The Wire

China answers the chip curbs by lashing 8,192 NPUs into one node

The same WAIC that staged China's compute answer seated a 29-nation AI-governance bloc under Beijing, while a benchmark of paid freelance work put the frontier at one project in six, more than quadruple its October mark.

4 min read2026-07-19

compute

Huawei's Atlas 950 SuperPoD ties 8,192 Ascend NPUs over one interconnect¹

Huawei used its WAIC 2026 stand in Shanghai to put the Atlas 950 SuperPoD on the floor: 8,192 Ascend 950DT NPUs joined over its proprietary UnifiedBus interconnect, pitched to train and serve trillion-parameter models with no US-origin parts.¹ Independent measurement puts the single SuperPoD at roughly eight FP8 exaFLOPS across 160 cabinets and about 1,000 square metres, with general availability set for the fourth quarter of 2026.² The chip is not the news. Huawei first announced the Atlas 950 line at its Connect conference in September 2025; what moved this week is the public display and the framing hung on it.¹ The read off the show floor is that China's reply to the export controls has turned from chasing single-chip parity with NVIDIA toward lashing thousands of home-grown processors into one system-level node, trading transistor density it cannot buy for the power and floor space those figures spend.³

policy

Twenty-nine nations sign China's founding charter for a rival AI body⁴

On July 16, the eve of WAIC, 29 countries signed an agreement in Shanghai establishing the World Artificial Intelligence Cooperation Organization, an intergovernmental body headquartered in Shanghai, with Foreign Minister Wang Yi signing for China.⁴ Xi Jinping's first WAIC keynote framed the launch as multilateral cover rather than a Chinese project: AI development, he said, "should not be a solo performance by a single country, but a symphony of international cooperation."⁵ The founding roster leans to the Global South, with Indonesia, Brazil, Pakistan, and Russia among the signatories, and WAICO seats China at the head of a governance track running parallel to the Western frameworks even as Washington steps back from the same norm-setting rooms.⁵

research

Fable 5 finishes 16.1% of real paid freelance jobs, up from 2.5% in October⁶

On the Remote Labor Index, which grades agents on 240 freelance projects that real clients actually paid for, Claude Fable 5 now completes 16.1% to a standard that human graders judge at least as good as the original worker's, roughly double Opus 4.8 at 8.3% and GPT-5.5 at 6.3%.⁶⁷ That is up from the 2.5% the best agent managed when the index launched in October 2025,⁸ a frontier the CAIS and Scale team say has "more than quadrupled in under eight months."⁶ The figure is the July leaderboard's, not tonight's news, and it arrives with its limits attached: Fable 5's access was cut off before every one of the 240 projects could be scored, though even assuming it failed each unscored job its rate would still be 14.6%, higher than any other model.⁹ When the researchers swapped the human graders for an automated LLM judge, it still ranked the models in the right order but badly overshot the absolute rates, by roughly 2.9x for GPT-5.5 and 2.3x for Opus 4.8, so automated grading tracks relative progress but cannot replace human evaluation, because judging an RLI deliverable is itself the kind of demanding agentic work the index is measuring.⁶

engineering

NVIDIA ships 13 DeepStream skills a coding agent can drive from a prompt¹⁰

NVIDIA's DeepStream 9.1, out July 15, ships 13 modular "agentic skills" written to be driven by coding agents such as Claude Code or Codex rather than by a human editing configuration files.¹⁰ The headline pair does the setup a vision team usually does by hand: MV3DT projects detections from several calibrated cameras into one shared 3D coordinate system and holds a single globally consistent ID on each object as it crosses views, and AutoMagicCalib performs the camera-network calibration that step depends on.¹⁰ From a natural-language prompt the agent clones the repo, pulls the models, and wires the pipeline, so standing up a multi-camera 3D tracker no longer means writing the orchestration itself.¹⁰ The skills run down to Jetson Orin edge hardware under JetPack 7.2, which is where most of this is meant to deploy.¹¹

Huawei's Atlas 950 SuperPoD ties 8,192 Ascend NPUs over one interconnect1

Twenty-nine nations sign China's founding charter for a rival AI body4

Fable 5 finishes 16.1% of real paid freelance jobs, up from 2.5% in October6

NVIDIA ships 13 DeepStream skills a coding agent can drive from a prompt10

Sources

Huawei's Atlas 950 SuperPoD ties 8,192 Ascend NPUs over one interconnect¹

Twenty-nine nations sign China's founding charter for a rival AI body⁴

Fable 5 finishes 16.1% of real paid freelance jobs, up from 2.5% in October⁶

NVIDIA ships 13 DeepStream skills a coding agent can drive from a prompt¹⁰