Move Fast and Break Buyers: Everyone is Tired

OpenClaw, NemoClaw, Mythos, Spud, Gemma 4, half a dozen new MCP servers, three competing agent protocols, a fresh wave of image models, voice models, video models, vector stores, fine tuning platforms and eval harnesses, most of them shipped or announced in the last few weeks, some of it serious infrastructure built by serious people, some of it complete nonsense, all of it impossible to properly evaluate at the rate it arrives. I work in this stuff for a living and I am tired. Everyone I talk to inside an actual business who is meant to be making sense of any of it is tired too. The pace at which AI is being released has decoupled from the pace at which any normal organisation can adopt it, and the gap is widening every month.

Enterprises are slow. They have always been slow, and the slowness is not a bug, it is the entire point of how risk and procurement and security and training and change management are supposed to work. A reasonable mid sized organisation that decides today to seriously adopt some piece of AI tooling is looking at a process that runs through legal review, security review, data protection assessment, vendor due diligence, integration scoping, a pilot, a wider rollout, training for the people who are actually going to use the thing, and a support model for when it inevitably breaks. That is a year, if everything goes well. It is eighteen months if it does not. And the question nobody at the labs seems to want to answer with a straight face is whether the model the enterprise is buying into today will still exist in a recognisable form by the time that process completes. The honest answer, based on the evidence in front of us, is that it almost certainly will not. OpenAI deprecates models on a rolling basis. The thing you trained your team on in spring has been pulled by autumn, and the replacement has different modes that nobody has had time to document yet, never mind train against.

Mobile in the early 2010s is a pretty good example of this. Anyone who was building apps then remembers that the chaos of early Android trained an entire generation of developers to ship iOS first and treat Android as the version they would get round to eventually, on the basic and unromantic grounds that at least one of the two platforms could be relied on not to break its own APIs every release. Apple won the developer mindshare it still has today by holding still, by being predictable, by being the boring grown up in a room full of people moving fast and breaking things. Android eventually stabilised and recovered, but the head start it gave away in those early years has never really been clawed back. Holding still is a competitive strategy. Most of the AI labs do not currently have one.

Which brings us to the obvious question, the one nobody in the labs is keen to engage with directly, which is what any of this is actually for. The official answer is AGI, or ASI, or transformative AI, or whichever term the lab in question has settled on this quarter for the thing that is supposedly just over the horizon. And here is where the whole story splits in two, and both halves are bad for the buyer.

If the labs genuinely believe that AGI is coming in the next two or three years, then the current release cadence is incoherent. You do not ship five point releases of your flagship model in seven months if you sincerely think the next pre-training run is the one that ends it all, because none of the integration work anyone is doing against your current API matters. The rational play, if you actually believe your own roadmap, is to slow down, stabilise, and let the world build something durable on top of what you already have, so that when the big one lands there is an ecosystem ready to absorb it. The behaviour we are watching is the behaviour of companies who do not really believe their own pitch, or who believe it but have decided it is more important to look like they are winning the quarterly benchmark race than to act like a company with a serious destination in mind. Either way the buyer is being asked to commit real money and real organisational change to a technology whose own creators are visibly hedging.

If the labs do not believe AGI is coming in any concrete near term sense, which is the position I suspect most of the people actually doing the work would land on if you got them drunk enough, then the release cadence makes perfect sense, but only as marketing. It is a way of demonstrating to investors and to the press and to each other that a given lab is still in the race and still worth its valuation, and the cost of that demonstration is being absorbed downstream by the people who have to make sense of the resulting noise. In this version of the story, the buyer is being marketed at by people who are essentially playing for the next funding round, and the buyer’s actual operational needs are not really part of the calculation at all. 

If there is a partial exception in the current field it is probably Anthropic, and only because they are the lab whose product surface looks like somebody actually drew it on a whiteboard once and then stuck to the plan. There is no image generator burning compute on viral demos. There is no video generator chasing whatever Sora was supposed to be before OpenAI cancelled it. The product is very much focussed on software development, and whatever you think of Claude Code on the merits, it at least reads as a deliberate bet rather than a hedge against six other bets. You can disagree with the strategy on every level. You can think they have picked the wrong vertical, that coding agents are going to plateau, that focusing this narrowly leaves them dangerously exposed if the centre of gravity in the industry moves somewhere else. None of that is really the point. The point is that focus, of any kind, is starting to look pretty rare, and the rarest things in any market are usually the ones worth paying attention to.

I think the next two years of enterprise AI adoption could be substantially worse than the consensus expects, not because the technology is bad, but because the buyers have been trained, by everything that has happened in the last six months alone, to assume that anything they commit to today will be embarrassing by Christmas. Perhaps some of the labs will figure out before the others do that, that boring in some cases, is a competitive advantage, and that the audience worth winning is the one that has been waiting, very patiently and very sceptically, for somebody to give them something they can actually plan around. There are not many candidates for that role at the moment.