Lead storyThe Information
DayOne Data Centers, a Singapore-based operator spun out of China's GDS Holdings in late 2024, filed a prospectus last week for a New York IPO that could raise as much as $5 billion, with part of the proceeds funding new data-center projects. The filing shows ByteDance, TikTok's parent, was 69.2% of revenue in the first half of 2026, and names Oracle as the second-largest customer at 15%. DayOne is pitching growth beyond that one relationship: ByteDance has committed to 838 megawatts of capacity across its sites in Malaysia, Thailand, and Finland, and the company is chasing Western AI business, including a Batam, Indonesia project with Nvidia-backed Firmus Technologies using more than 170,000 Nvidia chips, expected to serve Meta.
Why it matters
One customer paying seven of every ten dollars is a real risk, and the filing admits it. The deeper signal is timing: Firmus just scrapped its own $5.5 billion IPO and AirTrunk postponed its Singapore listing, so data-center money is harder to raise than a month ago. For founders and investors, watch whether a $5 billion listing clears. If it does, it reopens the door for every other AI infrastructure name waiting in line.
Story 0220VC
On Thursday's 20VC news episode, Jason Lemkin and Rory O'Driscoll walked through the week's money moves: NVIDIA is reportedly eyeing a roughly $30 billion deal for Reflection while Amazon circles chip startup Decart; Isomorphic Labs is reportedly eyeing a $40 to $50 billion valuation; and 42 new decacorns were minted in 2026, passing 2021's record. On the other side: Anthropic reportedly cut Haiku pricing by 90%, Cloudflare cut 20% of staff despite strong growth, Flock Safety cut 270 jobs, and nuclear startup Valar Atomics sued its pre-seed investor Day One Ventures over pro-rata rights. The NVIDIA/Reflection, Isomorphic, and Haiku pricing figures are claims discussed on the show, not confirmed official announcements.
Why it matters
The same week that minted 42 companies worth $10 billion or more also saw a 20% layoff at a healthy company. That is the market in one frame: private AI valuations are still running hot while operators cut costs and slash prices. For founders, two lessons. Price pressure is real even on frontier products. And investor rights like pro-rata are now worth fighting over in court, so read term sheets closely.
Story 0320VC
LangChain co-founder and CEO Harrison Chase joined 20VC today for an hour-long interview on where value accumulates in the agent stack. The agenda: why building reliable AI agents is still hard, whether model labs will eat the agent stack, whether model routing is a real business or a commodity, why no one has solved AI memory, and why AI bills keep exploding even as tokens get cheaper. He also weighed in on whether companies should fear Chinese AI models, where he would not invest today, whether founders should take valuation discounts for tier-1 firms, and a rapid-fire 'OpenAI at $1.5T or Anthropic at $2T' question. LangChain's scale anchors the discussion: roughly $160 million in announced funding with the last round at a $1.25 billion valuation, 35% of the Fortune 500 as customers, more than a billion open-source downloads, and over a billion events a day through LangSmith.
Why it matters
These are the questions every AI startup should be able to answer in plain language: if tokens keep getting cheaper but your bill keeps rising, your value is leaking somewhere in the stack. Chase's bet, and the bet behind vertical agent companies, is that the real moat sits in the agent layer, not the model. Watch where customers keep paying as tokens approach zero. That is where the next big software companies get built.
Story 04Anthropic
Anthropic opened its fifth Asia-Pacific office in Singapore and signed a memorandum of understanding with the Ministry of Digital Development and Information covering AI safety, model evaluation, and responsible adoption. The post carries new traction numbers: Singapore ranks second of 121 countries in per-capita Claude.ai usage, at nearly six times what its population alone would predict. More than 1,500 GovTech Singapore engineers now use Claude Code to modernize government systems, and DBS, GIC, Grab, Sea, and Supabase are all live customers. Grab's BriX sales assistant alone saves its teams roughly 40,000 hours a quarter, and regulated firms can now run Claude with in-region inference on Amazon Bedrock.
Why it matters
Read this as enterprise adoption data wearing an office-announcement costume. When a country's government engineers, its biggest bank, and its sovereign wealth fund all run on your product, the sale is no longer about demos. It shows how fast frontier labs are converting Asia-Pacific demand into locked-in contracts, and it sets the enterprise bar for 'safe and reliable' at government-grade vetting. Any AI tool selling into regulated industries is now competing with that standard.