Most AI teams now run more than one compute provider, and a funded category of tools exists to move a training job between them. None of those tools move the data. The job lands on a new cluster and the dataset is still sitting in the bucket it was in last quarter, in the region it was in last quarter, billing egress on every read.
That gap was survivable while compute was the scarce thing and storage was cheap. Neither is true in 2026. Chips, power and floor space now bind together, contract minimums are lengthening, and enterprise SSD prices are up 80% on the year.
Here is what changed underneath the storage line, and what to redesign before the 2027 budget locks.
Trend 1: Compute is decentralised now, multi-vendor and multi-cluster
SkyPilot launched as a company with $20M to run AI workloads across fragmented compute, and CoreWeave added SkyPilot support for multi-cloud orchestration. A funded category exists because the pattern is already how teams work. SkyPilot's own framing of the problem: neoclouds solved the hard infrastructure problems and left the last mile unsolved.
Running one provider stopped being the default. Teams hold capacity across a primary, a secondary, and whatever spot inventory clears, and they expect to move a job between them. Anything that has to follow the job needs to be reachable from every cluster on the same interface.
This looks like a buyer's choice. The next three trends are why it is not.
Trend 2: Chips, power and floor space all bind at once
The constraint did not move off the chip. It got a second axis.
Getting a fleet is still hard. NVIDIA allocation is reserved through 2026 into 2027. Blackwell is effectively reserved for hyperscalers and large enterprises. TSMC packaging capacity remains sold out. Secondary-market H100 lead times from resellers run 36 to 52 weeks against CoWoS packaging and HBM limits. Smaller pre-configured systems in the 8-GPU class ship quickly and are limited in size. A larger allocation takes money down and a place in a queue, with no guarantee at the end of it.
TSMC expanded packaging capacity. Grid capacity did not scale with it, so a second constraint arrived on top of the first rather than replacing it.
The line from Inflect that carries the section: a facility with confirmed power and a later chip delivery date will come online faster than one with chips in hand and no confirmed power capacity.
The two constraints compound rather than trade off. Chips gate whether you can get the fleet at all. Power and floor space gate where that fleet can physically land, and how long the operator can afford to keep it there. A team can clear one gate and still be stuck at the other.
When the footprint of your compute is set by silicon allocation and grid interconnection together, neither of which you control, the data layer cannot be pinned to any one footprint.
Trend 3: Contracts lengthen, and buyers get pushed onto the newest silicon
Compute Exchange finds reserved conversations now start at six months. Providers prefer one to two years. Contracts beyond a year carry roughly 20% prepayment. Historically the market ran from monthly agreements to three years. The short end has largely disappeared for premium GPUs.
The structural reason is financing. Neocloud operators use customer prepayments to fund their own infrastructure purchases.
What Akave sees across neocloud partner conversations is worth stating plainly. Operators are running at full capacity, and one year is now the practical contract floor rather than six months. Smaller pre-configured systems in the 8-GPU class ship quickly, while a larger allocation requires money down and a place in a queue with no guarantee. Some operators are not renewing existing contracts because they need the rack space back for newer parts. The tenant is not being repriced. It is being displaced.
The corroborating pattern for that displacement: B200 and B300 supply ramped through 2026, and H100 moved from top tier to mid tier as providers routed new demand to Blackwell.
The buyer's exposure stopped being price and became tenure. A team can be current on its bill, happy with its provider, and still lose the cluster at renewal for reasons that have nothing to do with them. If tenure is not guaranteed, the data cannot live inside the tenancy.
One clarification is worth making explicit. Renting eight GPUs is still easy. On-demand H100 provisioning runs in under two minutes with meaningful spot inventory available. What is gone is reserved multi-node capacity on a term. That distinction is where the risk sits.
That would be survivable if the data were cheap to hold and cheap to move. Neither is still true.
Trend 4: Storage prices are climbing on the same curve as GPUs
The reader came for compute and learns their storage line is next. This is the shift almost nobody is writing from the buyer's side.
NVIDIA raised Vera Rubin and Grace Blackwell system prices 15% to 17% for shipments from early 2027, driven by memory content rather than GPU silicon. Demand has not slowed through any of it.
The storage a buyer can actually get is also not the storage they were sold. Saturn Cloud's June 2026 comparison found parallel filesystem tiers are backend-only or partner-contract at Nebius, Lambda, GMI Cloud, Hyperstack and Voltage Park, with CoreWeave and Crusoe the exceptions. Self-service tenants land on the provider's default object store and discover the gap at the first multi-terabyte checkpoint write.
Storage stopped being the cheap part of the bill and started following the GPU curve with a lag. A 2027 budget built on 2025 storage rates is already wrong.
Price predictability became an architectural property rather than a procurement outcome. Akave Cloud is priced at $14.99/TB flat-rate, zero egress across the neocloud footprint. A flat rate is not a discount. It is a hedge against a supply shock when the underlying media is up 50% to 80% on the year.
Rising cost per terabyte is painful. Rising cost per terabyte on data you have to move between vendors every quarter is the actual problem.
Trend 5: The job moves across vendors. The data does not.
Compute orchestration is a funded, shipping category. Data orchestration has not followed it across vendors.
SkyPilot's own post on running AI jobs across GPU neoclouds mentions data locality once, in the introduction, as a pain point facing infrastructure engineers. The article then covers job scheduling, GPU allocation and multi-cluster resource management, and never returns to data movement, egress, or checkpoint synchronisation. The leading compute orchestrator demonstrates the gap in its own material.
The honest competitive picture is worth stating. Alluxio has held the term "data orchestration" for years, defining it as a layer between computation frameworks and storage systems, with unified access across multiple storage systems and caching for GPU utilisation. It sits as an acceleration layer above object storage. That means it inherits whatever the underlying storage is: its jurisdiction, its egress terms, its audit properties. Orchestration built on top of someone else's bucket cannot fix what the bucket does not do.
Checkpoints are the concrete example. Frontier-scale training snapshots can exceed 500GB, which makes them the artefact that most obviously will not follow a job that moved providers overnight. Move a checkpoint across compute vendors once a quarter and the egress bill is the surprise line item. Move it twice and the design is the surprise.
Teams optimised the compute layer for portability and left the data layer pinned. The result is either egress on every rotation, or a workload that quietly cannot move.
A storage layer reachable from every cluster on the same S3 API, where moving data with the job is not a per-gigabyte event, is the precondition any cross-vendor data movement requires. That is the layer this piece is about.
Portability answers where the data can go. It does not answer where the data is allowed to go.
Trend 6: Sovereign deployment decides where the data is allowed to sit
AWS launched its European Sovereign Cloud in January 2026 with expansion announced across Europe. Other US providers brought sovereign offerings to market over the same period.
NTT DATA's 2026 research on sovereignty demand puts numbers on the shift. 96% of organisations are weighing relocation of AI infrastructure on geopolitical and supply chain pressure. 95% call private or sovereign AI important to their strategy. 35% of Chief AI and Information Officers name enabling it as their biggest barrier to adoption. The report splits sovereignty into mandated, regulated and strategic autonomy, and notes EU investment is largely regulation-driven while Middle East demand follows national strategy. The 35% barrier figure is the most useful in the set.
What Akave sees on the ground is worth adding. Local governments are driving independent sovereign AI programmes, which is pulling deployments into specific jurisdictions as a requirement rather than a preference.
The regulatory picture is worth being precise about, because a lot of vendor content got the date wrong. The AI Digital Omnibus entered into force in July 2026 and moved Annex III high-risk obligations, which carry the Article 10 training-data governance requirements, from 2 August 2026 to 2 December 2027. What took effect on 2 August 2026 was the Article 50 transparency obligation. Penalties for high-risk non-compliance run to EUR 15M or 3% of global turnover. The EUR 35M and 7% tier covers prohibited practices under Article 5 and does not apply.
Sovereignty stopped being a compliance cost absorbed after the architecture was chosen and became a deployment driver that decides the architecture. The jurisdiction question now precedes the vendor question.
Isolated endpoints, customer-held keys, and the ability to store where the law allows and query from anywhere, independent of which provider is running the job this quarter, is what the shift requires.
The identity picture matters here too. Machine identities now outnumber human identities 109 to 1, per Palo Alto Networks' 2026 Identity Security Landscape Report. Forrester ranks AI agent threats among the top 2026 CISO threat categories, and IBM's 2026 Cost of a Data Breach report found 92% of organisations with AI-related breaches had no proper AI access controls. When the identities reading the data outnumber the people by two orders of magnitude, and the compute underneath rotates quarterly, the only place an audit record can live is the layer the enterprise owns. Any modification is independently detectable through content-addressing and Merkle integrity, and the immutable storage ledger records that integrity across independent attestation nodes, so no single operator, including Akave, can rewrite the record.
The deferral to December 2027 is a hook, not relief. It gives you a reason to act inside the 2027 budget cycle rather than a reason to feel late. US buyers with EU-touching pipelines are already being asked for equivalent evidence in procurement.
How we think about this at Akave
We built Akave Cloud on the premise that compute would keep rotating and storage could not. Every design decision follows from that.
The data has to be reachable from every cluster the workload might land on. Akave Cloud is S3-compatible across the neocloud footprint, so the storage layer does not have to be renegotiated when the compute contract ends or the provider reclaims the rack for newer silicon. That answers Trends 1 and 3.
The price has to hold while the media does not. Akave Cloud is priced at $14.99/TB flat-rate, zero egress. With enterprise SSD up 80% and hard drives up roughly 50% on the year, a flat rate is a hedge against a supply shock rather than a discount. It is the number that survives the 2027 budget conversation. That answers Trend 4.
Movement between providers cannot be a per-gigabyte event. Zero egress is what makes moving data with the job economically possible at all, and every orchestration layer that sits on hyperscaler buckets pays the tax on every rotation. That answers Trend 5.
Residency has to sit at the storage layer, not in a promise the operator makes. Isolated endpoints and customer-held keys hold the jurisdiction and access policy with the enterprise, regardless of where the GPUs sit. Store where the law allows, query from anywhere. That answers Trend 6.
The audit record cannot depend on trusting the operator to preserve it. Any modification to a stored artifact is independently detectable through content-addressing and Merkle integrity, and the immutable storage ledger records that integrity across independent attestation nodes, so no single operator, including Akave, can rewrite the record. That is the property that makes the storage layer defensible as evidence in six months, when the compute vendor underneath has already changed twice.
The name is the argument, for what it is worth. Aka is root or origin. Ve is verification. Akave is meant to be the verification root for the data an AI system acted on.
The pattern underneath the six shifts
Read together, the six shifts are one shift asked six ways. Every layer of the AI stack that a buyer assumed was durable turned out to be rented. Compute is rented, and now rented on longer terms than the buyer wanted. Storage is rented at prices moving with the memory market, on tiers most buyers do not actually get access to. Jurisdiction is rented from whoever owns the region.
Storage is the only layer that can be held long enough to be governed. That is the property 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
Is there still a GPU shortage in 2026?
Yes, on premium reserved multi-node capacity. NVIDIA allocation is reserved through 2026 into 2027, TSMC packaging is sold out, and Blackwell is effectively booked by hyperscalers and large enterprises. What is not scarce is on-demand H100 rental in small configurations, which still provisions in under two minutes at most specialist providers. The shortage moved shape, it did not lift.
Why are GPU contract minimums getting longer while some prices fall?
Because neocloud operators use customer prepayments to fund their own infrastructure purchases. Longer contracts with roughly 20% prepay make the financing math work, so the short end of the reserved market has largely disappeared. What is falling is the on-demand rate for older Hopper silicon, which is unrelated to the reserved-tier terms buyers actually sign for at-scale training.
What is actually constraining AI infrastructure in 2026, chips or power?
Both, and they compound rather than trade off. Chips gate whether you can get the fleet at all. Power and floor space gate where that fleet can physically land, and how long the operator can afford to keep it there. A team can clear one gate and still be stuck at the other, with US interconnection queues running 24 to 72 months in most markets and longer in constrained regions.
Why are enterprise storage prices rising in 2026?
Because memory has become the swing factor in AI infrastructure cost. Enterprise SSD is up 80% on the year, hard drives up roughly 50%, and DRAM and HBM contract prices moved 80% to 90% quarter over quarter in Q1 2026. Memory now represents roughly a quarter of a high-end AI server rack's cost, up sharply, and the flow-through is showing up in system prices from early 2027.
Can my GPU provider decline to renew and take back the cluster?
Yes, and this is starting to happen. Some operators are not renewing existing contracts because they need the rack space back for newer parts. The tenant is not being repriced. It is being displaced by Blackwell demand that pays better. If data lives inside the tenancy, the displacement is also a data event.
What is data orchestration, and how is it different from compute orchestration?
Compute orchestration moves training jobs between GPU providers. SkyPilot and similar tools handle scheduling, allocation and multi-cluster resource management. Data orchestration is the equivalent layer for the data those jobs read and write, and it is a less mature category. Alluxio holds the term and operates as a caching layer above object storage, which inherits whatever the underlying storage is. The gap is that no shipping tool moves the underlying data across compute vendors without egress charges.
How do I move training data between compute providers without paying egress every time?
You cannot, if the data sits on a storage layer that charges egress. The workaround is a storage layer that charges zero egress across the neocloud footprint, so moving data with the job is a rebalancing event rather than a per-gigabyte tax. Akave Cloud is one implementation of that layer.
When do the EU AI Act's training-data rules actually apply, and what are the penalties?
The AI Digital Omnibus moved the Annex III high-risk obligations, which include Article 10 training-data governance, from 2 August 2026 to 2 December 2027. Article 50 transparency obligations for GPAI-generated content did take effect on 2 August 2026. Penalties for high-risk non-compliance run to EUR 15M or 3% of global turnover. The EUR 35M and 7% tier covers prohibited practices under Article 5 and does not apply to training-data governance.
Does storing data in Frankfurt make my architecture sovereign?
Not by itself. Regional storage locates the data, but the entity holding it still answers to a jurisdiction. Sovereign architecture requires isolated endpoints, customer-held keys, and provable residency at the storage layer, so where data physically lives is a property of the system rather than an assurance in a document. The AWS European Sovereign Cloud, launched January 2026, is one attempt at solving this from the hyperscaler side.
References
- SkyPilot, AI Job Orchestration Part 1: GPU Neoclouds. https://skypilot.ai/blog/ai-job-orchestration-pt1-gpu-neoclouds
- SkyPilot, company launch. https://skypilot.ai/blog/skypilot-the-company
- CoreWeave, SkyPilot integration announcement. https://coreweave.com/blog
- Inflect, data center power shortage 2026. https://inflect.com/blog
- Compute Exchange, reserved GPU contract length. https://compute.exchange/blogs
- Thunder Compute, AI GPU rental market trends. https://thundercompute.com/blog
- Spheron, NVIDIA H100 news 2026. https://spheron.network/blog
- Astute Group, enterprise SSD up 80%. https://astutegroup.com/news
- Data Center Disk, hard drive prices 2026. https://datacenterdisk.com/news
- Data Storage, State of GPU Cloud August 2026. https://datastorage.com/articles
- Shattered.io, NVIDIA AI server price hike. https://shattered.io
- Saturn Cloud, GPU Cloud Comparison Report, June 2026. https://saturncloud.io/reports
- Alluxio, why data orchestration whitepaper. https://alluxio.io/whitepapers
- NTT DATA, sovereign AI research, via Help Net Security, May 2026. https://helpnetsecurity.com
- AWS European Sovereign Cloud launch. https://press.aboutamazon.com
- European Commission, AI Omnibus enters into force. https://digital-strategy.ec.europa.eu
- EU AI Act Article 99, penalty tiers. https://artificialintelligenceact.eu/article/99
- Palo Alto Networks, 2026 Identity Security Landscape Report, via Help Net Security, 14 May 2026. https://helpnetsecurity.com
- Forrester, Top Cybersecurity Threats for 2026. https://forrester.com/blogs
- IBM, Cost of a Data Breach Report 2026, via Help Net Security, 30 July 2026. https://helpnetsecurity.com