Passer au contenu principal

SAFOZi

GPU as a Service in the Mediterranean — SAFOZI

GPU as a Service in the Mediterranean: Power, Sovereignty and the Compute Divide

Summary

The Mediterranean is one sea but three GPU markets. The northern shore has public AI factories and listed prices, the eastern shore is split by politics and export rules, and the southern shore has big plans with most capacity still to come.

  • Supply: France, Italy, Spain and Greece hold the region’s clearest GPU capacity. Israel and Turkey have the strongest positions outside the EU. Several smaller countries show no public GPU service.
  • Buyers: researchers get public compute, while enterprises and governments are the likeliest paying customers. Inference and fine-tuning fit most demand better than large-scale training.
  • Investment: at $60,000 a GPU, a rental price of $2.11 to $2.99 an hour needs roughly 46% to 65% utilisation to repay the capital in five years, before running costs.
  • Sovereignty: the EU is writing cloud sovereignty into law, but the chips remain imported. US export rules also treat countries on the same sea very differently.
  • Outlook: most hosters and cloud providers should survive, but in fewer, more specialised forms. Pure GPU resellers carry the highest risk.
  • Opportunity: the gap between the shores is the opening. Trusted providers can link northern capacity with southern and eastern demand, if the law allows it.

The Mediterranean has the cables, the capital and the demand to become a compute region, but those assets sit on different shores. The EU is moving up to €10B of public money into AI gigafactories, Israel has opened a national B200 cluster, and Egypt, Morocco and Tunisia have announced large campuses. Yet in several coastal states this review found no GPU service at all.

This article maps who supplies GPU as a service (GPUaaS) in all 21 Mediterranean countries, who buys it and why the business is hard. The core argument is that this is not one market. Power, law, capital and export rules change at each border, so a provider’s real competition depends on which shore it stands on.

The Mediterranean at a glance: three shores, three economies

Compute in the region follows capital and power more than growth. Morocco grows fastest among the countries below, but France has the largest confirmed private GPU build.

ShoreCountriesGrowth signalCompute position, November 2026
NorthSpain, France, Monaco, Italy, Slovenia, Croatia, Bosnia and Herzegovina, Montenegro, Albania, Greece, Malta, CyprusSpain 2.5%, Italy 0.8%, France 0.4%; eurozone 1.1%Public AI factories, listed GPU prices, private campuses
EastTurkey, Syria, Lebanon, IsraelTürkiye 2.7%Israel and Turkey active; Syria and Lebanon constrained
SouthEgypt, Libya, Tunisia, Algeria, MoroccoMorocco 4.5%Large plans, with most capacity due from 2027

Growth figures are IMF forecasts for 2026 from the October World Economic Outlook, which puts the world at 3.1%.

The numbers that matter most for a GPU provider:

Bar chart of 2025 business electricity prices, from 0.142 euros per kWh in Malta to 0.280 in Cyprus, with an EU average of 0.207.
Eurostat NRG_PC_205 via Apiar Data. EU average is the source’s simple average of 27 countries. Large data centres usually pay less than this band.

The power gap shapes the map. By my calculation, a constant 1 MW load at these listed rates would cost about €1.4M a year in Spain and €2.3M in Italy. Large operators negotiate lower prices, but the ranking still favours Iberia over Italy and the islands.

The provider landscape: public clusters, private clouds and paper campuses

No region-wide GPU count exists, so the picture below is built from individual disclosures. The table separates what runs today from what is planned, because the two are often reported together.

Provider or systemCountryHardware or scaleStatusAccess
Mistral AIFrance13,800 GB300 GPUs; 44 MW total; $830M debtFunded; go-live date not statedEnterprise commitments
OVHcloudFranceH100, A100, L40S; about $2.99 an hour for H100Available at GravelinesSelf-service
ScalewayFranceH100 SXM at about $2.87 to $3.31 an hourAvailable in Paris and WarsawSelf-service
SeewebItalyH100 SXM at $2.11 and H200 at $2.90 an hourAvailablePay per use
Fastweb + VodafoneItaly248 H100 GPUsOperatingEnterprise and sovereign AI
Leonardo, CINECAItaly13,824 GPUs based on NVIDIA AmpereOperatingResearch and innovation
MareNostrum 5, BSCSpainHopper and GB200 partitionsOperatingResearch and innovation
Daedalus, GRNETGreeceGH200; 85.69 petaflops; €58.9MInaugurated; full access expected in DecemberCalls for researchers, startups and SMEs
Vega, IZUMSlovenia240 A100 GPUsOperatingResearch and innovation
Nebius national systemIsrael1,000 B200 GPUs out of 4,000 deployedPhase one liveAllocated by the Innovation Authority
GlassHouseTurkeyH100, hosted in TurkeyListed; no public priceManaged, pay as you go
OneCloudMoroccoL40S nodes on Oracle infrastructureListed on provider’s websiteSales-led

Global clouds cost several times more than local specialists. One vendor comparison puts H100 on AWS at about $12.30 an hour, Google at $12.49 and Azure at $14.00 in European regions. Those figures come from a competitor and are derived from 8-GPU instances, but the gap is too wide to ignore. Buyers pay it for platform breadth, not for the GPU alone.

The same comparison lists hyperscaler H100s in France but not in Italy, Spain or Greece. It also reports no direct on-demand B200 at the big three in the EU. Treat that as a mid-2026 snapshot and check live listings before relying on it.

Northern shore: public money meets private clouds

The northern shore is the region’s most developed market, but capacity is concentrated in four countries. The EU’s AI Factories network gives them public compute that smaller neighbours reach through partnerships.

CountryMain developmentStatus
FranceMistral’s data centre near Paris; Alice Recoque at Bruyères-le-Châtel; OVHcloud and ScalewayStrongest private market
ItalyLeonardo, Fastweb, Seeweb, Aruba; Scaleway has opened two zonesBroadest local offer
SpainAI gigafactory bid at Móra la Nova and San Fernando de Henares; MareNostrum 5Bid awaiting EU selection
GreeceDaedalus at Lavrio, the core of the Pharos AI FactoryInaugurated 19 October 2026
SloveniaVega at MariborOperating, older GPUs
CroatiaPantheon campus at Topusko, €12BConstruction planned for early 2027
Cyprus, MaltaAI Factory antennas linked to PharosAccess through Greece
Monaco, Albania, Montenegro, Bosnia and HerzegovinaNo public GPU service found in this reviewLikely served from abroad

France leads on private capital. Mistral’s site takes its capacity to 44 MW, with 200 MW targeted across Europe by the end of 2027 and up to 1 GW by 2030. It is also the rare AMD story in this region: Alice Recoque will use MI430X GPUs.

Spain is the region’s clearest state bet. Its gigafactory bid has €720M from the state holding SETT, with Telefónica, ACS and Santander as partners, and it could mobilise up to €5B. The EU call for up to seven gigafactories closes on 12 November 2026, with awards expected by early 2027. Up to €10B of public funding is meant to unlock at least €20B of private money.

Italy has the widest commercial choice. Seeweb lists an H100 at $2.11 an hour, and a June review of the market found offers from Aruba and Servereasy as well. High electricity prices remain its main weakness.

Greece shows what an AI factory can do for a smaller state. Daedalus ranks 31st on the TOP500 and will serve researchers, startups, SMEs and the public sector.

Croatia shows why a headline should be read with care. Pantheon’s backers hope for more than €50B in total, but the published detail names no chip vendor and no IT load, and full operation is not due until 2029.

Eastern shore: two active markets, two constrained ones

The eastern shore has the widest gap in the region. Israel and Turkey have real activity, while Syria and Lebanon are still rebuilding basic digital infrastructure.

Israel has the most concentrated national programme. Nebius won a tender in May 2025 and now runs 1,000 B200 GPUs for the national system. The Innovation Authority gives 70% of resources to high-tech companies and 30% to early-stage research groups, with usage periods of one to six months. This is a subsidised market-builder, not a commercial price list.

Turkey is building sovereign cloud around a domestic telecoms group. Turkcell and Google Cloud plan a cloud region of three data centres in Ankara, with full capacity planned for 2028, data kept in the country and encryption keys held by customers. Turkcell has about 50 MW today. Local providers already sell H100 access hosted in Turkey, but GlassHouse publishes no price. Reported national targets include $10B of private investment and 1 GW by 2030, which need official confirmation. A 36% policy rate is the main obstacle to any lira-funded GPU business.

Lebanon is at an earlier stage. A $150M World Bank-backed programme will finance digital infrastructure, including data centres. No public GPU service was found.

Syria faces legal limits as well as physical ones. The US has ended its comprehensive sanctions programme and eased licensing for civilian goods, and Congress has since repealed the Caesar Act. But most items on the US control list were still reported to need a licence, and Syria remained a designated state sponsor of terrorism in that guidance. No public GPU service was found.

Southern shore: big plans, later delivery

The southern shore is planning some of the region’s largest campuses, but most capacity arrives in 2027 or later. Each country is taking its own route.

CountryLead projectPlanned scaleTiming
EgyptAfrica Data Centers Egypt with Vodafone, Elsewedy and Cassava200 MW and $1B target; first phase 20 MWFirst phase over three years
MoroccoNexus AI Factory; Oracle’s Casablanca region16 MW, then 20 MW more, for 36 MW by end-2027Earlier 40 MW date has slipped
TunisiaSoleCrypt AI Campus, Bizerte40 MW; first 7 MW with 1,728 Rubin GPUsEarly 2028
AlgeriaENSIA academic GPU system; Akid Lotfi centre in OranENSIA reported to serve 52 universitiesState-led
LibyaNo public GPU service foundConnectivity first, including a network across more than 200 oil-sector sitesNo date found

Egypt shows that compute has become geopolitical. Huawei reportedly bid with 2,008 Ascend chips, while US officials were preparing a counteroffer. No award has been announced.

Morocco has the broadest set of options, with a hyperscaler region, a renewable-powered campus plan and local providers. Tunisia has talent and a strategic position: the Bizerte site sits close to the Medusa cable landing, which supports a role facing Europe.

Timeline grid of stated GPU and data centre milestones on the northern, eastern and southern Mediterranean shores from 2026 to 2028 and later.
Stated or reported dates, November 2026 · not to scale

The cross-sea link is the region’s most underused asset. Medusa, 2Africa and similar cables tie the southern shore directly to Marseille, Barcelona and Genoa. That makes a split model possible, with training or overflow in the north and regulated inference hosted at home.

Training, fine-tuning and inference: three different GPU businesses

GPUaaS is three products with different hardware, customers and economics. The north can sell all three, but most of the region should focus on inference and fine-tuning.

The global market points the same way. Gartner forecasts AI-optimised IaaS spending on inference at $23.3B in 2026, against $19B for training, and Deloitte forecasts that inference will take about two-thirds of AI compute in 2026.

WorkloadWhat it doesTypical hardwareRevenue patternRegional fit
TrainingBuilds a model from scratchMany linked H100, H200, B200 or GB300 GPUsLarge, bursty projectsFrance, Spain and Israel, plus public AI factories
Fine-tuningAdapts a model to local data or languagesOne or a few L40S, RTX PRO 6000 or H100 GPUsShort, repeated jobsStrong in every language market
InferenceRuns a trained model for usersMemory-rich GPUs, often shared across workloadsSteady, recurring usageStrongest where data must stay local

Language gives this region a natural fine-tuning market. Arabic dialects, Turkish, Greek, Hebrew, Italian and Spanish all need adapted models, and the data is often sensitive. Mistral’s regional endpoints show the direction: customers now choose where inference runs.

GPU models running in the region

The region spans five chip generations. Public clusters still rely heavily on older GPUs, while the newest hardware is concentrated in a few funded projects.

GPUStatusWhereBest workload
A100In useLeonardo, Italy; Vega, Slovenia; SeewebResearch and inference
L40SAvailableOVHcloud, Seeweb, OneCloudInference and small fine-tuning jobs
H100AvailableOVHcloud, Scaleway, Seeweb, Fastweb, GlassHouseFine-tuning and training
GH200In useDaedalus, GreeceResearch training and inference
H200AvailableSeewebLarge-model inference
B200In useNebius, IsraelLarge-scale training
GB200In use or plannedMareNostrum 5, Spain; Nexus plan, MoroccoLarge-scale training
GB300FundedMistral, FranceFrontier-scale training
RubinPlannedBizerte, TunisiaNext-generation workloads
MI430XUpcomingAlice Recoque, FranceResearch at scale
Ascend 950, 910BProposedHuawei’s reported Egypt bidTraining and inference

Who makes the chips

No GPU in this article is designed or made in the Mediterranean region, even in France, Italy or Israel. NVIDIA dominates, with about 81% of AI data centre chip revenue in a widely cited estimate attributed to IDC. AMD estimates range from 5% to 13%.

For a provider, this is also a software and geopolitical choice. NVIDIA offers the deepest ecosystem. AMD now has a European flagship in Alice Recoque. Huawei offers a second supply route on the southern shore, but one that could complicate relations with US suppliers. The safest plan is to buy for contracted workloads and keep software portable.

The AI users: who actually buys compute

Demand is real on every shore, but public compute serves much of the research market. That leaves enterprises and governments as the most reliable paying customers.

BuyerTypical workloadWhat they valueNamed examples
Researchers and universitiesModel training and scienceFree or subsidised accessEuroHPC users, Daedalus calls, ENSIA
StartupsFine-tuning and inferenceLow price, instant accessIsraeli high-tech firms, Multiverse Computing
EnterprisesCustomer service, search, industrial AIResidency, contracts, supportFastweb’s MIIA users; Mistral’s anchor enterprises
GovernmentsPublic services, sovereign AIControl, local hostingPharos, Israel, Egypt, Spain

Public programmes are also market-makers. Israel’s scheme sets minimums of 16 B200 GPUs for industry and eight for academia, which trains local firms to use serious clusters. Daedalus opens its calls to startups and SMEs. A commercial provider must work out which of those users will graduate to paid capacity.

Enterprise commitments may become the stronger model. Mistral now turns multi-year enterprise commitments into access to infrastructure it builds, through what it calls European Compute Units. In effect, customers help finance the capacity before it exists.

Why GPUaaS is not an easy investment

GPUaaS is hard to underwrite on every shore, but for different reasons. The north faces price competition and public alternatives, the east faces capital and legal limits, and the south faces immature demand.

Bar chart: a $60,000 H100 breaks even at 65% utilisation at Seeweb prices, 48% at Scaleway, 46% at OVHcloud and 11% at AWS.
Author’s illustration. $60,000 per GPU comes from one $480,000 quote for an 8-GPU HGX H100. Prices checked between July and November 2026; the AWS figure is a competitor’s estimate. Excludes power, staff and financing.

The arithmetic is unforgiving. At $60,000 per GPU and a five-year life, a $2.11 rental price needs 65% utilisation just to return the capital. At $2.99, it needs 46%. Hyperscalers need far less, but they also carry the costs of a full platform. Power, staff and interest push the real break-even higher, and cheaper hardware quotes would push it lower.

Market signals by shore

  • North: listed prices are real, so margins can be compared and competed away. Public AI factories also reach startups and SMEs, which may narrow part of the commercial market.
  • East: Israel’s subsidised programme sets expectations for local firms, while Turkey’s cost of money makes domestic financing very expensive.
  • South: most training-class pricing remains private, and investors have little public data on demand.

Structural costs

  • Power: listed 2025 business rates range from €0.142 per kWh in Malta to €0.280 in Cyprus.
  • Grid: Spain’s 3.8 GW allocation shows the scale of demand, and its delivery remains uncertain.
  • Currency: GPUs are priced in dollars, while buyers may earn euros, lira, shekels, dirhams or Egyptian pounds.
  • Refresh: the region already runs five chip generations, so a new cluster can lose its pricing power within a few years.
  • Tight real estate: Paris colocation vacancy was 8.0% in Q2 2026, so even mature markets are not simply waiting for tenants.
  • Policy risk: the January 2025 AI diffusion rule has been rescinded, no replacement had been issued by September, and US officials have signalled further action.

Hosting equipment: sovereignty changes at each border

Local hosting gives data residency, not full sovereignty. In the Mediterranean, the meaning of sovereignty also changes with the legal system on each shore.

BlocMain rule or pressureWhat it means for GPU hosting
EU membersProposed Cloud and AI Development ActPublic bodies could face four assurance levels, with EU providers as the general rule
TurkeyKVKK data protection lawLocal hosting is a selling point for domestic providers
IsraelUS-linked hardware supplyNational compute is run by Nebius, an international operator
Southern shoreSector and national rules, including Morocco’s decree 2-24-921Demand forms around regulated data

The EU proposal is significant but unfinished. Published in June 2026, it aims to triple EU data centre capacity within five to seven years and would generally limit public bodies to EU-established cloud providers. Non-EU providers could qualify only under strict conditions. It still has to pass through Parliament and Council.

A European flag is not the same as a sovereign GPU service. OVHcloud holds SecNumCloud qualifications for some products, but its Public Cloud GPU instances were not qualified as of July 2026. Buyers should check the exact product, not only the company.

US export rules cut across the sea. Under the country groups in the now-rescinded AI diffusion framework, France, Italy and Spain sat in the most favoured tier, while Cyprus, Lebanon, Libya and Syria sat in the most restricted. That rule is gone, but the map shows how Washington may sort the region. A Cypriot project and an Italian one should not assume the same access. Providers need legal advice on current rules.

The US CLOUD Act adds another layer, because it reaches US providers wherever the data sits. Turkcell’s model tackles this by giving customers sole control of encryption keys, though the platform remains Google’s. Sovereignty often means choosing a dependency rather than removing one.

Before calling a hosted setup sovereign, buyers should ask:

  • Who owns the GPUs, and can they be remotely restricted?
  • Who holds root access and encryption keys?
  • Which foreign laws reach the operator or its parent?
  • Can the service run if vendor support or licences stop?
  • Is the specific GPU product covered by the certification being claimed?

Will web hosters and cloud providers survive?

Most will survive, but fewer will stay independent, and many will sell something different by 2030. This section is my forecast, not a measured result.

Provider typeForecast to 2030Main reasonBest survival move
Commodity web hostersSurvive through consolidationPrice pressure and SaaS site buildersSell to a regional group, or move into managed services
European sovereign cloudsLikely to growProcurement rules and customer cautionWin qualifications for GPU products, not only general cloud
Telco cloudsLikely to surviveNetwork reach and enterprise accountsPartner on sovereign regions, as Turkcell has
Colocation operatorsLikely to growEvery model still needs power and interconnectionAdd high-density capacity near cable landings
Pure GPU resellersHighest failure riskThin margins and price transparencySecure contracts first, or specialise
Model companies with computeNew competitorsMistral sells both models and capacityPartner with them, or serve the niches they ignore
Global hyperscalersWill stayScale and platform breadthLocal partners and sovereign structures

The northern shore will consolidate first, because prices are public. Seeweb’s $2.11 rate and OVHcloud’s $2.99 sit in one comparable market, so an undifferentiated reseller has little to stand on. Gartner expects more than 80% of neoclouds to lose independence by 2030.

Local cloud has a niche, not a takeover. European providers have held about 15% of their home market. The EU’s cloud proposal could lift that share in public procurement, but only if it survives negotiation.

On the eastern and southern shores, hosters have a longer runway. They face less direct price competition from local rivals, and local billing, language and support still matter. Their risk is being bypassed when a telecoms group or hyperscaler arrives.

The practical advice is the same on every shore. Treat GPUs as one product, buy only against signed demand, and build value in support, compliance and local relationships.

The real opportunity: bridging the shores

The opportunity is not another large campus. It is trusted capacity that crosses the region’s legal and economic lines.

Three demand signals exist independently of any AI hype cycle:

  • Public procurement: the EU’s proposal could turn sovereignty from a preference into a buying rule.
  • Regulated data: banks, health groups and public bodies across the region have data they cannot freely move.
  • Language: models for Arabic, Turkish, Greek, Hebrew and other regional languages need local data and fine-tuning.

The barriers also protect the market. A credible provider must secure GPUs, power, licences, compliance and support, and must do so under several legal systems. Few companies can do that across more than one shore. This is my reading rather than a measured finding, but it suggests that trust will matter more than scale in much of the region.

Speculative signalReal signal
Gigawatt targetsLive capacity and a funded first phase
Memoranda of understandingSigned anchor customers
« Sovereign » labelsCertification of the specific GPU product
A long-term investment totalNamed chip vendor and IT load
Grid power assumedContracted power with a backup plan
One subsidised launchPaying customers after the subsidy

The most durable models will start small and work across borders: northern capacity for training, local hosting for regulated inference, and pooled access for researchers. Whether that model can cross the sea depends on export rules as much as on cables.

Five developments will show which way the region moves:

  • The EU’s gigafactory awards, expected by early 2027, and whether Spain’s bid succeeds.
  • The final form of the Cloud and AI Development Act.
  • Any US replacement for the AI diffusion rule.
  • The outcome of Egypt’s data centre bid.
  • Whether Mistral, Turkcell and Nexus deliver capacity on their stated dates.

The Mediterranean will not reward the largest announcement. It should reward the operators that make capacity usable, lawful and trusted on both sides of the sea.

Notes on data and sources

Figures reflect sources available on 2 November 2026. Many project figures are company or government claims, and capacity totals differ by definition. Hourly prices were checked on different dates and change often. Some comparisons come from Spheron, a vendor, so treat them as indicative. Country groupings by shore are the author’s.

These facts could not be verified: IMF 2026 forecasts for Greece, Israel and Egypt, Pantheon’s IT load and chip vendor, Mistral’s go-live date, Türkiye’s national targets from an official source, and the absence of a GPU service in countries where none was found.