Abstract

On September 12, 2026, three venues thousands of miles apart reached decisive moments on the same day: the Inclusion·Bund Summit in Shanghai entered its final day, ECCV 2026 — Europe's flagship computer vision conference — closed in Malmö, Sweden, and the 18th BRICS Summit opened in New Delhi, India. On the surface these were routine entries on the industry, academic, and geopolitical calendars. Read their agendas side by side, however, and they point at the same proposition: artificial intelligence is moving from generating content to executing tasks — out of the virtual world and into the physical one.

This article argues that the second half of 2026 marks the beginning of AI's "execution era." The claim rests not on a single breakthrough but on the convergence of three independent threads: in research, world models have displaced language modeling as computer vision's central agenda; in industry, AI agents are beginning to take over transactions and operations while embodied AI reaches real production lines; and in geopolitics, the BRICS mechanism is writing digital infrastructure and "resilience" into its agenda, attempting to stand up a non-Western foundation for that physicalization. Below I reconstruct the key facts from each summit, explain why they are the same story, analyze them at three levels — technical paradigm, industrial meaning, and geopolitical meaning — and close with takeaways for practitioners.

1. What Happened at the Three Summits

1.1 Inclusion·Bund Summit, Shanghai: AI moves "from generation to execution"

The 2026 Bund Summit ran September 9–12 at the Expo Park in Shanghai's Huangpu district under the theme "Co-creating the New AI Economy." It was the latest edition of a conference launched in 2020 (officials describe it as having been "held four consecutive times" before), and the first to make "the new AI economy" its core subject — moving past demonstrating model capabilities to asking directly how AI converts into productivity, economic opportunity, and shared growth.

The program comprised one main forum and more than 40 insight forums with over 600 speakers from China and abroad; a 15,000-square-meter technology expo drew more than 300 companies, registered attendance surpassed 50,000, and participants came from more than 50 countries and regions. The main forum lineup spanned academia and industry, including 2011 Nobel laureate in economics Thomas Sargent, 2025 laureate Philippe Aghion, and Wang Jian, director of Zhejiang Lab and founder of Alibaba Cloud.

The most telling detail is the framing the organizers chose: the summit "focuses on the critical leap of AI from 'generation' to 'execution,'" organized around three tracks — "intelligence leap," "economic reconstruction," and "human-centered future." Inside the expo, pavilions were given titles such as "AI payments and the agentic commerce ecosystem" and, remarkably, "bringing artificial intelligence into the physical world." Concrete scenarios included AI payment agents moving "from being able to handle tasks to being able to transact"; more than 30 agent applications entering work and "life beyond the eight-hour day"; and over 40 embodied-AI vendors — Unitree, AgiBot, Sudo Technology, Lingbo Technology among them — showing robots working in pharmacies, manufacturing, parcel sorting, inspection, and eldercare. The robots were no longer walking; they were working.

1.2 ECCV 2026: world models become "the research agenda itself"

ECCV (the European Conference on Computer Vision) is one of only three A*-ranked venues in the field worldwide, alongside CVPR and ICCV, and because it meets biennially each edition is a two-year snapshot. The 2026 edition, held September 8–12 in Malmö, Sweden, was the 19th and the largest in the conference's history: 10,473 submissions, 2,883 accepted at a 27.5 percent acceptance rate, proceedings published by Springer, and a record 86 workshops.

What truly defined the meeting, though, was its intellectual organization. Three threads interlocked:

Three keynotes mapped this trajectory precisely: Kristen Grauman (UT Austin; former director at Meta FAIR) on first-person embodied perception (September 10); Yann LeCun with the keynote "World Models: Enabling the Next AI Revolution" (September 11, 3 p.m.); and Jamie Shotton (Wayve) on "From Visual Recognition to Embodied AI," treating autonomous driving as embodied AI's first proving ground (September 12).

One technical judgment is worth flagging. Researchers have identified a structural flaw in the analogy "world models ≈ large language models": language has a universal token (the word or sub-word), while the physical world has no equivalent. Two trends converging at ECCV 2026 implicitly answer that problem — treating 3D Gaussians as the physical world's token, and world models as the prediction-and-planning layer. No one declared the problem solved, but the conference's papers collectively described its architecture.

1.3 The 18th BRICS Summit: laying the foundation for "physicalized intelligence"

On the same day, the 18th BRICS Summit opened at Bharat Mandapam in New Delhi, running September 12–13. It marked the 20th anniversary of the grouping and India's fourth chairship (after 2012, 2016, and 2021), under the theme "Building for Resilience, Innovation, Cooperation and Sustainability."

Today's BRICS has grown from its original four members to 11 full members (Brazil, China, Egypt, Ethiopia, India, Indonesia, Iran, Russia, Saudi Arabia, South Africa, and the UAE) plus 10 partner countries (Belarus, Bolivia, Kazakhstan, Cuba, Malaysia, Nigeria, Thailand, Uganda, Uzbekistan, and Vietnam), covering roughly 49.5 percent of the world's population and about 40 percent of global GDP. Chinese President Xi Jinping attended, his first visit to India in seven years.

What matters for this article is not the language of the declaration but the structure of the agenda: India placed Digital Public Infrastructure among its showcase cooperation tracks on the official summit site, and paired "Resilience" with "Innovation" in the theme itself. When AI moves from software into the physical world, it necessarily depends on compute, energy, networks, payments, and supply chains — the physical substrate — and that is precisely what the BRICS mechanism is willing and able to discuss.

2. The Convergence: Why This Is One Story

Three summits, three worlds — industry, academia, geopolitics. Why claim they answer one question? Because the way AI creates value is undergoing a fundamental shift: from emitting information to changing physical states.

Chained together, they form a complete value chain: research supplies the ability to understand and predict the physical world (world models); industry supplies the ability to turn that into deliverable services (agents plus embodied AI); geopolitics supplies the infrastructure and governance framework that determines how far, and under whose rules, that ability runs (digital infrastructure plus resilience). Remove any one link and AI's physicalization stalls.

This also explains an apparent paradox: on a single day, the most abstract academic meeting (ECCV) discussed the most practical matters of prediction and action; the most commercial industry gathering (Shanghai) discussed a paradigm shift; and the most political summit (BRICS) discussed infrastructure. Different vocabularies, same destination: intelligence must land in the physical world before it produces a genuinely new economy.

3. Deeper Analysis

3.1 The technical paradigm: from "predict the next word" to "predict the next state"

The dominant narrative of the past three years has been the scaling law — more data and more parameters to predict the next token. The signal out of ECCV 2026 is that computer vision's central question has been replaced by "predict the next state."

That turn has a hard constraint behind it. Language is unifiable in large models because it has a universal token. The physical world does not: pixels in an image, frames in a video, and joint angles in a robot carry different dimensions and semantics. ECCV's convergence of "3D Gaussians as a universal 3D substrate plus world models as the prediction-and-planning layer" is, in essence, a search for the missing token of the physical world. If it works, the payoff is not merely better video generation but a representation layer that can be planned over, controlled, and verified — the very thing robotics has sought for decades.

LeCun's own trajectory is a footnote to this turn. He left Meta in November 2025 to found AMI Labs, arguing that world models should be learned from physical reality rather than text, and raised $1.03 billion in March 2026. A Turing laureate best known as a critic of the language-model orthodoxy staked his reputation on a "physical-first" route — and ECCV 2026 handed that route an institutional stage.

3.2 Industrial meaning: agents take over execution, embodied AI takes over work

The industrial signal from Shanghai can be summarized as two takeovers.

The first happens in the digital world: agents take over transactions and processes. The summit put "when agents become transaction subjects" on the agenda, with Ant Group, Alibaba, OPPO, and Mastercard on stage discussing payments, trust, and collaboration infrastructure. This is not a concept but an operating-model migration backed by survey data: IBM's May 2026 report The blueprint for agentic operations found that 60 percent of enterprises plan to adopt next-generation delivery structures in which AI agents coordinate integrated workflows across finance, supply chain, HR, procurement, physical operations, and customer service; the same report notes that 55 percent of organizations are actively developing or deploying an agentic AI operating model. In other words, what enterprises are buying is not "a smarter assistant" but "an operating system that finishes the work itself."

The second takeover happens in the physical world: embodied AI takes over the work site. Shanghai's expo placed robots in pharmacies, factories, sorting centers, inspection routes, and eldercare settings, with more than 40 vendors on display — and the evaluation criterion has shifted from "can it walk" to "can it work, and is it worth it." Capital markets expect the same order of magnitude: Grand View Research estimates the global embodied AI market will grow from $4.6 billion in 2025 to $6.5 billion in 2026 and reach $67.63 billion by 2033, a compound annual growth rate of about 39.7 percent. (A caveat: because "embodied AI," "physical AI," and "robotics" are defined differently, estimates diverge widely — Data Bridge projects roughly $70 billion by 2033, while MarketsandMarkets sees about $23 billion by 2030. Agreement on direction with disagreement on magnitude is itself evidence that the field is still early.)

Placed side by side, the industrial meaning is clear: the center of gravity in AI commercialization is shifting from selling capability to selling completion. Customers used to pay for model capability; now they pay for outcomes — a process carried through, a transaction settled, a workstation replaced.

3.3 Geopolitical meaning: physicalization pulls AI competition back to hard resources

Physicalizing AI has an easily overlooked consequence: it drags competition from the software layer back to the hard-resource layer.

In the pure-software era, AI's bottlenecks were data and algorithms, marginal costs could approach zero, and a single company could reshape the global landscape. But once AI must drive robots, autonomous vehicles, and industrial lines, the bottlenecks become compute, energy, chips, networks, sensor supply chains, and manufacturing capacity — physical factors that cannot be replicated by downloading a model.

This is precisely the strategic meaning of the BRICS summit. Eleven full members plus ten partner countries, covering roughly 40 percent of global GDP and nearly half its population and including many resource and manufacturing powers, have both the need and the means to build alternatives. When India places Digital Public Infrastructure on the cooperation agenda and writes "Resilience" into the theme, it is in effect pre-positioning a non-Western substrate and rulebook for this physicalization path — cross-border payments, digital identity, compute, and energy cooperation. This need not mean technological decoupling, but it does mean that physicalizing AI will simultaneously be a contest over infrastructure primacy.

A telling contrast: on the same day, Shanghai talked about landing scenarios for the new AI economy, Malmö talked about how machines understand the physical world, and New Delhi talked about building the foundation for all of it. All three, without coordination, moved their attention from "how smart the model is" to "how cooperative the world is."

4. Implications for Practitioners

  1. Change your metrics. If your AI project still uses "generation quality" or benchmark scores as its core KPI, you may already be behind the paradigm. In the execution era the core metrics are completion rate, closed-loop rate, and cost per task — whether AI actually finishes something, and at what price.

  2. Recenter data assets from text to trajectories. Training world models and embodied AI requires data with physical causality: manipulation trajectories, sensor time series, failure cases. Whoever accumulates high-quality real-world interaction data holds the scarce resource of the next cycle. Such data is hard to obtain publicly, which makes the moat deeper.

  3. Restructure the organization around a delivery structure, not stacked tools. IBM's research points not to "buying a few more agents" but to "adopting a delivery structure in which agents coordinate workflows." That means redesigning organization, permission systems, auditing, and governance. Sixty percent of enterprises plan to do this; the other forty percent face not a technology gap but an organizational one.

  4. Put geopolitical variables on the technology roadmap. The availability of compute, energy, networks, and supply chains will directly determine how large a market your physicalization plan can cover. Treating "resilience" as a technical constraint rather than a political slogan is the more pragmatic posture.

  5. Stay clear-eyed that world models are not large language models. The physical world has no universal token, so the successes of language modeling cannot simply be extrapolated. Real breakthroughs may come from representation-layer innovation (3D Gaussians as a physical substrate) rather than pure parameter scaling.

5. Conclusion

On September 12, 2026, three summits in Shanghai, Malmö, and New Delhi each reached a milestone on the same day. Read separately, they are three news stories; read together, they are a joint statement about where AI is heading:

Academia has confirmed that understanding the physical world is the core problem of next-generation intelligence (world models); industry has confirmed that AI's value will be redeemed through execution rather than generation (agents plus embodied AI); and geopolitics has begun laying out the substrate and rules for that physicalization path (digital infrastructure plus resilience).

AI's execution era is not a forecast but a set of facts already unfolding. The real dividing line is not whose model is smarter, but who can reliably and at scale land intelligence in the physical world. The next round of competition will not be decided by IQ. It will be decided by whether you can get things done.

References

  1. ECCV 2026 official schedule (including LeCun, Grauman, and other keynotes). European Conference on Computer Vision. https://eccv.ecva.net/virtual/2026/calendar
  2. ECCV 2026 conference page and proceedings information. https://eccv.ecva.net/Conferences/2026
  3. Chan, W. "ECCV 2026 Preview: Computer Vision Converges on World Models and Embodied AI." TechTimes, September 5, 2026. https://www.techtimes.com/articles/326764/20260905/eccv-2026-preview-computer-vision-converges-world-models-embodied-ai.htm
  4. Lund University. "Conference brings AI researchers from around the world together." September 3, 2026. https://www.lunduniversity.lu.se/article/conference-brings-ai-researchers-around-world-together
  5. LeCun, Y. "A Path Towards Autonomous Machine Intelligence." OpenReview, 2022. https://openreview.net/forum?id=BZ5a1r-kVsf
  6. BRICS India 2026 official site (theme, members and partners, statistics). https://www.brics2026.gov.in/about-us/
  7. "18th BRICS summit." Wikipedia. https://en.wikipedia.org/wiki/18th_BRICS_summit
  8. China Daily. "2026外滩大会在上海举行." September 10, 2026. https://cn.chinadaily.com.cn/a/202609/10/WS6aa25e88e4b09a165c78934a.html
  9. China Daily. "聚焦'共创AI新经济' 2026外滩大会9月9日在上海开幕." September 7, 2026. https://cn.chinadaily.com.cn/a/202609/07/WS6a9e70e1e4b09a165c788764.html
  10. Sina Finance / Cover News. "共创AI新经济，2026 Inclusion·外滩大会今日在沪开幕." September 10, 2026. https://finance.sina.com.cn/tech/roll/2026-09-10/doc-inirinuu5322971.shtml
  11. Shanghai Municipal Financial Regulatory Bureau (Shanghai Fabu). "牢牢把握前沿科技发展机遇……朱忠明考察外滩大会展览." September 11, 2026. https://jrj.sh.gov.cn/ZXYW178/20260911/fc3c2d95ec6143a7926a57dae044f314.html
  12. Eastmoney / China Business Journal. "聚焦'共创AI新经济' 2026外滩大会9月9日在上海开幕." September 4, 2026. https://wap.eastmoney.com/a/202609043865397585.html
  13. IBM Think. "How to manage AI agents without wrecking your business." August 18, 2026. (Citing IBM IBV, The blueprint for agentic operations.) https://www.ibm.com/think/news/how-to-manage-AI-agents
  14. IBM. "The blueprint for agentic operations." IBM Institute for Business Value. https://www.ibm.com/thought-leadership/institute-business-value/en-us/report/agentic-ai-enterprise-operations
  15. IBM Newsroom. "Think 2026: IBM Delivers the Blueprint for the AI Operating Model as the AI Divide Widens." May 5, 2026. https://newsroom.ibm.com/2026-05-05-think-2026-ibm-delivers-the-blueprint-for-the-ai-operating-model-as-the-ai-divide-widens
  16. Grand View Research. "Embodied AI Market Size & Share, Industry Report, 2033." https://www.grandviewresearch.com/industry-analysis/embodied-ai-market-report
  17. Data Bridge Market Research. "Global Embodied AI Market … Forecast to 2033." https://www.databridgemarketresearch.com/reports/global-embodied-ai-market
  18. MarketsandMarkets. "Embodied AI Market Size, Share and Trends." https://www.marketsandmarkets.com/Market-Reports/embodied-ai-market-83867232.html