Tuesday, January 6, 2026

Benedict's Newsletter: No. 624

NO. 624   FREE EDITION   SUNDAY 4 JAN 2026
SPONSORED BY LINEAR
Meet the coordination layer for human and AI product development

Linear helps you deploy, track, and orchestrate AI agents that ship features, fix bugs, and build products alongside your human team.

Learn more

My work

AI eats the world

Twice a year, I produce a big presentation exploring macro and strategic trends in the tech industry. New in November 2025, 'AI eats the world'. LINK

Podcast: how does OpenAI compete?

OpenAI has all the mindshare and 800m weekly active users, but the models remain commodities, and platforms with their own distribution are coming up fast. How will it compete? How will any of this work? How can you differentiate AI? LINK

News

End-of-year M&A

There was a final flurry of AI M&A at the end of the year. Alphabet will pay $4.75bn for Intersects Inc., which provides data centre electricity systems, helping Google get power to new DCs quicker. Meta bought Manu, which makes enterprise research agents that got a lot of attention a few months ago: the price was undisclosed but apparently 'more than $2bn'. And most dramatically, Nvidia did an acquihire of Groq, paying $20bn for a non-exclusive licence, with the key staff moving to join Nvidia and the investors getting a payout. Groq was founded by members of Google's TPU AI accelerator chip team nine years ago, and raised at only $7bn in October. ALPHABETMANUSNVIDIA

In non-tech news, Larry Ellison, who founded Oracle (ask your parents), made a personal guarantee of $40bn towards his son's bid to buy Warner Bros. Netflix remains the company's preferred buyer, though. LINK

The week in AI

The ACCA, which is the largest chartered accounting body, is ending remote exams because AI makes it too easy to cheat. LINK

Google is suing SerpAPI, a company that scrapes its search results and sells the data to third parties, most obviously to Perplexity and other AI companies that want to do their own search results without having their own search engine. LINK

OpenAI published an engineering post on how it tries to limit prompt injection in its Atlas web browser (remember that?). 'Prompt injection' here means that there could be text hidden in any web page you visit that the ChatGPT agent embedded in the browser would interpret as instructions - to email your bank details somewhere, for example. It's not at all clear whether this is a solvable problem. LINK

New cars

Waymo will extend its autonomous taxi service to London in 2026. Meanwhile, Uber and Lyft are partnering with Baidu to trial its AV service in London as well. Back in the USA, Tesla's robotaxi project in Austin is now experimenting with removing the human safety drivers (Tesla is currently running 30-40 cars there, with humans sitting in the car ready to take over for the 'full self-driving'). LONDONTESLA

Meanwhile, Tesla sales continue to slide and BYD is now outselling it on pure battery-driven cars. Some of this Tesla sales slump is about Elon Musk's behaviour (especially in Europe, and in the US demographics that most like EVs), but it's hard to  see that being as visible in China, and meanwhile the product line-up is stale too. It seems pretty clear that Elon Musk wants to pivot away from EVs, where there's little new to say and competition will be ferocious, and upwards to both autonomy (sigh) and now robots. LINK

About

What matters in tech? What's going on, what might it mean, and what will happen next?

I've spent 25 years analysing mobile, media and technology, and worked in equity research, strategy, consulting and venture capital. I'm now an independent analyst, and I speak and consult on strategy and technology for companies around the world.

Ideas

The Information has more gossip about OpenAI's plans for advertising. I'm old enough that this reminds me of MySpace after Rupert took over. LINK

Huawei now makes the best-selling car in China in the over-$100k segment, the Maestro S800, outselling any foreign car in that range. Meanwhile, BYD is making a car with a frame cast in a single piece. HUAWEIBYD

The FT has a long interview with Yann LeCun talking about his new venture, and AI at Meta - the old team was indeed 'selective' in its use of benchmarks for the flopped Llama 4, which lost Mark Zuckerberg's confidence, and most of that team has either left or (he thinks ) will leave. The new team, headed by billion-dollar hire Alex Wang, is good, but as he keeps saying, he thinks LLMs are a dead-end and more progress needs real-world data (cause-and-effect, video). LINK

Coupang, the 'Amazon of South Korea', has had a breach, in which a former employee stole user data for 34m customers. The company post-mortem is wild, including a MacBook thrown into a river. LINK

Outside interests

Relevant to AI, but interesting in its own right: how far can animals be tricked by conjuring tricks? Here, the Jay. LINK

The NY Times process holding the front page. LINK

Mark Curtis, co-founder of Fjord, on the journey of design from 'interactive media' in the early 90s to generative AI today. LINK

Data

Analysing how developers used AI coding in 2025. LINK

Preview from the Premium edition

Big tech's 2026

There's an obvious set of questions for AI in 2026, at each level of the stack - does the rate of model improvement change, do we have some new breakthrough (continuous learning?) or change in cost, does Nvidia get more competition, and what happens to the data centre boom / bubble? Here, though, I'm considering the strategies of the big tech companies themselves, as three years into the AI cycle they pivot themselves around this.  

Google had an innovator's dilemma question (can they do this if it threatens the legacy business?) and a Steve Ballmer question (can Sundar Pichai do this?) and now seems to have passed both of those. The models have caught up and the product teams are doing pretty well both in making this a feature (AI Answers, and a lot of other things) and trying to find new experimental products. Meanwhile, the fortress balance sheet and the decade of investment in TPUs are a good setup for a few hundred billion dollars of data centre deployment. They do, however, have to get the API and pricing story right for GCP. But on the other hand, we still don't know if someone (else) will work out the real 'step two' for how LLMs can change search, and make another attempt to unbundle Google. 

Apple has time. Siri will probably get fixed, and be a lot better, or at

Upgrade to Premium
You're getting the Free edition. Subscribers to the Premium edition got this two days ago on Sunday evening, together with an exclusive column, complete access to the archive of over 600 issues, and more.
 

No comments:

Post a Comment

The Sunday Morning Wrap-Up

Here’s the highlight reel of the past week’s posts and events. ͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­...