The instrument you installed to prove your AI spend is out of control is now owned by the company that sends you the cloud bill.
DoiT has acquired Attribute — the Israeli startup whose eBPF sensor traces every token, inference and GPU cycle back to the team, product or agent that burned it — for an estimated $65 million, CTech reported on 17 August. DoiT is a channel partner as well as a software vendor: on its own procurement page, "You buy AWS, Google Cloud, or Azure services through DoiT as your channel partner," and DoiT then consolidates that spend into a single invoice. It says it manages more than $20 billion in cloud spend across 4,500 customers in 27 countries, on top of a five-year Strategic Collaboration Agreement signed in November 2023 to drive $5 billion in business with AWS.
The meter and the metered are now the same company. And the meter runs inside your Linux kernel.
What DoiT Actually Bought
DoiT bought a runtime attribution engine, not a dashboard. Attribute was founded in 2023 by Izhak Zimmermann and Liad Tropp, both alumni of the Israeli military's Unit 81, and raised roughly $13.5 million in seed funding before the sale, per CTech. The standalone company is already gone: attrb.io now issues a 308 permanent redirect to doit.com/products/attribute.
The technical claim is specific and, if it holds, genuinely useful. AI cost attribution is the practice of tracing AI spend — GPU cycles, model API calls, token consumption — back to the specific customer, feature, team or agent responsible for it. Attribute does that without tags or SDKs by putting an eBPF sensor in the kernel. Per DoiT's product page, the "eBPF sensor reads traffic into/out of the gateway, tracing each inference call back to the workload, product, and customer that triggered it," breaking out input, output and cached tokens per request across OpenAI, Anthropic, Azure OpenAI and Amazon Bedrock. DoiT says the sensor installs in roughly 15 minutes with no application instrumentation or configuration changes.
That solves a real problem — though note who measured it. The figures DoiT leads with, 79% of enterprises hitting AI cost overruns and only 15% able to calculate AI ROI without significant bottlenecks, come from a survey DoiT itself commissioned: Sapio Research polled 500 finance leaders at US and UK organisations of 1,000-plus employees in February 2026. Independent work points the same way at a lower number — WitnessAI's July 2026 survey of 300 executives found 68% of AI initiatives ran over budget — so the problem is real even if the vendor's own figure is the most flattering version of it. Anyone who has tried to answer "which product line burned this month's inference budget" from a provider invoice knows why tag-based allocation fails — the same reason a RAG pipeline's real monthly cost only resolves when you decompose it per stage.
The technology is not the problem. The ownership is.
Your Cost Auditor Is Also Your Channel Partner
The conflict is structural: DoiT is the reseller of record on the cloud consumption that Attribute exists to reduce, and it does not publish how it is paid — every tier above the free one is listed as "Bespoke pricing". When your reseller also owns the kernel-level meter that decides what counts as waste, the instrument and the invoice come from the same P&L, on terms nobody outside the contract can see. That is a different relationship from buying attribution software from a party with no economic position in your bill — and it is the same shape as an AI red-teaming vendor being acquired by the company whose runtime it audits, or a payments network buying the fraud-scoring engine banks use to grade its own rails.
Here is the strongest version of the other side, because it is not weak. DoiT's business model is explicitly built against that incentive: its pricing page carries a savings guarantee — "We guarantee that we save you more than we charge" — and it markets optimization products whose entire value proposition is cutting the bill it invoices. A reseller that quietly lets customers overspend loses them at renewal, and DoiT has spent roughly $200 million on five acquisitions in two years (PerfectScale, LiveDiagrams, CloudWize, SELECT and now Attribute) building an optimization portfolio, not a billing one. Its pricing page also states you can use the software while procuring cloud through a different partner. Hyperscaler-native cost tooling has a sharper version of the same conflict and nobody rips out Cost Explorer over it.
All true. None of it changes what a procurement file needs to record. A vendor's incentive alignment is a claim; a contract term is a commitment. The reason to write it down is not that DoiT is acting badly — there is no evidence it is — but that "our interests are aligned" is not a clause you can enforce in eighteen months when the account team, the comp plan and the margin structure have all changed. This is the same lesson as a data-retention toggle surviving a gateway acquisition: the setting is a product decision, and product decisions are revocable by whoever owns the product.
Nobody Re-Reviewed the Kernel Sensor
The security question is sharper than the commercial one, because an eBPF sensor is one of the most privileged things you can install. eBPF is a Linux facility that runs verified programs inside the kernel; loading and attaching one requires elevated capability, and Red Hat's engineering team notes that using eBPF in applications "may require escalating pod privileges to CAP_SYS_ADMIN or CAP_SYS_BPF level, which can compromise security." Those capabilities are not decorative. The eBPF verifier bug tracked as CVE-2021-31440 allowed unprivileged users inside a privileged container to escalate to the node and break out of the container — a demonstration of what the blast radius looks like when kernel-level code goes wrong. DoiT's own explainer makes the opposite case, that eBPF programs "run in a verified, sandboxed environment, so they can't crash the kernel or introduce security holes". That is the standard defence and it is mostly right — but CVE-2021-31440 was a bug in the verifier, which makes it an argument for patched kernels and enumerated privileges, not for skipping the review.
Now read DoiT's own explainer on what its sensor does. It "observes system activity directly from the Linux kernel, without tags, SDKs, or code changes" and "streams what it observes to a collection layer outside the kernel." The product page says the sensor reads traffic into and out of your LLM gateway. That is the exact position in the stack where prompts, retrieved context and completions transit — and DoiT's published material does not state whether payload content is parsed or discarded, or whether observed data stays inside your environment. Those are answerable questions. They are just not answered on the marketing pages, which is where most security reviews stop.
Whoever signed off on that sensor signed off on a seed-stage startup with roughly $13.5 million raised and two founders. The entity operating it today is a global company invoicing 4,500 customers across 27 countries, with a different security program, a different sub-processor chain and a different set of commercial incentives. That is not an accusation — DoiT publishes SOC 2 and ISO 27001 evidence through its trust centre, and a startup agent absorbed into a certified organisation may well end up better governed. It is simply a change of control on a privileged host agent, and change of control is the trigger condition your third-party risk process already has a clause for. The same reasoning applies to any component holding standing privilege in your environment, which is why standing IAM permissions granted to agents keep producing authorized-but-unintended escalation.
DoiT Launched the Product. It Never Announced the Deal.
The disclosure sequence is the part worth putting in front of your vendor-management lead. Attribute shipped as a DoiT product on 7 July 2026 — the changelog entry, the launch post and the accompanying announcement all present it as a DoiT launch. CTech reported the deal and its estimated price on 17 August, six weeks later. As of publication DoiT has issued no acquisition announcement of its own.
That silence is the tell, because DoiT is not a company that stays quiet about acquisitions. It published a press release for PerfectScale, for LiveDiagrams and for CloudWize. For Attribute there is a product launch instead, and the launch announcement introduces Izhak Zimmermann only as "General Manager of Attribute" — accurate, and silent on the fact that he founded and ran the company DoiT had bought. No reasonable buyer reverse-engineers a change of control from a job title in a launch quote.
None of this is unusual or improper; private acquirers announce on their own schedule, and there is no obligation to press-release a deal. It is a practical problem all the same, and it is the same failure mode as an AI vendor joining OpenAI in March and customers finding out in August. A vendor-security review is triggered by an event. If the event is never announced, the review never fires — and for six weeks the trigger for re-reviewing a kernel-level agent existed without anyone downstream being able to see it.
Assume your vendor inventory did not catch this one. Then ask how many others it also did not catch.
The Measurement Standard Is Being Written Now
There is a live venue for fixing this, and it opened two weeks ago. On 4 August 2026 the Linux Foundation launched the Tokenomics Foundation with 30 initial members to "establish open, vendor-neutral industry standards, benchmarks and best practices for the economics of AI", covering tokenomics definitions, a reference model for the full cost of AI, cost-to-serve measurement methods and token cost telemetry in the FOCUS specification. J.R. Storment runs it alongside the FinOps Foundation. Members include JPMorganChase, IBM, SAP, Oracle, ServiceNow, Accenture and Flexera.
DoiT is one of them. It became a founding member with Field CTO Amit Kinha on the governing board, announced the same day. Thirty members and a Linux Foundation charter is a genuinely vendor-neutral structure — one seat is participation, not capture, and a reseller with $20 billion of spend under management has more real telemetry than most of the room.
The point is what buyers should ask that body for while the definitions are still soft. A cost-attribution standard that specifies token classes and cost-to-serve arithmetic but says nothing about who is allowed to publish the number has skipped the assurance question entirely. Financial audit resolved this decades ago with independence rules, not better arithmetic. If your organisation sits inside FinOps or Tokenomics working groups, the disclosure of an attribution vendor's economic interest in the spend it measures is a cheap thing to put in the spec now and an expensive thing to retrofit later.
What to Do Before Your Next Renewal
Treat this as a change-of-control event on two separate things: a measurement vendor and a privileged host agent. They need different owners and different questions.
This Week:
- Find the sensor. DoiT says it "deploys to your clusters" but does not publish the workload type, so search by image rather than by name, and record exactly what capabilities its pods hold —
privileged,hostPID,hostNetwork,CAP_BPF,CAP_SYS_ADMIN— and which nodes it runs on. You cannot review a privilege you have not enumerated. - Pull the original approval. Whoever signed off on Attribute signed off on a different legal entity. Put the ticket, the questionnaire and the date in front of the person who owns third-party risk today, and mark it stale.
- Send four written questions to your DoiT account team. Is payload content from gateway traffic parsed, stored or transmitted? Where does sensor telemetry land, in which region, and under which sub-processors? Has the sensor been in scope for a SOC 2 or penetration test under DoiT's program? What is the notice period before a new sub-processor is added?
This Month:
- Separate the meter from the invoice in your own reporting. Whatever Attribute reports, reconcile at least one month against raw provider billing exports and your gateway's own token counts. Any LLM gateway you already run emits an independent count — two sources that agree are worth far more than one source that is confident.
- Name the conflict in the risk register. One line: the vendor supplying cost-attribution data holds a resale margin on the spend being attributed. Rate it, accept it if that is the call, but stop carrying it implicitly.
- Check the rest of your observability chain for the same pattern. Cost and telemetry vendors are consolidating fast — Dynatrace buying Arize is the same quarter's version of this — and change-of-control clauses are worth exactly as much as the inventory that tells you when to invoke them.
Before Renewal:
- Get the independence commitments in writing. Attribution methodology disclosed and versioned; notice before it changes; a right to export raw attributed data in a portable format; sub-processor notice with a right to object; termination rights on a material change to the sensor's data collection. Ask specifically whether pricing for Attribute is tied to procurement staying with DoiT.
- Price the exit. Establish what replacing kernel-level attribution costs you in instrumentation work if you ever move procurement. If that number is large, it is not a tool — it is a switching cost, and you should know it before you sign, not after.
The Bottom Line
Every measurement layer in the enterprise AI stack is being bought by someone with a position in what it measures. The red teamer, the observability vendor, the fraud engine, the gateway, and now the cost meter — all acquired inside a single quarter, all by parties who are not neutral about the readings. Financial accounting spent a century learning that the auditor cannot work for the audited, and it learned it the expensive way, one scandal at a time. AI cost governance is currently repeating the entire syllabus in about eighteen months.
The instrument is probably fine. Verify it anyway.
Trust the number when someone with nothing to gain can reproduce it.
Continue Reading
- Stripe Bought OpenRouter. A Toggle Is Not a Contract.
- Visa Bought BioCatch. Get Neutrality in Writing.
- Anaconda Bought Your AI Red Teamer. Read the License.
- Your AI Vendor Joined OpenAI in March. You Heard in August.
- Dynatrace Bought Arize. Fix Your Spans Before It Closes.
- Best LLM Gateways for Cost Control: Self-Host First
- What RAG Actually Costs: $1,308 a Month at 10M Tokens/Day