$19.6M AI Contract, $93M Cash, No Debt, & 400MW of Power

🌾 Welcome to StableBread’s Newsletter!

DigiPowerX (DGXX) is a Miami-based AI infrastructure operator with ~400MW of secured power capacity across four U.S. sites.

The company started as a Bitcoin mining and energy operator on February 18, 2017, and is now repurposing that power into Tier 3 data center capacity sold to GPU-as-a-Service customers under its "NeoCloudz" brand.

The pivot has been in motion since November 2025, when the company filed an amended $200M at-the-market equity offering and started developing the ARMS 200 (its 1MW prefab GPU data center pod design) with Super Micro Computer.

This is about when I started following the company, but held off on the write-up until I saw some contracted revenues.

That changed on April 20, 2026, when DGXX produced its first piece of contracted evidence: a $19.6M, 24-month bare-metal GPU rental agreement with a stealth AI lab called SubQ. The deal is small relative to the company's stated 100MW AI plan, but it validates the thesis DigiPowerX has been building toward.

Beyond the SubQ contract, here's what DGXX has going for it:

DGXX trades around $3.50, giving the company a $244M market cap (69.8M shares × $3.50). After netting the $93M cash and digital currency position, enterprise value is $151M. The cash backing alone is $1.33/share.

Today DGXX trades at 4.4x EV/trailing sales on $34.2M of mostly legacy revenue. Management's FY 2027 capacity goal is 100 MW of AI infrastructure (90 colo + 10 GaaS), implying ~$282M of annualized revenue at full utilization. Our base case forecast lands at ~$197M after realistic haircuts.

Either way, today's $151M EV trades at 0.5-0.8x forward sales.

AI infrastructure peers trade at 4-9x forward EV/Sales (CRWV at 4.2x, NBIS at 6.3x, IREN at 6.9x). Even at the low end of the peer range applied to FY 2027 base case revenue, EV expands ~4x from current levels. At peer-mean (~7x), it's a 6-7x EV re-rating.

Per-share return runs lower than EV expansion because of dilution from ATM and lease financing funding the buildout. Our base case projects $15.62/share by FY 2027 (+346%). Our bull case scales to $26.28/share by FY 2027 (+651%).

The downside is bounded by ~$1.33/share of gross cash backing today, but that floor erodes as the buildout consumes the balance sheet through 2026.

Let's dive in deeper now to fully understand the business and the risk/reward scenarios.

DigiPowerX’s legacy business is built on three revenue lines, each at a different stage of repositioning into AI infrastructure:

Full year 2025 revenue came in at $34.2M, down 7.6% y/y. The decline is intentional—capacity that previously generated low-margin self-mining revenue is being redirected toward higher-multiple AI infrastructure.

Worth noting: on the April 1, 2026 earnings call, CFO Paul Ciullo specifically pushed back on the "legacy" framing for colocation and energy, calling them "the foundation upon which the company's AI data center build-out is anchored."

What's actually being wound down is Bitcoin self-mining.

AI colocation at $150/kW/month generates ~$1.8M/MW annually. GPU-as-a-Service at $3.50/hr/GPU generates ~$9.8M/MW annually. Bitcoin self-mining produces a fraction of either on the same power footprint.

Per the company's December 2025 shareholder letter, DigiPowerX has secured ~400MW of power across four U.S. sites:

Greenfield power capacity at this scale typically takes 24-48 months and $30-75M of upfront infrastructure spend per 100MW interconnect. DigiPowerX already has substations, transmission, and zoning in place across all four sites.

The cogen power advantage is structural and quantifiable. At ~$0.04/kWh vs. $0.07-0.10 for grid-priced peers, DigiPowerX saves $0.03-0.06 per kWh on every MW served by the NY cogen plant.

At full ramp of the NY site (~60MW operating today, scaling to 123MW secured), that's roughly $15-30M/year of operating margin advantage (60,000 kW × 8,760 hrs × $0.03-0.06/kWh) over competitors paying market rates.

DigiPowerX has been adding senior people across the disciplines needed to operate enterprise AI infrastructure:

DigiPowerX is hiring senior people from large-scale operators across the disciplines that matter: engineering and infrastructure from Oracle and Verizon, finance from Deutsche Bank, and institutional capital access through BlackRock.

That mix gives the AI pivot a credible execution bench beyond its crypto roots.

DigiPowerX partnered with Super Micro Computer (SMCI) to co-develop the ARMS 200, a prefabricated, modular data center pod.

SMCI provides the rack-scale GPU servers and reference architecture; DigiPowerX (via its USDC subsidiary) handles pod-level design, integration, and manufacturing.

Each unit delivers 1MW of compute and houses up to 400 NVIDIA B200/B300 GPUs (per the SubQ press release confirming 4,000 GPUs across 10 pods).

SMCI is one of NVIDIA's primary OEM partners and the leading supplier of AI-optimized GPU servers. Its reference architectures sit inside hyperscaler deployments at Meta, Microsoft, and OpenAI, and its rack-scale systems set the integration standard for Blackwell-class compute.

SMCI carries some accounting overhang from a 2024 Hindenburg short report and subsequent SEC investigation, but product delivery has continued on schedule.

For DigiPowerX, the partnership de-risks GPU integration and shortens time-to-market on each pod. ARMS 200 deployment time is ~180 days vs. 24-36 months for traditional AI-grade data center construction. The first pod was commissioned at Alabama in mid-March 2026 and entered GPU testing shortly after.

The ARMS 200 received Tier 3 certification under ANSI/TIA-942, requiring 99.982% uptime. That's the bar enterprise customers need before signing meaningful contracts.

Two monetization paths off the platform:

DigiPowerX is also developing ARMS 300 and ARMS 400 variants aimed at hyperscaler and government workloads. The ARMS 300 is being engineered around next-generation NVIDIA chips for a planned ~120MW deployment at the New York facility. The ARMS 200 is patent-pending.

In early March 2026, DigiPowerX publicly announced US Data Centers Inc. (USDC), a subsidiary focused on global deployment of modular Tier III AI data centers.

The entity itself was established December 31, 2025 per CFO commentary on the April 1, 2026 investor conference.

Hans Vestberg, former Verizon chairman/CEO and current BlackRock board director, joined DigiPowerX as senior advisor on February 2, 2026, then was named co-founder of USDC the following month.

The restructure rolled out with limited initial transparency. Concerns over dilution (DGXX shareholders effectively giving up 45% of a venture they might have expected to own outright) and unclear governance drove the stock down to recent lows around $1.89.

DigiPowerX's March 16 clarification press release and a follow-up interview with CEO Michel Amar walked back most of the concerns.

Management clarified that USDC owns no operating assets. It's a manufacturing and distribution business for the ARMS platform. All pods, GPUs, sites, and customer revenue stay 100% with DigiPowerX, which buys ARMS units from USDC at cost.

A single 100MW customer order would require $250-300M in capex. Funding that through a $200M-class market cap parent either wrecks the balance sheet or forces heavy dilution.

Spinning USDC out as a separate manufacturing entity puts the capex risk on a separate balance sheet. DigiPowerX gets upside through its 55% stake without owning the manufacturing risk.

Here’s DGXX’s current cap table:

Vestberg's BlackRock board seat gives DigiPowerX a direct line to hyperscalers and institutional capital.

Governance remains the open question. The March 16 release discloses that DigiPowerX does not have voting control over USDC despite owning 55% of equity, and the voting structure has not been publicly explained.

Until that's resolved, DGXX owns 55% of the economics but doesn't formally control the decisions.

Now let's talk about their first customer. The infrastructure, partnerships, and team only matter if real customers show up. April 20, 2026 was the day that started happening.

DigiPowerX signed a 24-month bare-metal GPU rental agreement with SubQ AI. Here’s the terms:

The math checks against management's stated rate: $3.50/hr × 400 GPUs × 8,760 hrs × ~80% utilization = ~$9.8M/year, matching the $19.6M / 24 months contract.

DigiPowerX CEO Michel Amar explained the deal here:

Sized against the company's stated 10-pod / 4,000-GPU GaaS footprint, this deal is roughly 1MW or ~10% of GaaS-specific capacity. Sized against management's full 100MW AI plan (90MW colocation + 10MW GaaS), it's only about 1% of the total target.

The $2.95M upfront covers ~15% of the per-pod buildout cost ($18-20M). Monthly Net-15 billing keeps cash collection tight, and the 24-month term locks in $9.8M of annualized contracted revenue.

Management is clear that this is the first customer, not the destination:

The CTO's comments the same day point in the same direction:

A Silicon Valley office isn't the kind of decision a company makes if SubQ is one-and-done.

The customer matters as much as the contract. SubQ AI is a Miami-based AI lab that rebranded from "Aldea" in early 2026.

The team is small but credible. Justin Dangel runs the company as CEO. Cofounder Alexander Whedon previously worked at Tribe AI and Meta. Headcount sits at roughly 15-30, with active hiring in Miami and the Bay Area.

SubQ is building a proprietary "post-transformer subquadratic" architecture aimed at scaling more efficiently than today's quadratic-cost transformers. That means longer context windows at meaningfully lower training and inference cost.

The current product is a Speech-to-Text API at $0.09 per hour with 6.1% WER and sub-100ms latency, positioned as Deepgram-compatible. A Speech-to-Speech product is in alpha. An LLM platform is on the roadmap.

The closest comp is Cartesia, founded by the authors of the Mamba paper, which has raised ~$91M.

Training a post-transformer foundation model from scratch requires bare-metal Blackwell GPUs with root access, which is exactly what DigiPowerX is offering through NeoCloudz.

If SubQ ramps to even 1,000 GPUs (~2.5MW) over the next few quarters, that's ~$24.5M of annualized revenue from one customer at SubQ-contract economics. The full multi-thousand-GPU target could fill the entire 10MW GaaS footprint over the next 4-6 quarters.

FY2025 results were reported on March 31, 2026: