The US lifted its export controls on Anthropic’s latest models. I’m ambivalent as to how much to analyse this. The security issues are real and the culture clash between Anthropic and the Trump admin is also real, but the ‘process’ of first restricting and then un-restricting was a mess with no due process or repeatability and there's a limit to how much value I can get for reading off-the record briefings. Meanwhile, OpenAI’s new model is still under some kind of restriction… and the real lesson is that access to cutting edge US models seems to be at the whim of whoever’s on top in the White House this week. LINK
Apparently, Sam Altman has suggested that the best way to clear political obstacles and give the public a stake in AI is for the US government to have a 5% stake in all the AI labs. This makes very little sense except as lobbying. Even if some level of public ownership (in other words, ‘socialism’) solved whatever problems you think there are, those problems are not limited to the USA on one side (are you going to give every country on earth stakes?), and on the other who qualifies a a lab? What about DeepSeek? OpenSource? JEPA? LINK
Bloomberg reported that Meta is planning to build an enterprise cloud business, so that it can rent out AI compute capacity if it turns out that it’s over-built. Three observations.
First, the stock went up 10% on the news, which tells you something about the febrile state of the market given that Mark Zuckerberg said this was on the table at the AGM in May, as widely reported at the time. In frothy markets, news (good or bad) can get priced in more than once.
Second, product-market fit in AI has come mostly in software development and (at a smaller scale) in the enterprise. The other companies investing hugely in AI have businesses there, but Meta does not, so immediate returns can only come from optimising its ad and recommendation systems (which it’s doing, with great success). That might now change, though of course it will take a long time to build an org that can do anything more for enterprises than just resell raw GPU.
And third, as I wrote when Mark Zuckerberg suggested this in May, if it turns out that Meta has overbuilt, that might be because there are enterprise but not consumer use-cases, but it might also be that everyone else will have overbuilt as well, and so Meta would be reselling into a glut. JULY, MAY
Earlier this year OpenAI and Anthropic announced AI deployment companies, in partnership with PE, strategy consultant and IT services companies. This week Microsoft, Amazon and WPP all announced their versions of the same. See this week’s column. MICROSOFT, AMAZON, WPP
AI data centre demand for memory has driven a crunch across the entire tech industry, with Apple forced to increase Mac and iPad prices by 20-25% last week and memory prices overall increasing 5-10x. This week Samsung and SK Hynix, which with Micron dominance the memory industry, announced a plan to invest about $520bn over the next 15 years in a new chip hub. (NB: this sounds like a stupendously large number until you divide it by 15 and compare it to Samsung and Hynix’s current capex, on which basis this implies a 5-10% annual growth rate.) LINK
More circular revenue (AKA vendor-financing) - apparently Nvidia has been offering neoclouds a deal wherein it will rent back any unused GPU capacity at an agreed price. In theory that backstops their revenue and makes it easier to get financing - Nvidia has $120bn of free cashflow in the last 12 months. More broadly, this is price-support for compute capacity: Nvidia is setting a price floor. However, if we do get a glut and massive downwards pricing pressure, then Nvidia won’t be able to buy all of that - especially since demand for its own chips and hence its cashflow will be affected too. LINK
Anthropic finally joined the ASIC club, with reports that it's talking to Samsung about making its own AI accelerator chip. LINK
Anthropic is also trying to close the loopholes that let Chinese developers use Claude, against its ToCs. LINK
The (presumably hostile) OpenAI documentary that Amazon dropped has been picked up by another studio. LINK
Bending Spoons is a fascinating model: they buy out tech companies that were cool a decade ago, ran out of energy, north and drive, but still have a big and committed or at least locked-in user base (and a good idea at the kernel). Then they lay off all the existing staff and point a turnaround engine at it - buildings full of keen and energetic young people in Milan. Acquisitions include Vimeo, Eventbrite, Evernote and AOL, once a $164bn company. LINK, INTERVIEW
Google just lost the last leg of a legal battle over a €4.1bn fine the EU levied in 2018 over supposedly anti-competitive measures in Android. Close to a decade later, Google (and Apple) have mostly done what the EU wanted and that had zero effect on competition, because the EU’s analysis of the market had no connection to the reality of consumer behavior or platform needs. Meanwhile, 2018 was far to late to intervene in app stores even if the analysis has been correct, and by 2026 we’re deep into the next platform shift and the whole argument is irrelevant. This is the basis operational challenge in regulating tech: the cycle time is faster than regulatory due process, so do you intervene early, consciously and necessarily speculating about what might happen, or wait to intervene until everything is clear, by which time it’s too late? LINK
All the way back in 1978, before some of you were born, when the Apple II was a year old and Microsoft had 13 employees (yes, that photo), Nicholas Negroponte proposed the idea of ‘convergence’ - that the telecoms, media and technology industries would converge on the same common systems. That was crazy talk, and it took 25 years to start looking real, and then for another couple of decades people tried to persuade themselves that there are were synergies between telecoms, media and technology companies. There aren’t. One of the less value-destructive deals that came out of that idea was Comcast buying NBC Universal back in 2009: I know exactly what the Powerpoint would the said, it was wrong, and now, 17 years later, Comcast is unwinding the deal. The stock spiked as the conglomerate discount is removed. LINK, 1978
While the AI companies build DeployCo services units, consultancies, both strategy (Bain/BCG/McKinsey) and IT (IBM, Accenture, Cognizant, the consulting arms of the Big Four accountancies any a bunch of others) are working out both what they will be selling (and who they’ll be hiring) and what they charge for it. Pricing has historically been denominated in bodies and hours, but if you now deliver the same work with fewer people and more software (and token costs, if that remains a thing), how do you charge? Fixed pricing? Outcome pricing (which presumes you can identify a specific return)? Success fees? SaaS companies are wondering the same thing from the other end, since they charge by seat and now have marginal cost for tokens. Law firms and ad agencies will have to grapple with this too. McKinsey says that ‘more than 30%’ of its fees are already priced on client outcomes.
FWIW, my suspicion is that a lot of cases, clients will pay the same absolute dollar fee for a project, but it will be labelled in different ways, and the consultants will work the same hours, using AI to produce more for the same fee. After all, there are more accountants now than there were before PCs, and they’re doing entirely different things. LINK
Remember Jio, the mobile arm of the Indian conglomerate Reliance? A decade ago it blew the market open with radically low prices, taking India to the highest data traffic and lowest prices in the world (and also bankrupting the CEO’s estranged brother). 524m customers, 42 gig/month average use per customer, $2.45 monthly ARPU. Last week it filed for IPO. There was something of an idea that somehow Jio was more than just another deep-discounting telco, with all sorts of ‘smart’ pipe’ and data service ideas (this some around in mobile every few ideas), and it does claim to have a proprietary 4G/5G cellular network stack. but reading the filing… nope, it’s a telco. LINK
Bloomberg has a big piece on the Amazon insiders taking bribes to help shady merchants. They reach out to merchants with a problem over WeChat, giving screenshots to show they have access. In 2020 the US indicted a ring that had allegedly made $100m. LINK, 2020
Remember GoPro? The category is growing fast - for DJI and Insta360, which have 90%of the market. There was no software and no network effect. LINK
Sriram Krishnan, briefly AI advisor to Trump (and formerly a partner at Andreessen Horowitz, joining after I left), gave an interview to the FT saying that he thinks Trump will oppose an ‘FDA for AI’. LINK
Netflix’s head of ads on the future of attention. LINK
One of many American cost diseases - subways are over-specified. LINK
The Venetian tradition of bridge brawls. LINK
Sony will no longer sell Playstation games on physical disks. LINK
Ramp combined its invoice payment data (showing spend on AI) with Revelio payroll data to show that for their user base (skewing to tech and startups), higher AI spend correlates with more hiring. I think the only thing that’s clear about the broader impact of AI on jobs is that there’s no consensus amongst economists that the data shows anything much so far. LINK
The first step in AI enterprise deployment was giving everyone Copilot, which mostly failed and is continuing to fail. The next step was pilots and trials to find and build point solutions and individual use cases. Almost all big companies are doing this, and it can work up to a point, but outside of some very specific industries (say, law), buying or building pieces of software one at a time is unlikely to move the needle for the entire company. To do that, you would need to rethink and reengineer whole workflows, processes, and, indeed, entire functions. You have to move beyond using the new thing as just doing more of the old thing.
That isn't something that any big company does quickly - it takes time even to work out what you might want to do, and it involves a lot of work in working out what the process should be, and then going out and plugging it all together. That creates a paradox, if you like, that automation requires a lot of manual labour.
This means business for all of the systems integrators on one side and for strategy consultants on the other. (Both of these are called ‘consultants’, and indeed a good sign that someone has no idea what they’re talking about is for them to talk as though McKinsey, PwC and Accenture are in the same business.) In parallel, some PE investors base their model on re-engineering their portfolio companies, which is why Anthropic and OpenAI partnered with PE firms as well as consultants as route to market.
Then, of course, there are the vendors themselves. There used to be a joke that a machine learning scientist was a statistician who lived in Silicon Valley, and perhaps now a forward-deployed engineer is an Accenture employee who lived in San Jose. After all, this isn’t that different a problem to what used to be called digital transformation: slow, painful rebuilding of corporate systems. I wonder how far the new AI Deployment units coming out to the Valley really want to do that kind of work (or are ready for those kinds of margins).
The other approach that's emerging, in reaction to that, is the ‘AI-enabled’ or ‘AI-accelerated’ company. Instead of trying to sell software to accountants or law firms, and dealing with the pain involved in transforming a company to really use that software, you create your own accounting service company, that's built from scratch, presuming AI, with all the new processes and architectures and agility. Of course, that raises some very old kinds of question - how much do you really understand that industry, and is software (even AI software!) really the key point of leverage?
Ironically, perhaps, one of the easiest places to do this is in software itself. Find a space where the incumbent enterprise software vendors have been around for 15 or 30 years, and maybe rolled up a bunch of companies that aren’t well-integrated. Then use AI to build it again, better, and cheaper, using AI to build it, and using AI as a core part of the product, exactly as we saw with the move from on-prem to SaaS.
Some of this reminds me of the old line that incumbents always make the new thing a feature: consultancies are making AI a feature. But it also reflects the reality: changing big companies is slow and hard for good reasons, and nothing can ever be done for the first time. If that wasn't true, none of these consultancies would exist in the first place.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。