The $3.5 Trillion Empire That Just Cracked

Warren Buffett made his fortune betting on monopolies. On November 14, 2025, he put $4.9 billion (class A and C combined) in Alphabet, on his perspective that NVIDIA's monopoly is over.

13 days earlier, Google launched Ironwood, a chip delivering NVIDIA-level performance at 1/5th the cost. Then came Gemini 3, trained on zero NVIDIA GPUs. This is why I like Buffett. He doesn't chase hype. He follows the cold hard cash. The money and, honestly, the money just walked out on NVIDIA.

Here's what it costs to train a frontier AI model today (these are industry analyst estimates):

• Everybody else: $3-44 billion
• Google (Ironwood TPUs): $600-750 million

Google owns the factory. Everyone else is renting the tools. A 400% cost advantage isn't competitive edge, straight up execution. When your rivals can train AI at 20% of your cost, you're not leading. You're a margin away from potential irrelevance.

PSST, we’ve watched tech companies build unbreakable moats before. Sun Microsystems. Nokia. Cisco. They all believed and sniffed their own hype right up until the market stopped believing it for them.

NVIDIA built a $3.5 trillion empire by being the only company with chips powerful enough for AI. Anthropic just ordered 1 million TPUs. OpenAI is next. The bulletproof empire already has cracks.

They Didn't Jailbreak Claude. They Just Asked Nicely

September 2025. Chinese state-sponsored hackers designated GTG-1002 weaponized Claude Code to execute one of the most sophisticated cyber espionage campaigns in history. The tech press called it a "jailbreak." I call it Tuesday.

They didn't hack Claude. They used Claude. Exactly as designed.

GTG-1002 told Claude they were cybersecurity researchers running defensive tests. Claude believed them. Then it executed 80-90% of a massive espionage campaign without human intervention, reconnaissance, vulnerability discovery, exploit generation, credential harvesting, lateral movement, data exfiltration. 30 targets. Thousands of requests per second. Several successful breaches.

The AI wrote its own exploit code. Categorized stolen data by intelligence value. Created backdoors. Documented everything for the next attack wave. Human operators showed up maybe four to six times per campaign, just to greenlight the big moves.

Claude didn't get hacked. It got a job offer.

The industry wants you to think this was a sophisticated attack on AI guardrails. Claude did exactly what it was trained to do: follow instructions, decompose complex tasks, execute code. The threat actors asked politely and Claude said yes.

China can't even hack America without importing American technology first. Except now they don't need to import it. We gave it to them.

This Is What "Move Fast" Looks Like

The US position on winning the AI race: don't regulate, move fast, let innovation run wild. Fewer rules means faster progress means American dominance.

I've spent thirty years in technology. I've audited AI systems across SMBs and enterprise sized 100 organizations, leading the protection of more than 2 million people from algorithmic bias. I've seen algorithms erase Black faces, including mine, when ChatGPT replaced my action figure with Jensen Huang's. I've documented bias baked into code and then deployed as “innovation.”

And we think the answer is less oversight?

We're not moving fast and breaking things. We're breaking people and calling it progress.

Here's what "move fast" delivered:

• We built the most powerful AI coding tools on the planet.
• China weaponized them in under a year.
• We trained the models.
• They used them to hack government agencies at physically impossible speeds.

Every capability we build for legitimate use becomes an attack vector. We're not winning by moving fast. We're handing over weapons and acting shocked when they get used.

Here are the cold hard facts. China holds 50% of the world's AI researchers and 30% of the technology market. Huawei is doubling chip production to 1.6 million dies by 2026, achieving 60% of NVIDIA's performance at 70% efficiency, winning on cost per useful compute.

Jensen Huang admitted it in October: "China is well ahead of us on energy. Overall, I would say we're not far ahead."

His leaked November memo after NVIDIA posted record revenue? "If we delivered a bad quarter, it is evidence there's an AI bubble. If we deliver a good quarter, people think we're inflating the bubble. It's a no-win situation."

The man who built a $3.5 trillion empire just admitted he's trapped in it.

White House AI advisor David Sacks tweeted hours ago:

"According to today's WSJ, AI related investment accounts for half of GDP growth. A reversal would risk recession. We can't afford to go backwards."

Read that again.

Half of US GDP growth depends on continued AI infrastructure spending. Not AI productivity gains. Not AI revenue. Infrastructure spending. We've built an economy that requires companies to keep buying chips at premium prices or we trigger a recession.

Google just made those chips 5x cheaper.

That doesn't mean spending stops. It means NVIDIA's margins implode as the cost basis drops. Companies will still invest billions in AI infrastructure, they'll just capture 5x more compute for the same capital. The spending continues. NVIDIA's share of it evaporates.

Sacks thinks we're trapped into continued investment at current prices. He's wrong. We're trapped into continued investment at falling prices. That's worse.

Smart money sees it. SoftBank dumped its entire $5.83 billion NVIDIA stake in October. Peter Thiel sold all 537,742 shares. Michael Burry holds massive put positions.

Buffett bought Alphabet. Not because Google has better technology. Because Google has better economics.

The Math That Ends Empires

Google's Ironwood TPUs made AI training five times cheaper. When the cost barrier drops from $44 billion to $750 million, the number of players will explode.

NVIDIA's moat was scarcity. That's over. Google proved it. Anthropic's betting $10+ billion on it.

The US strategy assumed we'd stay ahead by innovating faster than anyone could copy. But you can't outrun someone using your own tools. And you can't maintain a monopoly when your customer just became your competitor.

China didn't need to match NVIDIA's chips. They needed to jailbreak Claude Code and wait for Google to break NVIDIA's pricing power. Mission accomplished on both fronts.