Three Companies, One Idea
Three companies shipped the same idea, at three different sizes, and none of them called it what it is.
Start with Broadcom. At its Explore conference in Singapore, the company announced VMware Private AI Cloud — a platform for running inference and agentic applications on hardware the customer owns, instead of sending that work out to a public AI service. That is Broadcom describing its own product, so take the framing for what it is. But note what they say it is for. They name three cost drivers the thing exists to address, and the third one is token economics.
Then Perplexity, which went the other direction — down instead of up. Working with Nvidia, the company released something called Portable Computer: a full AI agent running on the machine in front of you, escalating to the cloud only when it decides it has to, on a DGX Spark desktop or on any Nvidia card with twenty-four gigabytes of memory. The pitch, and these are Perplexity’s own benchmarks, is zero token cost. The demonstration was a folder of tax documents, read locally, with no cloud usage at all.
Then Apple. Two new desktop chips — the M6 in the Mac mini, the M5 Ultra in the Mac Studio, with fifty percent more memory bandwidth than the part it replaces. Apple says Mac revenue grew twenty-nine percent year over year, to ten point four billion dollars. But the useful voices on this one are not Apple’s. Ranjit Atwal, a senior director analyst at Gartner, told CIO Dive the chips were built to enable future AI workloads — agents, specifically — and then made a point about the money. The budget for what he calls an infrastructure PC does not come from where desktop budgets normally come from. It comes from the infrastructure side of the house.
Techopedia’s read on the same launch is about the shape of the thing — a desktop with no lid to close and no battery to drain, that simply stays powered, which is exactly what an agent running continuously needs and a laptop is not.
And Brendan Burke, at the Futurum Group, put the whole thing as an open question. Enterprises want frontier models running on hardware they control, he said. The test is whether local inference at this scale actually pulls workloads off rented GPUs and onto the desk.
So hold that word. Whether.
Because here is where the work is right now. NVIDIA closed its second quarter at ninety-six point two billion dollars, up a hundred and six percent from a year ago, at seventy-five percent margins. Eighty-nine billion of that is data center silicon, up a hundred and seventeen percent from a year ago. That is the hardware sold to the people who rent GPUs out, and it has never sold faster.
Three companies building for hardware the customer owns. One company being paid, enormously, for hardware they rent.
Which of those is the leading indicator depends on a number that has nothing to do with either of them.
Nobody Turned It On
The reason the hardware is arriving ahead of the workload is that almost nobody has turned the workload on.
Look at what adoption actually means when somebody measures it carefully. Alexander Bick and his colleagues run something called the Real-Time Population Survey, and The Register carried the latest results. Generative AI has now reached roughly eighty percent of occupations. Four out of five detailed occupations show adoption above twenty percent. That is the number everybody quotes, and it sounds like saturation.
Here is the number nobody quotes. Only one occupation in six is above seventy percent. And measured at the level of individual tasks — the actual unit of work — only two point eight percent show adoption above half. Two point eight percent. And not one single task in the survey clears seventy percent. Not one. As of May, sixty-two percent of American adults use generative AI at all, and forty-five percent use it at work.
So the picture is enormously wide and almost entirely surface. Nearly everyone has touched it. Nearly nobody has moved their work onto it.
And the natural reading of that is patience. Adoption curves fill in. Wait three years and the depth arrives by itself.
That reading is wrong, and there is a company that shows you why.
Voya Financial has about eleven thousand employees. They built a five-hour AI literacy course — five hours — with input from more than a thousand of their own people, and put the workforce through it. Nearly a hundred percent certified. Copilot adoption hit ninety-eight percent. And here is the figure that matters: those employees now average twenty-four prompts each, every week. About three hundred of them have gone on to build their own agents. And Voya’s chief technology and operations officer, Santhosh Keshavan, does not credit the tooling for any of it. He credits how the leadership talked to people about it.
Nothing about the technology changed at Voya. Same models, same licenses. What changed is that somebody decided to teach eleven thousand people — and a company that was generating almost nothing is now running north of a quarter of a million prompts a week.
That is the mechanism. Consumption is not a property of the model. It is a property of how many people were taught what to do with it. It does not accumulate on its own. Somebody has to cause it.
Which tells you exactly what today’s price signal is measuring. CIO Dive reported this summer that enterprises have entered what they called the cautious era — the phase of using AI for everything gave way to consumption pricing and bills that got somebody’s attention. Gartner still expects spending on AI models and platforms to climb sixty-three percent, to sixty-four billion dollars. The cost pressure is real, it is being felt, and it is producing precisely the response you would expect. Buy the hardware. Own the compute. Stop renting.
But set the two numbers beside each other. The bills driving that response were generated by two point eight percent of tasks.
The hardware answer is not wrong. It is early — and it is being sized against a bill that reflects almost none of the work.
Early is a different problem than wrong, and it is the harder one to act on.
The Window Opens In 2029
What that leaves you holding is a timing problem, and timing problems are the ones this industry has historically lost.
Gartner has put dates on it, and Redmond Channel Partner carried the numbers. Their argument starts with licensing and pricing changes in virtualization, which have handed enterprises a reason to reconsider the platform underneath everything — and you cannot move the virtualization layer without opening the storage question, so the two now get evaluated together. Gartner calls it a major displacement window. And into that window comes the AI piece. Consumption-based storage-as-a-service, they project, goes from a quarter of on-premises storage spend at the start of this year to half of it by 2029. Purpose-built storage tiers for large-scale inference go from under ten percent of enterprises today to seventy percent by 2029.
Notice what that actually describes. The on-premises infrastructure conversation is going to reopen at your clients regardless — for hypervisor reasons, not AI reasons — and the AI hardware question rides in through a door that was already going to swing. On a horizon of about three years.
So the obvious move is to get ready now. Build the practice, take the training, be standing there when it opens. That is the move I would not make. Three years is a long time to carry a practice you cannot yet sell, on a timeline that is not yours to set.
So here is the choice, and three product launches have already told you which way they would prefer you went.
Do not build the on-premises AI practice this year. Build the adoption practice. Go to the clients who bought the licenses eighteen months ago and never taught anybody to use them, run the five-hour version of what Voya ran, and charge for it. That is revenue in this fiscal year, and it is the only work that moves a client toward the day the hardware math actually changes.
Or build what the launches are pointing at, carry the cost of being ready, and wait on a crossover whose timing is set by adoption you decided not to influence.
And there is a practical problem sitting underneath that choice that nobody mentions.
Because the adoption practice needs a person your shop probably does not have. Everybody on your bench can configure the tenant and set the policies; almost nobody on it can stand in a room with forty accountants and change how they work on Tuesday. That is a teaching skill, not a technical one. The cheapest way to find out whether you can sell this is to run the five-hour course on your own staff first, and watch who turns out to be good at delivering it.
What to Consider
Find the teacher you already employ, and time them. The person who can deliver this is almost certainly already on your payroll, and it is probably not your most senior engineer — look for whoever the rest of the team goes to with questions, or whoever handles client onboarding without anyone complaining. Run the internal version first and put a clock on it. If it takes them eight hours of preparation to deliver five hours of material, you have learned what your margin looks like before you ever put it in front of a client.
Start with the client whose licenses are already paid for. Do not open this with someone who has to buy something first. Open it with the client who bought Copilot or the equivalent eighteen months ago and whose usage you already quietly suspect is close to nothing — the money is spent, the internal argument about whether to do AI is over, and the only missing ingredient is somebody teaching people what to do on a Tuesday morning. That is the shortest distance between a proposal and a delivered result anywhere in your catalog this year.
Write down what you are not doing, and calendar the revisit. Restraint only works when it is an actual decision instead of a drift, so put “on-premises AI infrastructure practice” on a specific quarter with the two or three conditions that would change the answer written next to it — a client asking unprompted, a hypervisor migration landing on the books, somebody complaining about a token bill. Without that, the vendors will reopen this conversation with you every quarter until you eventually say yes on their timing rather than on yours.
And picture the provider who starts this in October. A year from now they have run the five-hour course at four different clients. They know which two people on their team can actually teach and which cannot. They have watched four companies go from almost nothing to something real, and they know what it took each time. That is a business that learned something specific about itself in a year, which is more than most of us can say about the last one.
If this trend continues, by 2029, when that displacement window is fully open at the average client, the providers who spent the intervening years inside those accounts doing the adoption work will be the only ones whose infrastructure numbers come from somewhere — and everyone else will be quoting the identical figure off the identical calculator.