Zelos builds a 15,000-GPU cluster to train its self-driving brain

Share
Zelos builds a 15,000-GPU cluster to train its self-driving brain

An autonomous-driving company answers the long-tail problem with compute, while the agent layer gets a gate that rules before the action runs and communications teams get a way to watch what AI says about them.


Zelos, the Suzhou RoboVan maker founded by ex-Baidu autonomous-driving engineers, says it has built the industry's first "10,000-card" L4 compute cluster — nearly 15,000 accelerators in total — built to train its APEX multimodal foundation model. CTO Zhuang Li disclosed the figure on September 17, a week after CEO Kong Qi announced the company's shift toward what it calls city-level physical AI. APEX is described as moving from the tens of billions of parameters toward the hundreds of billions, fed by 270 million kilometers of real L4 operating mileage from more than 30,000 vehicles across 300-plus cities in over 20 countries. It is the first time an L4 autonomy company has put its training infrastructure at that scale, and it reframes where the driving race is being fought: not in the vehicle's sensor stack but in the data center behind it.

The argument Zelos is making is a flywheel, not a flex. Real urban freight throws up long-tail conditions that rules cannot enumerate and simulation cannot reproduce — unmarked rural roads, temporary construction barriers, irregular obstacles, pedestrians and vehicles mixing without order, lighting that flips under bad weather. A stronger base model cuts operating cost, cheaper operations win more customers, and more running vehicles return more of the scarce physical-world data that trains the next base model. That loop only closes if a company owns both the real fleet and the foundation model, which is exactly the position Zelos is claiming: general model labs have no physical-world data, and most autonomy companies have too little real operating data to feed one.

The deployment architecture is worth a look, because it is a bet that a car alone cannot be smart enough. On 5–30 km/h complex roads, an on-device VLA model handles real-time inference while a cloud VLA model is called in for understanding; at 0–5 km/h — the last-meter mess of blocked lanes and wrong-way traffic — a third layer, a "safety agent," takes over decisions like rerouting, turning around or pulling over, with the distillation path running a better on-device model back down after every base-model upgrade. The company's tell is that its VLA is not forced to render decisions in human language before acting; it treats the step as a pre-verbal "intent," closer to a flinch than a sentence. Hold the scale numbers at company-reported, not audited — the cluster size, fleet count and mileage all come from Zelos itself.


Arcjet launched Agent Runtime Security, which checks each agent action against policy before it executes and logs the decision as audit evidence. Platforms can feed the system through the OpenTelemetry tooling they already run, or via Anthropic's Compliance API for Claude customers, and each agent then appears in an inventory with its prompts, tool-call parameters and security decisions stitched into one workflow per session. Policies cover prompt-injection detection, personal-data redaction, bot detection and rate or spend caps, run on Open Policy Agent, and can be changed without redeploying the application; Arcjet returns a decision your code can use to halt the call, ask a human, or explain the refusal to the agent, with a second check after the call to verify the result. Founder and CEO David Mytton's framing is that a dangerous outcome usually builds out of "a series of steps that look perfectly reasonable on their own" — which is the same case for keeping the gate in the call path rather than in the prompt, as we argued in How to — decide what an AI agent may do. It lands next to StackGen putting one governance plane under every production agent: the observability vendors are converging on the same architecture, and the differentiator will be whether the enforcement point sits inside the agent's process or outside it.


PeakMetrics launched AI Perceptions, a service that regularly asks ChatGPT, Gemini, Claude, Grok and Perplexity the questions a company cares about and scores how the answers portray it. Communications teams define prompts covering their brand, products, competitors and industry issues, then grade responses on favorability, trust, competitive positioning, message adoption and purchase intent — the generative-engine-optimization problem stated as a reputation metric rather than a traffic one. The platform also tracks which pages a model cites when it invokes web search, and folds the results into PeakMetrics' existing narrative monitoring, which assigns bot-detection scores to social accounts; the company says automated activity typically accounts for 20% to 30% of the content in a customer workspace and can spike during a crisis. Pricing starts at $99 a month. The honest limit: you can see the answer and sometimes the citation, but not the weighting that produced it — which makes this a monitoring tool that tells you a narrative shifted, not yet a control.

What to watch: whether other autonomy firms disclose comparable training scale, and whether Arcjet's pre-execution gate holds up in production agents that already call dozens of tools per run.

Would you rather own the base model or the fleet, if you could only pick one in the autonomy race? Write to us via our contact page.

Source: Leiphone — 九识建成首个L4万卡集群,无人驾驶进入多模态大模型新范式

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

arXiv just funded its independence — as AI floods its queue

arXiv just funded its independence — as AI floods its queue

arXiv just stopped being a university project. On September 23 the preprint server announced $17.2 million in multiyear commitments from the Simons Foundation International, XTX Markets, and the Siegel Family Endowment, funding its first years as an independent nonprofit. The timing is not a coincidence: the same platform now absorbs more than 28,000 submissions a month, and computer science — AI work above all — is nearly half of them. The money is the smallest part of this story; the load is…

Black Forest Labs' FLUX 3 Klein runs on an 8 GB consumer GPU

Black Forest Labs' FLUX 3 Klein runs on an 8 GB consumer GPU

Black Forest Labs shipped the open-weights half of its FLUX 3 family, and Nvidia put a 120-billion-parameter agent model in the open. In research, a Goodfire essay argues the field misread its own evidence about what lives inside a network. Black Forest Labs has released FLUX 3 Image Klein, a 5.6-billion-parameter image model that runs on consumer hardware — an INT8 build fits in 8 GB of VRAM and produces an image in under nine seconds on an RTX 4070.

Brain organoids classified Braille at 83% — in a trio, not alone

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol.

AI agent ports 250k-line weather sim to GPU with 5.1x speedup

AI agent ports 250k-line weather sim to GPU with 5.1x speedup

A new arXiv paper shows an AI agent modernizing one of the hardest kinds of legacy code — a 250,000-line Fortran weather model — and the validation-first workflow that kept the port trustworthy. Researchers used an AI agent to port CReSS, a legacy Fortran weather simulation code with more than 250,000 lines, to run on GPUs — and the ported version ran a real typhoon simulation 5.1x faster while staying numerically faithful to the original. The workflow, described in a paper submitted to the…

Long context costs 3.6x to 7x more per token. Compaction holds up

Long context costs 3.6x to 7x more per token. Compaction holds up

Two numbers a buyer never sees on a pricing page: what a long prompt actually costs the seller, and how little accuracy the buyer gives up by not sending one. Serving an open-weight model at a 256K-token context instead of 16K costs 3.6 times more per generated token for GLM 5.2, 5.6 times more for Kimi K2.6 and 7 times more for Qwen 3 Coder Next, according to a 31-page NVIDIA technical report posted to arXiv on November 3.