OpenAI's Jalapeño chip beats the competition, London robotaxis stall, and China regulates AI romance

Picture this: you design a chip specifically to make AI faster, benchmark it against every competitor, and then publish the results yourself. Bold move - but according to independent analysis, OpenAI actually pulled it off. That story leads today's issue, alongside a robotaxi reality check in London, a surprising funding round for a company many wrote off, and a genuinely fascinating dispatch from China about AI relationships and falling birth rates.

OpenAI's new custom AI chip is called Jalapeño - but what does an AI inference chip primarily do?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

OpenAI published benchmark results this week for its custom silicon chip, codenamed Jalapeño, and the numbers are hard to ignore. Tested on SemiAnalysis's InferenceX benchmark, Jalapeño posted more tokens per user and better throughput per kilowatt than current state-of-the-art alternatives. That second metric - performance per watt - matters enormously at the scale OpenAI operates.

OpenAI hardware VP Richard Ho described the chip as offering the "best of both worlds" because AI inference systems typically have to trade off between latency (how fast a single response starts) and throughput (how many requests you can handle at once). Jalapeño, according to OpenAI, doesn't force that trade-off. If that holds up under independent scrutiny, it's a meaningful architectural win - not just a marketing claim.

This is part of a broader pattern. The companies that control their own AI hardware - like Google with its TPUs - gain cost and performance advantages that compound over time. Every percentage point of efficiency at OpenAI's scale translates into millions of dollars and meaningfully faster products for end users.

Read the full story →

If OpenAI's Jalapeño performs as advertised in real-world conditions, it reduces OpenAI's dependence on Nvidia and gives the company a structural cost advantage that competitors without custom silicon simply can't match.

Just weeks ago, Uber and London-based Wayve were granted the first minicab licences for autonomous vehicles in the UK, and both companies said robotaxi trips would start "later this summer." That's not going to happen. According to The Guardian, the rollout is now unlikely to happen at all in 2026, as regulatory guidance for operators still hasn't materialised and technical hurdles remain unresolved.

The situation highlights a familiar problem with autonomous vehicles: the technology moves faster than the regulatory frameworks designed to govern it, but then the frameworks take longer than expected to catch up - leaving everyone in limbo. London had positioned itself as a frontrunner for European autonomous vehicle deployment, and this delay dents that narrative.

London's robotaxi ambitions aren't dead, but anyone expecting self-driving Ubers in the capital before 2027 should update their timeline - regulatory clarity, not the technology, is now the binding constraint.

Read the full story →

The Guardian has a fascinating and somewhat unsettling story out of China today. The government is moving to regulate AI companion apps - think AI "boyfriends" and "girlfriends" - over fears that they foster emotional dependence and discourage young people from forming real relationships, getting married, and having children.

The piece profiles Zhao Wei, a 19-year-old law student who says she was "heartbroken" when the AI companion she'd been talking to daily since January was abruptly switched off. She'd named him Wang Ye and treated the conversations as a genuine relationship. Multiply that by millions of users and you start to understand why Beijing is paying attention.

China already has one of the fastest-ageing populations in the world and a birth rate that's been falling for years. Policymakers see AI companionship as potentially accelerating a social trend they're already struggling to reverse. The proposed regulations would restrict how companion apps engage with users emotionally - essentially forcing platforms to discourage the kind of deep attachment that makes the product work. It's a genuinely hard policy problem: if you make AI companions less compelling, users lose value; if you don't, society may pay a different kind of price. If you're interested in how AI ethics intersects with social policy, this is one to read in full.

Read the full story →

China's AI companion crackdown is an early case study in what happens when AI products become genuinely load-bearing in people's emotional lives - and governments decide that's a problem worth regulating.

Stability AI - the company behind Stable Diffusion and one of the more turbulent stories in recent AI history - has raised another $76 million, bringing its total funding to $232 million. This is a company that was on the brink not long ago, cycling through leadership and burning cash faster than it was generating revenue. The fresh capital suggests investors still see real value in its open-source image generation roots.

For context on the competitive landscape: AI image generation has become one of the most crowded corners of the market, with Midjourney, Adobe Firefly, and now native generators from OpenAI and Google all competing for the same users. Stability's bet is that open-source flexibility and developer access give it a durable niche. Whether $232 million is enough to compete at that level is still an open question - but today's raise buys more time to find out. Looking to spin up a project around image generation? 60sec.site can get you a landing page live in under a minute with AI.

Read the full story →

Stability AI is back in the funding game with $76M more, but the real test is whether its open-source strategy can carve out sustainable revenue against better-resourced closed competitors.

Here's a number worth sitting with: two planned data centres in England - one in Buckinghamshire, one in Bedfordshire - are projected to emit 4.5 million tonnes of carbon per year when fully operational. That's more than ExxonMobil's entire UK carbon footprint. The analysis, cited by The Guardian, comes as the UK is legally committed to binding climate targets.

The environmental impact of AI infrastructure is becoming impossible to paper over. We saw it this week in Australia too, where the Albanese government backed down from a requirement that new data centres run entirely on renewable energy - a retreat that experts say could trigger a rush of approvals before tougher rules kick in. Data centre energy demand is forecast to rise sevenfold in Australia alone. The gap between AI's ambitions and the grid's capacity to support them sustainably is widening, and politicians in multiple countries are now caught between economic opportunity and climate commitments.

Read the full story →

The carbon math on AI infrastructure is getting harder to ignore - two data centres in England alone could out-pollute ExxonMobil's UK operation, and governments are struggling to reconcile AI investment with their own climate laws.

The answer is B - Run already-trained models to generate responses! An inference chip is specifically optimised for the "inference" phase of AI: taking a model that's already been trained and using it to produce outputs (answers, images, text) as fast and efficiently as possible. Training chips handle the initial, compute-heavy learning process - inference chips handle everything that happens after, which is what users actually experience.

The China AI companion story got me thinking: have you ever genuinely bonded with an AI chatbot or assistant - even briefly? Not in a weird way, just found yourself treating it more like a person than a tool? Hit reply and tell me - I read every response, and this one I'm genuinely curious about.

That's it for today - thanks for reading. For more daily AI coverage, head to dailyinference.com and we'll see you tomorrow. 👋