My work

How will OpenAI compete?

OpenAI has some big questions. It doesn’t have unique tech. It has a big user base, but with limited engagement and stickiness and no network effect. The incumbents have matched the tech and are leveraging their product and distribution. And a lot of the value and leverage will come from new experiences that haven’t been invented yet, and it can’t invent all of those itself. What’s the plan? LINK

AI eats the world

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

Another quarterly surge

All of the big tech companies reported Q1 earnings this week, and all of them showed dramatic AI revenue growth and further increases in capex plans for the rest of this year. As of today, Amazon, Meta, Microsoft and Alphabet plan to spend about $700bn on data centres this year (Amazon doesn’t explicitly break out AWS and logistics capex but says AWS will be the vast majority), up from close to $400bn in 2025 and $225bn in 2024. 

There is nuance within this if you’re interested (GCP revenue is surging, for example), but the interesting thing stepping back is the market reaction: where Alphabet and Amazon rose and Microsoft was flat (see below), Meta fell 10%. It doesn’t have an enterprise cloud business, so the ROI has to come purely from better ad yields (which is already happening) and… more ‘stuff’ - people are still upset about the ~$100bn that Zuck has spent so far on ‘Metaverse’. LINK

Apple stands aside

Apple still stands outside this story - revenue also grew strongly (the latest iPhones are a hit and it’s doing fine in China), but it’s between cycles: we don’t know how well the Siri reset is going, or when the glasses are, nor whether John Ternus, the new CEO, thinks he needs a frontier mode of his own with associated investment, nor how much investment will be needed to run Apple’s upcoming AI stuff in the cloud versus on the device. LINK

Microsoft’s OpenAI breakup continues

The slow-motion divorce of OpenAI and Microsoft got a little more resolution this week: Microsoft now only has first refusal for OpenAI hosting, giving up exclusivity. Amazon immediately announced that it will offer ChatGPT APIs on AWS, giving enterprises and developers a choice. There’s a layer of politics here, and obviously this is partly about OpenAI’s hunger for capacity and Microsoft’s need to manage its own resources, here, but I think the underlying story is the ongoing commoditisation of foundation models. LINK, AMAZON

The great 2026 AI pricing squeeze

Agentic coding is all anyone can talk about in tech, and it comes with orders of magnitude more token use and hence compute capacity, which is more than anyone can really handle, especially Anthropic. It has become very clear that you could use a $20/month Claude subscription to consume hundreds or thousands of dollars of underlying capacity, and this week it changed a bunch of pricing, as part of a general move away from flat-rate monthly pricing towards explicit usage charges, tying spend to marginal cost. 

This all looks a lot like mobile data pricing in the early days of smartphones, 15 and 20 years ago. We have all of the same problems: no one knows what tokens or bits really mean (especially consumers), use cases that look similar can drive radically different amounts of capacity, and then new use-cases unlock orders of magnitude more consumption (the iPhone and especially video in the late 2000s, agents and especially coding now). AT&T launched the icon with unlimited flat-rate data, but the network couldn’t cope (cellular networks have significant marginal cost) and by 2010 it had to switch to capped bundles: model providers are doing the same now. LINK

This is a big question for SaaS companies building products on these models - do they pass on variable cost or try to wrap it into bundles, or charge by action or even (speculatively) by outcome? Salesforce keeps changing its mind on this. LINK

China kills Meta’s Manus deal

This has been trailed for months: the Chinese government, in a one-line press release, ordered the reversal of Meta’s ~$2bn acquisition of Manus, an agentic AI company. Manus had moved to Singapore before the deal, but narrowly. China clearly sees this as ‘Singapore-washing’ of a Chinese company - more broadly, there is clearly reluctance to allow leading companies in a crucially strategic sector to go to foreign owners. Of course, that makes investing in startups harder, especially for foreign VC funds (but are there any left?). Meanwhile, unwinding the deal is easier said than done - there are no physical assets to hand back, just a bunch of IP, know-how, and people that have already been integrated into Meta for months. In particular, while Manus got attention last year for a virtual assistant, Meta has been using this to optimise the ad-buying flow. It does now have decent models of its own, so it could rebuild that (indeed, it could ‘just’ rewrite the Manus code with AI!), but what does the government want? There’s no due process here. LINK

OpenAI revenue wobbles

The WSJ reported that OpenAI missed internal revenue and user growth targets this year, as competition from both Google’s Gemini and then Anthropic started to bite. Shares of its publicly traded infra partners (especially Oracle) fell sharply. The news isn’t a huge surprise given that the company hasn’t given an update to either number for a while even as Anthropic surged (though it did point out that Anthropic is giving gross revenue numbers where OpenAI’s are net). Meanwhile, it’s notable how often we see the CFO saying slightly different things to Sam Altman - this is not a company known for its politics but with the CEO out sick there might be more to come here. OPENAI, PARTNERS

The week in AI

Like everyone else in chips (see Intel last week), Qualcomm is reweighting to AI for datacentres to take advantage of the crushing capacity shortage. Nvidia is best, but people people will take whatever they can get right now, and Qualcomm, coming from ARM and mobile, does know low-power. LINK

Remember Deepseek? It released a new flagship model this week, and it’s… OK, landing at the bottom of the top dozen or so, depending on on which benchmarks you look at. It’s significantly cheaper than the US models at the top of the boards, but there isn’t the same shocking (but mostly misunderstood) cost saving that freaked people out a year ago. More notable, perhaps: it runs on Huawei chips as well as Nvidia. Meanwhile, There are lots of Chinese models now - Xiaomi and Alibaba are also in the top 10 by some measures. LINK

China claimed to have built a new supercomputer using home-growth chips. Experts are … sceptical. LINK

Tesla has been promising autonomy without actually delivering for over a decade, and as part of that it sold cars with the claim this would be added in a software update. Now it’s had to admit that the V3 hardware in 3-4m cars (about half the fleet) isn’t able to run even Tesla’s current semi-autonomous system, and will need either a trade-in or a hardware upgrade. Oops. LINK

Lyft expands 

Lyft continued its strategy of buying ‘not-Ubers’ and taxi apps in cities around the world, adding Gett UK.  LINK

Ideas

Outside interests

Data

The WSJ says TikTok Shop’s US sales were $4.9bn in Q1, up 46% Y-on-Y. LINK

Apollo points out that call centre and BPO employment in the Philippines continues to rise, despite all the talk of automation. This could be a trailing indicator, of course, but it reflects a lot of data that early AI deployment is focused more on analysis and productivity than cost savings or revenue, which will come later. LINK

Yet another national US survey, this time from Marquette, showing that the general public is very sceptical of AI in general and data centre construction in particular. LINK

Preview from the Premium edition

I got married this week, so I haven’t had time to write a column. Normal service resumes next week, but for today I’ve included a column I wrote about ChatGPT in January 2023.  Meanwhile, I’ll be speaking on a panel at the Milken conference in LA this week - come and say hello if you’ll be there. 

Generative ML questions 

The wave of energy around generative ML just keeps coming. Microsoft has grabbed it with both hands and Google has brought Larry and Sergey back from retirement to help work out what to do. And out in the broader tech community, products and companies are popping up so fast that I half-wonder if people are using ChatGPT to suggest ideas. 

Some of this enthusiasm feels like a relapse of pent-up energy. We’ve been looking for the ‘next big thing’ after smartphones for a long time: machine learning had become boring, VR/AR is years from taking off (if ever), and crypto has crashed, while a lot of people never believed in it in the first place. Now, very suddenly, there is a new substrate to build on - a new Petri dish - and anyone can get to very cool results very quickly. 

We’re also at the point that we’ve digested just enough of this to start asking questions. Some of these are re-runs of machine learning questions - should this be build by Google or a startup? Horizontal or vertical go-to-market? How specialised and hard-to-get is the data you need? How do you get data with no customers, and vice versa? By extension, are there areas where these questions look structurally different for generative ML? 

We’re also re-running the ‘what are the use cases?’ question. Machine learning started working with image recognition, but it generalised much more broadly very quickly - really, it was better to think of if as pattern recognition, but then the question became “what things do we not realise are actually pattern recognition problems?”, or, “what could we turn into pattern recognition problems?”  Indeed, “what can we turn into image recognition problems that we were previously solving in some other way?” It’s like asking “what can we do with databases?” in 1960. 

We’re still deep in that discovery process for ML, and we’ve barely started for GenML. What kinds of problems can be solved with “hey computer, make me something that matches this pattern”? You can make text, but what can you turn into a text generation problem? It’s quickly become apparent that ‘writing code’ is an text generation problem. Then you can also use it to design a UI, or a simple web layout. How else can we understand this? The same for making images or video  - where is it useful to generate these, but then, what can you turn into those kinds of problem now that this might be possible? 

The other axis is the accuracy question, which increasingly I think is as much about perception as technology. If a given model can make something “96% accurate” in a given domain, is that good or bad? In some domains, no human will be able to tell the difference, but in others that 4% could get someone killed. In principle, GenML models make something that looks like what you asked for, but they do not, in principle, make something that is what you asked for. So where is that useful and where is it dangerous?

Extending this, I suspect there are a lot of use cases in the middle where the answer is ‘helpful, but I’ll need to check (and you can learn from what I check)’ - rather like asking an intern to produce a document for you. Hence Github calls its GenML tool ‘copilot’ - you’re there flying the plane as well. And perhaps one criteria for assessing domains is whether, as a human, you can spot the mistakes. I can ask an intern to make a spreadsheet for me because I can check it, but I can’t read software code myself, and so I wouldn’t be able to use CoPilot: I couldn’t spot the mistakes. 

I mentioned interns, and I used to describe machine learning as doing anything that an intern could do (or a dog), except that now you had infinite interns. You could tell an intern to listen to a call and tell you if the customer was angry, but you didn’t have enough interns to listen to every call coming into your call centre, and ML let you do that - it automated interns, not experts. I think this remains a useful metaphor. But there’s another useful comparison, and that’s to think about previous waves of automation. 

When big companies first bought big computers, in the 1950s and 1960s, they were automating away another kind of machines - adding machines and ‘electromechanical calculators’ - and indeed the people who operated these were called ‘computers’. Every office building had floors and floors full of these ‘computers’ where each person was a cell in a spreadsheet and the whole building was a giant manually-operated Excel file. A generation earlier, those machines themselves had automated pen and paper and mental arithmetic. And there’s an argument that Excel in turn generated far more analysts and hence more analysis, because it was possible for them to do so much more - Jevons’ paradox in action. Now we get more automation, and it’s not at all clear yet that these new systems will be different - yes, they can pass the Turing Test, but perhaps that just shows that the Turing test (or MBA essays) are subject to Goodhart’s law.