10 May 2026

News

Deals for AI deployment

How do big complex companies with thousands of people and workflows, and thousands of applications, work out how LLMs will be useful for them. The dumb answer is to give everyone CoPilot and wait for them to work it out, which is rather like giving everyone a web browser in 1997 and saying “There you are! Be more productive!” A better answer is that it’s slow, hard and complicated, and they tend to rely on professional services for help, whether it’s for the tech itself (Accenture, Deloitte etc) or the deeper strategic question of what the code should actually be doing (Bain, BCG, McKinsey). They often also lack the margins to experiment with Claude for millions of dollars, and would rather pay someone who’s worked it out (as much as anyone has so far). But the other side is that many PE investors are an intermediate layer of go-to-market, management support and, of course, relentless and focused shareholder pressure for efficiency, all of which means that the big AI labs have announced partnerships with both BBM and, this week, PE. Anthropic announced that it’s forming an AI Series company in partnership with one group of PE firms plus Goldman Sachs, with something over $1bn of funding, while OpenAI is apparently working on a $10bn venture to do the same, to be called ‘the Deployment Company’. As the name suggests, for the labs this is a GTM plan; for PE, their model rests on delivering increased operating efficiency at their portfolio companies, and if AI can’t do that, what else is it? ANTHROPIC, OPENAI

Meanwhile, Long Lake Capital, another PE fund, bought out American Express Global Business Travel for $6.3bn aiming to use AI, again, to transform the operating efficiency of a complex product full of friction and manual systems. LINK

The White House swings back to AI caution

Trump’s administration came to office full of loudly-expressed contempt for ‘AI Safety’ and the doomers, seeing their concerns as mostly nonsense that would serve only to slow down US competition with China. Now, though, they’re shifting to a more cautious approach, as a bunch of much more specific concerns are tangible, around things like energy consumption, education, privacy and (in a few edge cases) mental health. The cyber threat discussion around Anthropic’s Mythos perhaps catalysed the change - this is real and important, and also has geopolitical implications. What if AI is an order of magnitude change in how many countries can cyber to cripple their opponents infrastructure, and how far does that expand beyond nation states? More cynically, this is clearly emerging as an issue in US local politics: there are votes in attacking data centres. Meanwhile, David Sachs, the Silicon Valley scenester who was ‘AI Tsar’ and strongly opposed to regulation, left the White House. So, now there are rumblings of new executive orders, oversight of new models, and perhaps more. LINK

Anthropic and the SpaceX shuffle

Anthropic’s CEO Dario Amodei says that the company is running at 80x annual growth, having planned for ’only’ 10x growth, leaving it (as everyone can see) very short of compute capacity to serve Claude Code.  The FT reports that it’s raising at a $1tr valuation and expects to hit a $45bn annualised revenue run rate imminently, compared to $9bn at the beginning of the year. 

Part of Anthropic’s scramble for more capacity, it took a lease on the 300 megawatt ‘Colossus’ datacentre that Elon Musk’s xAI (now part of SpaceX) built in huge haste last year. Cue general puzzlement: it’s obvious why Anthropic is buying, but why is xAI selling? The Information claims that shortcuts in the hasty construction, and the mix of different chips used, mean the centre isn’t very useful for training, but remains useful for inference (which is what Anthropic needs). More cynically, this is more revenue to load into the SpaceX IPO filing, as it presents itself as a cloud AI company as much as it is about buckets. Certainly, xAI itself is flailing, with all the founders quitting and Elon Musk saying he needed to start again from scratch. But what else? DEAL, SHORTCUTS, REVENUE

Meanwhile, SpaceX itself filed to build a semiconductor fab in Texas for an initial $55bn and potential up to $119bn. The plan has been reported previously at $20bn. This is part of an overall ambition to manufacture a terawatt of compute annually that Musk has suggested previously (though that would probably require trillions rather than billions of coax). Someone needs to make the chips for the orbital data centres, after all… LINK

Indexing AI

Back on Earth, one of the mechanics of this year’s proposed IPOs of OpenAI, SpaceX, and perhaps Anthropic is whether these companies would be included in the stock market indices used by index funds to determine their investment. As a reminder, OpenAI and Anthropic are both valued at (say) $1tr in private markets, and SpaceX is reportedly looking at a $1.75tr valuation. If they are, index funds by definition will have to buy the stock proportionate to the company’s share of the overall market, and index funds are now around 25% of US ownership by market cap, so that would be a significant support for the share price. The S&P 500 used to require a one-year wait and profitability, but suggests that if all three companies are excluded on that basis then the index wouldn’t really be tracking the market since (at probably over a trillion dollars each) they’ll be such a large share of it. 

It’s also worth noting that in 1999  and 2000 combined, total VC investment was only about $150bn and venture-backed IPOs only raised about $45bn at a valuation of $270bn (both in real terms) - each of these three IPOs is presumably looking for more than that. LINK

Apple partners with Intel 

Part of Intel’s turnaround is a foundry business, and Apple likes to have two sources but relied on TSMC for the SoCs that drive its devices. So far, TSMC has a generational lead in process, and Intel’s foundry is unproven… but now the WSJ reports that Apple and Intel have reached a preliminary agreement, though it’s not clear for what or how much. Intel stock went up 10%, either way. It seems a long time since the days when ‘WinTel’ ruled the tech industry - now Intel is jumping when Apple calls. LINK

Remember Boston Dynamics? 

Boston Dynamics, home of those viral/terrifying robot dog videos, never found an actual business and was bought by Hyundai back in 2021. Now with ‘physical AI’ exploding, you’re thinking this would be their moment, but most of the leadership is bailing. LINK

The week in AI

OpenAI continues to build out its ad product. LINK

DeepSeek is raising $5-10bn (from Chinese investors - cf Manus) at a $50bn valuation. LINK

Google continues to improve and iterate AI overviews. LINK

Amazon opens logistics

Amazon is opening its US logistics operation, competing directly with FedEx, UPS, and the USPS. You no longer need to be selling on Amazon for Amazon to deliver your package— now the full stack of freight, warehousing, and last-mile delivery is open to any other company. Just on last-mile, according to ShipMatrix Amazon is already the largest single carrier at 28% of US volume, and this looks like a pretty obvious move to load up the asset base with more volume and liquidity. LINK, SHIPMATRIX

In other news

IAC is renaming itself People Inc and focusing aggressively on using its brands and content to drive products and services, as AI changes what the content itself is worth. LINK

The meme stock GameStop made an unsolicited offer to buy eBay for $56bn. It has a line of credit for $20bn, and the rest would have to be in stock. Ebay is certainly very wilted, and due for a turn-around of some kind, but no-one can really see why this deal would make sense, though, of course, that may not be the point. LINK

Ideas

Apollo asks the macro AI question - is AI the new ‘China shock’? LINK

Sequoia held an AI event with a bunch of important speakers from the frontier - Andrej Karpathy’s spot is especially worth watching. KARPATHY, LINK

Google held an ‘AI for the economy’ day with lots of interesting panels. Bias to optimism, obviously. LINK

Mozilla explains how it used Mythos to find a huge number of new bugs. LINK

As more countries consider adding age restrictions to social media, kids get around this by doing anything from using their parents’ ID to, well, drawing on a moustache with a pen. LINK

There’s a huge surge in AI-generated podcasts - perhaps 40% of new shows. LINK

When the headline says it all - ‘the devout Muslim making a living from Islamophobic AI slop’. Partly about AI, partly about Facebook and social networks with rev-shares. LINK

The continuing existence of Claris and Filemaker hidden somewhere inside Apple is a weird anomaly. It’s a little bit like WPP still (until recently) owning the original WPP. LINK

Interesting BLS essay on the evolution of tech and productivity in grocery stores. LINK

Outside interests

Why is US healthcare so expensive? LINK

Data

McKinsey analysis on AI shopping and omnichannel. LINK

Microsoft’s new ‘AI at work’ report. LINK

OpenAI plans to spend $50bn on compute this year. LINK

As part of last week’s quarterly announcement, the hyperscalers all announced huge jumps in future revenue commitments (client contracts signed but not delivered), mostly reflecting the surge in AI demand, especially all the deals announced with Anthropic and OpenAI, which lack their own infrastructure. The Information put numbers on that - for example, 280bn of Microsoft’s $627bn (which also includes Office 365 and everything else) is OpenAI, and $200bn of the $468bn that Google reported is Anthropic. LINK

Column

Tokens and bits 

In the last 20 years, mobile network traffic has risen by several thousand times, and yet the stocks have been flat. The industry has revenue of over a trillion dollars a year, and capex of $200bn or so, and the technology is amazing, but this is low-margin commodity infrastructure. These companies imagined everything we do with smartphones today, but when it happened it was built by other companies, further up the stack, and they didn’t get any of the value from it. 

Meanwhile, when that data explosion began, back in the late 2000s, they had all of the same questions that model labs face today. New use cases and applications (smartphones and video then, agentic coding now) unlocked unimaginable growth in demand for raw capacity that the infrastructure could not keep up with - bits then, tokens now. That capacity had substantial marginal cost, for model labs just as for mobile networks, but how to pay for it? What should you charge?

Bits and tokens are the unit of cost and usage, but they’re also opaque and map very badly to use-cases and value. Tasks that look the same can use radically different amounts of capacity, and things that are very valuable can use far less capacity than things that are worth very little. Just as the AI industry today plays with task-based or value-based pricing, telcos back then tried to segment the pricing curve, mapping price against value rather than use, but none of that really worked, and meanwhile it opened arbitrage opportunities all over the place (and that’s not to mention SMS). 

In the end, telcos got price, demand, supply and capacity into equilibrium with flat-rate capped bundles. 15 years later, we pay more or less the same that we paid then, but over time that has given us steadily faster speeds and bigger bundles, powered by that 20% of revenues on capex plus successive waves of new and more efficient technology. 

And meanwhile, to repeat, they got a trillion dollars of revenue, but at commodity margins,  and all the applications were built by other people. 

Why should AI be different? We are now in a moment of extreme disequilibrium, with demand far ahead of available supply and the supply chain for more capacity backlogged by years. Anyone who can make tokens can price them at ROI. But there’s no good reason to presume that this is permanent. Demand is infinite for tokens, but demand is often infinite at the right price, and normally we get to equilibrium - that’s what telcos managed. This should be temporary. We should get to equilibrium, and over time, it’s hard to see why this technology should have pricing power, and why tokens should not sell, so to speak, at marginal cost. So far, we do not see a sign of network effects or any other dynamic that would give foundation models a winner-takes-all effect - there’s no equivalent of the way developers had to build for Windows or iOS. Equally, it remains very unclear that these models can by themselves serve most use-cases without being wrapped in tooling and GTM one problem and one industry at a time. We do know that the experience is bound to change a great deal, and can the labs invent all of that? And all of this presumes that it remains those giant foundation labs, and doesn’t fragment massively into large and small models on one side and edge on the other. Indeed, edge is another parallel to mobile, where WiFi offload solved half of the capacity problem. 

The underlying issue, of course, is that we don’t really know what AI infra supply or demand will look like in a few years, despite how many young and keen analysts build spreadsheets. There will be, say, 2 to 10 companies spending, say, $200bn to $2tr every year. Or less. Or maybe more! But will that infra do everything, or will it be a low-margin commodity, where billions of people pay a flat rate subscription, maybe directly or maybe bundled into something else, but everything interesting is built on top?