weeks is as slow as using a telegraph for an Agent team that can make several requests per second and never stops working 24 hours a day.

There is a real case that has been reposted by many people: Developer Peter Steinberger (known as the "father of the lobster") commanded 100 Codex Agents to run simultaneously by himself. He burned 1.3 million dollars' worth of tokens in 30 days and ran 603 billion tokens. One person did the work of an engineering team.

In the past, the software industry competed on "who can write code".

Now it competes on "who can dispatch an army of Agents that never gets tired, interrupted, or lazy".

03 A "list of collapsing premises": From chips to software, it's the same thing

Old world vs. new world: A list of six collapsing premises

If you compare the "chip premises" in the Huawei paper with the "development premises" of the software industry side by side, you will find that they point in the same direction -

Premise 1 | Computation and storage must be separated → They are being re-integrated (integrated computing and storage, 3D stacking).

Premise 2 | The smaller the transistor, the faster and cheaper it is → After 2nm, the budget for a single chip exceeds 1 billion dollars, and "small" no longer means "cheap".

Premise 3 | The most expensive thing is computing power → The most expensive thing is moving data, and 80% of the energy consumption is on the way.

Premise 4 | Efficiency comes from spatial organization → Efficiency comes from time structure.

Premise 5 | Output = number of people × time → A new formula has emerged:

The generational leap of the output formula

Output = carbon-based time + (silicon-based time × parallel coefficient)

The old formula assumes that "people = time". In the new formula, time can be bought, parallelized, and can operate independently of people. One person plus 100 Agents directly breaks through the variable of "number of people".

Premise 6 | Moat = advanced manufacturing process / scale → The moat has become the ability to schedule time. The conclusion the paper gives to capital is: The next round of funds should follow τ, not the node.

All these six premises point to the same thing:

We are moving from a "space civilization" to a "time civilization".

04 Not just theory: A set of real data from CSDN's internal experiment

This "time-first" logic is not just a theoretical deduction. In the past month, I personally carried out an AI-driven R & D experiment inside CSDN. Here is a set of real data -

This is the R & D efficiency dashboard for a certain week: 24 active R & D personnel, 309 commits in a week, 69,000 lines of changes, among which 60,000 lines were generated by AI - the comprehensive AI proportion is 86.5%.

86.5%. Chew on this number: When code is no longer "written" but "adjusted", the ability to "write" is depreciating, and the ability to "adjust AI to write" is appreciating.

What's even more counterintuitive is the quality data:

The number of test bugs: 186 in January → 62 in April. During the same period, the AI usage rate continued to rise, while the team size decreased.

Intuitively, you might think that the more code AI writes, the more bugs there should be. But the data shows the opposite. Why? There are three reasons:

AI doesn't slack off - people may skip unit tests and forget to write boundary checks, but AI won't. AI won't be interrupted - when people write code, they may be interrupted by DingTalk, meetings, or colleagues, but AI can complete the task in one go. And the most crucial reason: We wrote the constraints into the Git Hook - any commit that is not associated with Jira or has a commit message inconsistent with the code changes cannot be merged into the main branch.

This is not a specification, but a physical barrier.

Relying on slogans and regular meetings, everyone will go back to the old way after three days; when written into the tool, non-compliance is physically impossible.

This is what we call "base tableization" internally - making the whole company queryable, and every important action produces an artifact that AI can read. The essence of the change from 186 to 62 is not that AI writes better than people, but that base tableization forces the whole team to upgrade their work attitude.

05 Not just software: Management, and the disappearing industries

If you think that the "collapse of premises" only happens in the code-writing industry, you're underestimating it.

Let's first talk about something that shakes the foundation more than software - management.

Since Drucker, the foundation of the entire modern management theory can be summed up in one sentence: Improve people's efficiency. How to recruit people, how to arrange shifts, how to assess, how to motivate - all these are about managing "people's time".

But now, what a technical leader really needs to dispatch is no longer 50 engineers, but 5,000 Agents.

To manage people, you need to hold meetings, align, take care of their emotions, and wait for them to wake up. To manage Agents, you need something else: How to break tasks into parallelizable units, how to prevent Agents from waiting for each other, and how to use a "base table" to let them understand what the system is doing at any time.

In the past, management asked: How to prevent 1,000 people from being idle?

Now it needs to ask: How to prevent 10,000 Agents from idling?

These are two completely different disciplines. And business schools are still teaching the first one.

Now let's take a broader view - this kind of thing where "the premise collapses but people don't notice" has happened many times in history.

Every time, the script is the same: It's not being defeated by opponents, but not realizing that the premise under your feet has changed.

Kodak: Its premise was that "people want printed photos". When "photos" became a string of unprinted data in mobile phones, this company that first invented the digital camera was dragged down by digital technology. It wasn't that its technology was bad, but that it didn't realize that the "premise" of what it was selling had disappeared.

Nokia: Its premise was that "the core of a mobile phone is calling and durability". When mobile phones became "computers in your pocket", its most proud signal and quality suddenly became irrelevant advantages.

Neither of these two companies is stupid.

They just believed too much that "what has always held true will always hold true".

There is another way of failure that is even more unexpected.

The instant noodle industry has long relied on the premise of "being cheap, fast, and able to fill your stomach". It has always been on guard against other instant noodle brands. But what ended up squeezing it out was takeout - when a person working overtime can order a hot meal in 30 minutes, the reason for the existence of instant noodles was taken away from the side by an industry it never regarded as an opponent.

Kodak and Nokia "failed to keep up with the changing premise".

The instant noodle industry "faced a changing premise, and the opponent wasn't even in its own track".

The software industry is now in the same position as Kodak, Nokia, and the instant noodle industry back then - The premise is quietly changing under your feet, but most people still think the foundation is stable. And what we should be most vigilant about is the latter: What really changes "how code is produced" is not another more competitive software company, but AI, which was originally not in your track.

06 What does it mean to you?

Let's get back to the judgment that programmers should pay close attention to.

Are the two iron laws of the software industry - "code is the core asset" and "agility is the development method" - really iron laws, or just the most cost-effective compromise in the carbon-based era?

My judgment is the latter. And once the premise changes, the value ranking of the entire profession will be rearranged.

In the industrial era, the most valuable ability was "managing people".

In the AI era, the most valuable ability is "scheduling time" - some people are starting to call it silicon-based scheduling ability.

But don't misunderstand, this doesn't mean that "programmers are going to lose their jobs". On the contrary -

Being able to write code won't make you obsolete,

but "only being able to write code" will keep you in place.

What really appreciates in value are those who move from "writing code" to "scheduling coding".

So if you're a front-line programmer, you can do something very specific and quite interesting this week -

Pick the most annoying and repetitive small task you have at hand and throw it all to Cursor or Claude Code to see how it "runs". Don't read the documentation first, just get the feel. You'll intuitively feel for the first time that "writing" and "scheduling" are two completely different ways of working. This feeling is more valuable than reading ten trend articles.

After taking this step, the next step is