Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more When DeepSeek released its R1 model this January, it wasn’t just another AI announcement. It was a watershed moment that sent shockwaves through the tech industry, forcing industry leaders to reconsider their fundamental approaches to AI development. What makes DeepSeek’s accomplishment remarkable isn’t that the company developed novel capabilities; rather, it was how it achieved comparable results to those delivered by tech heavyweights at a fraction of the cost. In reality, DeepSeek didn’t do anything that hadn’t been done before; its innovation stemmed from pursuing different priorities. As a result, we are now experiencing rapid-fire development along two parallel tracks: efficiency and compute.  As DeepSeek prepares to release its R2 model, and as it concurrently faces the potential of even greater chip restrictions from the U.S., it’s important to look at how it captured so much attention. Engineering around constraints DeepSeek’s arrival, as sudden and dramatic as it was, captivated us all because it showcased the capacity for innovation to thrive even under significant constraints. Faced with U.S. export controls limiting access to cutting-edge AI chips, DeepSeek was forced to find alternative pathways to AI advancement. While U.S. companies pursued performance gains through more powerful hardware, bigger models and better data, DeepSeek focused on optimizing what was available. It implemented known ideas with remarkable execution — and there is novelty in executing what’s known and doing it well. This efficiency-first mindset yielded incredibly impressive results. DeepSeek’s R1 model reportedly matches OpenAI’s capabilities at just 5 to 10% of the operating cost. According to reports, the final training run for DeepSeek’s V3 predecessor cost a mere $6 million — which was described by former Tesla AI scientist Andrej Karpathy as “a joke of a budget” compared to the tens or hundreds of millions spent by U.S. competitors. More strikingly, while OpenAI reportedly spent $500 million training its recent “Orion” model, DeepSeek achieved superior benchmark results for just $5.6 million — less than 1.2% of OpenAI’s investment. If you get starry eyed believing these incredible results were achieved even as DeepSeek was at a severe disadvantage based on its inability to access advanced AI chips, I hate to tell you, but that narrative isn’t entirely accurate (even though it makes a good story). Initial U.S. export controls focused primarily on compute capabilities, not on memory and networking — two crucial components for AI development. That means that the chips DeepSeek had access to were not poor quality chips; their networking and memory capabilities allowed DeepSeek to parallelize operations across many units, a key strategy for running their large model efficiently. This, combined with China’s national push toward controlling the entire vertical stack of AI infrastructure, resulted in accelerated innovation that many Western observers didn’t anticipate. DeepSeek’s advancements were an inevitable part of AI development, but they brought known advancements forward a few years earlier than would have been possible otherwise, and that’s pretty amazing. Pragmatism over process Beyond hardware optimization, DeepSeek’s approach to training data represents another departure from conventional Western practices. Rather than relying solely on web-scraped content, DeepSeek reportedly leveraged significant amounts of synthetic data and outputs from other proprietary models. This is a classic example of model distillation, or the ability to learn from really powerful models. Such an approach, however, raises questions about data privacy and governance that might concern Western enterprise customers. Still, it underscores DeepSeek’s overall pragmatic focus on results over process. The effective use of synthetic data is a key differentiator. Synthetic data can be very effective when it comes to training large models, but you have to be careful; some model architectures handle synthetic data better than others. For instance, transformer-based models with mixture of experts (MoE) architectures like DeepSeek’s tend to be more robust when incorporating synthetic data, while more traditional dense architectures like those used in early Llama models can experience performance degradation or even “model collapse” when trained on too much synthetic content. This architectural sensitivity matters because synthetic data introduces different patterns and distributions compared to real-world data. When a model architecture doesn’t handle synthetic data well, it may learn shortcuts or biases present in the synthetic data generation process rather than generalizable knowledge. This can lead to reduced performance on real-world tasks, increased hallucinations or brittleness when facing novel situations.  Still, DeepSeek’s engineering teams reportedly designed their model architecture specifically with synthetic data integration in mind from the earliest planning stages. This allowed the company to leverage the cost benefits of synthetic data without sacrificing performance. Market reverberations Why does all of this matter? Stock market aside, DeepSeek’s emergence has triggered substantive strategic shifts among industry leaders. Case in point: OpenAI. Sam Altman recently announced plans to release the company’s first “open-weight” language model since 2019. This is a pretty notable pivot for a company that built its business on proprietary systems. It seems DeepSeek’s rise, on top of Llama’s success, has hit OpenAI’s leader hard. Just a month after DeepSeek arrived on the scene, Altman admitted that OpenAI had been “on the wrong side of history” regarding open-source AI.  With OpenAI reportedly spending $7 to 8 billion annually on operations, the economic pressure from efficient alternatives like DeepSeek has become impossible to ignore. As AI scholar Kai-Fu Lee bluntly put it: “You’re spending $7 billion or $8 billion a year, making a massive loss, and here you have a competitor coming in with an open-source model that’s for free.” This necessitates change. This economic reality prompted OpenAI to pursue a massive $40 billion funding round that valued the company at an unprecedented $300 billion. But even with a war chest of funds at its disposal, the fundamental challenge remains: OpenAI’s approach is dramatically more resource-intensive than DeepSeek’s. Beyond model training Another significant trend accelerated by DeepSeek is the shift toward “test-time compute” (TTC). As major AI labs have now trained their models on much of the available public data on the internet, data scarcity is slowing further improvements in pre-training. To get around this, DeepSeek announced a collaboration with Tsinghua University to enable “self-principled critique tuning” (SPCT). This approach trains AI to develop its own rules for judging content and then uses those rules to provide detailed critiques. The system includes a built-in “judge” that evaluates the AI’s answers in real-time, comparing responses against core rules and quality standards. The development is part of a movement towards autonomous self-evaluation and improvement in AI systems in which models use inference time to improve results, rather than simply making models larger during training. DeepSeek calls its system “DeepSeek-GRM” (generalist reward modeling). But, as with its model distillation approach, this could be considered a mix of promise and risk. For example, if the AI develops its own judging criteria, there’s a risk those principles diverge from human values, ethics or context. The rules could end up being overly rigid or biased, optimizing for style over substance, and/or reinforce incorrect assumptions or hallucinations. Additionally, without a human in the loop, issues could arise if the “judge” is flawed or misaligned. It’s a kind of AI talking to itself, without robust external grounding. On top of this, users and developers may not understand why the AI reached a certain conclusion — which feeds into a bigger concern: Should an AI be allowed to decide what is “good” or “correct” based solely on its own logic? These risks shouldn’t be discounted. At the same time, this approach is gaining traction, as again DeepSeek builds on the body of work of others (think OpenAI’s “critique and revise” methods, Anthropic’s constitutional AI or research on self-rewarding agents) to create what is likely the first full-stack application of SPCT in a commercial effort. This could mark a powerful shift in AI autonomy, but there still is a need for rigorous auditing, transparency and safeguards. It’s not just about models getting smarter, but that they remain aligned, interpretable, and trustworthy as they begin critiquing themselves without human guardrails. Moving into the future So, taking all of this into account, the rise of DeepSeek signals a broader shift in the AI industry toward parallel innovation tracks. While companies continue building more powerful compute clusters for next-generation capabilities, there will also be intense focus on finding efficiency gains through software engineering and model architecture improvements to offset the challenges of AI energy consumption, which far outpaces power generation capacity.  Companies are taking note. Microsoft, for example, has halted data center development in multiple regions globally, recalibrating toward a more distributed, efficient infrastructure approach. While still planning to invest approximately $80 billion in AI infrastructure this fiscal year, the company is reallocating resources in response to the efficiency gains DeepSeek introduced to the market. Meta has also responded, With so much movement in such a short time, it becomes somewhat ironic that the U.S. sanctions designed to maintain American AI dominance may have instead accelerated the very innovation they sought to contain. By constraining access to materials, DeepSeek was forced to blaze a new trail. Moving forward, as the industry continues to evolve globally, adaptability for all players will be key. Policies, people and market reactions will continue to shift the ground rules — whether it’s eliminating the AI diffusion rule, a new ban on technology purchases or something else entirely. It’s what we learn from one another and how we respond that will be worth watching. Jae Lee is CEO and co-founder of TwelveLabs. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Google has released an updated preview of​​ Gemini 2.5 Pro, its “most intelligent” model, first announced in March and upgraded in May, as a preview, intending to release the same model to general availability in a couple of weeks.  Enterprises can test building new applications or replace earlier versions with an updated version of the “I/O edition” of Gemini 2.5 Pro that, according to a blog post by Google, is more creative in its responses and outperforms other models in coding and reasoning.  During its annual I/O developer conference in May, Google announced that it updated Gemini 2.5 Pro to be better than its earlier iteration, which it quietly released. Google DeepMind CEO Demis Hassabis said the I/O edition is the company’s best coding model yet.  But this new preview, called Gemini 2.5 Pro Preview 06-05 Thinking, is even better than the I/O edition. The stable version Google plans to release publicly is “ready for enterprise-scale capabilities.” The I/O edition, or gemini-2.5-pro-preview-05-06, was first made available to developers and enterprises in May through Google AI Studio and Vertex AI. Gemini 2.5 Pro Preview 06-05 Thinking can be accessed via the same platforms.  Performance metrics This new version of Gemini 2.5 Pro performs even better than the first release.  Google said the new version of Gemini 2.5 Pro improved by 24 points in LMArena and by 35 points in WebDevArena, where it currently tops the leaderboard. The company’s benchmark tests showed that the model outscored competitors like OpenAI’s o3, o3-mini, and o4-mini, Anthropic’s Claude 4 Opus, Grok 3 Beta from xAI and DeepSeek R1.  “We’ve also addressed feedback from our previous 2.5 Pro releases, improving its style and structure — it can be more creative with better-formatted responses,” Google said in the blog post.  What enterprises can expect Google’s continuous improvement of Gemini 2.5 Pro might be confusing for many, but Google previously framed these as a response to community feedback. Pricing for the new version is $1.25 per million tokens without caching for inputs and $10 for the output price.  When the very first version of Gemini 2.5 Pro launched in March, VentureBeat’s Matt Marshall called it “the smartest model you’re not using.” Since then, Google has integrated the model into many of its new applications and services, including “Deep Think,” where Gemini considers multiple hypotheses before responding.  The release of Gemini 2.5 Pro, and its two upgraded versions, revived Google’s place in the large language model space after competitors like DeepSeek and OpenAI diverted the industry’s attention to their reasoning models.  In just a few hours of announcing the updated Gemini 2.5 Pro, developers have already begun playing around with it. While many found the update to live up to Google’s promise of being faster, the jury is still out if this latest Gemini 2.5 Pro does actually perform better.  Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Last year, China saw a boom in foundation models, the do-everything large language models that underpin the AI revolution. This year, the focus has shifted to AI agents—systems that are less about responding to users’ queries and more about autonomously accomplishing things for them.  There are now a host of Chinese startups building these general-purpose digital tools, which can answer emails, browse the internet to plan vacations, and even design an interactive website. Many of these have emerged in just the last two months, following in the footsteps of Manus—a general AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March.  These emerging AI agents aren’t large language models themselves. Instead, they’re built on top of them, using a workflow-based structure designed to get things done. A lot of these systems also introduce a different way of interacting with AI. Rather than just chatting back and forth with users, they are optimized for managing and executing multistep tasks—booking flights, managing schedules, conducting research—by using external tools and remembering instructions.  China could take the lead on building these kinds of agents. The country’s tightly integrated app ecosystems, rapid product cycles, and digitally fluent user base could provide a favorable environment for embedding AI into daily life.  For now, its leading AI agent startups are focusing their attention on the global market, because the best Western models don’t operate inside China’s firewalls. But that could change soon: Tech giants like ByteDance and Tencent are preparing their own AI agents that could bake automation directly into their native super-apps, pulling data from their vast ecosystem of programs that dominate many aspects of daily life in the country.  As the race to define what a useful AI agent looks like unfolds, a mix of ambitious startups and entrenched tech giants are now testing how these tools might actually work in practice—and for whom. Set the standard It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Effect. The company raised $75 million in a funding round led by the US venture capital firm Benchmark, took the product on an ambitious global roadshow, and hired dozens of new employees.  Even before registration opened to the public in May, Manus had become a reference point for what a broad, consumer‑oriented AI agent should accomplish. Rather than handling narrow chores for businesses, this “general” agent is designed to be able to help with everyday tasks like trip planning, stock comparison, or your kid’s school project.  Unlike previous AI agents, Manus uses a browser-based sandbox that lets users supervise the agent like an intern, watching in real time as it scrolls through web pages, reads articles, or codes actions. It also proactively asks clarifying questions, supports long-term memory that would serve as context for future tasks. “Manus represents a promising product experience for AI agents,” says Ang Li, cofounder and CEO of Simular, a startup based in Palo Alto, California, that’s building computer use agents, AI agents that control a virtual computer. “I believe Chinese startups have a huge advantage when it comes to designing consumer products, thanks to cutthroat domestic competition that leads to fast execution and greater attention to product details.” In the case of Manus, the competition is moving fast. Two of the most buzzy follow‑ups, Genspark and Flowith, for example, are already boasting benchmark scores that match or edge past Manus’s.  Genspark, led by former Baidu executives Eric Jing and Kay Zhu, links many small “super agents” through what it calls multi‑component prompting. The agent can switch among several large language models, accepts both images and text, and carries out tasks from making slide decks to placing phone calls. Whereas Manus relies heavily on Browser Use, a popular open-source product that lets agents operate a web browser in a virtual window like a human, Genspark directly integrates with a wide array of tools and APIs. Launched in April, the company says that it already has over 5 million users and over $36 million in yearly revenue. Flowith, the work of a young team that first grabbed public attention in April 2025 at a developer event hosted by the popular social media app Xiaohongshu, takes a different tack. Marketed as an “infinite agent,” it opens on a blank canvas where each question becomes a node on a branching map. Users can backtrack, take new branches, and store results in personal or sharable “knowledge gardens”—a design that feels more like project management software (think Notion) than a typical chat interface. Every inquiry or task builds its own mind-map-like graph, encouraging a more nonlinear and creative interaction with AI. Flowith’s core agent, NEO, runs in the cloud and can perform scheduled tasks like sending emails and compiling files. The founders want the app to be a “knowledge marketbase”, and aims to tap into the social aspect of AI with the aspiration of becoming “the OnlyFans of AI knowledge creators”. What they also share with Manus is the global ambition. Both Genspark and Flowith have stated that their primary focus is the international market. A global address Startups like Manus, Genspark, and Flowith—though founded by Chinese entrepreneurs—could blend seamlessly into the global tech scene and compete effectively abroad. Founders, investors, and analysts that MIT Technology Review has spoken to believe Chinese companies are moving fast, executing well, and quickly coming up with new products.  Money reinforces the pull to launch overseas. Customers there pay more, and there are plenty to go around. “You can price in USD, and with the exchange rate that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even if we’re only operating at 10% power because of cultural differences overseas, we’ll still make more than in China.” But creating the same functionality in China is a challenge. Major US AI companies including OpenAI and Anthropic have opted out of mainland China because of geopolitical risks and challenges with regulatory compliance. Their absence initially created a black market as users resorted to VPNs and third-party mirrors to access tools like ChatGPT and Claude. That vacuum has since been filled by a new wave of Chinese chatbots—DeepSeek, Doubao, Kimi—but the appetite for foreign models hasn’t gone away.  Manus, for example, uses Anthropic’s Claude Sonnet—widely considered the top model for agentic tasks. Manus cofounder Zhang Tao has repeatedly praised Claude’s ability to juggle tools, remember contexts, and hold multi‑round conversations—all crucial for turning chatty software into an effective executive assistant. But the company’s use of Sonnet has made its agent functionally unusable inside China without a VPN. If you open Manus from a mainland IP address, you’ll see a notice explaining that the team is “working on integrating Qwen’s model,” a special local version that is built on top of Alibaba’s open-source model.  An engineer overseeing ByteDance’s work on developing an agent, who spoke to MIT Technology Review anonymously to avoid sanction, said that the absence of Claude Sonnet models “limits everything we do in China.” DeepSeek’s open models, he added, still hallucinate too often and lack training on real‑world workflows. Developers we spoke with rank Alibaba’s Qwen series as the best domestic alternative, yet most say that switching to Qwen knocks performance down a notch. Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that gap will close: “Building agentic capabilities in base LLMs has become a key focus for many LLM builders, and once people realize the value of this, it will only be a matter of time.” For now, Manus is doubling down on audiences it can already serve. In a written response, the company said its “primary focus is overseas expansion,” noting that new offices in San Francisco, Singapore, and Tokyo have opened in the past month. A super‑app approach Although the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also become household names. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap in the local market has pushed China’s big tech firms to roll out their own user-facing agents, though early versions remain uneven in quality and rough around the edges.  ByteDance is testing Coze Space, an AI agent based on its own Doubao model family that lets users toggle between “plan” and “execute” modes, so they can either directly guide the agent’s actions or step back and watch it work autonomously. It connects up to 14 popular apps, including GitHub, Notion, and the company’s own Lark office suite. Early reviews say the tool can feel clunky and has a high failure rate, but it clearly aims to match what Manus offers. Meanwhile, Zhipu AI has released a free agent called AutoGLM Rumination, built on its proprietary ChatGLM models. Shanghai‑based Minimax has launched Minimax Agent. Both products look almost identical to Manus and demo basic tasks such as building a simple website, planning a trip, making a small Flash game, or running quick data analysis. Despite the limited usability of most general AI agents launched within China, big companies have plans to change that. During a May 15 earnings call, Tencent president Liu Zhiping teased an agent that would weave automation directly into China’s most ubiquitous app, WeChat.  Considered the original super-app, WeChat already handles messaging, mobile payments, news, and millions of mini‑programs that act like embedded apps. These programs give Tencent, its developer, access to data from millions of services that pervade everyday life in China, an advantage most competitors can only envy. Historically, China’s consumer internet has splintered into competing walled gardens—share a Taobao link in WeChat and it resolves as plaintext, not a preview card. Unlike the more interoperable Western internet, China’s tech giants have long resisted integration with one another, choosing to wage platform war at the expense of a seamless user experience. But the use of mini‑programs has given WeChat unprecedented reach across services that once resisted interoperability, from gym bookings to grocery orders. An agent able to roam that ecosystem could bypass the integration headaches dogging independent startups. Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China’s AI race but has been slower to release consumer-facing products. Even though Qwen was the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch—a new mode that marks its most agent-like effort to date.  ByteDance and Alibaba did not reply to MIT Technology Review’s request for comments. “Historically, Chinese tech products tend to pursue the all-in-one, super-app approach, and the latest Chinese AI agents reflect just that,” says Li of Simular, who previously worked at Google DeepMind on AI-enabled work automation. “In contrast, AI agents in the US are more focused on serving specific verticals.” Pei, the researcher at Stanford, says that existing tech giants could have a huge advantage in bringing the vision of general AI agents to life—especially those with built-in integration across services. “The customer-facing AI agent market is still very early, with tons of problems like authentication and liability,” he says. “But companies that already operate across a wide range of services have a natural advantage in deploying agents at scale.”

The US Department of Commerce has renamed its AI Safety Institute to the Center for AI Standards and Innovation (CAISI), shifting its focus from overall safety to combating national security risks and preventing “burdensome and unnecessary regulation” abroad. Secretary of Commerce Howard Lutnick announced the change on June 3rd, calling the agency’s overhaul a way to “evaluate and enhance US innovation” and “ensure US dominance of international AI standards.” The AI Safety Institute was announced in 2023 under former President Joe Biden, part of a global effort to create best practices for governments mitigating AI system risk. It signed memorandums of understanding with major US AI companies, including OpenAI and Anthropic, to get access to new models and suggest improvements before release. Near the end of Biden’s term in early 2025, it released draft guidelines for managing AI risks that included using systems to create biological weapons or other clear threats to national security, but also more common categories of harmful content like child sexual abuse material (CSAM). Lutnick’s statement says that the new institute will “focus on demonstrable risks, such as cybersecurity, biosecurity, and chemical weapons” in its evaluations. It will also investigate “malign foreign influence arising from use of adversaries’ AI systems,” a category that likely includes DeepSeek, a Chinese large language model that shook up the American AI industry earlier this year. The move is part of a larger Trump administration effort to accelerate the expansion of American AI companies. On his first day in office Trump rescinded a Biden executive order that ordered new safety standards for large AI systems and a report evaluating the potential risks for US consumers and the labor market. His own executive orders have encouraged increasing generative AI adoption in fields like education and promoting coal as a source of power for energy-hungry AI data centers. And the current Republican budget bill includes a 10-year moratorium on state-level AI regulations — a provision even some in Trump’s party have come to oppose.

The transformation of Hangzhou’s old steelworks into a park is a tribute to China’s industrial past in a city of the future The congressional hearing about Chinese AI engine DeepSeek held in the US this April has propelled Hangzhou, the heart of China’s new digital economy, to the headlines. With companies such as DeepSeek, Unitree and Alibaba – whose payment app allowed me to get on the metro without needing to buy a ticket – headquartered in Hangzhou, China’s future in AI, robotics and automation is emanating from this city. Getting off the metro in the suburban area of Gongshu, the sun was shining on an old steelworks, overgrown with vines and flowers now that it is being transformed by Jiakun Architects and TLS Landscape Architecture into the Grand Canal Steelworks Park. The unfolding trade war might help to accelerate China’s journey into an automated future, leaving the world of factories behind, yet this new public space shows an impulse to commemorate the country’s economic history, and the forces that have shaped its contemporary built environment. Starting in Hangzhou and travelling more than 1,700km to Beijing, the Grand Canal is an engineering project built 2,500 years ago to connect the different regions of eastern China. The country’s geography means rivers flow from west to east: from higher elevations, culminating in the Himalayas, to the basin that is the country’s eastern seaboard. Historically, it was difficult to transport goods from mercantile centres in the south, including Hangzhou and Suzhou, to the political centre in Beijing up north. As a civil engineering project, the Grand Canal rivals the Great Wall, but if the Great Wall aims to protect China from the outside, the Grand Canal articulates Chinese commerce from the inside. The historic waterway has been an important conduit of economic and cultural exchange, enabling the movement of people and goods such as grain, silk, wine, salt and gravel across the country. It became a UNESCO World Heritage site in 2014. The state‑owned enterprise collective was founded, and the physical facility of Hangzhou steelworks built, in the 1950s during the Great Leap Forward, when China strove for self‑sufficiency, and wended its way through the country’s economic trajectory: first the economic chaos of the 1960s, then the reforms and opening up in the 1980s. Steel remains an important industry today in China, home to more than half of the world’s production, but the listing of the Grand Canal enabled city leaders to move production to a new site and decommission the Hangzhou steelworks. External mandates, including entry into the World Trade Organization, the Beijing Olympics and UNESCO listings, have been instrumentalised in the country to pursue a range of internal interests, particularly economical and real estate ones.  In 2016, the factory was shut down in 150 days, in what the company describes as a ‘heroic’ effort, and the site attracted tourists of industrial ruins. In the competition brief, Hangzhou planners asked for ‘as much of the existing blast furnaces and buildings’ as possible to be preserved. When I arrived in China in 2008, Chinese cities were notorious for heritage demolition, but today urban planners and architects increasingly work to preserve historical buildings. Just like several industrial sites in Beijing and Shanghai have been transformed into major public and cultural spaces in the past decade, in the Yangtze River Delta – of which Hangzhou is a major hub – several industrial sites along the Grand Canal’s course are being given a new lease of life. Today, the three blast furnaces of Hangzhou steelworks remain, with the silhouettes of their smokestacks easily recognisable from a distance. The project preserves as much as possible of the aesthetics of a steel mill with none of the danger or dust, ready to welcome instead new community facilities and cultural programmes in a vast and restored piece of landscape. Situated in a former working‑class district that has been gentrifying and welcoming young families, the new park is becoming a popular venue for music festivals, flower viewing in springtime and year‑round picnics – when I visited, parents were teaching their children to ride a bicycle, and students from Zhejiang University, about a kilometre from the park, were having lunch on the grass. New programmes accommodated in the old coke oven and steel mills will include a series of exhibition halls and spaces welcoming a wide range of cultural and artistic workshops as well as events – the project’s first phase has just completed but tenant organisations have not yet moved in, and works are ongoing to the north of the park. On the day of my visit, a student art exhibition was on display near one of the furnaces, with works made from detritus from the site, including old packing containers. The rehabilitated buildings also provide a range of commercial units, where cafés, restaurants, shops, a bookshop, ice cream shop and a gym have already opened their doors to visitors.  Several structures were deemed structurally unsafe and required demolition, such as the old iron casting building. The architects proposed to partially reconstruct it on its original footprint; the much more open structure, built with reclaimed bricks, now houses a semi‑outdoor garden. Material choices evoke the site’s industrial past: weathered steel, exposed concrete and large expanses of glazing dominate the landscape. The widespread use of red, including in an elevated walkway that traverses the park – at times vaguely reminiscent of a Japanese torii gate in the space below – gives a warm and reassuring earthiness to the otherwise industrial colour palette. Elements selected by the designers underwent sanitisation and detoxification before being reused. The landscaping includes old machinery parts and boulders; recuperated steel panels are for instance inlaid into the paving while pipes for pouring molten steel have been turned into a fountain. The train tracks that once transported material continue to run through the site, providing paths in between the new patches of vegetation, planted with local grasses as well as Japanese maples, camphors and persimmon trees. As Jiawen Chen from TLS describes it, the aesthetic feels ‘wild, but not weedy or abandoned’. The landscape architects’ inspiration came from the site itself after the steelworks’ closure, she explains, once vegetation had begun to reclaim it. Contaminated soil was replaced with clean local soil – at a depth between 0.5 and 1.5 metres, in line with Chinese regulations. The removed soil was sent to specialised facilities for purification, while severely contaminated layers were sealed with concrete. TLS proposed phytoremediation (using plants to detoxify soil) in selected areas of the site ‘as a symbolic and educational gesture’, Chen explains, but ‘the client preferred to be cautious’. From the eastern end of the park, hiking trails lead to the mountain and its Buddhist temples. The old steel mill’s grounds fade seamlessly into the hills. Standing in what it is still a construction site, a sign suggests there will soon be a rowing centre here.  While Jiakun Architects and TLS have prioritised making the site palatable as a public space, the project also brings to life a history that many are likely to have forgotten. Throughout, the park incorporates different elements of China’s economic history, including the life of the Grand Canal and the industrial era. There is, for example, a Maoist steelworker painted on the mural of one of the cafés, as well as historical photographs and drawings of the steelworks peppering the site, framed and hung on the walls. The ambition might be in part to pay homage to steelworkers, but it is hard to imagine them visiting. Gongshu, like the other suburbs of Hangzhou, has seen rapid increases in its property prices.  The steelworks were built during the Maoist era, a time of ‘battling with earth, battling with heaven, battling with humanity’, to borrow Mao’s own words. Ordinary people melted down pots and pans to surpass the UK in steel production, and industry was seen as a sharp break from a traditional Chinese way of life, in which humans aspire to live in harmony with their environment. The priorities of the government today are more conservative, seeking to create a garden city to attract engineers and their families. Hangzhou has long represented the balmy and sophisticated life of China’s south, a land of rice and fish. To the west of the city, not far from the old steelworks, are the ecologically protected Xixi wetlands, and Hangzhou’s urban planning exemplifies the Chinese principle of 天人合一, or nature and humankind as one.  Today, Hangzhou is only 45 minutes from Shanghai by high‑speed train. The two cities feel like extensions of one another, an urban region of 100 million people. The creation of the Grand Canal Steelworks Park reflects the move away from heavy industry that Chinese cities such as Hangzhou are currently making, shifting towards a supposedly cleaner knowledge‑driven economy. Yet the preservation of the steelworks epitomises the sentimental attitude towards the site’s history and acts as a reminder that today’s middle classes are the children of yesterday’s steelworkers, drinking coffee and playing with their own children in grassy lawns next to shuttered blast furnaces.  The park’s second phase is already nearing completion, and the competition for the nearby Grand Canal Museum was won by Herzog & de Meuron in 2020 – the building is under construction, and should open at the end of this year. It is a district rich in history, but the city is resolutely turned towards the future.  2025-06-02 Reuben J Brown Share AR May 2025CircularityBuy Now