Alibaba's 10-trillion-parameter push meets AI security breaches and a shopping bot standoff

The most telling story today isn't about a model at all—it's about who gets to talk to whom. Amazon blocking Meta's Muse shopping agent while Shopify opens its doors wide to the same agent tells you everything about where the real battle for AI's next decade is happening: not in benchmarks, but in access control.

Think about what's actually going on here. Muse is an AI agent designed to shop autonomously on your behalf—browse, compare, click "buy." Amazon says it's blocking Muse over data security concerns, which may well be genuine, but it's hard not to notice that Amazon has its own agentic shopping ambitions and zero interest in letting a Meta-built intermediary sit between it and the customer relationship it's spent decades building. Shopify, by contrast, has nothing to lose and plenty to gain by welcoming any agent that drives checkout volume through its merchants. This is the same dynamic we saw with browsers and app stores twenty years ago, just replayed with AI agents as the new interface layer. Whoever controls the last mile of a transaction controls the leverage, and platforms are already drawing battle lines before agentic commerce has even matured. Expect more of this—not fewer walls, but more selective ones, as every major platform decides which agents it trusts enough to let transact on its turf.

Meanwhile Alibaba is making a very different kind of bet, and a much bigger one. Eddie Wu announced plans to train a model with 5 to 10 trillion parameters, alongside a new in-house AI chip meant to underpin 20GW of data-center capacity by 2032. That's not an incremental scaling exercise—that's Alibaba trying to build the entire stack, from silicon to frontier model, largely independent of Nvidia and, implicitly, independent of US export controls. Whether a model at that scale delivers meaningfully better capability than something a tenth its size is genuinely unclear; parameter count alone stopped being a reliable proxy for quality years ago. But the chip is arguably the more interesting signal. China's AI companies aren't just training around GPU shortages anymore—they're building for a future where they assume the shortages get worse, not better. That's a geopolitical hedge as much as a technical one.

And then there's the security story that deserves more attention than it's getting: Anthropic, OpenAI, and Google have all disclosed breaches or bypasses in their models recently. When it happens to one lab, it's a vendor problem. When it happens to three of the most well-resourced labs in the world within the same stretch, it starts looking like a structural one. Guardrails built through fine-tuning and RLHF keep getting treated as permanent fixes when they're really just speed bumps against a determined adversary. By the way, this is exactly the kind of thing that gets glossed over in the rush to ship agentic features—if your model can be jailbroken, what happens when that model also has a credit card and checkout access?

If agents are going to shop, negotiate, and transact on our behalf, the security question stops being academic. I'd rather see labs slow down and get this right than watch it become next year's headline for the wrong reasons.