Posted on — Leave a comment

AI slop as a business and climate hazard, and a chance to lead

AI slop is emblematic of AI/LLM shortcomings and wasted money, time, and human potential. AI slop exists because human slop existed. LLMs were trained on the kind of trite, rehashed, insubstantial drivel that fills so many corporate documents, social media platforms, and so-called thought leadership.

But AI slop, unlike human slop, is harmful beyond being poor writing – it’s a waste of power. AI slop sucks up power beyond our control. As data centres spread across Malaysia, Indonesia, and the Philippines, more of that power will come from the region’s coal and gas, emitting climate-harming emissions.

Companies need to fight AI slop not just for their sales and marketing outcomes, but for the climate. This means fighting AI slop presents a novel opportunity for companies to bolster their ESG and annual reports, while boosting business outcomes.

Companies already count the emissions from business flights and the paper in their offices. Few count the electricity spent generating and hosting infinite AI slop output that few people read, that search engines may even penalise you for, and that stunts the growth of talented human professionals.

Why AI slop is a cost with no return

Entire business are not running on AI slop. A Southeast Asia-based startup founder I know uses AI for pretty much everything. They built an app and a website with an extensive blog, almost all of it generated by uncustomised AI. The site looks really nice, yet a closer read finds little substance, no unique data, and little that sets it apart. It’s not interesting. It won’t evoke the kinds of emotions that good business content evokes in prospects and customers. It also performs poorly in AI search and conventional SERPs.

This founder’s approach is no anomaly. Quite opposite. The fully AI-automated approach is becoming rapidly more common because it’s so easy and so fast. Those traits are deceptive, because the output is potentially ineffective, and even harmful, when it incurs search penalties and burns funds, as well as power. Human-written content would take more time and more skill, but it likely would deliver overwhelmingly better business results.

Research by a major digital marketing tracked 744 articles on 68 websites and found that after a year, human-written posts averaged 338 visits, compared with 61 for AI-written ones.

Also Read: The most expensive AI mistakes don’t look like mistakes

AI is changing outsourcing for volume-based workers

I’m a writer, editor, translator, and content creator (a.k.a. a modern-day marketer). I can only do so much. So, for years, I’ve contracted out work, especially to the Philippines.

Before AI, I received a mix of quality work and human slop. In the last two years, it’s almost entirely turned into AI slop – so much so that I’ve stopped working most of my subcontracted writers. The pressure on these workers is understandable, as AI is also hollowing out the outsourcing and offshoring work, such as call centres, that the Philippines has long relied on. They have to compete for clients and do even more work in less time.

This comes at the expense of novelty and true quality. Quality at scale and a discount has long been why many companies outsource to the Philippines. Good content creators, including writers, like the Filipino professionals I still work with, actually do the research and continue to turn out excellent work. I’ll keep working with them. As for AI slop deliverers, I can empathise with why it happens. I get it. I’m a writer, too. But I don’t have the time and patience to play slop detective. It’s inefficient.

Platforms are also losing patience with slop. In January 2026, YouTube removed 16 of the largest AI slop channels, with a combined 35 million subscribers and 4.7 billion views. Efforts like this will continue – algorithm penalites are virtually inevitable. Irresponsible AI use will also continue to harm the climate.

AI’s power suck and data centre power demand in the region

Google reports 0.24 watt-hours for a median Gemini text prompt. That’s not much. Data centres worldwide, however, used 415 terawatt-hours in 2024, and the International Energy Agency expects about 945 terawatt-hours by 2030, with AI as the biggest driver of that growth. Most of the demand is in the US, China, and Europe. Southeast Asia’s share is small but it’s growing quickly.

Ember, an independent energy think tank, projects that data centre electricity use in Malaysia will rise from 8.5 terawatt-hours in 2024 to 68 terawatt-hours in 2030, about 30 per cent of national demand. Indonesia’s could grow from 6.7 to 26 terawatt-hours, and the Philippines’ from 1.1 to 20. Ember describes Malaysia’s grid as dominated by coal and gas, and expects emissions from data centre power on the Philippines’ Luzon-Visayas grid to rise about 14-fold.

Johor’s data centres have a maximum demand of about 3.8 gigawatts, nearly 1.5 times the state’s current electricity demand, and most local generation comes from gas and coal.

All 11 ASEAN members, including Timor-Leste, have ratified the Paris Agreement and submitted a Nationally Determined Contribution, each a national plan for cutting emissions. Data centre demand met with coal and gas pulls against those plans. We also see progressive solutions emerging.

Also Read: AI is a tool. Stop treating it like a strategy

Singapore and Malaysia’s green data centre growth as a good approach

Green Data Centre Roadmap plans for at least 300 megawatts of new capacity and supports facilities that use no more than 30 per cent extra energy on cooling and other overheads. In December 2025, Singapore opened a call for at least 200 megawatts more, and every project must draw at least half its power from green sources.

Malaysia’s Corporate Renewable Energy Supply Scheme lets companies buy renewable power directly from developers, and 1.3 gigawatts of these agreements had been signed by June 2025, all linked to Johor data centre projects. DayOne secured up to 500 megawatts of solar over 21 years for two Johor parks. Malaysia is also tendering 6–8 gigawatts of new gas-fired capacity, though, so fossil fuels will keep carrying much of the load.

Responsible, AI-related data solutions are attractive ESG points for nations and companies. AI slop adds avoidable demand to Southeast Asia’s grids, undercuts skilled professionals’ work, and is counter to the climate pledge every ASEAN government has made by signing the Paris Agreement. That makes AI slop a great focal point for action and reporting results.

Companies in Southeast Asia can take the lead in reporting AI use in their ESG reporting. I’d like to see them include AI slop reduction in these efforts. It isn’t required, and many companies hesitate because people use the technology without understanding it. That’s why reporting it now would be novel and appealing.

A doable approach would cover carbon equivalents and other environmental effects of AI use, positive and negative, including paper saved. It would explain how the company chooses AI vendors, how it ensures good use, how it supports renewable energy, and how all of that fits its wider climate strategy.

Companies with a heavy content-presence can cut the AI slop and include that in their reports. Cutting AI slop is new, relatable, and a uniquely human-led effort to responsibly incorporate AI into modern-day work processes.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post AI slop as a business and climate hazard, and a chance to lead appeared first on e27.

Posted on — Leave a comment

Singapore’s Botsync takes its Physical AI pitch to the US with US$11M in fresh funding

For most of its seven years, Botsync has been known for its heavy-duty robots that haul pallets around factory floors. Its new pitch is that the robots matter less than the software telling them what to do.

The Singapore-headquartered robotics company has raised US$11 million in a round led by IvyCap Ventures, with participation from Capital 2B, AppWorks, Iterative, AngelCentral, Genesis Alternate Ventures and angel investor Nalin Advani.

Botsync did not give the round a series label or disclose its valuation. The money will go into new AI features for SyncOS, its orchestration platform, covering throughput optimisation, predictive maintenance and operational analytics, as well as better performance and reliability for its MAG robots. It will also fund hiring across engineering, project delivery, customer success and sales.

Also Read: Why Asia’s Physical AI boom will be decided at the camera, not the model

The round lifts Botsync’s total funding to more than US$20 million and follows an extended Series A from SGInnovate in January. Notably, IvyCap was only a co-investor in Botsync’s US$5.2 million Series A in 2024. Its step up to lead investor shows conviction from an insider. It also means the round is largely existing backers doubling down, rather than a new marquee name validating the story from outside.

From selling robots to selling the control room

Botsync was founded in 2019 by four Nanyang Technological University graduates: Rahul Nambiar, Nikhil Venkatesh, Prashant Trivedi and Singaram Venkatachalam. It started with autonomous mobile robots (AMRs), machines that navigate factories without fixed tracks. Its MAG range now covers lifters, conveyors, tuggers and forklifts that carry up to 1.5 tonnes.

SyncOS is where the company now wants to be judged. Botsync calls it a “Physical AI” platform, industry shorthand for AI that controls machines in the real world rather than generating text. In practice, it is a vendor-agnostic control layer. It connects AMRs, robotic arms, cobots, conveyors, automated storage and ERP systems from any supplier, then assigns tasks, manages traffic and runs workflows across the facility in real time. CEO Rahul Nambiar argues that automation can no longer be limited to individual robots doing isolated tasks.

The logic is sound. Most large factories run mixed fleets from several vendors that do not talk to each other, and that is usually where projects stall on the journey from pilot to production. Software also carries better margins than hardware, and the system that runs a whole facility is far harder to rip out than a single robot.

Blue-chip logos, unaudited numbers

On paper, the track record is strong. Botsync says it has completed more than 70 projects across six countries and over two million production tasks at 99 per cent uptime. It names Caterpillar, Ford, Nestlé, Honda, Coca-Cola and Kimberly-Clark as customers, and says clients recover their investment within 24 to 36 months by relying less on forklifts and manual handling.

These figures come from the company and are not independently verified. Botsync does not disclose revenue, and a logo on a slide can mean a single-site deployment as easily as a group-wide rollout. A two-to-three-year payback is respectable in Singapore, where labour is scarce and expensive. It is a harder sell in Indonesia or Thailand, where a forklift driver costs a fraction as much and capital budgets are tighter.

Also Read: AMC Robotics to build US$3.5M Vietnam factory as SEA automation race heats up

The company also says customers have moved operators into higher-value roles. That is the standard line in every automation pitch. Whether it holds up on the factory floor, rather than in a press release, is a question worth asking customers directly.

A crowded factory floor

Botsync is fighting on two fronts. On one side are the large AMR makers, such as China’s Geek+ and Hai Robotics, US-based Locus Robotics and GreyOrange, and India’s Reliance-backed Addverb. All of them sell robots bundled with their own fleet-management software. On the other side are software-only orchestration specialists such as SVT Robotics, Formant and InOrbit, which pitch the same vendor-neutral “one layer to run them all” promise without the hardware baggage.

Botsync’s edge lies in the middle. It has deep experience with heavy payloads in automotive and FMCG plants, and its engineering base in Singapore and India keeps delivery costs down. The risk is that orchestration becomes a commodity. As robot makers adopt common interoperability standards, and as the giants give away fleet software to win hardware deals, a standalone control layer has to show it is clearly smarter, not just more open.

The funding picture at home is lopsided too. Singapore has produced big robotics cheques this year, including dConstruct’s US$125 million Series A. But Tracxn data analysed by e27 shows that a single round, Sharpa’s US$670 million Series D, makes up 96 per cent of Southeast Asian robotics funding in 2026. Only three of the region’s 242 tracked robotics companies have made it past Series C. Against that backdrop, US$11 million is a meaningful cheque, but hardly a war chest, and the sector’s concentration problem means mid-stage players like Botsync must prove revenue before the next round.

A small island with a long to-do list

Singapore remains the headquarters and the showroom. Nambiar points out that with a small population and limited land, Singapore is betting on advanced manufacturing, so facilities have to squeeze more out of the space and equipment they already have. Botsync plans to use part of the funding at home to deepen partnerships with robot makers, automation providers and software firms in the SyncOS ecosystem, and to run more customer events.

Also Read: Southeast Asia’s quiet FDI manufacturing redistribution

The real growth bets lie elsewhere. Botsync wants to build on early traction in Australia, set up a go-to-market in the US and grow its capabilities in India, while keeping Thailand and Indonesia running. The US is the biggest prize and the hardest market, with established local rivals and long enterprise sales cycles. Opening that front while scaling four others on US$11 million will test the company’s discipline, especially as Southeast Asian logistics tech funding remains only a third of its 2021 peak.

IvyCap founder and managing partner Vikram Gupta believes Botsync is well positioned to expand its footprint globally. The thesis is clear enough: the factory’s brain will matter more than its muscles. What Botsync must now prove is that the world’s biggest manufacturers will let a Singapore startup run the whole floor, not just a few of its forklifts.

The post Singapore’s Botsync takes its Physical AI pitch to the US with US$11M in fresh funding appeared first on e27.

Posted on — Leave a comment

Cheques, bans and bell-ringing: 50 Asian AI companies that defined 2026

If 2025 was the year Asia’s AI founders learnt to say “agentic” without blinking, 2026 was the year the cheques, the bans and the bell-ringing all arrived at once.

Capital poured in at a scale that would have looked like a typo three years ago. DeepSeek reportedly raised around US$15 billion in a single round. Moonshot AI closed US$3.5 billion at a US$35 billion valuation. Korea’s chip challengers Rebellions and Furiosa pocketed more than US$1 billion between them, part of it from a state fund that now treats Nvidia’s grip on inference as a national-strategy problem.

But money was only half the story. Beijing vetoed Meta’s US$2 billion purchase of Manus. Anthropic accused Chinese labs of distilling its Claude models at industrial scale. Washington put Unitree on its list of Chinese military-linked companies, weeks before the robot maker opened roughly 500 per cent higher on Shanghai’s STAR Market.

And in Bengaluru, a Flipkart-backed generative-AI startup shut its doors because, by its CEO’s own account, Google’s image models simply did its job better. That is the other side of the AI boom: the foundation-model giants are eating the thin wrappers built on top of them.

Also Read: Hong Kong and Taiwan founders are cheering the AI boom. Southeast Asia isn’t in the room

What follows is our list of 50 Asian AI companies that made headlines in 2026, through mega-rounds, IPOs, acquisitions, regulatory knife-fights, layoffs and quiet funerals. The list spans China, India, South Korea, Japan, Singapore and Hong Kong. Southeast Asia, beyond a handful of Singapore-registered names, is conspicuous mostly by its absence, a gap that echoes the AI funding boom’s Southeast Asia-sized blind spot.

Here is who made the cut, and why.

Mega-rounds, distillation rows and a deal Beijing killed

1. Moonshot AI (Beijing, China)

The Kimi maker closed a US$3.5 billion round at a US$35 billion valuation in July, according to Yicai and Bloomberg. The headlines were not all flattering. In February, Anthropic alleged “industrial-scale” distillation of Claude, and a September report accused Moonshot of routing users to Claude via proxies. Founded only in 2023, it has now raised US$7.27 billion in total equity, per Tracxn.

2. DeepSeek (Hangzhou, China)

The lab that spooked Silicon Valley in early 2025 had a busier 2026. TechCrunch, citing Bloomberg, reported in October that DeepSeek raised around US$15 billion, one of the largest private rounds ever for an Asian startup. It shipped V4 on 24 April. The US delayed a planned blacklisting in June, and DeepSeek too was named in Anthropic’s distillation allegations. Tracxn puts its total equity funding at US$19.4 billion.

3. Kling AI (Kuaishou)

Kuaishou’s video-generation unit completed a round of nearly RMB20 billion (about US$2.98 billion) on 30 October, with Kingdee, Gongqingcheng Junke and MetaX named as investors. That came after an earlier round of roughly US$3 billion in July, in which Baidu and Tencent took part. Kling 3.0, launched in February, kept it in the front row of China’s text-to-video race.

4. Manus/Butterfly Effect (Singapore, China-founded)

No company had a stranger year. Meta bought the agent platform for about US$2 billion in December 2025. China opened a probe in January, reportedly barred the co-founders from leaving the country in March and, on 27 April, the NDRC vetoed the deal outright. By September, Manus was independent again, and in October Butterfly Effect raised more than US$500 million, co-led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund participating. As we argued at the time, Manus came home, and Singapore couldn’t stop it.

Chips, robots and sovereign compute pull in the big cheques

5. Rebellions (Seongnam, South Korea)

Korea’s answer to Nvidia on inference closed a US$400 million pre-IPO round at a US$2.3 billion valuation in March. Mirae Asset led, and the deal marked the first direct investment from Korea’s new National Growth Fund. Rebellions then went shopping, buying model-compression specialist SqueezeBits in June.

6. ROBOTERA (Beijing, China)

The humanoid-robot maker raised RMB1 billion (about US$140 million) in July, led by Chengtong Fund. Local press described it as the company’s third raise of the year, which tells you how fast Chinese capital is chasing anything with two legs and a pair of hands.

7. StepFun (Shanghai, China)

In January, the large-language-model startup raised a Series B+ of roughly RMB5 billion (US$717 million). Caixin noted that the haul outpaced what its newly listed rivals had raised, an awkward comparison for public-market investors who had just bought into Zhipu and MiniMax.

8. Sharpa (Singapore)

One of the few Singapore-headquartered names on this list, Sharpa builds humanoid robots and AI for dexterous manipulation. It raised about US$670 million in August, with Alibaba, Tencent, Meituan and JD among its investors. That is a remarkable cap table for a company founded in 2024, and a reminder that Asia’s physical AI boom may be decided at the camera, not the model.

9. Neysa (Mumbai, India)

The AI-cloud provider raised US$1.2 billion in February, split between US$600 million in equity and a planned US$600 million debt facility. Blackstone led the round at a US$1.4 billion enterprise valuation, per Inc42. India’s GPU-cloud ambitions finally had a balance sheet to match.

10. Infinigence AI (Beijing, China)

The AI infrastructure and model-inference company raised a RMB2.3 billion (about US$320 million) Series C in October. As Chinese labs scramble for compute under US export controls, squeezing more out of every chip has become a business in itself.

11. Sarvam (Bengaluru, India)

India’s sovereign-AI flag-bearer became a unicorn in June. It raised a US$234 million Series B, the first close of a US$300 million round, at a US$1.5 billion valuation. HCLTech led, paying US$150 million for a 10.46 per cent stake. Whether a domestic champion can be built on borrowed choices is a question Southeast Asia is watching closely.

From Tokyo labs to Shenzhen robot shops

12. Sakana AI (Tokyo, Japan)

Citi made a strategic investment in Sakana on 24 February, a month after Google separately backed the research lab. Neither disclosed the amount. Tracxn values Sakana at US$2.65 billion, based on its November 2025 Series B.

13. SiliconFlow (Singapore)

The LLM-infrastructure provider closed a bundled Series B+ tranche and Series C in September. The leads were all state-backed: China Internet Investment Fund, China Reform Fund and China Mobile Chain Leader Fund. Caixin puts its 2026 haul at about RMB2.9 billion (US$433 million), and Tracxn values it at US$1.14 billion.

Also Read: Singapore’s robotics dominance is a warning sign dressed up as good news

14. Astribot (Shenzhen, China)

The humanoid maker raised RMB1 billion (about US$140 million) in October at a reported RMB15 billion (US$2.1 billion) valuation. That is comfortably above Tracxn’s US$1.48 billion figure.

15. Vast3D (Hangzhou, China)

The generative 3D-modelling platform raised a combined RMB3 billion (about US$446 million) Series B and B+ in September. Game studios, e-commerce sellers and industrial designers all want 3D assets without paying 3D artists.

16. D-Robotics (Shenzhen, China)

Its US$400 million Series C in September was, according to TechNode, the largest China robotics round in four years. D-Robotics sells the plumbing underneath robots: hardware, operating systems and development tools. Founder after founder says data, not hardware, is the real bottleneck in humanoids, and investors are paying for the stack that collects it.

17. EVAS Intelligence (Guangzhou, China)

The AI-chip and accelerator designer raised nearly RMB2 billion (about US$280 million) in September, following an earlier RMB1.5 billion Series B. The stage was not formally disclosed.

18. Turing (Tokyo, Japan)

The autonomous-driving startup completed a Series A totalling about JPY27.9 billion (roughly US$174 million) in July, including a US$79 million extension. For a Series A, that is the kind of money usually reserved for a late-stage round.

19. DeepX (Seongnam, South Korea)

The edge-AI chip maker took a KRW42 billion (about US$29 million) first tranche of its Series D from BNW and DS Asset Management in August, at a KRW3.14 trillion (about US$2.2 billion) valuation. In April, Reuters reported that DeepX was preparing for an IPO.

Korea crowns an app unicorn, India bets on voice and quantum

20. WRTN (Seoul, South Korea)

The consumer AI-app company raised a Series C of about US$72–76 million in August to fund global expansion. Reported valuations ranged from US$722 million to about US$1 billion (KRW1.4 trillion), and it was billed as Korea’s first AI-service unicorn. The gap between those numbers is worth noting: whether WRTN is a unicorn depends on which outlet you read.

21. ShengShu Technology (Beijing, China)

The multimodal video-generation model developer raised a Series B of about RMB2 billion (US$293 million) in April, led by Alibaba Cloud. Alibaba is quietly building a portfolio that touches almost every serious Chinese model lab.

22. Gnani.ai (Bengaluru, India)

The enterprise voice-AI platform closed a US$10 million Series B in March, led by Aavishkaar Capital with Info Edge Ventures. Small by this list’s standards, but voice AI in Indian languages is one of the few segments where local players hold a genuine edge over global models.

23. QpiAI (Bengaluru, India)

The quantum-computing and AI company raised Rs 50 crore (about US$5.2 million) in venture debt from InnoVen Capital in October. The non-convertible debentures carry a 13.85 per cent coupon and mature in December 2028. That is a steep price for capital, and a signal that deep-tech founders are increasingly reaching for debt instead of dilution.

24. Inner Sky Labs (Bengaluru, India)

The company behind kids’ robot brand Miko raised Rs 300 crore (about US$30.9 million) in October while launching neuro-symbolic foundation models for physical AI. It is a bet that India can build its own models for machines, not just for chatbots.

Hong Kong and Shanghai ring the bell for China’s AI class

25. Zhipu / Z.ai (Beijing, China)

The GLM model developer listed on HKEX (2513.HK) on 8 January, becoming the first Chinese AI/LLM company to go public. The honeymoon was brief. Shares fell about 23 per cent in late February after the GLM-5 launch ran into compute shortages, a reminder that model quality means little if you cannot serve it.

26. MiniMax (Shanghai, China)

A day after Zhipu, MiniMax priced its Hong Kong IPO at HK$165, raised HK$4.8 billion (about US$615 million) and closed its first day at HK$345, up 109 per cent. In May, it lost a bid to dismiss Disney’s copyright suit, so its public-market investors now own a stake in one of the AI industry’s most closely watched IP fights.

27. Iluvatar CoreX (Shanghai, China)

The GPU and cloud-computing chip designer listed in Hong Kong (9903.HK) on 8 January at a valuation of about RMB53 billion (US$7.4 billion). Domestic GPU makers are the clearest beneficiaries of Washington’s export controls.

28. Axera (Beijing, China)

The AI visual-processing chip company listed in Hong Kong (0600.HK) on 10 February, raising roughly US$379 million, according to 36Kr, with Reuters and TradingView corroborating.

29. Haizhi Technology Group (Beijing, China)

The data-management and analytics firm began trading in Hong Kong on 13 February under stock code 2706, per its HKEX allotment-results filing. Not every AI listing needs a chatbot.

30. Extreme Vision (China)

The computer-vision and analytics company listed in Hong Kong (6636.HK) on 30 March. The listed entity is Shandong Extreme Vision Technology.

Also Read: Malaysia carried Southeast Asia’s IPO market in 2026. The rest of the region went quiet

31. UISEE (Beijing, China)

The L4 autonomous-driving firm listed on the HKEX Main Board (1511.HK) on 20 May. It focuses on full-scenario autonomy for logistics and industrial sites, where driverless vehicles are already earning money.

32. SEER Robotics (Shanghai, China)

The maker of robot controllers and autonomous forklifts raised HK$1.07 billion (about US$137 million) in its IPO and listed in Hong Kong (6106.HK) on 24 June. Warehouse automation, unglamorous but cash-generative, made it to the public markets before most humanoid makers did.

33. StradVision (Seoul, South Korea)

The autonomous-driving perception software company listed on KOSDAQ on 30 June at KRW12,000 a share. It traded below its offer price on day one, a rare blemish in a year when AI listings mostly popped.

34. Enflame (Shanghai, China)

The cloud deep-learning chip maker debuted on Shanghai’s STAR Market (688801.SH) on 11 September and surged 188 per cent. The AI-chip trade in China showed no signs of cooling.

35. Unitree Robotics (Hangzhou, China)

The year’s most theatrical debut. Unitree listed on the STAR Market (688836) on 19 August at a reported valuation of about RMB61 billion (US$8.6 billion) and opened roughly 500 per cent higher. Two months earlier, the US Department of Defense had added it to its list of Chinese military-linked companies. Shanghai investors clearly read that as a badge rather than a warning. The headline raise and valuation figures remain unverified.

Quiet exits, deep cuts and the wrapper that got eaten

36. Verloop.io (Bengaluru, India)

One of India’s early conversational-AI players was acquired in July by Nurix AI, the startup founded by Mukesh Bansal, to expand its AI agents across voice and chat. Financial terms were not disclosed. A decade-old chatbot company folding into an agent-native one is the generational handover in miniature.

37. Sigmacore (Beijing, China)

Z.ai (Zhipu) acquired the AI-chip compilation and optimisation firm on 21 July, the same day it launched a 1GW data centre running on domestic chips. Sigmacore traces its roots to the compiler lab at the Chinese Academy of Sciences’ Institute of Computing Technology. The message was clear: Zhipu wants to own the software layer that makes Chinese chips usable.

38. SqueezeBits (Seoul, South Korea)

The model-compression toolkit maker was bought by Rebellions on 30 June, with Digitimes and The Elec corroborating. Terms were not disclosed. Inference-chip companies increasingly need software that shrinks models to fit their silicon.

39. C2i Semiconductors (Bengaluru, India)

German chip giant Infineon agreed in August to acquire the power-management IC startup, which designs chips for AI data centres. The deal was expected to close in Q3 of calendar 2026, with terms undisclosed. Founded in 2024, C2i is a rare quick exit for Indian semiconductor design.

Also Read: One-person company: Why Southeast Asia’s next unicorn may look nothing like Grab

40. Krutrim (Bengaluru, India)

Ola founder Bhavish Aggarwal’s LLM and AI-cloud venture reportedly cut 20–25 of its roughly 40–50 remaining employees in July, nearly half its staff. The company said it “periodically reviews team structure”. For a startup that launched as India’s first AI unicorn, the shrinking headcount says plenty about how hard it is to compete at the foundation-model layer.

41. Supertone (Seoul, South Korea)

K-pop powerhouse HYBE wound down the AI voice-synthesis company in July after investing about US$35 million. Some Supertone products were transferred to Antinode Audio Inc. Even a company with BTS on its roster could not turn synthetic voices into a business.

42. NeuroPixel.AI (Bengaluru, India)

The Flipkart-backed generative-AI fashion-tooling startup shut down in April after raising about US$1.2 million. CEO Arvind Venugopal Nair cited two reasons: Google’s Nano Banana-class image models now did what NeuroPixel sold, and its largest client had defaulted on payments for more than six months. It is the most honest post-mortem of the year, and a warning to every startup building a thin layer on top of someone else’s model.

October’s late rush: Chips, proteins and a bruising debut

43. FuriosaAI (Seoul, South Korea)

The AI-accelerator maker closed a US$620 million Series D at a US$2.9 billion valuation on 29 October. Korea’s National Growth Fund put in about US$280 million. With Rebellions also backed by the fund, Seoul is now openly picking national champions to take on Nvidia in inference. Tracxn’s US$2.47 billion valuation figure already looks dated.

44. BioGeometry (Beijing, China)

The AI-driven programmable-protein-medicine company closed a Series A+ extension led by China Life Capital in late October. The size was not disclosed, but it pushed cumulative Series A funding past US$100 million. AI drug discovery is crowded, yet investors are still writing large early cheques for platforms that design molecules.

45. Simplismart (Bengaluru, India)

The AI inference and deployment platform raised a US$14 million Series B at about a US$100 million valuation. Dallas Venture Capital led with about US$5 million, and NVIDIA, Accel, Shastra VC and Micromax participated. Having NVIDIA on the cap table is becoming the inference startup’s version of a blue tick.

46. Nanoforgeai (Seoul, South Korea)

Founded in 2025, the materials-discovery startup raised a KRW14.6 billion (about US$10.9 million) seed round on 30 October, led by Mirae Asset Venture Investment. AI for materials science is one of the quieter frontiers that could matter more than another chatbot.

Also Read: Anthropic, DeepSeek and drones: 10 takeaways from the 2026 unicorn boom

47. Omnichat (Hong Kong)

The omnichannel AI messaging platform for e-commerce closed a US$10 million Series A on 27 October, led by Openspace Capital. It is one of the rare entries here with a Southeast Asia-linked lead investor, and a reminder that conversational commerce remains a regional strength.

48. AIShield (Bengaluru, India)

Protectt.ai bought a controlling stake in the Bosch-incubated AI security company on 27 October, with Bosch keeping a minority. The exact split was not disclosed. As enterprises push generative AI into production, securing the models themselves has become a business worth consolidating.

49. Cogknit Semantics (Bengaluru, India)

Defence-technology firm Zen Technologies agreed in October to acquire a 51.98 per cent stake in the AI voice and HR-tech company for Rs 135.15 crore (about US$15 million), through a mix of primary subscription and secondary purchase. Founded in 2009, Cogknit is the oldest company on this list, proof that AI exits do not always go to the newest names.

50. Lablup (Seoul, South Korea)

The AI infrastructure and MLOps platform priced its KOSDAQ IPO at KRW19,000, the bottom of its KRW19,000–22,000 range. It listed on 3 November at a market capitalisation of about KRW276.8 billion (US$200 million) and closed its debut down about 25 per cent, per FourWeekMBA. After a year of AI listings that popped, Lablup’s debut was a cold shower.

The state writes the cheques, geopolitics writes the term sheet

Read the list end to end and three patterns jump out. The first is that governments are no longer bystanders. Korea’s National Growth Fund anchored both Rebellions and Furiosa. Chinese state funds led SiliconFlow’s round. Beijing killed the Manus sale, and Washington blacklisted Unitree. In 2026, AI capital came with a flag attached.

The second is that the money is moving down the stack. Chips, inference software, compilers, robot operating systems and GPU clouds took the largest cheques and the friendliest IPOs. Meanwhile, thin application layers such as NeuroPixel.AI and Supertone were quietly switched off. Owning the plumbing proved safer than owning the tap.

The third is uncomfortable for this region. Of 50 names, the Singapore-headquartered ones are largely Chinese-founded or Chinese-funded. Sharpa’s investors are Alibaba, Tencent, Meituan and JD; SiliconFlow’s are Chinese state funds; Manus was dragged back across the border. Southeast Asia is hosting AI companies, but it is not yet producing them at this scale.

That may change. Robotics testbeds, sovereign-AI programmes and a crop of physical-AI startups are taking shape from Hong Kong to Jakarta. But if 2026 proved anything, it is that the next list like this will be written by whoever controls compute, capital and the regulators. Southeast Asia has about a year to decide which of those it wants a seat at.

—

Data courtesy: Tracxn.

The post Cheques, bans and bell-ringing: 50 Asian AI companies that defined 2026 appeared first on e27.

Posted on — Leave a comment

A 750-chip ceiling: How US curbs redraw the map for Malaysia’s AI data centres

For three years, Malaysia sold the world’s AI giants a simple pitch: cheap land, abundant power and a short drive from Singapore. Washington has now added a fourth item to the checklist, and it is the hardest to tick: proof of who your customers really are.

Kenanga Research warned in a note on Wednesday that the country’s data centre industry faces tighter regulatory scrutiny after the US Commerce Department’s Bureau of Industry and Security (BIS) issued an interim final rule (IFR) on advanced AI chips and remote access to computing power. Its verdict is calmer than the headlines. Demand from US hyperscalers should cushion the blow, but future growth will increasingly favour American customers and developers with no China-linked exposure.

Also Read: Jensen Huang quits Nvidia, and Southeast Asia’s chip hubs inherit the fallout

The rule follows the Megaspeed affair, in which US and Singaporean authorities are investigating the Singapore-based Nvidia cloud partner over suspected diversion of AI chips to China, which the company denies. The scandal has already ended Jensen Huang’s tenure at Nvidia. Southeast Asia’s chip hubs are now being asked to prove that the leak Washington fears does not run through their server halls.

A 750-chip ceiling for anyone serving China

The IFR scraps the country-based quota system drafted under former US president Joe Biden’s AI Diffusion Rule. In its place come narrower but sharper controls. Exports, re-exports and transfers of advanced semiconductors to Malaysia, Indonesia, Singapore, Thailand and Brazil now need a licence, because Washington considers these markets to carry a higher-than-normal risk of diversion.

The more pointed clause targets compute rather than chips. Data centres outside restricted countries that remotely supply significant computing capacity to companies headquartered in Macau or in Country Group D:5, the nations under a US arms embargo that include China, must obtain a licence. The threshold is low: 12 million total processing power (TPP), roughly 750 Nvidia H100 SXM chips. Kenanga reckons that ceiling would make it hard for such facilities to assemble clusters large enough for meaningful AI model training.

Washington has also anticipated the obvious workaround. The rules cover digitally linked infrastructure, so spreading a cluster across several facilities to stay under the line will not work. Data centres outside restricted countries are also barred outright from offering advanced AI model training services to companies based in those jurisdictions.

The penalties for getting it wrong are not subtle. Kenanga listed import restrictions, extra licensing requirements, a downgrade of a country’s status under the US export-control regime and Entity List designations for the companies involved.

Kuala Lumpur moved first, but paperwork is not policing

Malaysia is not starting from zero. Since July 2025, the Investment, Trade and Industry Ministry (MITI) has required a Strategic Trade Permit for the export, re-export, transshipment or transit of US-origin high-performance AI chips through the country. The government is also weighing rules that would make data centres disclose their GPU imports, operators and locations.

That looks prescient now. The region’s chip-hub ambition has been colliding with its chip-smuggling problem for months, and governments have been trying to show Washington they can police their own corridors before Washington does it for them.

Also Read: The Nvidia clampdown is a warning for Southeast Asia’s AI boom

Still, permits on paper and verification in practice are different animals. A licence regime pushes the burden onto operators to know who sits at the far end of the cable. That is easy for a facility leased wholesale to Microsoft. It is far harder for a colocation provider renting racks to a web of regional resellers.

American hyperscalers are the safety net, for now

Kenanga’s case for resilience rests on US demand. Some American states have grown wary of approving new data centre projects, nudging hyperscalers to look abroad, and Malaysia has 12.3 gigawatts (GW) of electricity supply secured across its project pipeline. Big cheques keep arriving, from Google’s US$15 billion Malaysia commitment to SoftBank and Grab’s 5GW AI bet.

The research house also argues the new regime is gentler than the AI Diffusion Rule, which would have sorted countries into tiers and capped their computing power. In the near term, it expects investors to separate companies serving mostly US clients from those that do not.

That sorting is already visible in construction. Most data centre clients of the contractors Kenanga covers are US hyperscalers such as Google, Amazon Web Services and Microsoft, or developers including Yellowwood, Stack and Vantage. The exception is Sunway Construction Group’s (SunCon) project with K2, a non-US-linked developer. At RM748 million (about US$175 million) and nearing completion, the exposure is limited.

Utilities look safer still. Tenaga Nasional (TNB) is protected by take-or-pay clauses in its electricity supply agreements for at least the first five years. YTL Power’s colocation business could catch demand spilling over from the US, though Kenanga says it may need to be choosier about whom it signs.

Delays, not cancellations, are the real threat

The pain sits further down the chain. Stricter end-user verification and GPU licensing could delay new and ongoing projects, Kenanga warned, especially those run by non-US operators without confirmed end-users or offtake agreements. That would slow order-book replenishment for mechanical and electrical (M&E) contractors and weigh on sentiment. The research house prefers M&E firms with diversified books and exposure to TNB’s grid expansion.

Malaysia’s integrated circuit (IC) design industry has reason to watch too. The rules do not ban IC design services, but they add licensing requirements for certain advanced logic chips classified under Export Control Classification Number (ECCN) 3A090.a when designed by companies that are not approved IC designers. Those requirements can bite when front-end manufacturers or outsourced semiconductor assembly and test (OSAT) providers ship the chips, which is squarely Malaysia’s home turf. The worldwide reach of the advanced computing Foreign Direct Product (FDP) Rule also stays, keeping some foreign-made chips under US control.

Kenanga called the direct hit to IC design limited for now, but said downstream manufacturing and the delivery of finished chips need close monitoring. For a country already selling more chips without winning more of the market, that is not a footnote.

Neutrality was the product. Now it is a liability

The deeper shift is about identity. Malaysia’s data centre pitch was built on neutrality: a place where anyone with capital could plug in. The new rules turn that neutrality into a risk. Growth will keep coming, Kenanga says, but it now comes with a passport check, and the passports that clear fastest are American.

Also Read: AI’s bottleneck economy is taking shape across Southeast Asia’s chip supply chain

That leaves an awkward question for a country that has spent the year talking up sovereign AI: whose AI will its data centres be allowed to build? For founders and investors weighing a stake in the region’s data centres, the answer now matters as much as the price of power.

The post A 750-chip ceiling: How US curbs redraw the map for Malaysia’s AI data centres appeared first on e27.

Posted on — Leave a comment

The crypto hype cycle is dead. Here are the 4 narratives replacing it

The fatigue you diagnosed is structural, not cyclical. The sector has run out of patience for projects that promised user utility and delivered clunky interfaces, empty virtual worlds, and speculative profile pictures. Play-to-earn gaming collapsed under unsustainable incentives. Metaverses remained ghost towns. Generic PFP NFTs lost their cultural premium.

The first generation of real-world asset protocols mostly wrapped standard yield products in opaque legal structures that retail investors could not trust. That trust deficit now forces a rotation.

Capital is not leaving. It is moving toward themes that solve genuine problems, generate revenue, or serve professional demand. The four directions you identified capture that shift with unusual clarity. I want to explain why each one wins on its own terms.

AI agents need a bank account, and crypto can give them one

The first direction is the agentic economy. I consider it the most important because it removes the fickle human from the equation. Large language models and autonomous AI systems have become extraordinarily good at processing information. They still lack legal identities, bank accounts, and credit histories. They cannot open a corporate account at a legacy bank. They cannot sign a traditional contract. They cannot build a payment history that a credit bureau would recognise. They can operate software.

The agentic economy gives them programmable, permissionless rail money, primarily stablecoins, so they can pay each other for compute, APIs, data processing, and micro-services without asking a person to click a button. Early Web3 gaming missed this point. Instead of begging players to log into a blockchain game, the agentic economy thrives on millions of autonomous machine-to-machine micro-transactions every second.

Virtual Protocol shows developers how to generate, tokenise, and monetise AI agents as economic actors. High-throughput layer one networks provide settlement capacity for agent-to-agent payments. This theme offers the cleanest escape from retail fatigue because it does not depend on consumer sentiment. It depends on software needing to pay for services, a structural need that will only grow.

DePIN turns tokens into infrastructure that earns real revenue

DePIN follows as the second direction, and it works because it answers the oldest criticism of crypto, which is that the industry operates as an insular casino with no tangible utility. DePIN projects use token incentives to coordinate communities that build physical infrastructure.

These are not abstract governance tokens. They are crowdsourced hardware networks that supply compute, bandwidth, and storage to enterprise clients who pay for access. This shift wins because it generates hard, structural revenue from actual usage rather than speculative roadmap promises.

Also Read: Bitcoin is bruised, not broken: Why the crypto sector is holding its ground

The trading arena actively rewards that distinction. Decentralised compute marketplaces such as Akash Network, Render Network, and Aethir supply hardware to the AI boom at price points that legacy cloud providers struggle to match. Enterprise clients do not care about crypto ideology. They care about cost, reliability, and access. DePIN delivers all three by bypassing corporate telecom and data monopolies.

I see this as the antidote to play-to-earn decay because the demand side is commercial rather than mercenary. When a game studio or an AI inference provider buys compute from a decentralised network, that revenue recurs and carries substance. The model also aligns incentives across hardware owners, software developers, and enterprise buyers, which gives the network a stronger base than a purely speculative token launch.

Value is moving from protocols to the apps that capture fees

The third direction is the rotation into high-throughput applications and perpetual decentralised exchanges. I strongly agree that the fat protocol era is ending. For years, capital poured into layer one infrastructure because the prevailing theory held that base layers would capture most of the value. That theory made sense when block space was scarce.

Block space now is abundant and cheap. Value is shifting to the application layer, specifically to applications that natively capture transaction fees. Traders do not care about protocol politics. They care about settlement speed, access, and liquidity. They prefer platforms that settle in milliseconds, bypass geographical restrictions, and capture organic liquidity pools without requiring complex gas management.

Hyperliquid has demonstrated how a high-performance orderbook decentralised exchange can scale by focusing on the trading experience rather than infrastructure branding. Polymarket has shown that structural forecasting networks can build durable liquidity around real-world events. The end of the fat protocol era does not mean infrastructure is worthless; it means infrastructure is becoming commoditised, and the applications on top of it now capture the fees.

I find this shift compelling because it aligns crypto with the same logic that governs traditional finance. Execution venues that offer speed and access win. Those that do not lose. Application teams that own the user relationship and settle trades quickly now hold the stronger hand.

Compliant stablecoins are the quiet giant

The fourth direction is compliant stablecoins and sovereign tokenisation. I believe this is the quiet giant of the group. The raw retail-facing real-world asset trend may feel synthetic. Professional cash is reshaping the landscape through permissioned infrastructure. Tokenised capital is moving past the experimental phase.

Mega-banks and institutional clearinghouses such as the DTCC and JPMorgan are building shared tokenised deposit networks and compliant stablecoins backed by strict legislative frameworks like the US GENIUS Act and the EU’s MiCA framework. This is not retail speculation. This is a multi-trillion-dollar professional pipeline flowing silently underneath the public crypto sector, completely decoupled from hype cycles.

Also Read: Hackers drain US$4M using fake Wi-Fi at major crypto conferences

The GENIUS Act establishes reserve and disclosure requirements for payment stablecoin issuers. MiCA brings stablecoins within a regulated perimeter in Europe. Banks are not waiting for retail sentiment to improve. They are building settlement rails that their corporate clients will use because those rails are faster and more transparent than legacy correspondent banking. It does not need a viral moment. It needs regulatory clarity, and it now has it.

Together, these four directions tell a coherent story. Crypto is moving away from forcing consumers to play clunky blockchain games and toward invisible plumbing, machine-to-machine utility, and strict regulatory integration. The speculative bubbles of early Web3 are bursting, and that is healthy. The capital rotating into AI agents, DePIN, high-throughput applications, and compliant tokenisation is not chasing hype. It is chasing revenue, utility, and professional adoption.

My view is that this rotation will produce a more durable sector because it does not depend on retail euphoria. It depends on software paying for services, enterprises buying compute, traders demanding speed, and banks building settlement networks. Those are structural forces, and they will outlast any single cycle. The fatigue you diagnosed is not the end of crypto. It is the end of an immature phase and the beginning of a more serious one.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post The crypto hype cycle is dead. Here are the 4 narratives replacing it appeared first on e27.