Here is a number that should make every tech executive in Silicon Valley put down their coffee: 6%.

That is not a discount code. It is not a rounding error. It is the percentage of the industry‘s typical pre-training cost that Baidu’s brand-new ERNIE 5.1 model burned to achieve performance that now sits behind only three models on the entire planet in search capability. Only OpenAI‘s GPT-5.5 and Anthropic’s Claude Opus 4.6 and 4.7 rank higher. Everyone else — Google, Meta, xAI, Mistral — is looking up at a model that cost roughly the price of a nice office renovation to train, while their own models demanded the financial equivalent of a small nation‘s GDP.

This story broke at Baidu Create 2026, the company’s annual AI developer conference held in Beijing on May 13 and 14, and it landed like a thunderclap. Baidu‘s Hong Kong-listed shares surged more than 7% to a three-month high the morning after the keynote. The market was not just reacting to a single model. It was reacting to a thesis — one that has been quietly building for eighteen months and is now impossible to ignore: Chinese AI companies, hemmed in by American semiconductor export controls, are not collapsing. They are getting leaner, meaner, and far more dangerous.

The story of ERNIE 5.1 begins with a technical philosophy Baidu calls “Multi-dimensional Elastic Pre-training,” but the plain-English version is more provocative. Instead of training each model size from scratch — the brute-force approach that has turned AI development into an arms race of GPU hoarding — Baidu built a framework called “Once-For-All.” In a single training run, the system simultaneously optimizes a vast family of sub-models that share weights but differ in depth, width, and the number of activated expert blocks. Think of it as baking one giant cake from which you can slice dozens of perfectly formed smaller cakes, rather than baking each one in a separate oven.

From the massive ERNIE 5.0 architecture — roughly 2.4 trillion parameters, launched in January 2026 — Baidu extracted an optimized subnet to create ERNIE 5.1. The result is startling: total parameters compressed to about one-third, activation parameters cut to roughly one-half, and a pre-training compute bill that landed at approximately 6% of what comparable models cost elsewhere. This is not some marketing sleight-of-hand claiming “we built a Ferrari for the price of a bicycle.” It is a statement about efficiency: at the same parameter scale and performance tier, Baidu used only 6% of the compute that the rest of the industry burns. The word “efficiency” hardly does it justice.

The numbers on the leaderboard are what turn this from an engineering curiosity into a strategic event. On the Arena Search leaderboard — the global benchmarking platform that pits AI models against each other in blind, crowd-sourced evaluations — ERNIE 5.1 scored 1,223 points, securing fourth place globally and first among all Chinese models. It is the only Chinese model on that list. The three models above it — GPT-5.5, Claude Opus 4.6, and Claude Opus 4.7 — are products of companies with essentially unrestricted access to the world‘s most advanced GPUs. Baidu operates under a different reality.

Since October 2022, the United States has imposed increasingly tight export controls on advanced semiconductors bound for China. The explicit goal was to slow down China’s AI progress by starving its labs of cutting-edge hardware. For a while, the assumption in Western capitals was that this would work — that Chinese models would inevitably fall behind as the compute gap widened. ERNIE 5.1 suggests something quite different happened instead. Facing hardware scarcity, Chinese labs turned necessity into a competitive advantage. They stopped trying to out-spend and started trying to out-think. The result is a model that performs at the frontier while using a fraction of the resources. If anything, the sanctions may have accelerated China‘s push toward algorithmic efficiency — and in doing so, may have produced a capability that is far more scalable and commercially threatening than a brute-force model could ever be.

The agentic capabilities of ERNIE 5.1 tell a parallel story. On the τ³-bench and SpreadsheetBench-Verified evaluation tasks, ERNIE 5.1 surpassed DeepSeek-V4-Pro, with agentic abilities now approaching the level of leading closed-source models. Its creative writing matches Gemini 3.1 Pro. On the AIME26 mathematics competition using tools, it scored 99.6, second only to Gemini 3.1 Pro. These are not niche metrics — they represent the core competencies that enterprises care about when deciding which AI to embed into their workflows.

And that is where Baidu Create 2026 got truly interesting. CEO Robin Li did not spend the keynote doing a victory lap on model benchmarks. Instead, he introduced a new metric for the AI era: Daily Active Agents, or DAA. His argument was blunt. Tokens measure input and cost; DAA measures output and value. How many AI agents are actually out there doing work and delivering results for humans? Li predicted that global DAA could easily surpass 10 billion in the near future. This is not a technologist obsessing over parameters. This is a CEO repositioning his company around outcomes.

Baidu rolled out four major agent products to back up the rhetoric. DuMate is a general-purpose AI agent that can autonomously handle emails, analyze sales data, generate procurement recommendations, create promotional materials, and even build mini-applications on the fly — it already hits state-of-the-art levels on multiple international agent benchmarks. Miaoda, a coding agent now in version 3.0, lets users build applications by simply describing what they want; during the keynote, an eight-year-old child built a working app on stage. Baidu Yijing upgrades the company’s digital human technology into what it calls the world‘s first full-scenario multi-agent digital human platform, now expanding internationally with multi-language support. And Famo 2.0, a self-evolving decision-making agent, has already delivered a 10.21% efficiency improvement for the Port of Qingdao’s automated terminal operating system. These are not demos. They are deployed products touching real economic activity.

The infrastructure underneath all of this is equally telling. Baidu‘s homegrown Kunlunxin P800 chip has completed scaled verification, with multiple ten-thousand-card clusters delivered since 2025. An entirely domestic Kunlunxin cluster successfully trained a key version of ERNIE 5.1 with a 97% effective training rate and linear scalability exceeding 85% at the ten-thousand-card level. The Tianchi 256-card supernode launches in June, promising a 25% throughput improvement and 50% better inference efficiency for mainstream models. The “chips, cloud, models, agents” stack — what Baidu calls its full-stack AI capability — is now demonstrably real.

To be clear, some skepticism is warranted. Industry analysts have noted that the 6% figure is likely an exaggerated marketing number, and there is a healthy debate about whether ERNIE 5.1’s performance on narrow benchmarks translates to broad real-world superiority. But even if the true cost figure is 10% or 15%, the directional signal remains the same: China has figured out how to do more with less, and that capability compounds over time.

For a global audience that has spent two years fixated on the OpenAI-versus-Google rivalry, Baidu Create 2026 should serve as a wake-up call. The AI race is not a two-player game. It is a multi-front competition in which resource constraints are breeding a different kind of innovation — one that prizes efficiency over scale, deployment over demos, and agentic autonomy over chatbot cleverness. The company often dismissed as “China‘s Google” just showed the world that it can train a top-four global model for pocket change. What happens when that efficiency advantage gets applied to the next generation of models, and the one after that?

The answer, increasingly, looks like a world where the AI frontier is not defined by who can afford the most GPUs, but by who can squeeze the most intelligence out of every single one. Right now, the company making the strongest case for that title is headquartered in Beijing.

If you had told me five years ago that a machine could read 50,000 scientific papers, design a flawless experiment, and deliver a commercially viable chemical formula — all in the time it takes me to finish my morning coffee — I would have laughed and called you a sci-fi dreamer. But here I am, standing inside Hall 1 of the China National Convention Center, watching exactly that unfold. And I am genuinely struggling to keep my jaw off the floor.

The 28th China Beijing International Science and Technology Expo, held in mid-May 2026, is not your typical trade show. There are no awkward product demos with glitchy prototypes. No corporate executives reading bullet points off teleprompters. What hits you the moment you walk in is something far more unsettling: a quiet, relentless confidence. China is not asking permission to lead. It is simply showing you what the future already looks like — on its own terms.

And the undisputed headline act? A technology that could fundamentally rewire how humanity does science.

Let me take you inside what I saw.

“Lights-Out Lab”: Where AI Becomes the Scientist

Tucked away in one of the exhibition zones, past the glossy quantum computing displays and the robot dogs performing backflips, sits a booth that could easily be overlooked. Beijing Dynaflow Experimental Technology is not a household name outside China. But after what I witnessed there, I suspect that is about to change.

They call it a “lights-out laboratory” — a fully autonomous research facility where no human being ever needs to set foot. Artificial intelligence generates the formula. AI designs the experimental protocol from scratch. Robotic arms execute every step, from measuring reagents to validating results. The data feeds back into the AI, which iterates and refines the next round. Wash, rinse, repeat. No coffee breaks. No sleep. No human error.

The company’s chairman, Chi Haipeng, explained it to me with the casual tone of someone describing a microwave oven: a surfactant formulation that traditionally took a full month of painstaking human trial-and-error was completed by this system in seconds. Not “faster.” Not “more efficiently.” Seconds.

Let that sink in. An entire month of labor — the kind of work that consumes PhD-level chemists, racks up laboratory costs, and burns through corporate R&D budgets — compressed into a few heartbeats. And the quality? More stable than human-generated results. Every single time.

This is not automation in the traditional sense. This is a paradigm shift. The machine is not following a human-designed protocol more quickly. It is designing its own protocols. It is making decisions about which experiments to pursue and which to abandon. It is, in every meaningful sense, the scientist.

And the implications extend far beyond surfactants. The company says its platform already serves industry giants across petrochemicals, drug discovery, life sciences, and food safety testing. Imagine a pharmaceutical company in Lagos or a materials startup in Jakarta accessing the same caliber of experimental capability that currently exists only in a handful of well-funded Western labs. Chi told me, without a trace of hyperbole, that this technology could “break geographic barriers and bring world-class research capability to developing regions.”

That is a bold claim. But after seeing the system in action, I am not inclined to dismiss it.

The Invisible Revolution Hiding in Plain Sight

For those following the global AI race, autonomous laboratories are not entirely new. Researchers at the University of Science and Technology of China have been making headlines with their “Robotic AI-Chemist,” a system that can autonomously design, execute, and refine experiments around the clock. Version 2.0 of that platform now operates on cloud-connected infrastructure, enabling scalable, reproducible scientific exploration across multiple sites simultaneously.

The Chinese Academy of Sciences has built MARS — a multi-agent system coordinating 19 large language model agents with 16 specialized tools — that achieved in 3.5 hours what would normally take a human team weeks: designing a biomimetic water-stable perovskite composite. In a separate demonstration, the same system optimized nanocrystal synthesis within just 10 iterations, a process that traditionally consumes months of labor-intensive tweaking.

These are not incremental improvements. They represent a fundamental reimagining of the scientific method itself — from trial-and-error guesswork to data-driven, AI-orchestrated discovery.

What makes the Beijing Expo stand out, however, is not the raw technology. It is the industrialization. This is not a university research paper or a proof-of-concept prototype locked away in a government lab. It is a commercial product, already serving paying customers, demonstrated openly at a trade show alongside graphene heating systems and AR glasses. China is operationalizing AI-driven science at a pace that should make every Western research institution profoundly uncomfortable.

More Than Just Lab Coats and Beakers

To be fair, the expo offers plenty of other marvels that demand attention. Liangliang Technology showcased its Hey2 AR translation glasses — a featherweight 49-gram device that supports real-time translation across more than 100 languages with latency under 0.5 seconds. I tried them on. The translation appears as floating text directly in your field of vision while you maintain eye contact with the person you are speaking to. It feels like magic. More importantly, it feels like the death of language barriers.

In another corner, Boson Quantum unveiled its latest specialized quantum computer, the “Yuliang Shanhai 1000,” capable of supporting over 3,000 computational qubits with an AI fault-tolerance system that has run continuously for 44 days without interruption. They claim it is already accelerating mRNA vaccine design and brain-computer interface decoding. Whether those claims hold up to scrutiny, I cannot independently verify. But the ambition is unmistakable.

Then there are the dexterous robotic hands equipped with AI tactile sensors — machines that can feel the texture and shape of an object and adjust their grip accordingly. A fully magnetic-levitation artificial heart. Graphene thermal management materials that can dissipate heat at rates that sound physically impossible. A C919 flight simulator with a queue of visitors so long I gave up waiting.

It is overwhelming, honestly. But the lights-out laboratory stayed with me in a way the flashier exhibits did not.

What This Actually Means for the World

I have spent the better part of my career covering technology across three continents. I have watched Silicon Valley promise revolutions that never quite materialize. I have seen European labs produce brilliant theoretical work that never reaches a factory floor. What I encountered in Beijing is different — and it demands a serious conversation that many in the West are not yet ready to have.

The scientific method has remained essentially unchanged for centuries. A human observes. A human hypothesizes. A human experiments. A human concludes. What the lights-out laboratory represents is the first genuine challenge to that paradigm since Francis Bacon. When AI can read the entire corpus of published chemistry literature, design a better experiment than any human could conceive, and execute it with robotic precision, the role of the human scientist shifts from operator to overseer. From doer to director.

This has profound implications for global competitiveness. If a Chinese AI platform can develop a new pharmaceutical formulation in days rather than years, what happens to drug companies in Switzerland, the United Kingdom, and the United States that still rely on traditional R&D pipelines? If a surfactant manufacturer in Guangdong can iterate through thousands of formulations in a single afternoon, how does a competitor in Germany using conventional methods possibly keep up?

These are not alarmist hypotheticals. The technology exists. It is commercially deployed. And it is accelerating.

The Quiet Confidence That Sticks With You

Walking out of the convention center into Beijing’s warm May evening, I found myself replaying a conversation I had with a young engineer at the Dynaflow booth. I asked him — perhaps too bluntly — whether he worried about AI replacing human scientists. He looked at me with genuine confusion, as if I had asked whether calculators had destroyed mathematics.

“Scientists will focus on creativity,” he said, shrugging. “The machine handles the boring part.”

There is a cultural difference embedded in that response that I am still unpacking. In much of the West, automation provokes anxiety. Here, it provokes pragmatism. The question is not whether machines will surpass humans at certain tasks — that is already taken as given. The question is how quickly you can harness that capability to solve real problems.

I left Beijing with a notebook full of observations and a head full of questions. The 2026 Science Expo will be remembered for many things: the quantum computing announcements, the AR breakthroughs, the parade of humanoid robots. But for me, it will be remembered as the moment I realized that the future of science is not a human in a lab coat squinting at a test tube. It is a dark room, somewhere in China, where robots work through the night — and nobody needs to turn on the lights.