Multi‑Cloud AI Architecture and Egress‑Aware Strategy under the EU Data Act

Executive summary

Artificial‑intelligence workloads are growing exponentially, and the cost curves are changing. The one‑off cost to train a large model—millions of GPU hours—remains substantial, but inference costs are now even more dominant. A 2025 FinOps report explains that training is a capital‑expenditure‑like hit (rent hundreds of GPUs for a few weeks), whereas inference is a utility bill that never stops; inference spend can dwarf the original training costfinout.io. Cloud usage surveys show that generative‑AI tools are the most widely adopted category (60 % of organisations) and also introduce the highest costs due to compute‑heavy inference, tokenised API pricing and retraining overheadcloudzero.com. Average monthly AI budgets rose 36 % between 2024 and 2025, from US $62,964 to US $85,521cloudzero.com.

A major source of hidden cost is data egress—the fee charged when data leaves a cloud provider’s network. Egress pricing on hyperscale clouds starts around US $0.09 per GB for internet egress and drops with volume; transferring data across availability zones or regions adds more costnops.io. In hybrid or multi‑cloud AI architectures where training happens on one platform and inference on another, egress charges accumulate quickly and can exceed training costsventurebeat.com. A prominent example is 37signals: the company faced a US $250,000 egress bill when leaving AWS; by moving storage on‑premises they expect to cut their annual infrastructure bill from US $3.2 million to well under US $1 millionawsinsider.net.

Regulatory pressure is changing this landscape. The EU Data Act, which came into force in January 2024 and becomes applicable on 12 September 2025, aims to eliminate data‑sharing barriers and strengthen customer rights. It mandates that cloud providers must allow users to switch services without functional loss and requires that any switching‑related charges reflect only the actual cost of egressskadden.com. Providers must remove unfair contract terms and ensure interoperabilityskadden.com. Article 29 eliminates switching fees entirely by 2027fivetran.com, and Article 34 allows providers to charge only cost‑based fees for in‑parallel use until thenfivetran.com. In response, Google introduced Data Transfer Essentials, a zero‑cost service for qualifying multi‑cloud transfers within the same organisationnetworkworld.com, Microsoft offers at‑cost transferstheregister.com, and AWS waives egress fees for customers leaving its platform (via support requests)infoq.com. Despite these programmes, egress waivers apply only to specific scenarios (exiting a cloud or internal multi‑cloud use) and do not eliminate fees for general internet egress; misuse triggers normal chargestheregister.com.

This white paper analyses the evolving economics of AI and cloud portability, examines the implications of the EU Data Act, compares hyperscaler responses, and proposes design strategies for multi‑cloud and hybrid AI architectures that balance cost, performance and compliance. It draws on public data, industry surveys and regulatory texts, with citations provided throughout.

1. The changing economics of AI workloads

1.1 Training vs. inference costs

Different financial profiles. Finout’s November 2025 analysis highlights that AI costs come in two flavours: training—a one‑off, GPU‑intensive marathon—and inference—a continuous meter that runs every time a model is usedfinout.io. Training is akin to buying a sports car: you pay millions up front for high‑end GPUs, but the bigger expense becomes the “fuel” (inference). Training bills can cross into millions of dollarsfinout.io, yet inference spends accumulate over time and often dwarf the initial training costfinout.io. Cost‑per‑model visibility, scheduling in low‑cost regions, and deciding when to rent versus own hardware are therefore criticalfinout.io.

Operational expenditure. CloudZero’s 2025 survey of 500 software engineers shows that generative‑AI tools are widely adopted (60 % of organisations) but introduce high costs due to compute‑heavy inference and tokenised API pricingcloudzero.com. Over half of organisations use public cloud (55 %) or hybrid setups (51 %)cloudzero.com. Without proactive optimisation, over‑provisioning and idle resources make cloud costs unpredictablecloudzero.com. Customer‑facing applications and data engineering tools are major cost centres because of large‑scale processing and inference workloadscloudzero.com.

1.2 GPU pricing and hidden costs

GPU compute dominance. GPU compute represents the largest infrastructure expense for AI startups, consuming 40–60 % of technical budgetsgmicloud.ai. Entry‑level training GPUs (NVIDIA A10/L4) cost US $0.50–1.20 per hour on specialised providers but US $1–2.50 per hour on hyperscalersgmicloud.ai. High‑end H100/H200 GPUs cost US $2.10–4.50 per hour with specialised providers or US $4–8 per hour on hyperscale cloudsgmicloud.ai. Hidden costs beyond compute—data transfer (egress fees), storage of datasets and model checkpoints, and network charges—add 20–40 % to monthly billsgmicloud.ai.

Training/inference separation. Finout recommends tiering models—using smaller, cheaper models for simple requests and routing to larger models only when necessary—and applying optimisation techniques such as distillation, quantisation and caching to reduce inference computefinout.io. Off‑peak training, spot instances and region‑hopping provide financial levers by exploiting price differentials across regionsfinout.io. For some workloads, on‑premises or specialised GPU clouds may offer better economics because idle GPUs become costly paperweights in the cloudfinout.io.

1.3 The egress cost trap

Pricing tiers. AWS’s standard egress price to the public internet is about US $0.09 per GB for the first 10 TB each month, decreasing at higher volumes. Transfers between availability zones cost around US $0.01 per GB, and inter‑region transfers range from US $0.02 to US $0.09 per GBnops.io. Other hyperscalers charge comparable rates. When large models and datasets (hundreds of terabytes) move between regions or clouds, these charges quickly outstrip compute costs. Cloudflare notes that egress fees strongly discourage multi‑cloud AI strategies; to avoid huge bills, providers encourage customers to keep data and models within their ecosystemcloudflare.com.

Vendor lock‑in. NGP Capital observes that discounts and proprietary services, combined with egress fees—charges for moving data out of the cloud—have created a strong lock‑in effect; hyperscalers are unlikely to eliminate these fees because they anchor customer retentionngpcap.com. This dynamic is pushing customers toward alternatives with clearer pricing and data portabilityngpcap.com. VentureBeat similarly reports that some organisations pay more to move data than to train models, prompting them to shift inference to colocation or on‑premises environments, cutting monthly bills by 60–80 %venturebeat.com.

Case study: 37signals. In May 2025, 37signals (operator of Basecamp and HEY) revealed that AWS waived a US $250,000 egress bill as part of its exit under AWS’s free egress programmeawsinsider.net. The company moved 18 petabytes of data to on‑premises hardware, expecting to reduce its annual infrastructure bill from US $3.2 million to well under US $1 millionawsinsider.net. The founder emphasised that owning hardware can be cheaper than staying in the cloud, and that hyperscalers have a vested interest in promoting the belief that operating your own servers is “too hard”awsinsider.net. This anecdote illustrates how egress fees and long‑term cloud charges can make on‑prem options attractive, especially when inference workloads dominate.

2. The EU Data Act and cloud portability

2.1 Overview of the EU Data Act

The EU Data Act is a horizontal regulation aimed at unlocking the value of data generated in the EU. It came into force on 11 January 2024 and becomes applicable on 12 September 2025skadden.com. Key objectives include:

• Data access and sharing: Grant businesses and consumers the right to access and share data generated by connected devices and digital services, limiting exclusive control by manufacturers or service providersskadden.com.
• Cloud service portability: Require data processing service providers (IaaS, PaaS, SaaS) to simplify switching between providers, ensuring interoperability and avoiding excessive contractual or technical barriersskadden.com. The Act obliges providers to remove unfair contract terms and make pricing for data transfer transparentskadden.com.
• Switching rights: Under Articles 23–29, users can retrieve exportable data during the contract term and switch providers without technical or contractual obstructionfivetran.com. The Act grants the right to functional equivalence at the new provider and prohibits lock‑in clausesfivetran.com.
• Fee restrictions: Article 29 eliminates switching fees entirely by 2027; until then, only cost‑based fees are permittedfivetran.com. For in‑parallel use of multiple providers, Article 34 allows providers to charge only the actual egress costs incurredfivetran.com. Fivetran notes that providers cannot impose premium surcharges or “data ransom” pricingfivetran.com.

By codifying these rights, the Data Act seeks to reset the balance of power between cloud providers and customers. As Fivetran summarises, customers gain the legal authority to control, extract and move their data without approval, and contract clauses that restrict switching become unenforceablefivetran.com.

2.2 Hyperscaler responses

Google Cloud. Google’s Data Transfer Essentials service, launched in September 2025, waives egress fees for qualified multi‑cloud transfers within the EU and UK. NetworkWorld reports that the service is designed to meet Article 34 requirements; it meters traffic separately and charges nothing for data movement between Google Cloud and another provider when it is part of the same organisation’s multi‑cloud architecturenetworkworld.com. The Register notes that Google positions this move as going beyond regulatory obligations, yet the service is limited to traffic deemed internal to the customer. If an audit finds misuse (e.g., transfers to a third party), Google reserves the right to bill the traffic at normal internet ratestheregister.com.

Microsoft Azure. Microsoft introduced an at‑cost data transfer programme: EU customers may request credits via support to move data to another provider, paying only the cost Microsoft incurstheregister.com. Azure already offers 100 GB/month of free egress globally, with additional waivers available through supportawsinsider.net.

Amazon Web Services (AWS). AWS increased its free egress allowance to 100 GB per month and launched a programme to waive data transfer fees for customers exiting its cloud. Customers must contact AWS support to request credits, demonstrating that the transfer is part of a legitimate migrationinfoq.com. The waiver applies globally and is framed as promoting customer choice and interoperability rather than a regulatory requirementawsinsider.net.

Although these initiatives are noteworthy, they cover only specific scenarios (migration or internal multi‑cloud transfer). Egress fees for general outbound traffic, cross‑region transfers or serving inference workloads across clouds remain chargeable. As The Register observes, the Data Act does not ban data transfer fees outright but requires that they reflect actual cost and not arbitrary mark‑upstheregister.com.

2.3 Implications for AI workloads

For AI teams, the Data Act makes multi‑cloud and hybrid strategies more practical by reducing financial and contractual barriers to switching. However, several caveats apply:

• Cost‑based, not free. Until 2027, providers can still charge actual egress costs for parallel usefivetran.com. Organisations moving large model checkpoints or streaming inference results between clouds should budget for these costs or minimise cross‑cloud traffic through caching, compression and careful architecture.
• Conditional programmes. Free egress programmes require support requests and apply only when migrating or combining services within the same organisationtheregister.com. Misuse (e.g., data export to external partners) can trigger normal egress chargestheregister.com.
• Contract diligence. Customers must review existing agreements to eliminate lock‑in clauses, push for interoperability commitments, and prepare procurement playbooks that enforce Data Act compliancefivetran.com.
• Regional differences. The Data Act applies to services offered in the EU; organisations operating globally must manage data sovereignty across jurisdictions. Providers may still charge standard egress fees outside the EU or for non‑switching use cases.

3. Designing egress‑aware multi‑cloud AI architectures

3.1 Align training and inference placement

1. Co‑locate data and compute. Train and serve models in the same region or on the same provider whenever possible to avoid cross‑region or cross‑cloud egress. When training on one cloud and serving on another (e.g., training on specialised GPU clouds and serving inference on hyperscalers), compress models and use efficient file formats to reduce transfer volume. Cache inference results locally so repeated queries do not trigger cross‑cloud transfersfinout.io.
2. Exploit specialised providers. Compare GPU pricing across providers. For high‑end training, specialised GPU clouds (e.g., GMI Cloud) offer H100/H200 clusters at US $2.10–4.50 per hour compared with US $4–8 per hour on hyperscalersgmicloud.ai. Startups can mix providers: train on the cheapest platform, then fine‑tune or serve inference on an edge or colocation provider closer to customers. However, account for hidden costs such as egress and storage that add 20–40 %gmicloud.ai.
3. Leverage spot and off‑peak pricing. Schedule training jobs during off‑peak hours and in regions with lower spot prices; this can yield significant savingsfinout.io. Use spot or preemptible instances for fault‑tolerant training; use reserved instances for steady inference workloads.
4. Tiered inference and model optimisation. Route simple requests to smaller models and reserve complex models for high‑value queriesfinout.io. Employ techniques such as quantisation and knowledge distillation to reduce model size and inference computefinout.io.

3.2 Minimise egress through architecture

1. Edge inference and caching. Deploy inference servers at the edge or on‑premises to serve requests locally. This reduces the need to send outputs back to a central cloud and lowers latency. Cache common responses and precompute results to avoid repeating the same inference across cloudsfinout.io.
2. Federated or split training. Use federated learning or split training to keep data within its jurisdiction. Only aggregate model updates across clouds, which are smaller than raw data. This reduces egress volume and supports data sovereignty.
3. Data compression and deduplication. Use efficient formats (e.g., FP16 or BFLOAT16 weights) and deduplicate checkpoints before transfer. When replicating data across clouds for redundancy, compress data and schedule transfers during low‑cost windows.
4. Monitoring and FinOps instrumentation. Tag resources by model, workload and environment; continuously monitor egress usage and set budgets. CloudZero’s survey emphasises that organisations lacking real‑time cost attribution risk overspendingcloudzero.com. Finout recommends measuring cost‑per‑model and cost‑per‑query to inform trade‑offs and avoid the “money pit”finout.io.

3.3 Contractual and governance strategies

1. Negotiate Data Act‑aligned contracts. Ensure contracts explicitly support data portability, functional equivalence at the new provider, and elimination of lock‑in clausesfivetran.com. Include provisions for cost‑based egress fees and define roles for personal data under GDPR.
2. Participate in standardisation. Support industry efforts like the Data Transfer Initiative (DTI) and open‑source model formats to improve interoperability. Under the Data Act, the European Commission will develop model contractual terms by September 2025skadden.com; organisations should adopt these templates when available.
3. Prepare for global compliance. Align internal processes with multiple regimes (GDPR, Data Act, national security laws). Use privacy‑preserving techniques (encryption, differential privacy) when transferring data across borders. Recognise that U.S. and other jurisdictions may not yet provide cost‑based egress protections.

4. Risks and mitigations

5. Strategic recommendations

1. Perform a comprehensive AI cost audit. Capture training, inference and egress costs separately; identify the proportion of AI spend allocated to GPU compute, storage, network and third‑party APIs. Use cost‑per‑model metrics to inform decisions.finout.io
2. Adopt a multi‑provider strategy with data localisation. Choose providers based on workload characteristics: specialised GPU clouds for training; hyperscalers for scale and ecosystem integration; and colocation or edge providers for low‑latency inference. Localise data and models to minimise cross‑cloud movement.
3. Plan for the EU Data Act’s timelines. Revise contracts by September 2025 to reflect switching rights and cost‑based egress. By January 2027, ensure that processes support zero‑fee switching (except actual costs for parallel use). Use the act as leverage when negotiating with providers.
4. Optimise inference to curb long‑term costs. Use model compression, quantisation and tiering; implement caching and precomputation; and align inference workloads with cost‑efficient hardware. This reduces compute demand and lowers the data transferred per queryfinout.io.
5. Implement strong data governance. Classify data (personal vs. non‑personal) and apply appropriate legal basis for transfersfivetran.com. Ensure encryption and key management remain under the customer’s control; avoid sharing sensitive data with providers outside the required scope.
6. Participate in industry consortia. Engage with European initiatives such as Gaia‑X and the European Cloud Federation to shape interoperable standards and contribute to model contractual terms. This fosters an ecosystem of providers adhering to Data Act principles and reinforces digital sovereignty.

6. Conclusion

The AI revolution is increasingly an economic one. Organisations must navigate skyrocketing inference costs, hidden egress charges and a rapidly shifting regulatory environment. The EU Data Act creates a legal foundation for data portability and cost‑based charges, but it does not abolish egress fees entirely until 2027 and applies primarily to services used in the EU. Hyperscalers are responding with limited free‑transfer programmes, yet the onus remains on customers to architect their AI systems to be portable, cost‑efficient and compliant.

Thorsten Meyer’s AI publishing network has emphasised private, responsible and sovereign AIdeepintellica.com. This white paper aligns with that vision: it advocates for transparent cost management, avoidance of vendor lock‑in, and adherence to regulatory obligations. By understanding the economics of training vs. inference, planning for the EU Data Act, and designing egress‑aware multi‑cloud architectures, organisations can unlock innovation without breaking the bank. The future of AI lies in flexibility—the ability to place workloads wherever price‑performance, capacity and sovereignty align. With careful planning, the combination of regulatory pressure and technological evolution can transform the cloud from a lock‑in engine into a truly open platform for AI innovation.