Where enterprise AI budgets go when spending discipline arrives
The defining tension in enterprise AI right now is not capability. It is cost control. In ETR’s July Macro Views Survey (n=1,636 technology leaders), 47.0% of organizations reported AI spending running above plan, yet only 17.0% pause or scale back initiatives when it happens. The rest seek supplemental budget, defer the reckoning to the next planning cycle, or harvest funds from elsewhere in the IT budget. That arrangement has a shelf life, and ETR’s monthly survey work suggests enterprises expect real consumption governance to arrive within a few months, not a few years.
So where does the pressure go when the discipline lands? ETR’s latest drill down survey, fielded to 100 technology leaders on open-weight AI models, points to an answer that should interest anyone underwriting the economics of the AI trade.
Among surveyed technology leaders, 78.0% are already piloting open-weight models or running them in production, with nearly a third in production for at least one workload. Only 11.0% have not evaluated them and have no plans to. Production users report that open-weight models’ share of their AI token consumption has climbed markedly over the past year, and expectations for the year ahead point in one direction: not a single production user in the survey expects that share to recede to trivial levels.
The driver is not ideology, and it is not primarily data privacy, though that ranks high. It is cost. Cost savings ranked as the number one driver of open-weight adoption among both pilots and production users, and among the small group of non-users, nearly all said demonstrated cost savings would move them.
The drill down asked a question we have not seen quantified elsewhere: assuming comparable performance, how large a discount would it take to move a workload from a proprietary model to an open-weight one? The full thresholds are reserved for ETR clients, but the shape of the answer is worth stating publicly: the discounts buyers require are modest, most named figures far smaller than the price gaps already emerging in the market, and buyers with production experience require even less than those still piloting. Familiarity lowers the switching bar; it does not raise it.
Read the last two weeks of market behavior against that finding. OpenAI cut prices on its newest model and publicized rapid usage growth in its agentic products. Anthropic extended access to its flagship model across paid plans and reset rate limits for all users. Moonshot released Kimi K3, an open-weight model that early third-party benchmarks place near the closed frontier at a fraction of the token price. Thinking Machines released Inkling, a US-trained open-weight model positioned explicitly as a base for enterprise customization. Nvidia continues to push its Nemotron family, with weights, data, and training recipes fully open. Capacity giveaways and price cuts are what repricing looks like when vendors are working to stay inside their customers’ switching thresholds.
If cost were the whole story, the migration would already be further along. It is not, and the reasons matter. Among surveyed pilots, the top obstacle to moving open-weight models into production is an incomplete security or compliance review, followed by organizational risk aversion. Performance concerns rank below both. Country of origin operates as a gating factor as well: nearly half of pilots call it a major or deciding consideration in model selection, which shapes how enterprises weigh Chinese-origin releases against US-trained alternatives.
Respondents also describe a hybrid end-state rather than a wholesale replacement of frontier models. Open-weight deployments concentrate in cost-sensitive, domain-specific, and privacy-sensitive workloads, while closed frontier models keep the work where raw capability matters. And self-hosting is not a free lunch: standing up inference in the tightest hardware market in years converts a metered software bill into either upfront hardware outlay or a rented-GPU bill priced off the same inflating components.
For investors and analysts, the survey suggests three threads to watch. First, model-layer pricing: if buyers’ switching thresholds are as low as our data indicates, sustained price competition at the model layer is structural, not promotional. Second, the budget donors: Macro Views shows AI overruns being funded by cuts to external contractors and consultants, legacy infrastructure, and non-AI software licenses, with security software comparatively protected. The categories on the donor side of that trade face a demand headwind that has already begun surfacing in vendor results this earnings season. Third, hardware pricing: Macro Views recorded hardware price inflation of 6.9% Y/Y, nearly double January’s reading, and purchasing pulled forward ahead of expected increases, a dynamic at least one major vendor has cited in its own guidance.
ETR clients can access the complete Open-Weight AI Models drill down, including the full switching thresholds, model family adoption shares, and expected token-mix trajectory, along with the July Macro Views Survey Findings Summary, on the ETR Platform. To learn more, contact service@etr.ai.