The neocloud belt goes east, and the bandwidth bill lands here

Something happened this summer that most of the trade press treated as separate stories. Firebird switched on a 300 MW AI factory in Hrazdan on 8 August (Firebird). HUMAIN went live in Riyadh on 31 August with AMD MI355X silicon and Cisco 800G optics (Cisco Newsroom). Nebius signed an 80 MW lease in Israel with a path to 280+ MW (DatacenterDynamics). DataVolt closed a 12-year non-recourse loan for TAS-1 in Tashkent (DataVolt). Silk Road Group broke ground in Tbilisi (Caucasus Business Journal). Azerbaijan started openly courting hyperscalers around the Alat Free Zone (Caucasus Business Journal).

Read them together and you get a belt. A rough arc from Yerevan through Ekibastuz, Tashkent, Baku and the Caspian, down through Israel to Riyadh and Abu Dhabi. AI compute is being poured into the geography between Europe and India, and it is being poured fast.

For a Balkan wholesale operator this is not an interesting industry news item. It is a demand curve.

Where the sites actually are

Firebird is the loudest of the group and probably the most consequential for our region. The Hrazdan factory is a real facility, not a rendering. It runs on Dell hardware and NVIDIA Spectrum-X Ethernet, opened with Prime Minister Pashinyan on stage, and lists Perplexity as one of its first customers (Firebird). The company is targeting more than 70,000 NVIDIA Rubin and Blackwell GPUs by end of 2027 and 2 GW of global capacity by 2028, most of it in Armenia and Kazakhstan.

Kazakhstan is where the scale goes vertical. Firebird holds a 125 MW allocation in Ekibastuz, operated by Kazakhtelecom, as the first phase of a Data Center Valley that the government has publicly framed as scalable toward 1 GW and up to USD 30 billion of investment (The Astana Times). The US Bureau of Industry and Security has already cleared export authorisation, which is the interesting detail. Ekibastuz is not going to be a domestic play. Whatever gets built there will be selling AI capacity across the region and into Europe, and it will need transit that does not route through Russia.

Uzbekistan is a smaller version of the same story. DataVolt raised up to USD 150 million in twelve-year non-recourse financing from DEG, EBRD, the OPEC Fund and Proparco for TAS-1 at Tashkent IT Park. First 6 MW ready by end of this year, second phase in late 2027 (DataVolt). Small in megawatts, but the financing structure tells you something: European development banks are now willing to underwrite Central Asian AI infrastructure on 12-year money. That was not true two years ago.

Azerbaijan has not landed an anchor tenant yet. The Ministry of Economy is openly recruiting hyperscalers to the Alat Free Economic Zone and the SABAH.city cluster, and Baku is being sold as an east-west routing hub anchored on the Digital Silk Way across the Caspian (Caucasus Business Journal). It is a positioning play for now. But given what is happening on either side of the Caspian, it is not a positioning play that will stay theoretical for long.

Georgia is quieter and more interesting than the numbers suggest. Silk Road Group started construction on a Tbilisi facility in September, explicitly built around Black Sea fibre landings and a move to stop routing Georgian hosting through Frankfurt, Istanbul or Yerevan (Caucasus Business Journal). No anchor customer named. No hyperscaler yet. But the intent is clear: keep local traffic local, and then sell what is left as regional capacity.

The Gulf is a different animal. Stargate UAE, led by G42 with OpenAI, Oracle, NVIDIA, Cisco and SoftBank, is aiming for 200 MW online this year as the first phase of what was originally pitched as a single 5 GW campus in Abu Dhabi. That plan is now being redrawn into a distributed network of sites, in part because concentrating that much compute in one place turned out to be a physical security problem after Iranian action earlier this year (Reuters). HUMAIN went live in Riyadh on 31 August with an AMD Instinct MI355X cluster interconnected on Cisco Silicon One and 800G optics, planning up to 250 MW next phase and 1 GW by 2030 (Cisco). Israel is the third leg: Nebius signed an 80 MW lease with Mega Or across Masmiyya and Beit Shemesh, with an expansion path north of 280 MW and roughly USD 880 million in construction commitments (DatacenterDynamics).

Add that up and you have somewhere north of 1.5 GW of AI compute either running or firmly under construction in the arc from Armenia to Saudi Arabia, with another several gigawatts announced. That is not a rounding error in anyone's forecast.

The traffic is not what you think

Here is the part most bandwidth planning gets wrong. Neocloud sites do not look like the colocation halls of the last decade. Two workloads dominate, and they behave nothing alike.

Inference is boring in the good sense. It scales with end-user demand, it is roughly flat over the day with predictable peaks, and you can sign a three-year commit against it and sleep well. Call it a few hundred Gbps for a 100 MW site, give or take, depending on how inference-heavy the tenant mix is.

Checkpoints are the problem. When a large model trains, its state has to be written out to remote object storage on a schedule, and those writes are not gentle. A trillion-parameter model producing a full training checkpoint dumps enough bytes in a short enough window to push the site's egress into the low single-digit Tb/s. For a few minutes. Then it goes quiet again.

Now put three independent tenants on the same 100 MW campus. Their checkpoints do not sum. If they did, the world would need a lot more submarine cable than it has. What actually happens is a coincidence factor: at three tenants running independently, roughly 0.70 of the theoretical sum. At eight, 0.40. Above that it keeps decaying. The peaks are real, but the aggregate peak is a lot less scary than the arithmetic suggests.

This has a concrete commercial edge that is worth spelling out. The single cheapest wavelength a neocloud operator can buy is the one they do not buy, because they staggered their tenants' checkpoint windows in the contract. No CFO signs off on that on their own. It has to come from the network side, and it has to come early, before the tenants have anchored their scheduling on habit.

We built the AI Neocloud Bandwidth Calculator partly to make this visible to the sales conversation. You put in the megawatts, model size and tenant count. It returns average egress, peak egress and the number of 400G lambdas the peak actually needs after the coincidence factor bites. The last number is what our customers eventually sign for. The first three are what tells them why.

The corridor

The rest is arithmetic. Model checkpoints land in object stores in Frankfurt, Amsterdam and Marseille. Inference traffic follows population, which for this belt means Istanbul, Athens, Bucharest, Sofia, and everything west of that. The lines between those endpoints and the compute cluster in the Caucasus and the Gulf run through us, or through routes we compete on.

That has implications the pricing spreadsheets have not caught up with yet. Two Bulgarian entry points matter more than they did a year ago. The Black Sea landing at Balchik, where Georgian traffic actually reaches European fibre, and the Komotini crossing into Greece, where the TANAP corridor pushes Caspian and Middle Eastern traffic west along the pipeline route. Both feed onward into Sofia, and from Sofia into the Balkans-to-Central-Europe backbone. Neither route is a headline product today. Both will be, once Caucasus and Gulf compute starts filling them.

The bigger shift, though, is not about which corridor gets used. It is about what a wavelength is being sold for. Neocloud tenants sizing burst egress do not care about the headline price of a single lambda. They care whether two lambdas on genuinely diverse fibre both survive a cable cut. They care whether the tail latency to Frankfurt is bounded, not whether the median is a millisecond faster than the competition. They care about a status page that does not lie to them at 3 AM.

That is a market a regional operator with owned duct, real diversity and a clean operations record can win. Not by underpricing tier-1 wholesalers, which is a losing game anyway, but by selling something the hyperscalers actually cannot get from a tier-1 wholesaler at any price: a route they can trust on the worst day of the year.

It is worth being direct about who else is playing. EXA Infrastructure has the deepest fibre footprint into the Balkans and just extended it further, buying all long-haul duct assets from Conexio in Serbia and launching Project Visegrád, the largest cross-border fibre backbone build in Central Europe in 25 years. Türk Telekom International sits on top of the natural Caucasus-to-Europe gravity path, with a terrestrial and subsea backbone that reaches from Western Europe through Türkiye into the Caucasus, the Middle East and Asia. iQ Networks is building the Silk Route Transit from Iraqi Kurdistan through Türkiye toward the Gulf, and Ooredoo Wholesale Services is the natural landlord of the Doha-to-Marseille corridor its parent operator uses, recently spun into a standalone fibre and subsea unit to chase exactly this AI-driven wholesale demand. Between them, they cover most of what a Caucasus or Gulf neocloud tenant needs. We do not pretend otherwise.

What Sofia Connect brings to the same conversation is Bulgaria. We run a latest-generation DWDM network, 400G today and 800G ready, and we terminate it at two places that matter for this belt. The first is the cable landing station at Balchik, where the fibre coming across the Black Sea from Georgia lands. The second is Komotini, where we are building the interconnection into the TANAP/SOCAR/TAP fibre pair that runs alongside the pipeline system from the Caspian, through Türkiye and Greece, into Western Europe. Once that Komotini leg is live, we can take neocloud traffic off either the Georgian Black Sea path or the TANAP corridor and move it onto our own network. From there it is a short ride: hand it off at the Bulgaria-Serbia border for onward transport west, or drop it in Sofia for local peering and inference distribution. If you want to see the shape of that traffic before you commit to it, run your numbers through our AI Neocloud Bandwidth Calculator and we will meet you at the answer.

The gigawatts are being built. The chips are landing. The question that gets decided in the next twenty-four months is whose fibre carries the traffic when it is running hot.

Sources

• Firebird, Grand opening of the Armenian AI factory
• The Astana Times, Kazakhstan emerges as Firebird AI market with Data Center Valley project in Ekibastuz
• DataVolt, USD 150 million non-recourse financing for TAS-1 in Tashkent
• Caucasus Business Journal, Azerbaijan courts hyperscalers as AI strategy targets data exports
• Caucasus Business Journal, Silk Road Group starts Georgia data centre build
• Reuters, UAE revises AI data center plan after Iranian attacks, sources say
• DatacenterDynamics, Nebius signs 80MW data center lease with Mega Or in Israel
• Cisco Newsroom, AMD, Cisco and HUMAIN expand Saudi Arabia AI infrastructure
• EXA Infrastructure, Purchase of long-haul duct assets from Conexio in Serbia
• EXA Infrastructure, Project Visegrád, largest cross-border fibre backbone in Central Europe in 25 years
• Türk Telekom International, Network map covering Europe, Türkiye, Caucasus, Middle East and Asia
• iQ Group, Silk Route Transit fibre project from Iraqi Kurdistan through Türkiye to the Gulf
• Ooredoo, Ooredoo Wholesale Services, international carrier arm and standalone fibre and subsea unit