Nine major GPU neoclouds hold SOC 2 today. Zero hold ISO 42001. That gap is about to become a procurement question, because the layer just above GPU cloud, the AI-native SaaS vendors and hyperscalers running on top of that hardware, is getting certified fast.
AWS certified in November 2024. Anthropic followed two months later. Intercom, Snowflake, Salesforce, and ServiceNow all certified through 2025. Microsoft has a set of Copilot services in scope. Meanwhile CoreWeave, the neocloud most exposed to enterprise AI workloads, says only that it's "looking at alignment" with the standard (CoreWeave engineering blog). If you're an enterprise buyer who just finished a SOC 2 review for your GPU cloud vendor, ISO 42001 is the next question your security or legal team is going to ask, and right now the honest answer from every neocloud is "not yet."
This guide covers what ISO 42001 actually certifies, how it differs from SOC 2 and the EU AI Act, who holds it today, and what to ask a GPU cloud vendor before you sign a contract that assumes it exists.
What ISO 42001 Actually Certifies (and What It Doesn't)
ISO/IEC 42001:2023 certifies that an organization has a documented, independently audited management system for governing how it builds, deploys, and monitors AI systems. It does not certify that any specific model is accurate, unbiased, or safe. Published by ISO in December 2023, it's the first international standard written specifically for AI management systems, distinct from general information security frameworks that predate the generative AI wave (Microsoft Learn, ISO/IEC 42001:2023 overview).
What an AI Management System (AIMS) Is, in Plain Terms
An AI management system, AIMS for short, is the organizational scaffolding around AI development and deployment: documented policies for risk assessment, defined roles and accountability for AI decisions, processes for human oversight and impact assessment, incident response procedures specific to AI failure modes, and a commitment to continual improvement of all of the above. It's a management-process standard, not a technical one. It doesn't specify which model architecture to use or set a minimum accuracy threshold. It specifies that you have to have a system, that the system is documented, and that an independent auditor checks the system actually operates the way you say it does.
That last part is what makes ISO 42001 meaningfully different from a company simply publishing an "AI principles" page. A principles page is a claim. A certification is a claim an outside party has tested.
Who Issues the Certification and How to Verify It's Real
ISO itself does not certify anyone. Certification is issued by independent, accredited certification bodies, and buyers should verify two things before trusting a badge: who did the audit, and who accredited that auditor. AWS's certification, for example, was verified by Schellman Compliance, LLC, which holds accreditation from the ANSI National Accreditation Board, ANAB (AWS blog, ISO/IEC 42001:2023 accredited certification). Intercom's certification was audited by the same firm, Schellman, under the same ANAB accreditation (Intercom, ISO 42001 certification announcement). Schellman shows up repeatedly in this space because it's one of the certification bodies that built out AI-specific audit practice early. When a vendor claims ISO 42001, ask for the certificate and confirm the issuing body carries ANAB (or an equivalent recognized) accreditation, the same diligence step that already applies to SOC 2 reports.
What ISO 42001 Does NOT Cover
It's worth being blunt about the limits, because the certification gets oversold in sales conversations. ISO 42001 does not test model performance, accuracy, or safety in any technical sense. AWS's own FAQ on the standard makes the scope explicit and adds a warning that matters for anyone building on top of a certified vendor: "Your organization is not automatically certified by association" (AWS ISO 42001 FAQs). If your SaaS product runs on a certified model provider, that doesn't make your product certified. You'd need your own audit for your own management system.
How ISO 42001 Differs from SOC 2 and the EU AI Act for AI Buyers
The three sit in genuinely different categories, and conflating them is the single most common mistake we see in vendor diligence conversations right now.
Comparison Table: Certifiable Standard vs Attestation vs Regulation
| ISO/IEC 42001 | SOC 2 | EU AI Act | |
|---|---|---|---|
| What it is | Certifiable management-system standard | Independent attestation report | Binding law (EU) |
| Issued by | Accredited certification bodies (e.g. Schellman, under ANAB) | CPA firms, under AICPA standards | No certification body; enforced by regulators |
| What it covers | Governance of AI development, deployment, and monitoring | Security, availability, confidentiality, and other Trust Services Criteria | Risk-tiered legal obligations for AI systems placed in the EU market |
| Legal weight | None on its own, voluntary | None on its own, voluntary | Mandatory for in-scope systems, with fines |
| Does it grant EU AI Act presumption of conformity? | No, not a harmonized standard | No | N/A, it is the regulation |
ISO 42001 is not a harmonized standard under the EU AI Act, so certification alone doesn't create a legal presumption of conformity. Only standards formally adopted by CEN-CENELEC and published in the EU's Official Journal carry that weight, and as of the standard's current status, no AI-specific harmonized standard has reached that bar yet (ISMS.online, ISO 42001 and EU AI Act presumption of conformity). An ISO 42001 badge is real evidence of governance maturity. It is not, by itself, a legal shield against an EU AI Act enforcement action. For the fuller regulatory picture, our EU AI Act compliance guide for GPU cloud covers risk classification and the data residency rules that sit alongside this certification question.
Why You Might Need All Three, Not Just One
These three answer different diligence questions, which is why mature enterprise buyers end up asking for all three rather than picking one:
- ISO 42001 answers: does this vendor have a real, audited process for governing AI risk?
- SOC 2 answers: does this vendor's infrastructure and access control actually work the way they say it does?
- EU AI Act compliance answers: is this vendor's product legal to deploy for my specific use case and region?
A vendor can pass all three, two, one, or none, and each gap tells you something different about where the risk sits.
Colorado's AI Act and the ISO 42001 "Safe Harbor" That Wasn't
This is worth getting precise about, because the original framing shows up in a lot of 2025-era compliance content that hasn't caught up with what actually happened. Colorado's original AI Act, SB 24-205, included a rebuttable presumption of reasonable care for organizations aligned with ISO/IEC 42001 or the NIST AI Risk Management Framework, a real safe harbor tied to these two governance frameworks. Then, in May 2026, Governor Polis signed SB 26-189, which repealed and reenacted the law entirely, shifting it from a risk-based, algorithmic-discrimination framework to a narrower automated-decision-making-technology (ADMT) disclosure model taking effect January 1, 2027 (Norton Rose Fulbright, Colorado enacts revised AI law). The rewrite dropped the duty of care, the algorithmic impact assessments, and, specifically, the safe harbor for NIST AI RMF or ISO 42001 alignment. The framework alignment is still useful operationally, just not as a Colorado-specific liability shield anymore ("SB 26-189 removed Colorado's specific mandate and the associated safe harbor; it did not eliminate the underlying compliance utility") (Carpe Datum Law, Colorado's AI reset). If a vendor pitches you on ISO 42001 as a Colorado legal safe harbor today, that pitch is a year out of date.
Who's Actually Certified Today: The GPU Cloud Gap
AWS certified first among hyperscalers, in November 2024, and a wave of AI-native SaaS vendors followed through 2025. No major GPU neocloud has certified as of mid-2026. That's the gap this post exists to name.
AWS Was First; AI-Native SaaS Vendors Followed
AWS's November 2024 accredited certification, audited by Schellman under ANAB, covers Amazon Bedrock, Amazon Q Business, Amazon Textract, and Amazon Transcribe specifically, not every AWS service (AWS blog announcement). AWS has since kept the certification current, completing its first surveillance audit for ISO/IEC 42001:2023 with no findings, published in early 2026 (AWS security blog, first surveillance audit).
Anthropic certified in January 2025. Intercom certified in April 2025, one of the first AI customer service platforms to do so, audited again by Schellman under ANAB, and stacked alongside its existing ISO 27001, ISO 27701, ISO 27018, HIPAA, SOC 2, and HDS certifications (Intercom, ISO 42001 certification announcement). Intercom's own framing: "Achieving certification demonstrates that Intercom's practices meet the highest standards of AI governance." Salesforce certified its AI Platform, Agentforce, and Slack AI in October 2025, extending the scope to eighteen products running across first-party infrastructure and AWS Hyperforce (Salesforce Compliance Site, ISO 42001:2023 certification). ServiceNow and Snowflake certified in 2025 as well (AI Compliance Vendors, ISO 42001 certified companies list). Microsoft is running the same playbook for a defined set of Copilot services, with independent third-party audits and certificates published on its Service Trust Portal (Microsoft Learn, ISO/IEC 42001:2023 overview).
Notice the pattern in every one of these: certification is scoped to named products, never a blanket "the whole company is certified" claim. That's the same scope discipline enterprise buyers already had to learn with sector-specific frameworks like DORA for GPU cloud vendors serving EU banks, where the audit boundary matters as much as the badge itself.
No Major GPU Neocloud Holds ISO 42001 Yet
Here's the gap. None of the nine GPU neoclouds most commonly compared for enterprise AI workloads, Lambda, CoreWeave, GMI Cloud, Vast.ai, Runpod, Crusoe, Nebius, Voltage Park, and Together AI, publicly claim ISO 42001 certification as of mid-2026. All hold SOC 2 and/or ISO 27001 instead (Spheron's SOC 2 GPU cloud comparison).
CoreWeave, the neocloud with the most enterprise AI exposure, is explicit that it hasn't crossed the line into certification: "We are also looking at alignment with ISO/IEC 42001, the new global framework for responsible AI management, to ensure that our governance model evolves with the technology itself" (CoreWeave engineering blog). "Looking at alignment" is materially different from "certified." It's the same gap Spheron has flagged before in FedRAMP for GPU cloud, where CoreWeave Federal is also pursuing, not holding, a target certification. The infrastructure layer keeps landing behind the application layer on every major AI governance framework.
That gap makes sense mechanically, even if it's inconvenient for procurement timelines. ISO 42001 shares clause structure (clauses 4 through 10) with ISO 27001 and ISO 27701, so a company pursuing it can build on an ISMS or privacy program it already has running rather than starting from zero (Schellman, AI governance and ISO 42001 FAQs). Intercom is a clean example: its ISO 42001 certification sits alongside ISO 27001, ISO 27701, ISO 27018, HIPAA, SOC 2, and HDS, all held before or alongside the AI-specific badge. GPU neoclouds have spent the last two years building out SOC 2 and ISO 27001 programs. ISO 42001 is the logical next step, not a parallel effort, and it's reasonable to expect the first neocloud certification within the next year or two rather than never.
Gartner's read on why this matters beyond any one framework: fragmented AI regulation will cover 50% of the world's economies by 2027, driving $5 billion in AI compliance investment (Gartner, via Network World). That regulatory sprawl is exactly what's pushing SaaS vendors toward ISO 42001 as a single, portable answer to a growing pile of jurisdiction-specific questions. GPU clouds sit one layer below that pressure today, but the pressure doesn't stay contained to one layer for long.
Questions to Ask a GPU Cloud Provider Before You Sign
Since no neocloud holds ISO 42001 yet, the useful diligence right now isn't "are you certified," it's confirming what a vendor actually means when ISO 42001 comes up in a sales conversation, and what they can show you in the meantime.
The Vendor Diligence Checklist
- Certified, or "aligned"? Ask the vendor to state plainly whether they hold an active ISO 42001 certificate or are describing a roadmap intention. "Looking at alignment," "working toward," and "planning to pursue" are not certifications, and a vendor should say so directly when asked.
- Name the certification body and its accreditation. If a vendor does claim certification, get the auditor's name and confirm ANAB (or equivalent) accreditation, the same check that matters for SOC 2 reports.
- Get the exact product scope. ISO 42001 certifications are scoped to named products or services, never a whole company. Confirm the tier or product you'll actually deploy on is named in the certificate, not just implied by association.
- Check the audit date. ISO 42001 certificates, like SOC 2 reports, are point-in-time with periodic surveillance audits. A certificate with no recent surveillance activity is a weaker signal than a fresh one.
- Confirm it's actually AIMS, not a relabeled ISO 27001. ISO 42001 and ISO 27001 share structural DNA and some vendors blur the two in marketing copy. Ask specifically whether the AI-management-system clauses (risk assessment for AI systems, AI impact assessment, human oversight processes) were in the audit scope, not just general information security controls.
- Ask what technical controls back the paperwork. A management-system certification is about process, not encryption strength or hardware isolation. If your workload needs confirmed technical safeguards, like encrypted memory or remote attestation, ask about those directly rather than assuming ISO 42001 covers them; our confidential GPU computing guide covers what NVIDIA TEE and encrypted VRAM actually verify, a different and complementary layer of evidence.
What to Do When No Vendor in Your Shortlist Is Certified Yet
That's the real situation for GPU cloud procurement right now, and it doesn't mean you're stuck. Treat ISO 42001 as a forward-looking differentiator to track rather than a current gate, and lean on what does exist today: SOC 2 Type II status, ISO 27001 certification, documented incident response, and the specific technical controls your workload actually needs. Deloitte's research on enterprise AI governance found that 87% of executives report having an AI governance framework, but fewer than 25% have actually operationalized it across their organization (Deloitte, ISO 42001 standard for AI governance and risk management). That gap between claimed and operationalized governance is exactly what a real certification is designed to close, and it's a useful filter even before ISO 42001 shows up on a neocloud's compliance page: ask every vendor to show you the operational evidence behind whatever governance claims they're already making, not just the badge.
Spheron's own compliance posture follows the aggregation model covered in our HIPAA-compliant GPU cloud guide: we pool capacity from 5+ providers, including partners that already hold SOC 2 Type II, ISO 27001, and HIPAA certifications, rather than presenting a single unified audit. ISO 42001 isn't part of that picture yet, and we'd rather say so plainly than let a badge do work it hasn't earned. If your procurement checklist has an ISO 42001 line item today, expect "not yet, here's our SOC 2 and ISO 27001 status instead" from every neocloud on your shortlist, Spheron included.
No GPU neocloud holds ISO 42001 today, which makes SOC 2, ISO 27001, and documented technical controls the real diligence trail for enterprise AI procurement right now.
Frequently Asked Questions
No. ISO/IEC 42001 is a voluntary standard. Certification is issued by independent accredited certification bodies, not by any government or by ISO itself, and no jurisdiction currently mandates it. Some regulations, like Colorado's original AI Act, once pointed to it as a way to demonstrate reasonable care, but that specific provision didn't survive the law's 2026 rewrite.
No. ISO 42001 certifies that an organization has a documented, audited management system for governing how it builds, deploys, and monitors AI, things like risk assessment, human oversight, and incident response. It says nothing about a specific model's accuracy, bias, or output quality. That's a common misread of the badge.
SOC 2 is an attestation about a company's security controls, audited against AICPA Trust Services Criteria. ISO 42001 is a certifiable management-system standard specifically about how AI is governed. A vendor can hold one without the other. Most GPU neoclouds hold SOC 2 today; as of mid-2026, none hold ISO 42001.
None of the major GPU neoclouds (Lambda, CoreWeave, GMI Cloud, Vast.ai, Runpod, Crusoe, Nebius, Voltage Park, Together AI) publicly claim ISO 42001 certification as of mid-2026. CoreWeave has said it's 'looking at alignment' with the standard, which is short of certification. AWS was the first major cloud provider to certify, in November 2024, for a specific set of AI services.
Not usually, today. Since no neocloud holds it yet, most enterprise buyers are asking about SOC 2, ISO 27001, and specific technical controls (encryption, access logging, incident response) as the current bar, while tracking ISO 42001 as a 2026-2027 differentiator rather than a hard gate. That's likely to change as more of the stack around GPU cloud (the SaaS and model layers) gets certified.