DataKrypto

• Google Cloud, NVIDIA, and other platforms could extend their own confidential-computing features to cover model and prompt encryption, narrowing DataKrypto's never-decrypts advantage.
• DataKrypto names no production customer, so a buyer cannot verify that its near-zero-overhead homomorphic encryption holds at production scale.
• The homepage advertises FIPS 140-3, but the NIST certificate is FIPS 140-2 Level 1 and sunsets in September 2026, so DataKrypto could lose its one verifiable federal credential unless it revalidates.
• Homomorphic-encryption specialists such as Zama, Duality, and Enveil could match DataKrypto's engineering, eroding its technical head start.
• On a roughly 3 million euro raise, DataKrypto may lack the capital to fund the enterprise and regulated sales motion its sovereign-AI positioning implies.

DataKrypto sells to the enterprise putting AI into production on data or models it cannot expose. The company frames the problem as the cleartext gap: during inference, model weights, prompts, and outputs sit in plaintext in GPU memory, even inside a hardware enclave, which is where an attacker or an untrusted operator could read them. A SecurityWeek account describes the same worry, that intellectual property and personal data fed to a local model could leak back to the model provider or beyond. The named buyers are regulated industries, healthcare, financial services, the public sector, and pharma.

The pain is real and vendor-framed rather than independently sized. No fetched source quantifies how many enterprises face this exposure or what it costs them. DataKrypto's framing and a supporting press account establish that the risk is credible, but the scale of the buying population stays unmeasured in the public record. [s1, s2, s3]

DataKrypto's product, FHEnom for AI, keeps every stage of the AI pipeline in ciphertext. The client encrypts the prompt at the source, an encrypted tokenizer processes it without producing plaintext tokens, the model computes on ciphertext on the GPU, and only the session-key holder decrypts the result. Enclaves are used only to hold keys, not to decrypt data for computing, which DataKrypto presents as the difference from confidential-computing rivals that decrypt inside the enclave. A separate engine, FHEnom, carries a NIST validation as the DataKrypto Fully Homomorphic Encryption Module.

The architecture is documented in detail, but its headline claim is not independently tested. DataKrypto states that encryption adds about 0.6 milliseconds of overhead for a 4,000-token batch with no loss in inference performance. Fully homomorphic encryption is normally far slower than plaintext computing, so a near-zero-overhead result on large language models is an extraordinary claim, and no fetched source provides an independent benchmark of it. [s2, s3, s4]

DataKrypto positions against every confidential-AI vendor that decrypts data to compute on it. Its argument is that enclaves isolate the computation but still expose model weights, prompts, and outputs in memory, while fully homomorphic encryption never decrypts. That is a genuine architectural distinction rather than a marketing one.

The threat is that the distinction may not hold as a durable advantage. Google Cloud and chip vendors sell their own confidential computing, and DataKrypto rides Google Cloud rather than displacing it. Other homomorphic-encryption firms, including Duality and Enveil, contest the same ground, so the harder-to-copy asset is DataKrypto's own cryptography engineering rather than the category position. [s1, s2, s8]

DataKrypto's public traction is channel and partner motion, not named customers. Its flagship product is listed on Google Cloud Marketplace after it completed Google Cloud's ISV Startup Springboard program, and it announced an encrypted-guardrails partnership with Tumeryk. These are real distribution and integration signals.

What the record lacks is a named deployment. No fetched source identifies a company running FHEnom for AI in production, so the marketplace listing and partnerships stand in for customer proof. For a product whose core claim is production-grade performance, the absence of a named reference is the gap a buyer would probe first. [s8, s10]

DataKrypto's founders are public and domain-relevant. Ravi Srivatsav, co-founder and CEO, previously founded ElasticBox and sold it to CenturyLink, and was earlier a partner at Bain and a product leader at NTT Research. Luigi Caramico, founder and chief technology officer, is described as a homomorphic-encryption pioneer with more than 25 years in Silicon Valley, and Carla Mascia leads cryptography research.

Their credibility is genuine but short of a category-defining record. The prior exit was in cloud infrastructure rather than cryptography, and the public record does not document a cryptography exit or a sustained research program of the founders' own. The cryptography depth rests on Luigi Caramico's individual reputation more than on an institutionally recognized track record. [s6, s7, s5]

DataKrypto holds one verifiable security credential and displays two others it does not substantiate. NIST's Cryptographic Module Validation Program lists the DataKrypto Fully Homomorphic Encryption Module at FIPS 140-2 Level 1, a real federal validation anyone can inspect in the registry, though it covers the earlier Fhenom engine and sunsets in September 2026. The homepage instead badges FIPS 140-3, a level the certificate does not support, and the technology page claims ISO 27001:2022 certification, which the reviewed record shows only as on-page badge text with no inspectable certificate.

A probe on 2026-07-03 of trust and security subdomains, the /security and /trust paths, and the homepage footer found no inspectable trust portal, no SOC 2 report, and no inspectable ISO 27001 certificate behind the badge. For a regulated buyer, the honest position is one entry-level FIPS validation on an older module, with the broader certification stack asserted rather than shown. [s4, s11, s2]

Add analyzed competitors to compare them side by side with DataKrypto.

DataKrypto targets the enterprise deploying AI on data or models it must keep confidential, and it names regulated verticals as the core. Its solutions pages list healthcare, financial services, the public sector, cloud and managed-service providers, and pharma, and add sovereign AI compute for organizations that cannot expose data to foreign jurisdictions. The buyer is the team accountable for both running AI and protecting regulated data.

The segment is credible and regulated-leaning, and the company's own investment announcement reports customers across telecommunications, healthcare, manufacturing, and SaaS. No fetched source names a customer in any of these verticals, so the segmentation reflects DataKrypto's intended market rather than a demonstrated one.

DataKrypto's advantage is a full-pipeline encryption design that never decrypts data to compute on it. Current vendor pages describe the prompt encrypted at the client, tokenized in ciphertext, run through an encrypted model on the GPU, and returned so only the session-key holder can read it, with enclaves reserved for key custody, while an earlier independent description had the tokenizer and embedding layer inside a TEE and the answer decrypted before delivery, a shift the record does not reconcile. DataKrypto contrasts this with confidential computing, which decrypts inside the enclave, and NIST has validated its encryption module against FIPS 140-3, a module-conformance credential that does not itself attest the proprietary FHE construction's security or performance.

The differentiator that would decide a purchase is unproven in public. DataKrypto claims about 0.6 milliseconds of overhead and no inference-performance loss on large language models, which for fully homomorphic encryption is an extraordinary result. No published benchmark confirms it. DataKrypto says a third-party chip maker verified the performance, but the results are not yet public, so the headline number rests on the vendor's word.

DataKrypto's documented channels are cloud-platform distribution and technology partnerships alongside a direct demo funnel. Its flagship product is available on Google Cloud Marketplace after it completed Google Cloud's ISV Startup Springboard program, it announced an encrypted-guardrails partnership with Tumeryk, and its pages route buyers to a demo request, though no source quantifies the channel mix.

The motion is real but unquantified. No named customer appears in the fetched record, so there is no account proof behind the partnerships. The Google Cloud channel is a credible path to regulated buyers, but the record does not yet show it converting into named deployments.

DataKrypto publishes no pricing on the cited pages, and the public surface points buyers to a demo and a sales conversation. No fetched source discloses the unit it charges by, whether per model, per workload, or per inference volume.

Without a published unit or figure, a buyer cannot infer what DataKrypto believes the product is worth or compare it against the cloud confidential-computing options in the same budget line. For an early company selling to regulated enterprises, undisclosed pricing is common, but it leaves the pricing thesis undocumented.

DataKrypto primarily describes deployable software and gateway infrastructure the customer integrates, and while its technology page says managed service providers deploy FHEnom for AI as encrypted inference-as-a-service, DataKrypto publishes no managed-service accountability terms of its own. FHEnom for AI is a framework applied to the customer's own models and inference pipeline, deployable across cloud, on-premises, and edge, with key custody configured to the deployment. In direct deployments the customer runs the encrypted pipeline and controls the keys, while the cited pages also describe MSP-operated inference-as-a-service, SI-operated blind training, split model and data owners, and neutral TEE custody in mixed-ownership setups.

The documented paths span customer-integrated deployment and MSP-delivered encrypted inference-as-a-service, so delivery is not universally customer-operated, though direct deployments place integration and key-management effort on the customer's team. The cited materials describe a framework the customer integrates rather than a managed service with published operational accountability, so what DataKrypto delivers is an encryption capability rather than an outcome.

DataKrypto holds two verifiable module validations and asserts more than it documents. NIST's Cryptographic Module Validation Program lists two DataKrypto validations: certificate 5175, FIPS 140-3 Overall Level 1 on the FHEnom module, validated March 2026 and active until August 2029, and the older 140-2 Level 1 certificate 4677. The homepage's FIPS 140-3 badge is therefore backed by an inspectable registry entry at the standard's entry level. The technology page claims ISO 27001:2022 certification, though the reviewed record includes no inspectable certificate behind it.

A probe on 2026-07-03 of trust and security subdomains, the /security and /trust paths, and the homepage footer found no inspectable trust portal, no SOC 2 report, and no certificate inspectable from the site itself, though the NIST registry independently backs the FIPS badges. The registry credentials are entry level and module-scoped, and the ISO badge remains uninspectable, narrower than the marketing implies.

DataKrypto positions FHEnom for AI as a foundational encryption layer beneath AI workloads rather than an application. It integrates into the AI pipeline and rides Google Cloud's infrastructure and enclaves, presenting itself as the encryption that makes a cloud a confidential place to run models.

The marketplace tie runs toward the platform, though the cited pages describe FHEnom as hardware-agnostic and deployable across cloud, on-premises, and edge, so the Google Cloud relationship is distribution rather than architectural dependence, and confidential-computing features of clouds and chips could still narrow its role. The record shows the Google Cloud infrastructure integration and one evidenced application-level integration, a joint offering with Tumeryk combining FHEnom's encryption engine with policy-aware prompt inspection, which leaves the ecosystem position at integration reach rather than platform leverage.

DataKrypto's team pairs a serial-entrepreneur chief executive with a career cryptographer and a research group in Italy. Ravi Srivatsav founded ElasticBox and sold it to CenturyLink before DataKrypto, and Luigi Caramico, the founder and chief technology officer, is described as a homomorphic-encryption pioneer with more than 25 years in Silicon Valley. Carla Mascia leads cryptography research, and the company runs an R&D team out of Rome.

The publicly evidenced reputation rests most visibly on Luigi Caramico, and while the company lists a broader cryptography and R&D team, the public record does not document their external track record. It also does not document a prior cryptography exit or a sustained public research program under the company's name, so the technical credibility rests heavily on Luigi Caramico's individual standing rather than an institutional one.