Google is burning $200 billion, betting on the biggest deadlock of the 21st century

  It's the first time in the history of the world.

It's getting close!

Recently, a prominent investor at Silicon Valley Buffett, Chamath Palihapitiya, predicted the ultimate roadmap for an "AI point" on X.

1. Humans construct AGIs.

2.AGI has become very good at doing research in the field of AI.

3. It designed a smarter AI.

4. The smarter AI designed an even smarter AI.

5. This cycle is repeated and accelerating.

And then he threw out a shocking conclusion.

Looking at the results and capabilities that have been released by the major AI labs over the past few weeks, I can safely say that we are already in this cycle.

Over the next 18 months, RSI will rapidly improve its AI capabilities.

In the end, the marginal cost of all AI models will be close to zero!

It's not just investors, the whole Silicon Valley today, who believe in the RSI.

Just recently, at a summit at UC Berkeley, Jasjeet Sekhon, chief strategist at Google DeepMind, cast a more weighty judgment:

The whole tech giant circle has an unprecedented crazy investment, essentially betting on the three-letter RSI!

This "greatest scientific bet on human civilization ever" has surpassed the scale of the Apollo moon, the Manhattan Project, and even the entire Internet.

Why are the giants willing to take the risk of breaking the capital chain and go through with this deadlock?

Today's AI is able to self-correct, work for hours and hours, and even generate and optimize underlying code like kernel and compiler components.

But to Jasjeet Sekhon, this is just a prelude to the RSI.

He said, don't think it's unbelievable that AI was designed with AI, because during the Industrial Revolution, humans used the first generation of steam engines to build the next generation of more powerful steam engines.

What does the real RSI look like?

A senior researcher defines real RSI as "the ability of an AI model to completely independently redesign its underlying architecture, or even create a whole new generation of models from scratch, without the intervention of human scientists".

Once this critical point has been crossed, the pace of AI evolution will be beyond human control.

Humanity's evolution lasted decades, and AI may take weeks or even days.

This is when the weird cycle will come true!

When you understand the potential of the RSI, you understand why Silicon Valley is burning like crazy.

This year alone, Google's total capital expenditure on AI data centers, chips and infrastructure has reached a staggering $195 billion to $205 billion, and will increase significantly next year!

Shareholders on Wall Street are starting to worry because, according to the financial statements, the growth in revenue from AI software or cloud services is temporarily unable to cover this amount of capital expenditure.

Sekhon, who was a professor of statistics at Yale, broke this dangerous path.

We are facing a potential "AI air hole" (air pocket, plummeting air pressure in flights, a huge underground bubble, which he describes as investing in a vacuum) where money is being spent and revenues are not showing up on schedule.

The data center is set up, the chips are on, the electricity is coming in, and it's really something that can deliver this input, and it won't be too late.

In fact, long before Sekhon, Wall Street used air pockets to describe this situation: capital expenditures run ahead of income, and investors buy a skyscraper.

If the risk is so great, why keep your eyes closed?

Sekhon's answer is: This is the biggest scientific bet on human civilization ever.

He said that the current global investment in AI has made the Manhattan Project, the Apollo moon, and even the development of the Internet seem similar.

This gambler's logic is simple: While existing technology is not yet a true RSI, "emptying it is obviously unwise".

Because once the RSI is actually realized, as mentioned above, the marginal cost of all AI models will be infinitely closer to zero.

By then, the companies that lead the RSI will have almost god-like productivity, a direct reduction in costs against all competitors, and a monopoly on global computing and intellectual resources.

This is no longer a loss issue for a company, but a "win-win, loser-retire" life-and-death situation.

For this ticket to the New Age, $200 billion is just a snack.

Today, all these data centers, TPUs, and electricity, if you can only rely on a generation of human engineers' manual training models, that's a huge, constantly depreciating fixed asset.

But once AI starts to accelerate AI research and development, it will directly change the logic of the ledger.

It's no longer a linear expense, it's a return on investment: this generation of AI helps you build the next generation faster, and the next generation helps you build the next generation.

This is why Sekhon says RSI is "the key part of investment logic".

Google, Anthropic, OpenAI, all three have a sample.

Look at Google first.

DeepMind's AlphaEvolve system is dedicated to optimizing algorithms using AI.

It accelerated a key core in Gemini training by 23% and reduced the training time by 1%; it modified a circuit design that was written directly into the chips of the next generation of TPUs.

It's now a regular tool in Google's infrastructure, like a caching strategy that takes two days to do what humans would have taken months to do in the past.

AI is helping to make chips, making faster AI, and it's become a reality.

See also Anthropic.

In April of this year, they conducted a more radical experiment: nine Claude intelligences were thrown into a limited coordinated research task, making their own assumptions, running their own experiments, exchanging their own findings.

Two human researchers spent a week recovering only 23% of the performance gap; nine smart objects accumulated 800 hours of work, costing about $180,000, and recovering 97%.

Does it look like AI is capable of doing independent research?

Don't rush, the same report contains an even more important rebuttal: when researchers moved this "most effective" approach to Claude's real production training environment, it didn't bring any statistically significant improvement.

This shows that "acceleration in limited tasks" and "the ability to truly reconstruct AI" have a gap in between.

And finally, OpenAI.

In July, OpenAI disclosed that their GPT-5.6 Sol had automatically rewritten the GPU kernel of the production environment through the Codex, reducing the cost of end-to-end reasoning by 20%; it also ran hundreds of architectural experiments on its own deductive decoding model, improving token generation efficiency by more than 15%.

OpenAI even published for the first time in its publication material an RSI Index, which evaluates self-improvement capabilities.

Do you think the closing rings are moving fast?

OpenAI says in its system card that GPT-5.6 Sol has not yet reached the company's "High" threshold of self-improvement, and that it can solve some of the real research problems but cannot reliably design and implement a complete post-training program.

With three examples combined, the current picture of the RSI is clear:

AI-assisted AI research and development is really happening; but real RSI AI is rebuilding its entire architecture independently, training and deploying stronger successors, a distance away from us.

However, there is also a shadow behind it.

At the summit, Berkeley CS professor Dawn Song, who has just joined the Meta "superintelligence" department, with Sekhon, turned the focus of the discussion to the issue everyone is avoiding: the extreme risks of AI.

The consensus between the two is that the most dangerous thing is not that the model suddenly becomes intelligent, but that the speed of attack and defense is imbalanced.

When AI has the capabilities of RSI, it can cure cancer and help humans explore the universe, but it can also be the ultimate weapon to destroy humanity.

The first is the "reduction of cybersecurity" dimension.

Dawn Song warns that in the short term, the development of AI will be "more favorable to attackers", which is a highly asymmetrical battlefield: hackers can only succeed if they find a vulnerability, while the defense must defend against all possible attacks.

When malicious attackers use intelligent AI to poison open source libraries, scan and exploit vulnerabilities left by human programmers, the days ahead will be very difficult:

Think of how fragile our existing defense systems are: the energy grid in the United States, the global hospital system, the financial network, in the face of super-AI, might be like a piece of paper.

But cybersecurity is also "children's school".

Sekhon was more concerned about the creatures, and he threw a terrifying scene at the meeting, where we're very close to a world where "anyone can design a deadly virus or protein using natural language and model conversations".

He proposed that future societies would have to license, monitor and track all "biologically relevant materials" as strictly as chemicals that make explosives are now tracked.

Currently, Google is trying to apply them to the field of biology using "watermark technology" ("SynthID") to identify AI-generated content.

This means that in the future, if companies synthesize DNA sequences for pharmaceutical companies, systems can scan those sequences to see if they contain "malicious code" that is secretly generated by AI and poses a lethal risk.

Do you think we're going to build faster defenses, or do you think AI will evolve faster?

These sounds like science fiction, but the Silicon Valley brains have given precise countdowns.

Chamath said: "The world is going crazy for the next 18 months.

According to Sekhon, the real RSI is "likely to emerge in the coming years".

Wojciech Zaremba, Vice President of DeepMind Research at Oriol Vinyals and OpenAI, has even reported on specific years: 1927 or 2028.

There is no other technology in human history that could fundamentally rewrite the trajectory of civilization, and the countdown is so close.

This time, being put on stage, is a dead end and a race against time.

Infrastructure investments are advancing at a computational pace, with chips, power, and data centers all in place within a few years.

But can the corresponding revenue and RSI catch up with the car?

In the middle of the gap, are we going to accidentally hit this "airhole"?

But at least here in Sekhon, the direction of Google's $200 billion fortune is clear: RSI is coming.

Steam engines have made better steam engines, and humanity has ushered in the industrial revolution.

Now, AI is making better AI and pushing our generation to an unknown dividing line.

Whether you're ready or not, the buzzing data centers won't stop, the self-evolution of AI won't stop.

The countdown has begun, and are you ready for a surprise?