Energy is abundant.
Delivery is not.
HAYSOLAR builds autonomous, behind-the-meter AI compute clusters — converting locally generated energy directly into compute capacity, without waiting on grid interconnection or exporting power at regulated tariffs.
Four things to know before you read the rest of this page
Power, not chips, is the constraint
~30–50% of new global data center capacity is now being built as on-site generation because grid interconnection queues run 4–8+ years in mature markets. IEA · Cleanview · LBNL, 2026
Armenia is already a proven AI-compute destination
A $4B, 50,000-GPU NVIDIA-backed AI megaproject (Firebird) is under construction in Armenia right now, with U.S. and Armenian government backing. HAYSOLAR is a different category — see below.
The unit economics are public and verifiable
Local consumption captures ~53 AMD/kWh in displaced retail tariff vs. 18–42 AMD/kWh for grid export — a 25–65% uplift, before compute monetization on top.
New category: distributed, autonomous, behind-the-meter
Not a hyperscale campus. A replicable, site-by-site model for solar assets too small or too grid-constrained for gigawatt-scale players to touch.
Why a solar company became a compute company
HAYSOLAR began as a renewable energy company developing grid-connected solar assets in Armenia. We hit a structural wall: the economics of renewable generation are set by regulated tariffs and utility policy, not by the cost of producing the energy. Pre-2020 stations sell under fixed feed-in tariffs; post-2022 stations sell into an open market with no price floor at all. Either way, someone else prices your export.
The highest-value electron is the one consumed locally.
A kilowatt-hour used behind the meter is worth the full retail tariff it displaces — not the export price it would otherwise fetch. So instead of exporting electricity, HAYSOLAR converts locally generated energy directly into AI compute, through autonomous, behind-the-meter clusters sited on the generation asset itself.
The spread is the business case
We're not asking investors to take a macro thesis about AI power scarcity on faith. The spread between what a business pays for grid electricity and what it's paid for exporting solar power is public, regulated, and verifiable today.
Two paths for the same electron
One path sells the electron into a market that prices it as a commodity. The other keeps it on site and converts it into a metered, monetizable unit of compute — at the full value of the tariff it displaces.
We sell compute capacity.
One programmable system, five layers
Generation, storage, compute and cooling are designed as a single site — not procured as separate contracts with separate counterparties.
Solar
On-site generation, sized to the compute load rather than to export capacity.
BESS
Battery storage to flatten intermittency and keep compute utilization high overnight.
GPU
Compute hardware sited directly at the point of generation — no transmission losses, no interconnection queue.
Cooling
Site-specific thermal management, sized for Armenia's climate and elevation profile.
Control Layer
Software that routes power and schedules compute jobs against real-time generation — in development; see roadmap in Executive Summary.
Power is the bottleneck, and it's measurable
This isn't a slogan. It's the operating reality reported by the IEA and grid operators tracking the AI buildout in 2026.
A grid that's already telling you the answer
A $4B, 50,000-GPU NVIDIA-backed AI cluster is already being built here
Firebird AI's data center in Hrazdan, Armenia — backed by NVIDIA, the Armenian government, the U.S. government, and a $300M syndicated loan from six Armenian banks — is scaling toward one of the world's top-five largest AI GPU clusters. Phase 1 (~100 MW, $500M) launches in 2026; Phase 2 ($4B, 50,000 NVIDIA GB300 GPUs) was announced in February 2026 during a U.S. Vice Presidential visit to Yerevan.
This validates the category, not our specific model — Firebird is a grid-connected, gigawatt-class hyperscale campus. HAYSOLAR is a different, complementary category. See Category below.
Solar buildout is real and recent
Installed solar capacity crossed 1,000 MW in 2025 and reached 1,141 MW by March 2026 — 662 MW of it autonomous/behind-the-meter capacity already, not utility-scale export plants.
The grid operator is asking for storage, not more export
Officials have publicly flagged that 1,000+ MW of solar creates management challenges for the grid, and subsidies are being redirected toward battery-paired systems rather than standalone generation.
Net billing already trains the market to think behind-the-meter
Armenia's existing net-metering regime for systems up to 500kW already conditions businesses to value on-site consumption over export — HAYSOLAR extends that logic to industrial scale and to compute.
Deployment and land costs remain lower than Western Europe
[insert specific comparison once site-level costing is finalized] — directionally true regionally, but this claim needs a cited benchmark before it goes in front of investors.
A new category: Distributed Autonomous AI Infrastructure
Not a bigger, cheaper version of a hyperscale campus. A different shape of business, built for a different part of the market.
e.g. Firebird
- One flagship site, gigawatt-scale ambition
- Grid-connected (plus on-site backup), state and hyperscaler-backed
- $500M–$4B capital intensity per project
- Wins where grid capacity and sovereign backing already exist
HAYSOLAR
- Many smaller sites, sized to existing solar assets
- Behind-the-meter by design — no interconnection queue to wait on
- Capital-efficient per site, replicable across underused solar land
- Wins precisely where grid capacity is constrained or uneconomical
The insight came from operating inside the constraint, not observing it from outside: while developing grid-connected solar in Armenia, it became clear that regulated tariffs — not generation cost — set the ceiling on returns. That's a structural problem for every grid-connected solar developer in the country, and it doesn't get solved by building a bigger version of the same asset.
[Add: specific unfair advantage — land relationships, engineering partnerships, or regulatory access that a well-capitalized competitor couldn't replicate quickly. Needs real input before this goes to investors.]
Your land is worth more as compute than as export
This is also how the network scales: HAYSOLAR replicates the Genesis model onto other suitable solar sites across Armenia, in cooperation with the landowner — not by buying land outright.
The commodity option
- Fixed rent, typically $500–1,000/acre/year internationally
- Priced off electricity export value — the lowest-value electron
- Landowner has no exposure to how well the asset performs
- Well understood, low-risk, low-upside
Priced off compute, not electrons
- Guaranteed minimum payment, structured like a standard land lease floor
- Plus a share of compute revenue on top — priced off a business that earns far more per hectare than electricity export
- HAYSOLAR funds, builds, and operates; landowner takes on no capital or operating risk
- Registered, long-term land use right — survives a change of ownership
The AI-compute category has a real fraud problem. Check everything, including us.
"Invest in GPU compute" is now also the pitch used by unregistered daily-yield platforms promising returns no real infrastructure business can generate. Before evaluating HAYSOLAR or anyone else in this category, verify these five things — we've listed exactly what we show for each.
Executive Summary
HAYSOLAR — Pre-Seed. Built to be lifted directly into a data room. Bracketed fields need a real number before this goes to an investor.
AI compute buildout is increasingly constrained by power delivery, not chip availability — grid interconnection queues and regulated tariff structures slow down or cap how much compute can be sited in any given market.
In Armenia specifically, solar generation economics are shaped by regulated feed-in tariffs and an open market with no price floor for newer stations, capping the return on grid-connected solar regardless of the underlying cost of generation.
HAYSOLAR builds autonomous, behind-the-meter AI compute clusters that convert locally generated solar power directly into compute capacity — capturing the full retail tariff value of each kilowatt-hour instead of the lower export/wholesale price.
The company sells compute capacity, not electricity, and is not a utility or a licensed energy trader.
Land: 2-hectare site in/near Aramus village, Kotayk Province — ~25 km north of Yerevan, elevation reported between 1,420–1,550 m depending on source (Wikipedia cites 1,420 m; Wikidata cites 1,550 m — needs on-site GPS confirmation). [exact plot coordinates / cadastral reference pending].
Engineering: Internal engineering-economics model (not yet a third-party EPC study) now benchmarked against a real Armenian precedent — Shtigen LLC's ArSun utility-scale plant (2 MW on ~4 ha, 0.5 MW/ha) — indicating the 2-hectare site can support ~0.85 MW solar nameplate, ~5 MWh BESS, and ~120 GPUs behind the meter. Shtigen (shtigen.com) is a real, active Armenian solar EPC with relevant utility-scale experience — a candidate contractor, [engagement not yet confirmed].
Regulatory: Structured as a compute-capacity sale, not a securities or crypto-asset offering — no CASP licensing or token issuance is part of this raise. (If that changes, it needs independent Armenian legal sign-off before any public claim of licensed status — see note below.)
Armenia's solar base has scaled past 1,141 MW while the regulator signals a shift away from standalone export generation toward storage-paired, locally-consumed models. A $4B NVIDIA-backed AI megaproject (Firebird) is already under construction in-country — proof the category is real here, not hypothetical. HAYSOLAR occupies the distributed, behind-the-meter tier that a hyperscale campus model doesn't serve.
High solar irradiation, a grid operator already grappling with export saturation, an existing net-billing culture among commercial consumers, and geography positioning between European and Asian markets.
Round size = total modeled CAPEX for one Genesis site (Aramus, 2 ha) — solar, BESS, hangar, and GPU hardware — per HAYSOLAR_Aramus_2ha_Project_Economics.xlsx, revised to use a real Armenian solar-density benchmark (Shtigen's ArSun plant) in place of a US empirical figure. [Decide: raised as full equity, or blended with debt/lease financing on the GPU hardware — GPUs are ~63% of this CAPEX and the most financeable component via hardware-backed debt.] Returns by scenario (8-yr IRR): Conservative −12.8% · Base 4.9% · Optimistic 25.1% — the Base case is materially weaker than in earlier drafts of this model, precisely because the Armenia-specific density figure yields less solar (and therefore fewer supportable GPUs) per hectare than the US benchmark did. See Executive Summary disclaimer above on what's modeled vs assumed.
The next generation of AI compute won't wait for the grid.
HAYSOLAR is building the infrastructure that doesn't have to.