Alibaba’s 2.4 trillion warning

The bottleneck moved. Alibaba says its Qwen3.8-Max release packs 2.4 trillion parameters, below Kimi K3’s 2.8 trillion, with weights due the following week. Its tests broadly matched or beat Claude Fable 5, but the model trailed Fable 5 and 3 Opus models on Arena. June’s deployment thesis starts lower: enterprises can carry 10 duplicate database fields for the same concept. Model scale is rising faster than the systems prepared to use it.

Scaling the model is also reshaping the machine under it. Etched’s inference bet reached a $10.3 billion valuation; TechRadar tallied 4 rounds at about $925.4 million, alongside an 80,000-square-foot, 10-megawatt facility. June arrived with $20 million, 4 founders, and a customer that had promised 100 agents before integration consumed weeks. Inference hardware and deployment software are attracting capital for the same reason: model capability is no longer the only scarce input.

A 2.4-trillion-parameter model, an 80,000-square-foot chip facility, and a 100-agent deployment promise describe 3 different scales—benchmark, hardware, and workflow—but they share one evaluation problem: each layer can look ready in isolation. June’s 4 founders launched Bonobo AI’s voice-to-text service in 2017; Salesforce acquired it 2 years later. Their new customer was trying to run 100 agents through systems carrying as many as 10 duplicate fields. Production readiness is becoming a stack-wide property, not a leaderboard position.

Latest from PulseMark

Why AI Coding Benchmarks Are Unreliable in Production Broken tasks, long-term regressions, and malicious issues expose what coding leaderboards leave out. Read the deep dive →