ByteDance competes with Cursor's AI Code Editor; The U.S. marks Singapore as Tier-2
January Week 3: Jan 21 - Jan 27
Hi friends 👋,
In this week’s edition of Coconut Capitalists, we’re diving into:
ByteDance’s AI-Powered Code Editor, the Cursor of Asia
America’s new AI chip export controls on China & Singapore
Quick-Fire Startup News from Around Asia
Let’s get into it.
ByteDance’s AI-Powered Code Editor, the Cursor of Asia
The Scoop
ByteDance has launched a competitor to Cursor’s wildly popular AI-powered code editor. Cursor, a San Francisco-based startup that’s worth over $2.6 billion, has over 30,000 enterprises as customers and has grown annual revenue from $4 million to $100 million in just eight months.
If you've never tried Cursor before, it's essentially like having ChatGPT built directly into your code editor. Cursor acts as an intelligent pair-programming assistant, analyzing your entire codebase & making improvements to your code's quality in real-time as you type. It's worth noting, that on average customers of Cursor have reported a 37% increase in software development speed & 22% fewer bugs.
However, while Cursor has taken off in Western markets (particularly in the US & Europe), many of the developers across Asia have never heard of the product - this gives ByteDance a massive opportunity. And they've taken it by introducing Trae AI.
AI-Native Products
On paper, ByteDance is the perfect company to build & sell this product in Asia:
First, they have Doubao-1.5 Pro, their in-house, lightning-fast language model that can power Trae AI. While Cursor lets you plug-and-play with different AI models like Claude-3.5-sonnet & OpenAI’s GPT-4o, ByteDance can take a different approach - reducing variable costs significantly by instead building a full-stack, vertically integrated, Cursor alternative with Doubao-1.5 Pro.
Second, ByteDance already has significant enterprise distribution. Their product Lark (which is very similar to Google Workspace or Microsoft 365) has over $400 million in annual revenue and is incredibly popular across Chinese, Japanese, and Singaporean enterprises. Packaging Trae with Lark would be a major upsell opportunity for ByteDance and would create a new revenue stream with their existing customer base.
But why can’t Cursor simply go after Asian enterprises? Why does it have to be ByteDance? Obviously, the Cursor team is very good - capturing $100 million in annual revenue & now with over 30,000 enterprises as customers - so the answer is yes - they definitely can capture this market too! However, for many enterprises operating across Asia - the issue is they may be hesitant to give direct codebase access (their most important intellectual property) to a “relatively new” American startup - and one that doesn’t have a local presence in the region. Because of this, ByteDance may have an advantage. Now, I do expect Cursor to eventually partner with a trusted technology vendor like Microsoft Azure, AWS, or GCP which all have extensive local operations across Asia, in order to gain access into these enterprises.
Why it Matters
ByteDance is not a new company. Founded in 2012, it already has a substantial portfolio of products from social media (TikTok, Douyin, etc.) to gaming (Mobile Legends) to enterprise software (Lark). Yet, even with their significant size, their product velocity is exceptional - in being able to ship dozens of new products every year. They move much faster than the other tech giants: Tencent, Alibaba, & Baidu. And their talent density rivals Western companies like OpenAI, Anthropic, & DeepMind. Cursor has the lead, but Trae has a real shot to be a worthy competitor.
America’s new AI chip export controls on China & Singapore
The Scoop
In Joe Biden's final week as US president, his administration signed a new piece of export control legislation, limiting access of Nvidia’s AI GPUs to non-American or non-allied nations. The idea behind this legislation is that it will help American AI companies & close allies of the US, to stay ahead of the competition by giving them access to the world's largest, leading-edge AI computing clusters.
The legislation (which is to go into effect in 100 days) places foreign countries—and companies based in those locations—into one of three tiers:
Tier 1, has almost no export restrictions. These are the US's 18 closest allies like Canada, Korea, Japan, and the majority of the European Union.
Tier 2, has limited access, allowing the equivalent of 50,000 Nvidia H100 GPUs over the next two years to each of the countries on the list (and the companies that are based there). This includes neutral countries (that seemingly weren’t considered allies or adversaries by the Biden Administration) such as Singapore, India, and Mexico.
Tier 3, has no access - it’s a complete ban. This includes leading-edge Nvidia AI chips such as the B100s & H100s, and surprisingly even 5-year-old & 8-year-old chips like the A100s & V100s. A few of the countries are China, Russia, and Iran.
It was expected by many that Donald Trump, who recently took office as the 47th US President, would rescind the Biden Administration's legislation on AI chip exports - in order to ease the very hot global trade tensions, particularly with China. However, this did not happen - The Trump Administration has so far kept the legislation (even though the new administration did repeal a few of Biden’s AI Safety Laws). It’s worth noting that Nvidia disavowed the administration’s decision with an extremely blunt press release, where their executive team called the legislation “misguided”.
Singapore & the Issue for Tier 2 Countries
We all know you can't have more than one best friend, right? Many of us learned this on the playground as kids. And for a country like Singapore, it's even more difficult to maintain close friendships with two countries that are economic enemies. Singapore's attempt to stay neutral (as an ally of both China & the US) is showing a few cracks.
While I’m not a foreign relations expert, it’s clear to anyone living in Singapore that over the past 70 years, the city-state has done a remarkable job of staying neutral on the world stage—collaborating with everyone from Taiwan to China, from North Korea to the US—and miraculously avoiding making enemies. However, for the first time, the US is signaling concern about Singapore’s technology ties with the Chinese government. In simpler terms, the US doesn’t like that Singapore has “two best friends”, and seems to be nudging them towards an ultimatum.
Why it Matters
It’s worth noting that the 168 page piece of legislation has one loophole that might just quickly turn it into Swiss Cheese. And that’s that there’s nothing really stopping an entity from a tier-2 or tier-3 based country from spinning up a US based Delaware C corp and renting leading-edge chips - essentially running a large model training process remotely (via ssh) but on US soil.
Now, for the three major international cloud providers (Microsoft Azure, AWS, or GCP) they have the ability to KYB (know your business/background check) every company that wants to run a large training job on their GPU clusters. And they have the 1000+ person legal/compliance departments to spot obfuscated connections between a US shell company & a tier-1 or tier-2 based entity. However, for newer AI infrastructure Together AI, Lambda Labs, and the San Francisco Compute Company this could create a massive headache (they have smart folks - so they’ll probably be fine!).
However, let's say that I happened to be a Chinese National - and I had a Beijing-based company called "Colin's Super Secret AI Lab" (sounds cool, right?). I could simply create a Delaware C Corporation (as a Chinese National), and rent massive clusters of Nvidia GPUs for training through one of these (IAAS) providers (without them probably ever noticing). And if this strategy doesn’t work out, there’s even a very popular “craigslist for GPUs” service called gpulist.ai where anyone can rent access to leading-edge Nvidia chips on-demand.
🇸🇬 Singapore News
Mistral has announced its plan to open a Singaporean headquarters in 2025. At the World Economic Forum event in Davos, Switzerland, CEO Arthur Mensch stated that the company is quickly expanding its presence in APAC. For many enterprises across Asia, they require custom or on-premise AI solutions, and Mistral's hands-on deployment approach has the chance to catch fire among this customer base.
🇰🇷 Korea News
Friendli AI, one of Korea's leading AI startups, which is founded by the famous Seoul National University Professor, Byung-Gon Chun, and provides “GPU infrastructure as a service” for AI model inference & fine-tuning workloads, has settled a lawsuit with Brooklyn-based Hugging Face over patent infringement. In the suit, filed nearly two years ago, Friendli AI alleged that Hugging Face violated its AI inference patent on "batching with iteration-level scheduling."
It's worth noting that the resolution, that occurred after the settlement was complete, ended up being incredibly positive for both companies. The next day, the Friendli AI CEO announced a strategic partnership with Hugging Face, now allowing one-click deployments of almost any Hugging Face model directly onto Friendli AI's infrastructure. This is incredibly valuable for both companies - Friendli AI gets western eyeballs experiencing their products for the first time, and Hugging Face gets introduced to a market (South Korea) where they're still relatively unknown.
Upstage AI has released a new & improved AI model for converting complex charts and graphs from PDF documents into raw HTML. This unlocks an incredibly important use-case in being able to feed scattered visual information into a language model. And the product has industry-leading processing time, of only 0.84 seconds on average, per PDF document. If you're new to Upstage AI, it is Korea's leading foundation model company & they raised $72 million in their last round of funding.
SK Hynix reported $49.6 billion in revenue for 2024, more than doubling its 2023 revenue of $22.81 billion. For context, SK Hynix produces high-bandwidth memory (HBM) chips, which are critical components for leading-edge GPUs from the likes of Nvidia, AMD, Intel, and Broadcom. It's worth noting that their flagship memory chips like HBM3 and HBM3E are reportedly already sold out for 2025.
🇯🇵 Japan News
OpenAI & SoftBank have formed Stargate - a joint venture between OpenAI, the Japanese conglomerate SoftBank, Larry Ellison’s Oracle, and the Middle Eastern sovereign wealth fund MGX. The group aims to spend up to $500 billion on the financing of AI data centers - based in the state of Texas (USA) - over the next four years. It's been widely reported that SoftBank and OpenAI will each commit $19 billion of funding to the joint venture.
The scale of this data center project is unprecedented. On Wednesday, OpenAI's CEO likened Stargate to a venture fund, with OpenAI and SoftBank being two of the general partners - where the two other general partners are Oracle and MGX. Essentially the GPs will commit $45 billion in total to the project (implying that $7 billion would come from Oracle and MGX combined). Moreover, the remaining $455 billion investment into Stargate will come at a later date from investors categorized as “LPs”, in addition to several types of debt & equity financing, as told by OpenAI’s CEO.
🇨🇳 China News
Researchers from the Shenzhen Institute of Advanced Technology & Nanjing University have introduced a new AI Model called VideoChat-Flash, which aims to increase processing speeds for video understanding by 100x. The team, which released an open-source model & research paper, states that they discovered a new compression technique (Pied Piper?!?) which shrinks long videos into shorter ones through a process called "HiCo" (Hierarchical visual token Compression). The model achieved a 99.1% accuracy on the NIAH (Needle-In-A-video-Haystack) benchmark (which is a test that challenges AI models to find specific details in very long videos - similar to finding a needle in a haystack).
An example use-case for the technology could be catching bad guys via CCTV cameras. Currently, security teams around the world spend countless hours manually reviewing CCTV footage to identify criminal activities. And still, far too often criminals slip through the cracks due to the sheer volume of footage that needs to be analyzed by humans. The team behind VideoChat-Flash (along with some great startups called Twelve Labs & Anduril) has the potential to finally solve this problem.
Elon Musk met with China's Vice President Han Zheng during the week of the US Presidential Inauguration. During the meeting, Vice President Han stated that he "welcomed US companies including Tesla to seize opportunities, share the fruits of China's development, and make new and greater contributions, with ever-closer economic & trade ties between China and the United States." Elon later recapped the conversation via a post on his platform X (fka Twitter) stating that: "I have been against a TikTok ban for a long time, because it goes against freedom of speech. That said, the current situation where TikTok is allowed to operate in America, but X is not allowed to operate in China is unbalanced. Something needs to change."
🇮🇩 Indonesia News
Apple and the Indonesian Government are nearing a deal to lift a 4-month ban on iPhone 16 sales. Indonesia enforces strict requirements for tech products, mandating that 40% of components be sourced or manufactured domestically.
Apple has struggled to comply with the law, leading to an iPhone 16 sales ban that came shortly after its global launch. Negotiations now center on Apple establishing local manufacturing facilities, likely in partnership with Indonesian firms. As Apple's international growth slows, access to Indonesia's 270 million-strong consumer market (with a 96.3% smartphone penetration rate over 16 years old) - and an ever-increasing middle class - will be critical for Apple's future.
BTW, congrats to the DeepSeek AI team (a few of you read this newsletter - thank you!) on achieving Number #1 in the U.S. App Store - it’s an incredible feat 💯