Post Title:
The chip export controls have been the centerpiece of U.S. AI policy for three years. This week Forbes reported that the input nobody thought to restrict has been flowing to Beijing the entire time — for hundreds of millions of dollars a year. Plus: Canva discovers AI economics the hard way, Meta enters the coding agent wars, and “We cannot drink DATA.”
Want more AI news without waiting for me to finish yelling at the latest development? Check out The SkyNet Tribune, my continuously updated AI news page featuring fresh headlines from across the web every hour. It is faster than a daily roundup, less exhausting than social media, and only slightly more likely to become self-aware.
🕵️ Silicon Valley’s Data Labelers Are Selling to Chinese Labs Too
Forbes reported Wednesday that Mercor and Surge AI — the startups supplying high-quality training datasets to OpenAI, Anthropic, and the U.S. government — have been selling the same kind of data to China’s leading AI labs. The reporting is based on documents and communications from Chinese lab buyers and employees at the U.S. platforms.
The scale isn’t marginal. Two data-labeling entrepreneurs, citing market estimates from Tencent and ByteDance executives, put Chinese labs’ spend with American data-labeling companies at roughly $500 million per year. Surge AI actively courted the business, with CEO Edwin Chen traveling to China to meet lab executives directly.
Here’s why it matters strategically: the U.S. has spent three years restricting China’s access to the chips that train frontier models, while leaving the expert human annotation those models are trained on completely unregulated. Four companies — Scale, Surge, Mercor, and Handshake — account for more than 75% of a training-data market worth about $8.5 billion in revenue. Export control has a shape, and this is the hole in it.
💸 Canva Cut Its Revenue Forecast Because AI Costs Too Much to Serve
This one deserves more attention than it’s getting. Canva slashed its annual revenue growth forecast from 30% to around 20% — a third of its growth target, gone — after finding that serving its AI suite at scale cost far more than projected. The $42 billion company deliberately slowed the rollout of Canva AI as a result.
CEO Melanie Perkins was unusually direct in her shareholder letter: the “average cost of serving an AI task was too high.” Quarterly revenue still climbed 25% year over year to $921.9 million, so this isn’t a demand problem. It’s a unit economics problem.
Every company shipping AI features is running the same calculation, and most are absorbing the cost quietly. Canva is the first at this scale to say out loud that the math didn’t work and pull back. Adding to the pressure: Canva has been a top-10 referred domain from ChatGPT, which means the assistant driving traffic to them is also increasingly a substitute for them.
💻 Meta Jumps Into the Coding Agent Wars With Muse Code
Meta Superintelligence Labs released Muse Code in beta, a terminal coding agent powered by the new Muse Spark 1.2 model. It installs with a single command and takes on whole engineering jobs across large repositories — planning the change, writing the code, verifying the result. The architectural pitch is persistent async background agents that stay alive across a session instead of being spawned per task, which cuts down redundant context gathering.
On DeepSWE 1.1 it scores 59.3%, behind Claude Opus 5 at 65.0% and GPT-5.6 Terra at 64.8%. Solidly in the game, not leading it.
The pricing structure is where you should pay attention. Standard tier runs $1.25 per million input tokens and $4.25 per million output. The Contributor tier drops to $0.30 per million — but code submitted under that tier becomes training data for Meta’s models. If you’re working on anything proprietary, that discount is not a discount.
🔊 OpenAI’s First Device Is a $300 Donut
Bloomberg’s Mark Gurman reports that OpenAI’s long-teased hardware is a doughnut-shaped, hockey-puck-sized smart speaker with no display, arriving in 2027 at $300–$400. It’s battery-powered and portable, with a camera, sensors, lights, and moving parts intended to give it personality — OpenAI reportedly wants it to feel “more alive” than a stationary speaker.
It’s being developed with LoveFrom, Jony Ive’s design studio. Functionally it works like ChatGPT voice mode with more advanced models behind it. Whether the world wants a $400 Alexa with feelings is the open question, but the moving-parts detail suggests OpenAI thinks the differentiator is presence rather than capability.
💧 “We Cannot Drink DATA”: Google’s $15B India Hub Hits Court
Google’s Visakhapatnam AI hub is facing a court challenge and street protests over water and wildlife. The Andhra Pradesh High Court has asked the state government to respond to allegations from activist group Jal Biradari that the gigawatt-scale campus — built with AdaniConneX and Airtel — will strain a nearby reservoir.
The numbers explain the anger. Visakhapatnam takes in about 410 million litres of water a day against demand of 480 million. A city of 2.5 million people already rations water, and a very large new customer is arriving. The petition also flags construction and noise near Kambalakonda Wildlife Sanctuary, which sits 860 metres from the site and is home to leopards and pangolins.
Activists and children have marched holding banners reading “We cannot drink DATA,” with handcuffs painted over the Google logo. Google says it will use advanced air cooling to protect local water and add sound-dampening to be “a quiet, unobtrusive neighbor.” The state government says no rural or residential water will be diverted.
📢 Somebody Just Raised $30M to Put Ads Inside Chatbots
Gravity, which places text-based ads within AI chatbot responses, raised a $30.5 million Series A co-led by Lightspeed and Committed, bringing total funding to $38.5 million. File this one away. The business model of the consumer internet has arrived at the assistant layer, right on schedule, and how that gets disclosed to users is going to be the fight of 2027.
🌟 What This Means
The Forbes story reframes three years of policy. Export controls treated compute as the bottleneck and assumed everything else was fungible. But frontier performance increasingly comes from expert human annotation and carefully built RL environments — and that market is four American companies deep, selling to whoever pays. You can embargo a chip at a border. Nobody built the machinery to embargo a spreadsheet of expert labels.
Canva and Google’s India problem are the same story told in different currencies. The cost of AI is showing up on real balance sheets and in real reservoirs, and both are places where somebody eventually says no. Canva’s shareholders got a letter about unsustainable serving costs. Visakhapatnam’s residents made a sign.
Meanwhile Meta is buying training data with a pricing tier, and Gravity is raising money to sell your attention inside a chat window. The infrastructure phase is ending. The extraction phase is well underway.
Blurb (for excerpt/preview):
Forbes found that Mercor and Surge AI — the data labelers supplying OpenAI, Anthropic, and the U.S. government — are selling training datasets to Chinese labs to the tune of $500 million a year, exposing the gap America’s chip controls never covered. Plus: Canva cuts its growth forecast a third because AI costs too much to serve, Meta ships Muse Code, OpenAI’s first device is a $300 donut, and Google’s India data center lands in court.
Sources:
- Silicon Valley’s Other China Problem: It’s Training Their AI — Forbes
- Canva cuts revenue forecast by a third as it tackles high AI costs — Startup Daily
- Canva Cuts Revenue Forecast Following AI Cost Blowout — B&T
- Introducing Muse Code and Muse Spark 1.2 — Meta AI Research
- Meta enters the AI coding wars with Muse Spark 1.2 and Muse Code — VentureBeat
- OpenAI’s ring-shaped smart speaker will reportedly cost between $300 and $400 — Engadget
- OpenAI’s ChatGPT Speaker Will Be Hockey Puck-Sized and Cost Over $300 — MacRumors
- Google’s $15bn India data centre is running into water and wildlife — TNW
- Google’s $15 billion India data center project battles water, wildlife concerns — BusinessWorld