is the CRM for the AI era. and Attio instantly builds your CRM - with every company, every contact, and every interaction you’ve ever had, enriched and organized.

You can also build AI-powered automations and use its research agent to tackle some of your most complex business processes, freeing you to focus on what matters the most: building your company.

Join industry leaders like Flatfile, Replicate, Modal and more.

Try the CRM that’s built for the future.

The week in AI capex

The bubble talk keeps building, and OpenAI didn’t help themselves this week when the CFO, grilled on stage, said the company would like a federal loan guarantee, allowing lower costs of capital. That obviously got a lot of people very upset, with Trump’s ‘AI Tsar’ David Sachs ruling it out, and Sam Altman then said that they were talking about loan guarantees for people building chip plans in the USA, not for OpenAI infrastructure itself. CFO, ALTMAN

Meanwhile, OpenAI announced a $38bn infrastructure deal with Amazon (when I saw the headline I thought it might be missing a zero - only 38bn?). LINK

And Meta said that it will spend ‘$600bn’ on AI in the USA over the next three years, on both infrastructure and jobs. Trying to analyse numbers like this might be missing the point, but Meta will probably have close to $400bn of COGS and capex in the next three years - is that included? The fact that Meta has neither a cloud business selling AI APIs nor a (paid) consumer chatbot app makes people even more nervous about its spending on AI, even if it should be self-evident that this will change everything about social as well. LINK

Finally, Google announced its latest ‘TPU’ custom AI accelerator chip. Google has been working on these fo a decade and it’s paying off. Amazon is the next closest with Trainium, Microsoft is on version two and Meta… seems to be a mess. OpenAI announced a deal with Broadcom (which makes most of these) but that will take time. And then there’s Apple. LINK

Apple partners with Google?

Bloomberg reports that Apple will power its much-delayed rewrite of Siri with a version of Google’s Gemini models, running in Apple’s cloud on Apple custom silicon (where Apple really does have best-in-class technology, though no one outside Apple knows how good it is for running LLMs at this scale). Apparently, it will pay Google $1bn a year for this (while Google will continue to pay Apple ~$20bn a year of rev-share to be the default search engine on iPhones). Apple will carry on trying to build its own models to the same level, though, and the whole thing is on track to launch next spring. LINK

Model numbers

The Information reports that Anthropic is on track for $5bn of revenue this year, and hopes for $75bn in 2028. Meanwhile, Sam Altman (in the post linked above) said OpenAI will be at a run rate of over $20bn by the end of the year. LINK

Meta’s revenue from scams

Self-service ad platforms, like self-service marketplaces, let millions of people build new businesses in ways that would never have been possible before. But they also, inherently, have a problem with spam, fraud, and abuse - when all of society is online, so are society’s problems (see the Shein story below). In the past, that conversation centred on Google, but now it mostly centres on Meta, and this week Reuters got hold of internal analysis where Meta’s abuse teams tried to quantify the problem. The headline is that they estimate that 10% of 2024 revenue came from ads promoting scams. One should be a little hesitant in judging whether that’s a good or bad number (it can never be zero, just as a city can never have no crime), and the Meta documents list lots of measures to mitigate this, but also talk of fraud teams being laid off and what sounds like very restrictive limits on revenue impact. LINK

What does your camera roll say about you?

Tinder wants AI to look at your camera roll to learn more about you. This is ironic - online dating companies always wanted to find ways for you to explain your personality so that they could find better matches, but that was always too hard, and Tinder short-circuited the whole industry with a radically different ‘just swipe’ approach. Now AI might be able to solve this - just scan the camera roll and find the patterns. Of course, this is applicable in much wider fields than dating. LINK

The week in AI

Getty lost its lawsuit against Stable Diffusion for using its stock images as training data, but on a technicality that the judge says is narrow and limited. LINK

There are so many new LLMs being released all the time that I don’t even try to track them all, but Kimi, just released by the Chinese lab Moonshot, has got a lot of buzz as the new open source leader. LINK

Perplexity will pay Snap $400m in cash and stock to integrate Perplexity into the app. LINK

Google has a research project for a space-based datacentre. Hmm. LINK

Amazon, Perplexity and Agents

Amazon sent a cease and desist to Perplexity, telling it not to configure its ‘Comet’ browser to make automated purchases on the Amazon website. Amazon sold $65bn of ads in the last 12 months, much of it on its core e-commerce site (there’s Prime TV as well), and it does upsells and recommendations, and it doesn’t want someone else to control or disintermediate that experience. This is a systemic issue with consumer agents using the web ‘for you’, which I pointed out last year when people first suggested this model: companies want to own and control their own user experience. Perplexity responded by complaining about ‘bullying’ and talking about the rights of the users, but it’s Amazon’s shop, not Perplexity’s. (This company is so full of shit). AMAZON, PERPLEXITY

Trillion Dollar Elon

Tesla shareholders approved a new share grant plan for Elon Musk that’s notionally worth a trillion dollars. To be fair, this plan only pays out in full if Tesla’s market cap reaches $8.5tr, up from its current $1.4tr. Given that Tesla’s car business is in structural decline, with no answer to accelerating competition from China beyond yet more empty promises about autonomy, hitting $8.5tr would clearly require a radical change in trajectory. It does sometimes appear that Musk is increasingly bored of cars, and might like to turn Tesla into an AI and robotics business instead, though of course whether he can and whether that could be worth trillions of dollars is a further discussion. But then, Tesla has always been founded on promises for the future. The current valuation is clearly not based on the performance or prospects of the car business, and anyone holding it at this price is betting on Elon anyway. LINK

Shein’s doll problem

Shein is in part a marketplace, and the French government found a third-party seller offering ‘child-like sex dolls’ for sale. Shein delisted its entire ‘adult products’ section and begged for forgiveness. (Amazon had this problem back in 2018.) Everything is content moderation. LINK

Bill Janeway gives a good overview of the economics debates around generative AI. LINK

Backing up this issue, here’s an interesting paper from the US Federal Reserve analysing the long-term impact of computers on productivity. People have argued about this for decades, but the current AI boom brings it back to the table. LINK

KKR’s head of infrastructure coined the wonderful term ‘bragawatts’ to describe the way AI companies are announcing massive ’gigawatt’ infrastructure plans that may never actually happen, as a way to grab mindshare and funding and intimidate competitors. LINK

Google analysis of ways that malware creators are now adopting generative AI. Sample subheading: “Experimental Malware Using Gemini for Self-Modification to Evade Detection”. LINK

Fascinating piece on the grassroots scamming economy in India. Some things are new, some very old, and apparently no-one feels guilty scamming ‘rich’ Westerners. LINK

The impact of the war on the coding scene in Gaza. LINK

A paper with some very preliminary analysis of how people perceive ads that use generative imagery. LINK

Meta did some research to work out that the human eye can distinguish far more ‘pixels per degree’ than previously thought - high-end VR hardware today can do 35-50, but apparently people can distinguish up to 100. LINK

McKinsey has a long piece collecting ideas and possibilities for agentic commerce. LINK

And so does the IAB. LINK

Apparently in Hollywood ‘Microdrama’ is the new Quibi. LINK

On sale at Sotheby’s this week, a diamond brooch that was looted from Napoleon’s carriage as he fled from Waterloo. LINK

So many tech execs are flying private to the WebSummit boondoggle in Lisbon that the airport has run out of landing slots and they’re having to fly to airports hours away in Spain, or even (gasp) fly commercial. LINK

Activate’s annual deck on the state of media and entertainment. SLIDES, VIDEO

Netflix is reporting a new metric: Monthly Active Viewers. The current number is 190m. LINK

Preview from the Premium edition

Mechanical turks and short circuits

Any consumer internet system with critical mass becomes, in part, a Mechanical Turk. It looks at what the users do and draws conclusions from that. Amazon knows that if you bought X, you might buy Y, because it’s seen what lots of people buy and saw that if they buy X, they’re likely to buy Y. Google’s dominance of search is based in part on seeing what people search for, and what they click on and what they search for next.

Of course, this is a network effect, and comes with a cold start problem. How do you provide those recommendations and suggestions and connections before you have users, and how do you get users when you don’t have those capabilities? (Part of the idea of Bitcoin and then Web 3 was that you could use money as an incentive system to jump-start the network, but that created its own inverse incentives as well.)

But an LLM, in principle, is, at a minimum, a step change in automated understanding of what those SKUs are. The model can look at those words, and those images, and those videos, and all that metadata, and connect them to models that have some kind of understanding, or at any rate, some vastly broader kind of correlation.