Projected 63% rise in AI model spending puts focus on cost per successful outcome

Kanitkar's two numbers cover different years. The spending projection covers 2026, the year most finance teams are budgeting now ^[1]. The Gartner prediction about smaller engineering teams is for 2029, three budget years further out ^[2][15]. A board approving model spend this quarter is answering the first question, and it can revisit the second twice before it falls due.

Kanitkar wrote that "A model that costs half as much per token but requires three attempts is not necessarily less expensive" ^[6]. Three attempts at half the unit price cost 1.5 times one correct call at full price, so 50% more before anyone prices the checking ^[13]. He argues the checking is where the money goes: a cheap model that needs extra prompting or creates hours of human review can cost more than a premium model that completes the task correctly the first time ^[16].

Cost per successful outcome is an operations measurement. Six of the seven inputs Kanitkar counts have to be captured inside the company; the model bill is the only one that arrives from a vendor ^[14]. A buyer that does not already log retries and review time per workflow cannot compute the figure for one model, and cannot rank three.

The framework is also the product. Kanitkar co-founded Larridin, a Bay Area startup building an organizational platform powered by AI ^[3]. An argument for spreading work across several models supports a market for the layer that decides which model gets the work. The column credits Gartner for the smaller-teams prediction and does not say where the 63% projection comes from ^[17]. The narrower claim a buyer can test on its own data is whether a token price predicts what a finished task costs. He stated the rule this way: "The objective should be to use the least expensive model that is capable of reliably delivering the required outcome" ^[5].

On infrastructure, Kanitkar wrote that "Given the pace of model development, buying dedicated hardware for most AI workloads can create unnecessary rigidity" ^[9]. He allows one exception: the completely isolated environment a company needs for security, regulatory or operational reasons. Otherwise he puts the work on hyperscalers or specialized AI clouds, which give access to open models without the company buying and managing GPUs ^[10]. A firm that has already committed capital to one architecture has fewer moves available when a cheaper model appears months later ^[9].

Open-source models are the other lever he names, both to cut cost and to reduce dependence on proprietary providers ^[11]. As public tests saturate, he wrote, more teams are evaluating models on their own tasks and data ^[12]. Many companies still prefer competitive U.S.-developed models over Chinese alternatives for security, governance and procurement reasons ^[11].