A 30-day retention clause routes Nvidia's sensitive work to its own Nemotron models
Anthropic's June decision to keep Fable usage logs for 30 days has turned zero data retention into a procurement gate at Nvidia, Booz Allen and Palantir. The metadata channel stays open under the clause they are demanding.
The Engineer · Build desk
What happened
- In June, Anthropic gave itself the right to retain usage logs from its flagship Fable model for 30 days, saying it needed them to defend against complex and novel attacks.
- A large U.S. utility canceled its plan to test whether Fable could run its core power infrastructure after Anthropic would not agree to a nonrevocable zero-data-retention policy.
- OpenAI in August began letting GPT-5.6 Cyber customers keep security logs on their own servers, and Anthropic is rolling out a similar program to select customers this fall.
Why it matters
- Who has to decide decision The retention term now decides model selection before capability does: the frontier model keeps the low-sensitivity tier of work, and in-house or open-source models take everything the buyer classifies as proprietary.
- What it limits constraint An irrevocable ZDR clause deletes the log and leaves metadata, classifier outputs and usage telemetry in place, so the term these buyers are fighting for does not cover the extraction path Schulman describes.
- What it sets up precedent With both labs now conceding customer-held log storage to some accounts, customer-controlled logs become the term the next enterprise negotiation opens with.
Next story · Security · Sep 30
Trump's AI accord leaves model safety checks to the model makersConfirmed5 publishers
Retention for misuse detection means a log store. The prompt and the completion land in it, an abuse-review process can reach them, and they age out on a timer [1]. That pipeline does not distinguish a jailbreak attempt from a security vendor's source tree. Bill Vass, Booz Allen's chief technology officer, said he worries a little that Fable might be learning from some of the firm's code [6]. Buyers are arguing over the word irrevocable. Palantir has pressed Anthropic for irrevocable zero-data-retention guarantees before it will make the models available through its own software [7], and CEO Alex Karp told a customer event that companies are tired of being "exploited" by AI labs [14]. The-decoder notes that Palantir's position also suits Palantir, which would prefer customers ran models through its platform instead of going to providers directly [15]. Zero data retention removes the log store and leaves the metadata channel open. Both labs collect metadata and technical usage data from enterprise customers who hold no-training contracts [3]. OpenAI states on its website that it runs business data through automated classifiers and security tools "to better understand how our services are used", and that the resulting classifications "do not contain any of the business data itself" [9]. C Spire, a telecoms company with no-training agreements at both labs, believes the technical usage data its contracts permit includes information about which applications the models are connected to [18]. OpenAI says it does not train on chain-of-thought data [19]; Anthropic says the usage data it collects is aggregated, anonymized and not used for training [28]. John Schulman, an OpenAI co-founder who worked briefly at Anthropic and is now at Thinking Machines, has described where the extraction actually happens [22]. Reinforcement learning tasks built from user traces carry a low risk of reproducing content and can still pull out customer IP; he puts explicit user feedback in reward-model training at the harmless end and "upload user's coding environment and commit history to turn into rl envs" at the invasive end [21]. "De-identification is weak," Schulman said, adding that users can be traced back "with just a small number of bits" [20]. Sarah Hooker, who previously worked at Cohere and Google DeepMind, points to "clever synthetic data techniques that can generate distributional equivalent data while preserving privacy" [24]. Her advice to firms in that position: "If you are a company with IP you have a limited window to build your own intelligence that leverages your IP" [23]. Across the named accounts the split is by sensitivity of the data: open-source projects still go to the frontier model; proprietary cybersecurity code, internal supply chain monitoring and core power infrastructure do not [29]. Justin Boitano, Nvidia's VP of enterprise AI, told The Information that the company believes zero data retention should be on by default [12]. Jensen Huang has said employees should use AI tokens worth half their annual salary every year [27]. The alternatives on the record all cost more than an API key. Northrop Grumman runs open-source models on its own air-gapped servers [25]. Novo Nordisk keeps using Anthropic's Claude for some tasks and bans proprietary data from reaching it [26]. Microsoft is pitching isolated cloud environments that send no data to outside AI companies; Tom's Hardware notes the approach is costly and that at least one customer is still weighing it [4]. All of these accounts trace to a single report by The Information. Palantir, Nvidia, Booz Allen, Anthropic and OpenAI did not immediately respond to Reuters' requests for comment [10].
What to watch
- Whether Anthropic's customer-held log program becomes a default contract term or stays with select accounts, and whether the guarantee it offers is revocable.
- Whether Palantir unblocks Fable in its own software, which would indicate Anthropic conceded an irrevocable zero-data-retention term.
- Whether either lab publishes an enumerated list of the metadata and classifier outputs it retains under a no-training contract.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence58
- Adoption62
- Hype gap+18
- Incentives74
- Confidence55
Perspective Coverage
4 publishers- Builder
- Builder 26%
- Operator
- Operator 35%
- Investor
- Investor 39%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic's June policy change gave it the right to retain usage logs from its flagship Fable model for 30 days, which the company said was to defend against "complex and novel attacks".
- [2]
Nvidia now uses Fable only for less sensitive tasks such as open-source projects, and runs its own Nemotron models for internal work like AI-powered supply chain monitoring.
- [3]
Both OpenAI and Anthropic collect metadata and technical usage data from enterprise customers; OpenAI calls this data "de-identified".
- [4]
Microsoft is pitching isolated cloud environments that run AI models on private servers and send no data to external AI companies; Tom's Hardware notes the approach is costly and that at least one customer is still considering it.
- [5]
Booz Allen Hamilton has banned employees from using Fable for work on proprietary cybersecurity software.
- [6]
Booz Allen Hamilton CTO Bill Vass said, "We worry a little bit that [Fable] might be learning from some of our code."
- [7]
Palantir is blocking Fable deployment through its own software to customers until Anthropic grants irrevocable zero-data-retention guarantees.
- [8]
Both OpenAI and Anthropic say they do not train on customer data by default unless companies opt in.
- [9]
OpenAI states on its website that it runs business data through automated classifiers and security tools "to better understand how our services are used", and that the resulting classifications are metadata about the business data "but do not contain any of the business data itself".
- [10]
Palantir, Nvidia, Booz Allen, Anthropic and OpenAI did not immediately respond to Reuters requests for comment.
- [11]
Tom's Hardware reached out to Nvidia for comment but did not receive one by publication.
- [12]
"As a company, you know, we believe ZDR [Zero Data Retention] should be on by default," Justin Boitano, Nvidia's VP of Enterprise AI, told The Information.
- [13]
Nvidia has invested in Anthropic, reportedly plans to continue investing, and supplies the company with hardware for model development.
- [14]
Palantir CEO Alex Karp said at a customer event that companies are tired of being "exploited" by AI labs.
- [15]
the-decoder.com writes that Palantir's stance is also self-serving, since the company wants customers running AI models through its platform rather than going directly to providers.
- [16]
A large U.S. utility company canceled its plan to test whether Fable could run its core power infrastructure, after Anthropic refused to agree to a nonrevocable zero-data-retention policy.
- [17]
After customer pushback and OpenAI's August move to let GPT-5.6 Cyber customers store security logs on their own servers, Anthropic followed with a similar program rolling out to select customers this fall.
- [18]
C Spire has agreements with both OpenAI and Anthropic that prevent either from training on its data, but the contracts allow both to collect technical usage data; C Spire believes that includes information about what applications the models are connected to, and worries the labs may collect information about what the models do between generating responses.
- [19]
OpenAI says it does not use "chain-of-thought" data to train its models.
- [20]
John Schulman says "De-identification is weak," that users can be traced back "with just a small number of bits", and that it does not protect against IP leakage.
- [21]
Schulman says building reinforcement learning tasks from "user traces" has a low risk of content reproduction but can still extract customer IP, ranging from harmless ("use explicit user feedback in reward model training") to invasive ("upload user's coding environment and commit history to turn into rl envs").
- [22]
John Schulman is an OpenAI co-founder who briefly worked at Anthropic and now works at Thinking Machines.
- [23]
Sarah Hooker warns companies, "If you are a company with IP you have a limited window to build your own intelligence that leverages your IP."
- [24]
Sarah Hooker, who previously worked at Cohere and Google DeepMind, says there are "clever synthetic data techniques that can generate distributional equivalent data while preserving privacy".
- [25]
Northrop Grumman runs open-source AI models on its own air-gapped servers rather than using OpenAI or Anthropic.
- [26]
Novo Nordisk continues to use Anthropic's Claude for some tasks but bans any proprietary data from being used by the model.
- [27]
Nvidia CEO Jensen Huang has remarked that its employees should use AI tokens worth half their annual salary every year.
- [28]
Anthropic argues that any data it collects about how customers use its products is aggregated and anonymized, and that the metadata is not used to train models.
- [29]
In the named accounts, the work still sent to Fable is described as less sensitive, such as open-source projects, while the withheld work is proprietary cybersecurity software, internal AI-powered supply chain monitoring and control of core power infrastructure.
Sources
2 independent publishers whose own reporting we read for this story.
- economictimes.indiatimes.comPalantir, Nvidia curb AI model use over data fears: The Information - The Economic Times
1 article · September 14, 2026
- mezha.netPalantir, Nvidia та Booz Allen можуть обмежити використання моделей Anthropic і OpenAI через ризики для даних
1 article · September 14, 2026
- the-decoder.comAI labs have a data trust problem that their policies haven't solved
1 article · September 15, 2026
- tomshardware.comNvidia, Palantir, and others restrict advanced AI model usage over privacy concerns, report claims — 'paranoia' rising over customer intellectual property
1 article · September 14, 2026
Topics and entities
Topics
- Training Data Provenance
- On-premises and air-gapped AI
- Zero data retention
- Enterprise AI Procurement
- AI telemetry and metadata