a16z: AI Has Rewritten 75 Years of Innovation Methodology

a16z partners argue that AI is rewriting the underlying economic logic of innovation: shifting from engineering constraints to capital constraints, where a 20-person team can now effectively deploy $1 billion. This judgment directly affects assessments of the startup vs. big tech competitive landscape, venture capital logic, and even the boundaries of AI capabilities. The three also admit that when it comes to model training at the scale of tens of billions of dollars, no one truly knows what will emerge—"this is a digital artifact humanity has never created before."

Recently, on a16z's latest podcast, general partners Martin Casado, Erik Torenberg, and board partner Steven Sinofsky engaged in a deep conversation: What does AI's breakthrough in mathematics mean? Is the underlying logic of innovation being rewritten? How will the competitive landscape between startups and big tech evolve?

The three discussants believe that the fundamental assumptions underpinning the entire computer industry over the past 75 years are being re-examined one by one by AI. The most critical one—that innovation is an engineering problem, not a capital problem—has already failed. AI is rewriting the underlying economic logic of innovation: shifting from engineering constraints to capital constraints, where a 20-person team can now effectively deploy $1 billion. This judgment directly affects assessments of the startup vs. big tech competitive landscape, venture capital logic, and even the boundaries of AI capabilities.

From Abacus to AI: 75 Years of Abstraction Layer Leaps

Sinofsky pulls out a 1953 IBM brochure.

On the cover, atomic orbits encircle a human head, and the opening line reads: "It took mankind millions of years to recognize the usefulness of the wheel."

In that brochure, IBM used an entire page to explain what a "digital computer" is—input, storage, computation, control, output.

Sinofsky says, "That's how we've understood computers for 75 years."

From the abacus to the slide rule, from the difference engine to the personal computer, from the graphing calculator to cloud computing—each technological leap is essentially an "upward shift in abstraction layers." Lower-level problems are encapsulated, and humans solve new problems at higher levels.

When the TI-85 graphing calculator appeared, math teachers collectively panicked: "Our profession is doomed." Sinofsky says:

But they didn't complain about the advent of calculus, because calculus was already a starting point for them. People react far more intensely to change than to the baseline.

The panic triggered by AI solving math problems is exactly the same as the panic over graphing calculators back then.

AI Solving Math: Breakthrough or "Game Master"?

Recently, someone had Claude attempt to solve the Riemann Hypothesis, sparking widespread discussion in the mathematics community.

Casado poured cold water on this directly.

Add up the salaries of all the postdocs who have worked on these problems over the years, and the total amount isn't large. This means the market has never really prioritized solving these problems. So I'm not sure that solving them proves AI has broken through some gate with significant economic value.

He compares AI's mathematical ability to a StarCraft champion: "This is the strongest StarCraft player in history—impressive, but I find it hard to directly link it to real economic value."

Sinofsky offers another perspective. He believes that breakthroughs in mathematics may be "spawning new abstraction tools."

Like the proof of the Four Color Theorem—not writing a beautiful mathematical derivation, but using computers to exhaustively enumerate all finite cases, trading computing power for the answer.

Once you have a new abstraction layer, everyone no longer has to start from scratch; they can build tools directly on that layer.

The Biggest Paradigm Shift: From Engineering Bottleneck to Capital Bottleneck

The core insight of this conversation comes from a thought experiment by Casado.

Twenty years ago, if you gave a 10-person startup $1 billion, they couldn't spend it. Buy some servers, and the money's gone.

Ten years ago, give them $1 billion, you hire engineers, write code, but the mythical man-month is real—more people actually slows things down.

Now, give 20 people $1 billion, and they can actually spend it effectively.

His conclusion:

We have transformed this industry from an engineering bottleneck problem into a capital bottleneck problem. That is a fundamental difference. We have never been in this state before.

Sinofsky adds that this is not the first capital bottleneck—the first 30 to 40 years of the computer industry were also an era of capital constraints. "If you wanted to do something with a computer, the first step was to get one."

Then came the engineering bottleneck era, and now we are back to a capital bottleneck.

Why Haven't Startups Been Crushed by Giants?

By common sense, Microsoft, Google, and Meta, with their capital, data, and distribution channels, should crush everything.

But the reality is: Cursor, Anthropic, and OpenAI are growing at "meteoric speed."

Casado gives two reasons.

First, AI has solved the distribution problem.

In the past, you didn't know whether your marketing budget was effective. Now, the demand for computing power and GPUs is unlimited, and you can directly decide how much to invest to drive growth.

Second, startups can now raise enough money to truly stand on the same starting line as the giants.

And the giants? Sinofsky puts it bluntly:

Microsoft is more worried about what Amazon and Google are doing, and doesn't care about any startup at all. Startups won't directly attack the giants, and the giants won't notice startups.

He also reveals a detail: some large companies, in order to prioritize computing power for enterprise customers, have internal product teams in a state of "AI famine," "while none of their competitors are in famine."

Sinofsky uses his own experience at Microsoft as an example. He took the first Surface to meet with Intel executives, and when they heard it contained an ARM chip, their attitude immediately turned cold.

They thought it was a chip for printers. They only do Moore's Law, just as Google only does hyperscale—if AI shifts to the edge, that's not their business.

Sinofsky concludes:

The essence of disruption is the cultural constants of large companies. Scorecards, sales systems, compensation structures, historical baggage, customer commitments—these cannot be changed. It's a law of physics.

$20 Billion In, No One Knows What Will Come Out

At the end of the conversation, the three touch on a deeper unknown.

Casado admits he made a mistake earlier:

I thought recursive self-improvement and rapid takeoff wouldn't happen, and that was correct. But I didn't expect that we could pour money in indefinitely, and the scaling laws keep holding.

You take $20 billion and pour it into a model. The two of us look at it and test it casually—I don't think we can understand what that means. With that much computing power and that much data, I don't know what it can do.

Sinofsky agrees and compares this to exponential growth: "No one can model exponentials."

He also gives a specific scenario: exhaustive enumeration of protein combinations was previously an "infinite problem," but now it can become a "capital problem."

We can turn previously infinite problems into finite ones through capital. That is a very strange thing.

Casado says this is why he no longer tries to predict the upper limits of AI capabilities:

I have decided that I cannot predict the capability boundaries of something created with $20 billion.