Benedict's Newsletter: No. 643

In this issue: A new AI presentation for Spring 2026. Plus, Google on the edge, cyber, satellites and the Eiffel tower.

NO. 643   FREE EDITION   SUNDAY 17 MAY 2026 SPONSORED BY SLACK Get Started with Agentforce in Slack AI isn't just reshaping technology, it's reshaping how businesses operate. With Agentforce in Slack, teams can put AI agents to work right where decisions are made. DM an agent, loop it into a channel, and let it pull data, surface insights, and move work forward automatically. No new tools. No new logins. Just AI that works inside the platform your organization already runs on. Read the guide. My work New presentation: AI eats the world, Spring 2025 Twice a year, I produce a big presentation exploring macro and strategic trends in the tech industry. The latest edition: ‘AI eats the world’. LINK News Google goes to the edge  Google announced a range of new AI features and capabilities for Android, some of which look good, though none of which are radically new. The most significant part of how much this relies on edge is running the models on the phone. You don‘t have marginal cost if your users bring their own compute. The catch is that the compute and memory requirements rule out the vast majority of the Android install base, which Google spins as ‘coming this summer to our most advanced devices’. Next month, Apple is expected to relaunch its rewritten Gemini-powered Siri, which, if it actually ships this time, will reach a much wider part of the base. LINK Meanwhile, Google will rebrand Chromebooks ‘Googlebook’ (reflecting the merger with Android), and add the same AI features. No specifics, but apparently, devices will be announced later this year (with OEMs hoping for an answer to the MacBook Neo). LINK The week in AI As reported last week, OpenAI launched ‘the OpenAI Deployment Company’. There used to be a joke that an ‘AI scientist’ is a statistician who lives in San Francisco - maybe now Accenture will rename all their staff ‘forward deployed engineers’. LINK SAP, the proverbial legacy IT company, is building its own AI story, buying Prior Labs, which had raised €9m, for (apparently) about €1bn. LINK Peter Steinberger, the creator of OpenClaw (now at OpenAI), posted a screenshot showing he spent a notional $1.3m on tokens in the last 30 days. Tokenmaxing is over: we have a winner. =LINK Meanwhile, Amazon is measuring staff’s token usage, so, Goodhart’s Law being what it is, people are deliberately inflating their ‘scores’. LINK With OpenAI’s CEO Fidji Simo still out on medical leave, Greg Brockman (CTO) is now head of product. Careful of the turnover - we don’t want the head of product at OpenAI to end up like ‘Drummer at Spinal Tap’. LINK Someone at OpenAI decided to complain to Bloomberg that Apple isn’t using OpenAI for every new AI feature, and hinted at lawsuits. Good luck with that. LINK The FT says Anthropic is raising $30bn at a $900bn valuation. As I noted in my new presentation, $900bn is more than the market cap at the issue of all the venture-backed IPOs in the USA from 1995 to 2000 combined. LINK Cerebras, which makes AI inference chips as big as possible (using the entire wafer for one chip instead of putting lots of chips on each wafer), IPOed this week, and the shortage of any AI stocks where you can go long instead of short led to a giant spike: it peaked at a $95bn market cap. LINK AI and cyber Ripples keep spreading out from the launch of Anthropic’s Mythos model, with by all accounts startling new abilities to find and exploit security holes in software.  This week OpenAI announced its own version, Daybreak. Where Anthropic says that Mythos is too dangerous to make available, OpenAI will sell it to approved clients. Too dangerous, or does Anthropic just not have enough compute? LINK Meanwhile, Microsoft launched its own agentic code scanning system, that chains together multiple third-party SOTA models (since Microsoft still doesn’t have its own), while Mistral (remember that?) is also apparently working on a cyber product. MICROSOFT, MISTRAL Finally, the threat intelligence team from Google says it’s now seeing real threat groups using AI. LINK Satellite to mobile  A month ago Amazon bought Globalstar, and this week SpaceX bought 50MHz of US cellular spectrum from Echostar for $17bn in cash and stock. Meanwhile, the three main US mobile networks, AT&T, Verizon, and T-Mobile, announced a JV to provide a unified way to use satellite to fill in rural dead zones.  Every few years a technology comes along that people in tech think will crush cellular, and so far they’ve always been wrong, because they miss three basic issues. First, it’s totally different to give a high headline speed to one user than to give good speeds to hundreds of thousands of people in the same place: the wireless network challenge is capacity, not speed. 5G was all about increased capacity for the same speeds. Hence, it’s vastly easier for Starlink, or any satellite tech, to fill in rural dead zones (and connect ships, airliners, and so on), than for it to connect lots of customers in one place, which is what you need to take on mobile. How many satellites would Starlink need to give 100 meg/second not to one person globally but to 100k people in a few square miles? How fast will that change? Second, there are physics problems - will this work if your phone is in your pocket indoors? In a car at speed? And third, if you solve the two previous questions, congratulations! You’ve just joined a low-margin ex-growth commodity industry.  All of that said - the challenge of knowing a lot about an industry is that you know all the reasons it can’t be disrupted, and you’ll be right for a long time, until something changes. So, let’s see. MOBILE, SPACEX 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. FIND OUT MORE Ideas Bain published a really interesting piece about ‘synthetic customers’ and their use in customer search. LINK Anthropic’s bankers will be pleased with this profile of the CFO in the WSJ. LINK Google finally released a guide to optimising for generative AI search. Mostly, it says to do the same as optimising for search. LINK The US university Princeton has ended an honour code for exams that goes back to the 19th century and will now monitor for students using AI to cheat. LINK The first post-acquisition interview with Alex Wang, Meta’s billion-dollar hire. They have a model again now, but let’s see. (Meanwhile, see the next item). LINK Wired says that morale at Meta is terrible: repeated layoffs and ever more AI pressure (including the new key tracking move). I hear this a lot too. LINK Corporate lawyers are very nervous about the disclosure implications of AI note-takers. LINK EY had to withdraw a report on loyalty systems that was full of hallucinations. We really should be past the point now that highly-paid professionals do not understand that LLMs are not databases. Also, doesn’t the content marketing team do fact checking? LINK With a torrent of free cashflow (more than its suppliers are willing or able to absorb), Nvidia has taken over $40bn of equity stakes so far this year, including, this week, $3.2bn in Corning. Depending on your perspective, you could call all of this ‘vendor financing’ or ‘circular investing’. LINK The new US crypto regulation (‘Clarity’) is slowly winding its way through the legislative process. As the name suggests, this is mostly positioned as a reaction to the complaint that the Biden financial services refused to consider or discuss a use of crypto more complex than a pure commodity. LINK Outside interests This week at auction: a segment of an original spiral staircase from the Eiffel Tower (removed and replaced in 1983). If you need this, you need this. LINK Hantavirus airdrop on Tristan da Cunha. LINK Data Gartner’s CMO survey says CMOs are putting 15% towards AI. LINK OpenAI released ‘Signal’, a fairly high-level report on how people use the product. LINK Ramp says that amongst its customer base, Anthropic spending has now overtaken OpenAI. LINK As part of the Musk/Altman/OpenAI lawsuit (others mostly a matter of gossip), Microsoft disclosed that it expects to have spent $13bn on OpenAI by this June, between both the $13bn in direct investment and everything it’s spent on infrastructure. LINK Gallup has another survey showing how unpopular data centres are in the USA. LINK The US Lawrence Berkeley National Laboratory published a study (at the end of 2024) of US datacentre energy and water use. Energy is obviously a real issue, but there’s growing panic about water. Hence, this report is useful in that it estimates all US datacentre water use combined is less than 0.02% of total US water use. Even if that triples in the next few years, as the most aggressive estimates suggest, it's a local planning issue, not a broader problem. LINK Andreessen Horowitz and its partners are currently the largest political donors in the USA, focusing on AI and crypto (disclosure: I worked at Andreessen Horowitz from 2014 to 2019). LINK Economists keep puzzling over whether we can see generative AI in productivity or employment stats. This study points out that the apparent decline in employment of young people in the US is much the same for graduates and non-graduates, and for fields that both are and are not obviously exposed to AI. LINK A Federal Reserve study comparing AI adoption in the USA and Europe. Unsurprisingly, like almost any other tech adoption metric, the large EU markets (France, Germany, Spain, Italy) are a year or two (or more) behind the US, UK and Nordics. LINK Preview from the Premium edition A foundational model thesis I've just published a new presentation, as I do twice a year, that tries to explain where I think we are in the emergence of AI, and what it means for everyone else. There are a lot of ideas and a lot of data, but I want to pull out one slide to think about, “A Provisional Thesis”, which puts together four linked ideas for the state of foundation models and model labs, and the role that they're likely to play. Some of this I’ve discussed elsewhere in the past, but I also wanted to put it all together in one place.  Firstly (like many people), I think chatbots are a really bad or, at best, a really limited interface. A graphical user interface is not simply a way of presenting all of an application's hundreds of features on the screen, where just asking is easier. Rather, a GUI is a way to help the user work out what they might want to do. It represents a lot of careful thought and painful learning about how this workflow might function, and what the choices should be. A GUI is institutional knowledge. And if you take that away, then suddenly the user has to screw up their eyes and think from first principles about how they might want to accomplish this task, or what the task even is, with no reference or scaffolding to help them. Moreover, I am not sure that a ‘generative UI’ can really solve them in general, because the model has the same problem - the model has to work out what the screen should be.  To put this more conceptually, building a tool is a completely different skill and probably a completely different person to using a tool. These are done by different kinds of people in different places, with different experience and different context.  Secondly, the model labs themselves are extremely unlikely to be able to create all of those tools themselves. You can see them groping towards 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. SUBSCRIBE OR UPGRADE THIS WAS SENT TO abdulkerim.bosnia.business@blogger.com ON 19 May, 2026, BECAUSE YOU SUBSCRIBED AT BEN-EVANS.COM. You can UNSUBSCRIBE, OR READ THE PRIVACY POLICY. OUR MAILING ADDRESS IS Benedict Evans, 2 Golden Square, London, W1, United Kingdom.   COPYRIGHT © 2013-2026 Benedict Evans, ALL RIGHTS RESERVED.