OpenAI's $120 billion funding round masks a strategic pivot that looks less like confidence and more like necessity. Internal memos reveal the company is consolidating around "main quests"—shuttering projects like Sora that burned millions daily—while doubling down on enterprise capture. The numbers expose the tension: ChatGPT reaches 900 million users monthly, yet only 5% pay. Anthropic, meanwhile, has tripled adoption metrics in the same window, clawing enterprise market share to near-parity with OpenAI at roughly 30% of American companies. Neither company has figured out how to extract value at scale from consumer dominance. This matters because platform lock-in—the harder it becomes to switch—determines long-term moat. Both are racing to deepen memory systems and expanding entry points to make migration costly. The winner takes the ecosystem.
But the sharper story is what happens when AI capability thresholds cross into genuine risk. Anthropic's recent Claude model—available only to vetted researchers—discovered critical vulnerabilities in 20-year-old banking infrastructure in two days on a $20,000 compute budget. The U.S. Treasury and Federal Reserve convened emergency meetings with major banks. This is not hypothetical. Legacy systems underpinning global finance run open-source protocols never interrogated by autonomous agents. The attack surface just expanded enormously. OpenAI responded with GPT-5.4 Cyber, released with slightly more access than Anthropic's restricted model, creating predictable friction over responsible disclosure.
Here is where the ethical frame breaks down into geopolitics. There are three schools of thought. The doomers believe the cat is out of the box; development is inevitable and humanity must adapt. The optimists, led by figures like Sam Altman, dismiss existential risk as overblown—AI is a tool, nothing more. The third school—increasingly dominant in policy circles—treats AI capability like nuclear weapons: development by the West is essential not as progress, but as deterrence. If China or Russia achieves this level of sophistication first, the advantage is asymmetric and irreversible. This framework reframes safety research and rapid scaling not as reckless but as strategic necessity. Mutual Assured Destruction logic applies. You must have the weapon before adversaries do.
This calculus is what animates the polarization. When Daniel Moreno, 20, threw a Molotov cocktail at Sam Altman's home last month—and police recovered a list of AI executives as targets alongside doomer manifestos—he was acting on the belief that development must be stopped. But stopping is asymmetric surrender. The geopolitical argument suggests only Western adoption of these models prevents worse outcomes. It is an uncomfortable bind with no clean resolution: proceed and risk accident, pause and risk subjugation. Neither option is safe.
Meanwhile, Meta's new Llama model and its aggressive capex spend into AI infrastructure suggests the company has chosen acceleration. The firm broke a record this year: estimated ad revenue will exceed Google's. The same week it lost a product-addiction lawsuit mirroring big tobacco precedent. Six million in fines cost Meta a day's revenue. What matters is the signal: the company is betting its future on AI being central to its product suite—video generation, feed optimization, personalized content—regardless of regulatory headwinds. The new model is closed, proprietary, and internal-facing. Lock-in by design.
Security, by contrast, is becoming a genuine venture category. As AI systems become sophisticated enough to discover vulnerabilities in systems considered mathematically secure, defense becomes an urgent discipline. Firms building adversarial testing, automated patching, and behavioral anomaly detection in financial infrastructure will not lack customers. The 50% of American companies now running paid AI subscriptions are also now responsible for defending against the models they're deploying. That gap closes only through investment and urgency.