Two Billion Against Your Labor
Three numbers describe the same pressure from three different directions.
Start with Trend Micro, which reported its second quarter results. The headline looked healthy — total net sales up thirteen percent year over year, and its AI security product, Vision One, growing annual recurring revenue forty-nine percent. That is the number a vendor puts on a slide. Then read the next line. Cloud costs nearly doubled in a year — from six point eight billion yen to twelve point seven billion — and the company named AI token costs directly. Operating income for the quarter fell fifty-four percent. Then it cut its full-year profit forecast by twelve billion yen, taking its expected operating margin from nineteen percent down to fifteen. Net sales guidance did not change at all. Only the profit did. So the AI product is selling exactly as planned, and the cost of running it is eating the margin underneath. That is not a company in trouble. That is a company disclosing, in public, that the thing customers want most now carries a cost that scales with how much they use it.
Now the second one, and it comes from the other end of the market. A firm called Thrive Holdings raised two billion dollars, at a twelve billion dollar valuation, from SoftBank and two other investors. Its purpose is buying service businesses and rebuilding how the work gets done with AI. It already owns more than seventy of them. One arm, called Current, holds over fifty accounting firms and more than two thousand professionals. The other arm is called Shield, and it holds about twenty information technology services companies. OpenAI owns a stake in the parent and sends its own people in to work inside them. Set aside for a moment whether anyone is knocking on your door. Two billion dollars is what it costs to state a conviction out loud, and the conviction is that a meaningful share of the labor inside professional service delivery can be removed without the customer noticing. That is not a prediction from a conference stage. That is professional investors putting a number on it.
And the third, which says the separation has already started. Techaisle’s global channel partner survey found that partners under ten million dollars in revenue project growth of eight point four percent. Partners above five hundred million project sixteen point eight. Same market, same vendors, same technology available to both. One group is growing at roughly half the rate of the other, and the gap is in the forecast, not the history.
A vendor’s costs going up. Two billion dollars betting service costs come down. And a growth gap that has already opened between the top of this channel and the bottom of it.
Those three things are not a coincidence. They’re the same shift arriving from opposite ends of your own books.
Software Got a Cost of Goods
Software has always had one economic property that made it different from every other business: once it is written, the next copy costs nothing. That single fact is why software companies carry the margins they do, and why reselling software has been a reasonable living for thirty years.
Inference breaks that. When a product answers a question using a model, that answer costs money — every time, for every customer, forever. Gartner captures where this is heading in its infrastructure forecast: AI-optimised cloud spending is climbing toward forty-two billion dollars, and the spend is shifting from training models to running them. Training is something you pay for once. Running is something you pay for on every use. And here is how large that appetite has become. The memory those systems need is in such short supply that it has repriced a market with nothing to do with AI. Asus and Gigabyte have now raised graphics card prices twice, on gaming cards, because AI data centers are consuming the memory supply those cards depend on. The most popular models are up roughly forty percent. Nobody buying a gaming card is buying AI. They are paying for it anyway.
Now the obvious objection, because it is a good one. Model prices are collapsing. OpenAI cut its GPT-5.6 pricing hard — one model by eighty percent, another by twenty. If the per-unit cost is falling that fast, where is the problem?
The problem is that per-unit price and total cost are different numbers. When something gets cheaper, products use more of it. The vendor’s bill is volume times price, and volume is winning. That is precisely what a software company watching its cloud costs nearly double is telling you, in a year when the unit prices it pays went down.
Now the other side of the ledger. Services never had software’s economics — the marginal cost of service delivery is a person, and that is what justified the price. Automation removes it. Halo announced an AI Studio letting providers build unlimited service-desk agents, alongside price reductions. Sophos announced it is putting OpenAI’s models directly into its security platform, sold to providers — that is the company’s own announcement, so weigh it accordingly.
Read who receives those releases. Everyone. Simultaneously. And when every competitor’s delivery cost falls in the same quarter, that saving does not become margin. It becomes somebody’s lower quote.
Which is when it stops being an industry trend and becomes a number on your own invoice.
So the question is which of your numbers still belongs to you.
Get On Their Income Statement
So look at what your client is actually watching.
Research from Mavvrik — a cost governance vendor, so weigh it accordingly — found that sixty-two percent of organizations have altered a business decision because of unexpected AI expense, forty percent escalated it to the board, and a third imposed emergency spending freezes. And here is the number underneath those: ninety-eight percent of them can see their AI bill. Eleven percent can predict it — down from fifteen percent the year before. Visibility is solved. Forecasting is getting worse. Set that alongside Gartner’s independent read on the same pressure: global IT spending revised up to six point three seven trillion dollars, growth of fourteen point two percent, driven by AI infrastructure. Both point at the same thing. AI cost has stopped being a technology topic and become a line item somebody’s board is asking about by name.
That is an opening, and it needs nothing you don’t already have. When a cost has a number, a governance process, and a nervous executive attached to it, a provider can take that number down and say so in the client’s own units — this much was being spent, this much is being spent now, here is the difference. That is not a promise about service quality. It is arithmetic on a statement the client already keeps, that their accountant already reads, and that nobody has to take your word for.
The second one is harder and worth far more. Semafor reported that the gap between corporate AI leaders and everyone else is widening sharply, with the heaviest users consuming vastly more than the median firm. The separation is not between companies that bought AI and companies that didn’t. It is between companies that pointed it at where they actually make money and companies that pointed it at whatever was easiest.
Knowing where a business makes money is not a technology skill. It is knowing which customers are profitable, which quotes get lost and why, where the delay between doing the work and getting paid actually sits. A provider who can name a revenue line and explain how they moved it has proven they understand the business — and that provider does not get replaced, because the knowledge lives in them and not in the tooling.
So here is the choice. Put your service lines on the client’s income statement — the cost you took out, and eventually the revenue you enabled, in their numbers on their statement. Or stay denominated in licenses and hours, and watch the price of both get set by a vendor’s inference bill and by every competitor who just bought the same automation you did.
Which sounds like an easy call right up until you count what you’d be giving up.
The fair objection is that the commodity tier is not dead weight — repeatable work at scale is where the leverage lives, and advisory work is headcount-bound and first to get cut in a downturn. But moving up does not mean giving that up. The automated tier is how you produce the cost reduction you now get paid for; you keep the leverage and stop selling it as the product. The provider who loses is not the one who kept the commodity work — it’s the one who kept charging for it as though the client couldn’t see what it costs.
What to Consider
- Find out whether you can even name the number, before you build a pitch around it. Pick your three largest clients and try to write one sentence for each: the cost that went down, in their dollars, because of something you did. Most providers discover they can describe the work in detail and cannot state the figure at all. That gap is the actual project — not a new service line, not a new stack, just the arithmetic you have never had to produce because the invoice never asked for it.
- Separate what automation saves you from what it saves the client, and decide deliberately who gets it. When your service desk agents take hours out of your delivery, that saving is real and it is yours until a competitor’s identical saving turns into a lower quote. Keeping it quiet buys you a quarter or two. Handing it to the client as a named, quantified reduction buys you the position of the provider who told them first. Both are defensible choices — but the default, which is letting it leak out slowly as discounting, is the only one that is not.
- Assign the client’s business model to a person, the same way you assign a stack to a person. Somebody in your shop should be able to answer, without looking anything up, how each major client makes money — which customers are profitable, where quotes get lost, where the lag between doing the work and getting paid sits. That is the knowledge revenue mapping runs on, it takes months to build, and it is the one capability in your business that is not arriving in anybody’s product release.
If this trend continues, within twelve to eighteen months, the first question in a competitive services bid stops being what you run and becomes which line on the client’s income statement you moved last year — and the provider who has never opened one will be answering from a stack sheet while somebody else answers from the client’s own books.