Attribute™ measures real-time consumption on the kernel and traces each token, mannequin request and GPU cycle to the client, characteristic and agent that drove it
DoiT, the corporate behind Cloud Intelligence™, launched Attribute™, a expertise that attributes AI spend throughout tokens, mannequin requests and GPU utilization with zero instrumentation. Now usually out there, Attribute™ requires no SDK, no tagging coverage and no code adjustments. A light-weight eBPF sensor installs in about quarter-hour and begins producing per-customer, per-feature and per-agent token economics the identical day.
The strain behind that is measurable. DoiT’s inner buyer information tasks month-to-month AI spend will triple within the subsequent 12 months. But in a latest survey of 500 leaders at giant enterprises, solely 15 % mentioned they may calculate AI ROI with out important bottlenecks, and the common value overrun was extraordinarily near the tolerance breaking level. Spending is climbing, returns are getting tougher to show and crucially, nobody can see the place the cash is definitely being spent.
Till now, attributing AI spend meant instrumenting it. Engineering groups wrapped each mannequin name in an SDK, threaded metadata via each request or enforced tagging requirements that shared GPUs and single-account mannequin APIs had been defeated by design. An API key or tag can not cut up a GPU operating many fashions directly, and an SDK solely sees the calls somebody remembered to wrap. The work falls on engineers, the protection is at all times partial and the numbers drift from actuality.
Additionally Learn: AiThority Interview with Matej Bukovinski, Chief Know-how Officer at Nutrient
Attribute™ takes the other method: it measures what truly runs. The sensor observes actual consumption contained in the working system kernel and maps each unit of GPU, CPU, API name, reminiscence, community and I/O again to the method, container, pod, and request accountable. It additionally identifies every outbound name to a managed mannequin API and joins it with supplier value information, so token spend on Anthropic, OpenAI, Google Gemini and AWS Bedrock lands on the workload, tenant, and agent that drove it reasonably than on a shared account nobody can break down. Cached, reasoning, enter and output tokens are cut up robotically. The result’s token economics that holds as much as audit: value per token, per request, per session, per buyer and per agent, with no engineering effort to provide or keep it.
The identical measurement extends past AI. Kubernetes clusters, multi-tenant databases, storage buckets and networking are attributed the identical approach, so the price of any shared useful resource lands on the workload that brought about it. Groups see the associated fee to serve every buyer, the gross margin on every account and the unit economics of every AI characteristic, generated robotically and repeatedly.
“For fifteen years the business handled value attribution as an instrumentation downside: tag it, wrap it, label it and hope protection holds,” mentioned Izhak Zimmermann, Normal Supervisor of Attribute, “Tokens are the atomic items of AI, and you may’t tag your approach to the reality inside a shared GPU or a single Bedrock account. Attribute measures what truly ran and what truly referred to as, on the kernel, with nothing to instrument. Fifteen minutes to put in.”
