Alibaba’s AI Ran 35 Hours Without a Human — Then Outperformed the Manufacturer

Executive Summary

• Fact: At the Alibaba Cloud Summit in Hangzhou on May 20, 2026, Alibaba unveiled Qwen3.7-Max — a model that ran autonomously for 35 consecutive hours, executed 1,158 tool calls, and achieved a 10× performance gain on a chip it had never encountered in training.

• Impact: The announcement came bundled with the Zhenwu M890 AI chip (3× the performance of its predecessor, 144 GB HBM3, already shipped to 400+ enterprise customers) and the Panjiu AL128 rack server — a vertically integrated stack that challenges Western AI infrastructure pricing at its foundation.

• Surprise: Qwen3.7-Max explicitly supports Anthropic’s Claude Code as an external harness — meaning the model China just fielded is compatible with the agent tooling Western enterprises are already deploying.

The headline number did its job. Thirty-five hours. That’s roughly a standard work week compressed into a single uninterrupted machine execution — no approvals, no Slack pings, no human checkpoint. At Alibaba’s annual Cloud Summit on May 20, the company’s flagship model Qwen3.7-Max didn’t just run for that long; it ran on hardware it had never been trained on, starting with zero documentation and zero reference code, and produced a software kernel that outperformed the chip manufacturer’s own official implementation by a factor of ten.

For a COO evaluating AI infrastructure right now, the question isn’t whether this is impressive. It’s what it changes.

What 35 Hours Actually Looks Like

The benchmark task was deliberately adversarial. Alibaba placed Qwen3.7-Max on the Zhenwu M890 chip — its own proprietary AI accelerator, not yet publicly documented — and asked it to optimize the kernel code for SGLang, a popular AI inference framework. The model had no access to hardware specs, no sample implementations, no training data specific to that chip.

Over 35 hours, it executed 432 separate kernel evaluations and 1,158 distinct tool calls. It diagnosed compilation failures autonomously, redesigned its approach each time, and iterated until it delivered a 10.0× geometric mean speedup over the Triton reference implementation. The comparison against peers is instructive: GLM 5.1 achieved 7.3× on the same task, Kimi K2.6 reached 5.0×, and DeepSeek V4 Pro managed 3.3×. Qwen3.7-Max wasn’t marginally better — it was operating in a different category.

Context window expanded to 1 million tokens (up from 256,000 in Qwen3.6-Max-Preview), and the model supports agent harnesses including OpenClaw, Hermes Agent, Anthropic’s Claude Code, Qwen Paw, and Qoder. That last detail is worth sitting with: the frontier agentic model Alibaba just released is explicitly compatible with the tooling Western enterprises are already standardizing on.

FREE DIAGNOSIS — Is your enterprise ready to authorize a 35-hour autonomous AI run? How Transform works → ✓ 30 min  ✓ No cost  ✓ Dedicated AI Director

The Full-Stack Bet That Reshuffles the Supply Chain

Alibaba didn’t just ship a model. The summit announced three interlocking products forming what the company calls an “AI factory” stack: Qwen3.7-Max for the software layer; the Zhenwu M890 from T-Head (Alibaba’s chip subsidiary) for the compute layer; and the Panjiu AL128 Supernode Server — a rack-scale unit packing 128 M890 accelerators with petabyte-per-second internal bandwidth — for the infrastructure layer. Cumulative Zhenwu shipments now stand at 560,000 units across 400+ customers in more than 20 industries. This is not a roadmap.

What’s striking about this move is how precisely it targets the dependency problem now visible in the West’s AI ecosystem. SpaceX’s recent IPO filing revealed that Anthropic is paying approximately $15 billion per year to lease compute — a figure that makes single-vendor compute risk impossible to ignore. Alibaba’s vertically integrated stack is both a product and a competitive argument: if you’re a Chinese enterprise, why route your AI workloads through foreign GPU supply chains that have already proven fragile under export controls?

For Western enterprises, the implication runs in a different direction. A credible full-stack Chinese AI competitor entering the market at scale will apply pricing pressure on every provider — and force a serious board-level conversation about what “AI vendor risk” really means across a 5-year horizon.

This is part of a broader pattern. Just days earlier, Google declared the agentic enterprise era at I/O 2026, and the question of who controls your AI agents is becoming the defining enterprise AI governance debate.

The Governance Gap Nobody Wants to Talk About

The 35-hour autonomous run is a capability story. What it exposes is an organizational one.

Epinium Data

In Epinium’s Transform engagements with brand and operations teams across Europe and LATAM, 68% of teams beginning their first agentic AI workflow have no defined approval protocol for tasks running longer than 60 minutes without a human checkpoint. The technology for 35-hour autonomous runs is now commercially available. The organizational frameworks to authorize, monitor, and audit them are still being written.

What we’re seeing at Epinium is a consistent pattern: enterprises underestimate not the capability of the models, but the internal change management required to deploy them at full autonomy. The hardest question isn’t “can our AI do this?” It’s “who in our org is authorized to say yes to a task this long, this autonomous, this consequential — and what happens when it goes sideways?”

Qwen3.7-Max’s benchmark makes that question urgent rather than theoretical. The model is available now. The governance frameworks, in most organizations, are not.

Five Questions About Long-Horizon Autonomous AI

What exactly is Qwen3.7-Max and how does it differ from previous Qwen models?

Qwen3.7-Max is Alibaba’s flagship model engineered specifically for agentic, long-horizon tasks rather than conversational AI. Its context window is 1 million tokens — nearly 4× the size of its predecessor Qwen3.6-Max-Preview. It handles multi-file software projects, orchestrates multi-agent workflows, and has been benchmarked on real-world chip optimization tasks rather than standardized academic tests. The shift is architectural: this model was built around enterprise autonomy, not chatbot responsiveness.

Can Qwen3.7-Max be integrated into Western enterprise AI stacks?

Yes, explicitly. Alibaba has listed Anthropic’s Claude Code alongside OpenClaw, Hermes Agent, Qwen Paw, and Qoder as supported external agent harnesses. That interoperability is a deliberate market signal: enterprises already standardizing on Claude Code or similar tooling can test Qwen3.7-Max without rebuilding their workflows. Whether compliance, data residency, and geopolitical policies permit that integration is a separate, organization-specific question.

Does the 35-hour autonomous run translate to practical enterprise workflows?

Not immediately, for most teams. The kernel optimization task was ideal: closed-loop, machine-verifiable success criteria, no ambiguous human judgment calls mid-run. Most enterprise workflows don’t have that structure yet. The value of the benchmark is directional — it shows that models can now handle multi-day, multi-step agentic tasks. The work for enterprise teams is redesigning their processes to match that capability profile, starting with workflows that have clear measurable outputs.

What is the Zhenwu M890 and why should non-Chinese enterprises care?

T-Head’s Zhenwu M890 is Alibaba’s most powerful AI accelerator to date, claiming 3× the performance of its predecessor with 144 GB HBM3 memory and 800 GB/s inter-chip bandwidth. Five hundred sixty thousand units have shipped to 400+ customers. For non-Chinese enterprises, the direct relevance is indirect but real: a viable domestically-produced Chinese AI chip at scale reduces China’s dependence on Nvidia export licenses, which changes the global GPU supply-demand dynamic — and ultimately affects pricing and availability for everyone.

When should an enterprise choose not to deploy a long-horizon autonomous agent?

When the task has ambiguous success criteria that require human judgment mid-process. When the downstream consequences of a wrong turn are difficult to reverse. When regulatory or compliance requirements mandate human review at specific decision points. And — perhaps most importantly — when the internal governance framework isn’t yet in place to authorize, monitor, and audit a task of that duration. Deploying autonomous agents without clear escalation and rollback procedures is how organizations lose trust in AI faster than they build it.

The pace is accelerating. Within the span of one week in May 2026, Google declared the agentic era at I/O, Walmart reported Sparky driving measurable order uplift, and Alibaba demonstrated an AI that coded its own chip’s software better than the chip’s designers. The market isn’t waiting for organizations to be ready. The question for every enterprise is how quickly they can close the gap between what the technology can do and what their governance can authorize.

Ready to close your organization’s AI governance gap? Epinium Transform helps brand teams and operations leaders build the frameworks, protocols, and agent oversight structures they need to move from AI pilots to production-grade autonomous deployment — starting with a free 30-minute diagnosis. Book free diagnosis →