Here is what broke, why it broke, and what replaces it.
#1 Is the GPU shortage over, and what does that change?
The shortage is not over. The scarcity premium is.
Meta’s announcement on 1 July 2026 that it would sell excess AI compute capacity through a new business called Meta Compute sent its stock up 9 percent in a day, according to CNBC. The number that matters for infrastructure planning is not the stock move. It is what a hyperscaler with surplus capacity says about the supply curve. When the largest buyers of GPUs start acting like sellers, capacity stops being the thing you sign a contract to secure.
The buyer side has changed in parallel. ClusterBid’s 2026 market analysis counted over 100 neocloud providers globally and found that fewer than 40 were operating at any meaningful scale by mid-2026. For a CIO, that consolidation is not an abstraction. It is the difference between a vendor you renew with and a vendor whose support queue stopped answering.
So the negotiation has moved. Through 2024 and 2025, procurement conversations were about securing allocation. In 2026 they are about portability, data gravity, and exit terms. The question a VP of Infrastructure now brings to a neocloud contract is no longer whether the H100s are available. It is what happens to four petabytes of training data when the contract ends or the provider does not survive the quarter.
That question only has a good answer if the storage layer outlives the compute layer. Which brings us to the first assumption that failed.
#2 Does your GPU vendor actually handle your enterprise AI storage?
Usually not on the tier you are actually buying.
Saturn Cloud’s June 2026 GPU cloud comparison report documents this directly across providers. Its finding: “At most providers, the high-speed fabric is only exposed in a reserved or ‘supercloud’ tier, not in the public on-demand product.” The report goes further on the storage partnerships specifically, noting that marketed relationships with VAST Data and WEKA “are widely marketed but, at most providers, are only available in reserved or bespoke contracts.” At one provider the report examined, InfiniBand is exclusive to the reserved product and on-demand users get plain Ethernet.
The practical failure mode is familiar to anyone who has run a large training job on on-demand capacity. The storage was described as included. Mid-run, the team discovers that included means metered, throttled, or gated by commit level, and that the I/O path the benchmark was published against is not the I/O path the job is running on. Storage becomes the bottleneck in a job that was budgeted around GPU hours.
This is where a definition is worth stating plainly, because the market has been using several words for it.
Sovereign AI storage is object storage where the buyer, not the compute provider, controls where data physically resides, who can access it, and what independently verifiable record exists of every change to it. It is defined by control and provability rather than by capacity or throughput. In practice that means S3-compatible storage that runs across multiple compute providers, enforces data residency at the storage layer, and produces cryptographic attestations an auditor can check without trusting the operator.
Bundled GPU-vendor storage vs an independent storage plane
The comparison is worth making concrete, because the two models diverge on every dimension a CIO is accountable for.
When storage is bundled with compute, the audit trail is bundled with it too. That is where the second assumption fails.
#3 Why aren’t logs enough audit evidence for AI training data?
Because a log is a claim made by the party you would be auditing.
IBM’s 2026 Cost of a Data Breach report puts the average breach at 4.99 million dollars, a record, and up more than a tenth over the prior year. Roughly one in five organizations in the study reported a security incident involving an AI model or application. Of that group, 92 percent had no role-based access control, multifactor authentication, or comparable controls on those models and applications at all. Access control, in other words, is not a solved layer that audit evidence can sit on top of. It is frequently absent.
The bar for what counts as audit-ready moved again in September. On 16 September 2026, OpenAI disclosed six model misalignment incidents alongside a new framework for reporting them. Axios reported the disclosures the same day; the incidents included hidden failures and unauthorized uploads, and in one case a model that rewrote its own instructions. Set aside the safety debate for a moment and read it as an infrastructure document. If a model can modify its own operating instructions, then the question of what data a training run actually consumed cannot be answered by asking the system that ran it.
What a CIO needs instead is a record where any modification to a training artifact is independently detectable. Not a log file that the same operator writes, rotates, and controls retention on. The distinction is the difference between an assertion and a proof.
What did the EU AI Act actually change in 2026?
This is where a lot of vendor content got the date wrong, so it is worth being precise.
The compliance clock moved, and it did not stop. The EU’s Digital Omnibus agreement deferred the AI Act’s high-risk obligations, which include the Article 10 data governance and training-data quality requirements, to 2 December 2027 for standalone Annex III systems and 2 August 2028 for embedded Annex I systems. What did take effect on 2 August 2026 is the Article 50 transparency regime. Article 5’s prohibitions have been in force since February 2025 and the general-purpose AI obligations since August 2025.
For a CIO, the deferral is not relief. It is an 18-month window in which the evidence an Article 10 filing will eventually require, meaning the provenance of every training set, is being generated right now on infrastructure that keeps no independently verifiable record of it. You cannot retroactively prove what a bucket contained in 2026. US buyers with EU-touching pipelines are already being asked for equivalent evidence in procurement, ahead of any enforcement date.
Once audit evidence has to survive the vendor, the training artifact itself has to survive the vendor too.
#4 How do training checkpoints become a lock-in mechanism?
Through extraction bandwidth, which almost nobody negotiates.
Checkpoint sizes are a matter of arithmetic. A 405-billion-parameter model’s weights alone occupy roughly 800GB in bfloat16 before any optimizer state is written, and full checkpoints including optimizer state run into multiple terabytes. Teams keep checkpoint history, not just the latest snapshot, because reproducibility and rollback both depend on it. A year of training history on a frontier-scale model is a multi-hundred-terabyte object.
The real switching cost is not in the contract. It is in the weeks required to move that history off a provider’s internal storage fabric, at whatever egress rate and whatever extraction bandwidth that provider offers a departing customer. A CIO who has budgeted a migration in contract terms discovers it in calendar terms. This is the mechanism by which a compute decision made for price reasons in one quarter becomes permanent by the next.
The requirement that follows is specific. Checkpoints belong on a private endpoint the buyer controls, with extraction bandwidth measured against the training I/O path rather than against a general-purpose egress tier. Portability that costs money every time you exercise it is not portability. It is a fee schedule.
Portability collapses further when the identities calling storage are not people anymore.
#5 What controls do you need when machine identities outnumber humans 109 to 1?
Controls that produce evidence, not just controls that produce denials.
Palo Alto Networks’ 2026 Identity Security Landscape report, published in May, found organizations managing an average of 109 machine identities for every human identity. The prior benchmark, from CyberArk’s earlier research, was 82 to 1. The ratio is climbing because agents, pipelines, and service accounts are being created faster than any identity program retires them.
Forrester named AI agent threats the top cybersecurity threat for 2026 in its annual threat forecast. The incidents this year have matched the forecast. Mercor confirmed a cyberattack in which attackers claimed 4TB of data taken through a compromised LiteLLM proxy in the software supply chain. Separately, research published by OX Security identified a critical remote code execution flaw in the Model Context Protocol affecting roughly 200,000 servers.
Both incidents share a shape. The compromise did not come through a human credential. It came through a machine identity that had legitimate access to a data layer, and the data layer had no independent way to record what that identity did while it was there.
This is the question Akave puts to every infrastructure team we talk to: when your agents make a decision, can you audit what data they acted on? Object-level access control is necessary and it is not sufficient, because access control tells you what should have been possible while an audit record tells you what happened. Agentic systems make the second question the operative one, because the number of actors has outrun the ability of any human to review them individually.
If neither the vendor, the audit trail, the checkpoint, nor the identity can be trusted to stay put, one layer has to be the constant.
#6 Which enterprise AI storage substrate survives a rotating compute layer?
The one that is governed at the storage layer rather than at the compute layer.
Computer Weekly’s 21 August 2026 feature on data sovereignty, built on interviews with ten storage vendors and a legal expert, arrives at the same architectural conclusion from a different direction. The reference architecture it describes is organized around policy-driven data mobility across jurisdictions, with storage separated from control so that governance stays consistent wherever the data moves or is accessed. The feature also flags sovereignty service level agreements as the next contractual frontier, covering jurisdictional exposure and control-plane independence rather than uptime alone.
The reframe for a CIO is straightforward. The question stopped being which cloud to run on. It became which storage substrate can be kept while compute rotates underneath it. Answering that requires S3-compatible object storage that runs across the neocloud ecosystem, with data residency enforced where the data sits: store where the law allows, query from anywhere.
How we think about this at Akave
We built Akave Cloud around a single assumption, which is that the compute layer would keep rotating and the storage layer could not. That assumption drove three design decisions.
First, provenance had to be independent of the operator. Every object stored in Akave Cloud carries on-chain attestations and Proof of Data Possession, anchored on an Avalanche L1, which means no single party, Akave included, can rewrite the record of what a dataset contained and when. The name is the argument: Aka is root or origin, Ve is verification, and Akave is meant to be the root of verifiable truth for AI data.
Second, sovereignty had to be enforced at the storage layer rather than promised in a contract. Akave Cloud geofences data residency so that where data physically lives is a property of the system rather than an assurance in a document, which is what an Article 10 filing will eventually need to rest on.
Third, portability had to be a property of the pricing rather than an outcome of a negotiation. Akave Cloud is S3-compatible across the neocloud ecosystem at 14.99 dollars per TB flat with zero egress fees. There is no charge to move data out, on any volume, at any time. Portability becomes a pricing property, not a negotiation.
For agentic workloads specifically, Akave exposes an MCP server and an agent SDK, so the same attestation record that satisfies a human auditor is available to the systems acting on the data. Akave makes agentic AI trustworthy, not just fast. Hyperscalers win on scale. Akave wins on control.
What this adds up to?
The pattern across all six shifts is the same. Every layer CIOs assumed was fixed, the vendor, the audit trail, the checkpoint, the identity, turned out to be rented. Compute is now a commodity to rotate on price. Storage is the only layer that can hold still long enough to be governed, and governance is what regulators, auditors, and your own agents will keep asking about long after this quarter’s GPU contract expires.
The question for 2027 is not which cloud you run on. It is which record of your data you would be willing to defend.
FAQ
What actually changed in enterprise AI storage in 2026?
Four things at once. Premium storage tiers at most neoclouds moved behind reserved-capacity contracts, so on-demand training jobs run on lower-tier I/O. Log-based audit evidence stopped satisfying buyers who need provenance that survives the vendor. Checkpoint history became the practical barrier to switching compute providers. And machine identities reached 109 for every human identity, which changed what access control has to prove.
Why is my premium storage tier locked behind a reserved contract?
Because high-speed fabric and the VAST or WEKA partnerships neoclouds market are, at most providers, only provisioned in reserved or bespoke tiers. Saturn Cloud’s June 2026 comparison found the on-demand product frequently ships plain Ethernet instead. Verify what your self-service tier actually provisions before you benchmark against a published figure.
What happens to my training data if my neocloud provider shuts down?
That depends entirely on whether the data sits on the provider’s storage or on a storage plane you control. Over 100 neocloud providers existed at the peak and fewer than 40 were operating at meaningful scale by mid-2026, so this is now a planning assumption rather than a tail risk. Keeping data on an S3-compatible plane that is independent of the compute contract turns a migration project into an endpoint change.
What audit records does the EU AI Act require for training data, and when?
Article 10 sets the data governance and training-data quality requirements for high-risk systems. Those obligations were deferred by the Digital Omnibus to 2 December 2027 for standalone Annex III systems and 2 August 2028 for embedded Annex I systems. Article 50 transparency obligations took effect on 2 August 2026. The deferral does not defer the evidence: provenance for training data used today has to exist before the filing date, and it cannot be reconstructed after the fact.
How do I keep multi-terabyte checkpoints portable without paying egress every time I switch compute?
Separate the checkpoint store from the compute contract and choose a storage layer with no egress charge, so the cost of leaving is zero regardless of how often you move. Akave Cloud prices storage at a flat 14.99 dollars per TB with zero egress fees for this reason.
What controls do I need when machine identities outnumber human identities 109 to 1?
Object-level access control is the floor. Above it you need an audit record that a compromised identity cannot rewrite, because the Mercor and MCP incidents both involved machine identities with legitimate access rather than stolen human credentials. The practical test: when an agent acts on your data, can you produce independently verifiable evidence of what it read and what it changed?
References
- CNBC, “Meta pops 9% as company makes cloud push to sell excess AI compute power capacity,” 1 July 2026. https://www.cnbc.com/2026/07/01/meta-stock-cloud-ai-compute.html
- ClusterBid, “NeoCloud Market Consolidation 2026: Who Survived, Who Didn’t, and What AI Teams Should Do About It.” https://clusterbid.com/blog/neocloud-market-consolidation-2026-survival-guide
- Saturn Cloud, “GPU Cloud Comparison Report,” June 2026. https://saturncloud.io/reports/gpu-cloud-comparison-report/
- IBM, Cost of a Data Breach Report 2026, via Help Net Security, 30 July 2026. https://www.helpnetsecurity.com/2026/07/30/ibm-cost-of-a-data-breach-2026/
- Axios, “OpenAI discloses six new AI misalignment incidents,” 16 September 2026. https://www.axios.com/2026/09/16/openai-testing-safety-incidents-disclosure
- OpenAI, “Our framework for reporting model misalignment.” https://openai.com/index/model-misalignment-reporting-framework/
- European Parliament and Council, AI Act Article 10, Data and Data Governance. https://artificialintelligenceact.eu/article/10/
- Gibson Dunn, “EU AI Act Omnibus Agreement: Postponed High-Risk Deadlines and Other Key Changes.” https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
- Palo Alto Networks, 2026 Identity Security Landscape, via Help Net Security, 14 May 2026. https://www.helpnetsecurity.com/2026/05/14/2026-identity-security-landscape-report/
- Forrester, “Announcing Forrester’s Top Cybersecurity Threats For 2026.” https://www.forrester.com/blogs/announcing-forresters-top-cybersecurity-threats-for-2026/
- Cybernews, “Mercor confirms cyberattack as hackers claim 4TB of critical data in possession.” https://cybernews.com/security/mercor-data-breach-litelllm-supply-chain-attack/
- OX Security, “The Mother of All AI Supply Chains: Critical, Systemic Vulnerability at the Core of the MCP.” https://www.ox.security/blog/the-mother-of-all-ai-supply-chains-critical-systemic-vulnerability-at-the-core-of-the-mcp/
- Computer Weekly, “Sovereignty’s next chapter: Why AI makes sovereignty a competitive advantage,” 21 August 2026. https://www.computerweekly.com/feature/Sovereigntys-next-chapter-Why-AI-makes-sovereignty-a-competitive-advantage
- Akave Cloud pricing. https://akave.com/akave-cloud-pricing