The short thesis:

10x is transitioning from a single-cell sequencing company into the dominant spatial-biology platform, while AI creates a second-order demand driver for its data. Atera could be the product that reaccelerates growth and turns TXG from a stagnant life-science tools company into a high-growth platform again

1. What TXG actually does

10x sells instruments + proprietary consumables to research institutions, pharma and biotech.

The basic model is:

Instrument → researchers run experiments → recurring consumables

That recurring consumables component is extremely important.

In 2025:

• $643M total revenue

• $507M consumables revenue

• ~70% gross margin

• 10,000 peer-reviewed papers using 10x-generated data

• 1,000 instruments sold during the year
  

And the instruments aren’t the real prize.

The prize is locking researchers into an experimental ecosystem that continually consumes reagents.

Think of it somewhat like:

NVIDIA → GPU → CUDA ecosystem

except:

10x → instrument → proprietary consumables → biological datasets

2. The original Chromium business is already huge

The core product is Chromium, which performs single-cell analysis.

Instead of looking at a tissue as:

“Here are 100,000 cells.”

you can essentially ask:

“What is each individual cell doing?”

That is incredibly useful for:

• cancer research

• immunology

• drug discovery

• neuroscience

• developmental biology

• gene therapy

• biomarker discovery

The problem is that single-cell sequencing is no longer a hypergrowth story by itself.

That’s why TXG’s revenue has been stagnant.

Products/services revenue was:

Year

Revenue

2023

$619M

2024

$610M

2025

$597M

So if the thesis were simply “single-cell sequencing grows,” I wouldn’t be particularly excited.

The real thesis is spatial biology.

3. Spatial biology is the big opportunity

This is where things get much more interesting.

Traditional single-cell sequencing tells you:

what genes a cell expresses.

Spatial biology tells you:

what genes a cell expresses + where that cell physically sits relative to everything else.

That matters enormously.

Imagine a tumor.

You don’t just want to know:

Cell A expresses gene X.

You want to know:

Cell A is sitting next to these immune cells, surrounding this tumor region, interacting with these other cells, and expressing these genes.

That spatial context can reveal how diseases actually work.

TXG has:

• Visium

• Xenium

• now Atera

And spatial consumables grew from $121M in 2024 to $144M in 2025, while single-cell consumables declined modestly.

That’s the signal I care about.

4. Atera is the potentially massive catalyst

This is the most important part of the thesis.

10x launched Atera in April 2026.

The company describes it as a spatial whole-transcriptome platform with single-cell sensitivity at substantially greater scale. Commercial shipments are expected to begin in H2 2026.

Why does that matter?

Because spatial biology has historically involved a tradeoff:

High resolution ↔ scale

Atera is trying to push both simultaneously.

If successful, the workflow becomes much more powerful:

Tissue → spatial transcriptome → enormous biological dataset

And that brings us to the most interesting part of the entire TXG thesis.

5. AI could actually be a huge tailwind

This is where I think TXG is being underappreciated.

AI models need data.

But in biology, the valuable data isn’t just:

“Gene X exists.”

The valuable data is:

Gene expression + cell identity + spatial location + morphology + disease state + treatment outcome

That’s dramatically richer.

10x has already announced partnerships designed specifically around creating AI-ready biological datasets, including collaborations with Cancer Research Institute, PharosAI and Bioptimus.

The Bioptimus collaboration is particularly interesting.

They’re working on STELA, a spatial-data-generation initiative intended to create foundational datasets connecting biological information with disease outcomes.

This creates a potentially powerful flywheel:

More instruments

↓

More experiments

↓

More biological data

↓

Better AI models

↓

More demand for high-quality spatial data

↓

More experiments

That’s the part of TXG I think is much more interesting than simply “life sciences tools.”

6. The potential moat

TXG has several layers of moat.

① Installed base

Once a research lab has invested in an instrument, switching isn’t trivial.

Researchers develop:

• protocols

• workflows

• expertise

• software pipelines

• historical datasets

around the platform.

② Consumables lock-in

The instruments drive recurring reagent consumption.

That’s the business model I like most.

③ Dataset moat

This could eventually become much more important.

10x says more than 10,000 peer-reviewed papers have been generated using its technology.

That creates an enormous installed knowledge base.

④ Ecosystem

Researchers aren’t buying an isolated machine.

They’re buying:

instrument + reagents + software + workflow + scientific ecosystem

That’s much harder to replicate than just making another sequencing instrument.

7. But here’s the problem

TXG’s current financial profile is not impressive.

2025:

• Revenue: $643M

• Operating loss: $61M

• Net loss: $44M

• Stock compensation: $109M

And management’s original 2026 guidance was only:

$600–625M revenue

which implied roughly 0–4% growth excluding the 2025 litigation-related revenue.

That’s terrible for a company that historically traded as a high-growth biotech tools company.

So the market’s skepticism is completely understandable.

8. But there’s an important inflection

Q1 2026 was actually better beneath the headline.

Revenue was $150.8M, down 3% reported, but up 9% excluding the $16.8M prior-year one-time royalty benefit.

More importantly:

• Single-cell consumable reaction volumes: double-digit growth

• Spatial consumables revenue: double-digit growth

• Gross margin: 70%

• Cash + marketable securities: $540M

That’s exactly what I want to see before an Atera launch.

The legacy business doesn’t have to explode.

It just needs to stabilize while Atera starts creating incremental growth.

9. The bull case

Here’s how I would model the big upside scenario.

Phase 1 — 2026

Chromium stabilizes.

Spatial continues growing.

Atera launches.

Revenue growth remains mediocre because Atera is initially small.

Market still thinks TXG is a broken growth story.

Phase 2 — 2027

Atera starts getting meaningful adoption.

Labs begin purchasing systems.

Each system creates recurring consumables revenue.

Spatial becomes a much larger percentage of TXG’s revenue.

Now revenue growth potentially accelerates toward 15–25%+.

That’s when the stock can rerate.

Phase 3 — 2028–30

If Atera becomes the standard platform for large-scale spatial biology:

TXG could potentially become a:

$1.5–2B+ revenue company

with substantially better operating leverage.

If it gets to ~75% gross margins and eventually 20%+ operating margins, the economics change dramatically.

At that point you’re no longer valuing TXG like a struggling $600M life-science tools company.

You’re valuing it like a mission-critical biological data platform.

10. The really interesting upside scenario

The blue-sky thesis is not:

“Atera sells lots of machines.”

It’s:

Atera becomes the standardized data-generation layer for AI-driven biology.

That would be much bigger.

Think:

Biology → massive multimodal datasets → foundation models → drug discovery

TXG sits at the data-generation layer.

If AI makes biological data dramatically more valuable, the companies producing the highest-quality biological data could capture a disproportionate amount of the value.

That’s why I think the Bioptimus/STELA partnership is more important than it initially looks.

11. What could kill the thesis?

There are several serious risks.

🔴 Atera doesn’t deliver

This is #1.

If Atera is technically impressive but:

• expensive

• difficult to use

• slow

• difficult to scale

• produces insufficiently differentiated data

then the whole reacceleration thesis falls apart.

🔴 Pharma adoption takes longer

The biggest opportunity is ultimately pharma/biotech.

But if pharmaceutical companies don’t rapidly adopt spatial workflows, TAM expansion could take much longer.

🔴 Valuation

This is important.

TXG isn’t necessarily cheap just because its historical price collapsed.

If you’re paying a high multiple for $600M revenue with minimal growth, you are already paying for some Atera success.

12. My probability-weighted view

Read more

Here’s the timeline. Onsemi announces a $7 billion all-stock acquisition of Synaptics on June 25. Stock craters 13% the next morning, then keeps bleeding, down as much as 21% at the low. Law firms start circling with “investor alert” press releases. Analysts split into two camps, some slashing price targets to $110-125, others raising theirs to $130-140 on the same news. When Wall Street can’t agree on whether a deal is good or bad within 48 hours of the announcement, that’s not a verdict. That’s confusion. And confusion is where mispricing lives.

The consensus view, in three sentences: $ON is now adding ~12% dilution and integration risk from a $7 billion acquisition that pulls it into consumer and wireless categories it has no track record in. The Synaptics deal muddies a clean automotive/industrial story right when investors wanted simplicity. Near-term earnings visibility just got worse, not better.

What they’re missing is the only semiconductor company on the planet that just finished assembling the complete stack for physical AI, power, sensing, compute, and connectivity, under one roof, while simultaneously sitting on the widest fab utilization gap in its peer group and a design-win moat in the highest-value socket in the EV.

Let’s start with the part of the story the market already sort of understands, because even that part is underpriced.

The Architecture Lock-In Nobody’s Pricing Correctly

Roughly 55% of the new EV models shown at the Beijing Auto Show in April 2026 were running on $ON’s EliteSiC. That’s not a stat you get from being a good vendor. That’s a stat you get from owning the traction inverter, the single highest-value silicon carbide socket in the entire vehicle, at the exact moment the industry is jumping from 400V/800V to 900V architecture.

Once an OEM qualifies your SiC into a traction inverter, they’re not swapping suppliers for 18 to 24 months. Thermal cycling, humidity testing, power cycling, full vehicle-level validation. Geely, NIO, and Xiaomi are running $ON’s SiC in their 900V platforms right now, and switching costs mean that socket is locked in for the life of the platform.

Compare that to $STM, the current SiC revenue leader. Their book is built on 400V/800V European OEMs, Stellantis, Tesla, Hyundai, and they’ve been cutting SiC prices to defend share while the Chinese 900V cycle runs right past them. $WOLF is a substrate supplier still stabilizing post-bankruptcy. $IFX is a fast follower with a slower SiC ramp. None of them own the architecture transition the way $ON does.

And supply is already tightening. Some SiC lines are moving into allocation before the EV volume recovery has even started. That means pricing power on the way up, for the company that already won the sockets.

The AI Data Center Business Nobody Modeled a Year Ago

$ON’s AI data center revenue more than doubled year over year in Q1 2026, and grew over 30% sequentially, beating management’s own number. They’re guiding to double it again for full-year 2026.

Why? Because $ON is inside Nvidia’s MGX reference architecture, and not as a logo on a slide. They’re providing power supply units, battery backup units, 800 VDC power distribution boards, power FETs, multiphase solutions, SiC JFETs, and GaN converters, grid-to-core coverage across three semiconductor materials in a single reference design. When an OEM builds a server to the MGX spec, $ON’s power tree comes with it, automatically, across hundreds of system builders.

Here’s the line from CEO Hassane El-Khoury on the Q1 2026 call that I think the market skimmed past: AI data center revenue growth came from “broader adoption across the power tree with multiple chip vendors and leading hyperscalers.”

Multiple chip vendors. Not just Nvidia. $GOOGL TPUs, $AMZN Trainium, $MSFT Maia, $AMD’s MI-series, they all need the same 800V-to-48V-to-12V-to-6V conversion chain regardless of whose accelerator sits on the board. $ON is getting specified at the rack infrastructure level, which means the revenue isn’t tied to any single GPU vendor’s roadmap. That’s a very different risk profile than the market is currently assigning it.

On June 9, 2026, $ON launched GaNEXUS, its branded gallium nitride portfolio built specifically for the 800V-to-48V bus conversion stage inside these racks. Sampling started in the first half of 2026. Revenue ramp begins in 2027. As of today, that’s essentially zero dollars in any analyst model.

I haven’t even gotten to the part of this thesis that I think actually moves the stock.

Below is where it gets interesting, and where most of the Street stopped paying attention.

[Subscribe to keep reading →]

Read more

At Investor Day 2026, management doubled the non-handset revenue target from ~$20B to $40B, with data center alone exceeding $15B by FY29. Combined TAM across all segments reaches ~$1.7T by 2030.

The data center piece is what re-rates the stock. It changes both the earnings growth profile and the multiple. Simple math: every incremental $1B of data center revenue at 40%+ operating margin adds ~$0.40-0.50 to EPS. $15B of data center revenue means $6-7.50 of EPS from a segment that contributes zero today.

Why Inference, Why Qualcomm

Training is capex-heavy and lumpy. Inference is volume-driven, latency-sensitive, power-constrained, and it scales with actual AI usage. Every agentic interaction, every multi-step reasoning chain generates inference tokens. Hyperscalers care more about cost-per-token and watts-per-token than peak FLOPS now, because inference is the majority of AI compute spend and growing faster than training.

Qualcomm spent decades optimizing Hexagon NPU architecture for sub-5W mobile workloads. That design discipline now targets the real constraint in new data center builds: power density. Facilities hit power limits before compute limits. Qualcomm optimizes at the architecture level, not just the process node. NVIDIA can shrink transistors on the same schedule. Replicating 20+ years of ultra-low-power tradeoffs across the ISA, memory subsystem, and compiler stack takes longer than a node shrink.

Inference accelerator spend runs

$50-80B

annually by 2028, growing to

$150-200B+

by 2030. Qualcomm needs 8-10% share to hit . With production contracts from two of the three largest inference buyers, that share looks conservative.

HBC Changes the Economics

HBC (High Bandwidth Compute) is the technical differentiator. A near-memory compute architecture that places a purpose-built accelerator die underneath LPDDR stacks via TSVs, attacking the core inference bottleneck: data movement. During autoregressive decoding, the chip reads the full weight matrix from memory for every token. The operation is memory-bandwidth-bound. HBC puts compute where the data lives.

Claimed specs: 6x bandwidth per watt vs HBM, 200x capacity per watt vs SRAM. AI250 with HBC Gen 1 delivers 133 TB/s per card, 18x over AI200 with LPDDR5X. AI300 with HBC Gen 2 pushes to 54x, with 4x-8x better performance-per-watt vs GPU architectures. These are Qualcomm’s own numbers, but Satya Nadella endorsed HBC by name and confirmed Microsoft is deploying HBC-based XPU solutions inside Azure. CEO-level endorsement before the chip ships collapses design-win risk substantially.

HBC also attaches optionally to the C1000 CPU, extending efficiency gains into CPU-based inference. Nobody else has anything like this.

Modular Fills the Software Gap

Every previous Qualcomm server attempt failed the same way: good silicon, no software. The ~$4B all-stock Modular acquisition addresses this directly. Chris Lattner, creator of LLVM and Swift, runs the company. He has the strongest compiler pedigree in the industry and his team built much of today’s AI infrastructure tooling.

Mojo and MAX provide a portable runtime across CPU, GPU, NPU, and custom ASIC without rewrites. The key difference from ROCm and oneAPI: Modular abstracts underneath CUDA rather than replacing it. Developers don’t abandon their existing stack. It targets the inference serving layer, not the training framework, which is why prior CUDA alternatives failed and why this approach has a real shot.

Qualcomm is keeping Modular silicon-agnostic. Hyperscalers deploying on Qualcomm hardware don’t get locked in, which makes the first deployment decision easier. Once production inference is running on Qualcomm with Modular optimizations, switching costs build through performance tuning and operational integration. Modular also eliminates the 6-12 month software lag on new hardware, delivering day-zero optimization on fresh Qualcomm silicon.

Customers

Meta: launch customer for Dragonfly C1000 CPUs, production starting H2 2028. Zuckerberg personally quoted in the press release. Microsoft: multi-generation commitments including HBC solutions for Azure, Nadella calling HBC “an innovative architecture with high memory bandwidth and integrated compute that unlocks significant improvements in cost and performance for the next generation of AI infrastructure.”

Two CEO-level, public, multi-year production commitments from two of the three largest inference spenders. Meta buys CPUs for its server fleet. Microsoft deploys HBC inference accelerators. Different products, which means the customer base isn’t concentrated on a single product line.

35+ ecosystem partners (Supermicro, Micron, Arista, Quanta, Lenovo, Foxconn, SK hynix) did supply chain work over years, not weeks before the event. Custom silicon co-design is also available for hyperscalers that want differentiated hardware without building a chip team.

Google builds TPUs, Amazon builds Trainium/Inferentia. Both invest heavily in custom ASICs and are unlikely customers. The thesis depends on hyperscalers that buy rather than build, plus enterprise and sovereign AI where custom silicon isn’t practical.

Hardware Stack

Four product lines over 24 months with annual refresh cadence:

AI accelerators: AI200 ships 2026. AI250 with HBC Gen 1 samples mid-2027. AI300 with HBC Gen 2 samples 2028. AI300 supports UALink, the open alternative to NVIDIA’s NVLink.

C1000 CPU: 250+ custom Oryon cores, >5 GHz, >2x performance per watt vs existing server CPUs, PCIe Gen 7 (>2 TB/s), CXL, optional HBC attach. Three SKU variants targeting agentic orchestration, general-purpose cloud, and AI head node workloads. Production H2 2028.

Connectivity: 800G and 1.6T across optical, copper, campus-reach up to 20 km.

Edge Diversification Is the Second Story

Qualcomm is framing robotics and industrial AI as “Physical AI”, the same low-power inference architecture from mobile, now targeting factory floors, autonomous machines, and edge infrastructure. NVIDIA’s Jetson/Isaac is the incumbent, but it was designed top-down from GPU compute. Qualcomm built bottom-up from power constraints. Battery-powered robots, fanless edge boxes, thermally limited industrial enclosures, these are sub-5W design problems. Qualcomm has been solving sub-5W design problems for twenty years. The integrated AI + connectivity stack is differentiated in a way NVIDIA can’t replicate without rebuilding its power profile. $8B industrial/networking/robotics target by FY29.

Automotive design-win pipeline expanded to $65 B. Revenue target $10B by FY29. NVIDIA DRIVE competes, but Qualcomm already has broader OEM penetration through Snapdragon Ride and the digital chassis. Automotive design wins lock in for 5-7 year production cycles.

Management expects agentic AI to trigger a device upgrade cycle across phones, PCs, and IoT. Agents need on-device reasoning, persistent context, low-latency local inference. Current devices can’t run them. If agents become the primary interface layer, the installed base refreshes. Qualcomm owns the silicon in the majority of Android phones and is gaining share in Windows PCs. No one else captures both sides of that upgrade wave. The data center re-rating is the thesis. An agentic upgrade cycle protects and grows the base.

Catalysts

Provide a caption (optional)

H2 2026: AI200 racks ship. First benchmarks against NVIDIA H100/H200 on inference workloads.

Mid-2027: AI250 + HBC Gen 1 sampling. First independent data on HBC efficiency. Few Street models incorporate memory architecture as a variable. If benchmarks confirm even half the claimed advantage, the inference TCO narrative shifts hard.

H2 2028: Meta C1000 production. AI300 + HBC Gen 2 sampling. Revenue starts showing up in quarterly results.

FY2029: >$18 Non-GAAP EPS. Handsets ~one-third of QCT. The multiple should re-rate from handset cyclicality discount toward data center growth premium.

Risks

If AI200 benchmarks disappoint on real workloads, the customer pipeline slows. Modular integration mid-ramp adds complexity, though Lattner’s record building LLVM and Swift from scratch helps. NVIDIA could narrow the power gap with Blackwell Ultra, but Qualcomm doesn’t need to beat NVIDIA everywhere, just win on TCO for inference-heavy deployments. Google and Amazon build their own silicon and are unlikely buyers. The all-stock Modular deal adds ~2-3% dilution. HBC performance claims are unverified until mid-2027 benchmarks.

These are execution risks. But the architecture, the customers, the software platform, and the developer channel have arrived at the same time for the first time. Prior data center attempts always had one or more of those pieces missing.

I’m a Buyer

Qualcomm at $197 is priced as a handset company running a data center experiment. The evidence from Investor Day says otherwise: doubled revenue targets, production contracts with Zuckerberg and Nadella attached, a full silicon-to-software stack shipping across four product lines, and the strongest compiler engineer alive building the inference runtime. The Street sees ~$15.26 EPS in FY29. Management sees >$18. The difference is whether you believe data center revenue reaches $15B at 60-70% gross margins. I do, because the customers are already signed, the power advantage is architectural not cyclical, and HBC gives Qualcomm a memory economics edge no one else is even attempting. If AI200 benchmarks land and HBC validates at mid-2027 sampling, the re-rating from handset discount to data center premium compresses fast. I’m long.

Overview

Soluna Holdings builds data centers adjacent to wind, solar, and hydroelectric assets, converting excess electricity into revenue streams such as Bitcoin mining, hosting, and emerging AI HPC workloads. By situating demand directly at the power source, Soluna enables plant operators to capture value from otherwise wasted capacity while securing some of the lowest-cost energy in the market for its own operations.

• Clean-energy positioning & pipeline — 2.8 GW pipeline and ~1 GW milestone matters in the energy-constrained/ESG-sensitive AI and crypto hosting market; long-term demand for renewably powered compute is real.

• Anchor customers / institutional deployment — Galaxy Digital and BitMine hosting customers help de-risk individual projects and validate the model (Galaxy’s commitment at Kati is notable).

• Capital-light growth model — Leveraging joint-venture structures and project-level financing, external partners fund the majority of construction and equipment costs, minimizing corporate capital needs while preserving upside via post-flip cash-flow participation.

• Management focus / strategic refocus — they voluntarily exited/moved away from the GPU-as-a-service model to their core behind-the-meter hosting model.

Origins

Before its 2021 transformation, Soluna Holdings operated for decades as Mechanical Technology, Inc. (MTI), a small publicly listed engineering firm known for precision-measurement and instrumentation products.

MTI Instruments, its principal subsidiary, supplied vibration analysis, balancing, and calibration systems to aerospace, semiconductor, and industrial customers—a steady but low-growth niche.

In the late 2010s the company began funding Soluna Computing, a venture to colocate renewable power with cryptocurrency mining and other high-density computing, and gradually shifted its strategic focus toward that higher-growth opportunity. By early 2021 MTI’s legacy instrumentation business had become secondary, prompting a corporate restructuring, the formation of a new Nevada holding company, and a rebrand to Soluna Holdings, Inc.

Business model

Soluna’s core model assembles four ingredients:

• Power — long-term access to renewable generation or curtailed energy.

• Site & build — develop modular “pods” or phases of data-center capacity adjacent to power.

• Customers — revenue from (a) Bitcoin hosting (hosting customers’ miners), (b) colocation contracts for AI/HPC customers, and (c) grid services (demand response, ancillary services).

• Project financing & waterfall — use third-party project capital and take management/development fees, while retaining some equity upside in projects.

They run a mixed revenue approach: some revenue is hosting-contract driven (customer pays to colocate equipment), some is spot or revenue-share (e.g., mining operations), and some derives from grid services.

Soluna’s “Flip” Joint-Venture Model

Soluna develops modular data-center projects—but it does not finance the entire build itself.
Instead, each site is placed in a separate project company (LLC) that Soluna owns about 25 % of.
The remaining ≈75 % is funded by a capital partner such as a renewable-infrastructure fund or private investor.

Step 1 – Pre-Flip Phase

• The outside partner supplies most of the construction equity and arranges any project-level debt.

• Revenue from hosting/mining begins to flow once the site is energized.

• Soluna earns:

  • Operations & Maintenance (O&M) margin (target ~20 %).

  • Developer profit of about 7.5 % of project cash flow.

The partner keeps the majority of cash until it reaches an agreed target return (often a set IRR or cash-on-cash multiple).

Step 2 – Post-Flip Phase

• When the partner’s target return is achieved, the economics “flip.”

• Soluna’s developer-profit share rises from 7.5 % to 50 % of all project cash flow, while its equity ownership remains 25 %.

• The project may still carry debt; repayment follows its own schedule and does not affect the flip trigger.

  

Why It Works

• Capital-light growth: Soluna can build multiple sites without funding the full capex.

• Risk-sharing: The partner takes most of the upfront construction cost and power-price risk.

• Upside leverage: Once flipped, Soluna captures a much larger share of free cash flow despite its minority equity stake.

Soluna states IRR hurdle rate is reached in typically 3-4 years, at which point the flip happens and the expected developer’s profit goes from about 7.5% to 50%.

Soluna pivot from GPU-as-a-Service to Pure Colocation

In 2025, Soluna had pursued GPUaaS through a partnership with Hewlett Packard Enterprise (HPE) but Soluna made a deliberate strategic decision to exit the GPU-as-a-Service (GPUaaS) model and refocus entirely on its core behind-the-meter colocation business.

The HPE arrangement was meant to let Soluna offer GPU-as-a-Service without building its own hardware stack from scratch.

Key points:

• Scope and size – A $34 million, 36-month agreement.

• Infrastructure – Hewlett Packard Enterprise would supply and manage liquid-cooled servers equipped with NVIDIA H100 GPUs, housed inside a Soluna data-center site.

• Service model – Soluna would market the compute power as a service to AI customers (for training, fine-tuning, and running large language models) powered by nearly 100 % renewable energy.

• Revenue sharing – HPE provided the hardware and operations; Soluna provided the site, renewable power, and customer sales, and the two companies would share the revenue from GPU rentals.

Shifting market dynamics—specifically, customer demand for larger GPU clusters and the upcoming NVIDIA H200 Blackwell architecture—led Soluna to terminate the agreement. This experience reinforced the company’s strategic focus on colocation, highlighting the operational and capital efficiencies of hosting third-party GPU workloads without owning the hardware.

The shift carries several positive implications:

1. Streamlined Operations & Reduced Complexity
GPUaaS required Soluna to manage large inventories of specialized hardware, navigate fluctuating AI workloads, and absorb operational risk from tenant GPU utilization. By moving to a pure colocation model, Soluna simplifies operations, reduces capital tied up in proprietary GPUs, and minimizes exposure to volatile AI compute markets.

2. Capital Efficiency & Risk Mitigation
Colocation allows Soluna to leverage project-level financing and joint-venture structures without committing to full equipment purchases. This capital-light approach limits corporate balance-sheet risk while maintaining upside via management fees and equity participation in each project. The company can now deploy more megawatts across multiple sites without increasing corporate leverage.

3. Strategic Flexibility for AI Opportunities
While Soluna no longer owns GPU infrastructure, its colocation model remains fully compatible with AI workloads. The company can host GPU-heavy tenants on a “bring-your-own-hardware” basis, backed by low-cost, flexible power and scalable infrastructure. This approach preserves the opportunity to participate in high-margin AI deals without the upfront risk and complexity of GPU ownership.

4. Focus on Core Strengths
By exiting GPUaaS, Soluna doubles down on its differentiator: building modular, behind-the-meter data centers adjacent to renewable energy assets. This focus strengthens operational excellence, reinforces credibility with capital partners, and accelerates the company’s ability to scale its 2.8 GW pipeline efficiently.

Financials

Revenue

Soluna’s revenue in Q2 2025 was $6.2 million, about $3.5 million less than the same quarter last year.
The decline came from four main areas:

1. Bitcoin economics – roughly $2.0 million of the decline came from the Bitcoin “halving” and lower hashprice, which reduced mining revenue.

2. Business-model shift – Soluna moved more customers to a profit-sharing hosting model (from fixed-fee hosting). While this lowered reported revenue, it also cut costs by about $0.8 million, leaving gross profit largely unchanged.

3. Lower proprietary mining – Soluna’s own mining output dropped due to site downtime and less efficient miners, reducing revenue by $0.6 million.

4. Grid-service revenue – Demand-response payments from the Texas ERCOT market fell by $0.1 million as more participants shared the pool.

Gross Profit by Quarter

Adjusted Ebitda per Quarter

Cash Reserves

Quarter-end unrestricted cash was $9.9 million, a $2.0 million rise since December 31, 2024.

Debt & Financing Overview:

As of Q2, Soluna Holdings had total debt of about $23.3 million, split between corporate-guaranteed and subsidiary/project-level obligations.

• GreenCloud debt – $9.1M is guaranteed by Soluna Holdings and amortizes over roughly two years.

• Subsidiary/project debt – $14.2M includes:

  • Spring Lane Equipment – $0.5M for Project Kati/Dorothy 2, non-amortizing and payable upon project completion.

  • Galaxy Digital – $4.5M, amortizing over five years and guaranteed by a Soluna subsidiary.

  • NYDIG – $9.2M, non-amortizing and tied to specific project structures, not corporate assets.

Key takeaway: Most of Soluna’s corporate-level exposure is limited to the GreenCloud note, while the rest of the debt is project-specific, repaid from project cash flows, keeping the overall corporate debt risk moderate.

Projects and Pipeline

Pipeline Scale:

• Total Long-Term Pipeline: 22 projects totaling 2,801 MW, representing Soluna’s long-term growth potential in the renewable-powered compute and hosting market.

• Curtailment Assessment Completed (YTD): All 22 projects (2,801 MW) have undergone initial curtailment and feasibility studies, indicating readiness for development.

Active Deals & Ready-to-Build Projects:

• Active Term Sheet Negotiations: 5 projects (~850 MW) are in active discussions with potential partners or anchor tenants.

• Shovel Ready / PPA or Signed Term Sheets: 6 projects (~698 MW) are “shovel-ready,” with power agreements or land secured.

Pipeline Development Stage Breakdown:

• Stage 1 – Curtailment Assessment: 2,103 MW

• Stage 2 – Term Sheet Signed: 75 MW

• Stage 3 – PPA & Land Acquisition: 220 MW

• Stage 4 – Engineering / Feasibility / Regulatory: 187 MW

• Stage 5 – ERCOT/ISO Planning Completed: — (number included in stage totals)

• Stage 6 – Shovel Ready: 166 MW (88 MW of that started construction as of September 11, 2025)

• Stage 7 – Capital Formation & Customer Development: 50 MW

• Stage 8 – In Construction: 50 MW

• Stage 9 – Energized / COD: 75 MW (current operating hosting and mining projects)

Project Kati – 166 MW Flagship Campus

Project Kati is Soluna’s largest development to date and the company’s primary near-term growth driver. Project Kati is a 166 MW renewable-powered campus in Texas designed for high-density Bitcoin hosting and future AI/HPC colocation. The site sits adjacent to curtailed wind generation in ERCOT, giving Soluna access to some of the lowest-cost power in the market.

Kati 1 – 83 MW (Bitcoin Hosting)

• Capacity: 83 MW of behind-the-meter clean compute.

• Anchor Customer: Galaxy Digital has contracted for 48 MW, relocating part of its mining fleet after converting its Helios site for AI/HPC.

• Soluna Allocation: The remaining 35 MW will support Soluna third-party Bitcoin hosting.

• Timeline: Site work began in late 2025 with energization targeted for early 2026. Financing is largely secured through GreenCloud and Spring Lane Capital under Soluna’s flip-JV model.

Kati 2 – 83 MW (AI & HPC Infrastructure)

• Capacity: An additional 83 MW, bringing the campus total to 166 MW.

• Purpose: Purpose-built for AI and HPC colocation, with electrical and cooling design ready for GPU-dense deployments supplied by tenants.

• Timeline: Permitting and interconnect approvals are underway, with construction expected to begin in late 2026 and first power anticipated mid-to-late 2027, contingent on customer commitments.

Partners

Soluna Holdings has strategically aligned itself with key partners to scale its renewable-powered compute infrastructure. These collaborations not only mitigate development risks but also validate the company’s business model in the competitive landscape of green data centers.

Generate Capital

In September 2025, Soluna secured a scalable credit facility of up to $100 million from Generate Capital, a leading infrastructure investment firm. The initial $12.6 million draw is allocated for refinancing and construction of active data center projects, including continued funding for Project Dorothy 2 and support for Project Kati 1. An additional uncommitted $64.5 million accordion facility would support Soluna’s growing renewable-powered pipeline and fund AI-related long-lead equipment procurement. Generate Capital also received warrants for the right to purchase 4 million shares of Soluna common stock and a board observer right, solidifying their role as a strategic partner.

Galaxy Digital

Galaxy Digital’s 48 MW hosting deal at Soluna’s Project Kati 1 stems from a broader strategic pivot: in 2025 Galaxy began converting its own Helios campus in Texas from a Bitcoin-mining facility into a high-margin AI/HPC data center, freeing capacity for new compute tenants such as CoreWeave and requiring it to relocate a large portion of its mining fleet. Rather than shut down profitable hashpower, Galaxy chose to move those miners to Soluna’s behind-the-meter, renewable-powered site, where low-cost curtailed energy and a capital-light hosting structure keep operating costs competitive and reduce the carbon footprint. The companies were already linked through a Galaxy loan to Soluna earlier in 2025, so the relocation both deepens that financial relationship and gives Soluna its largest single customer to date while letting Galaxy preserve its mining business and redeploy Helios for next-generation AI workloads.

BitMine Immersion Technologies

The partnership began with the deployment of 3 MW at Project Sophie, marking a significant milestone for both companies as they demonstrated the viability of colocated, renewable-driven compute infrastructure. Building on this early success, BitMine has significantly expanded its presence within Soluna’s network, committing to a larger-scale deployment of 13 MW at Project Dorothy 2, which is currently under construction.

Compass Mining

Compass Mining has continued to strengthen its partnership with Soluna Holdings. Initially deployed at Project Dorothy 1, Compass Mining expanded its footprint in 2025 with an additional 8 MW at the adjacent Project Dorothy 2, bringing its total contracted capacity across the Dorothy campus to approximately 13 MW. This expansion underscores Compass Mining’s confidence in Soluna’s ability to provide scalable and sustainable hosting solutions.

GreenCloud

GreenCloud has been instrumental in financing Soluna’s projects, including a $20 million investment for Project Kati 1. This partnership facilitates the construction of the first 35 MW of the 83 MW phase, with plans to extend up to $100 million in additional capital.

Spring Lane Capital

The $20 million funding from Spring Lane Capital is structured to fully cover Project Kati 1’s construction and working capital needs. The financing follows a “waterfall” model, allowing Soluna to benefit from enhanced management and development fees during both the construction and operational phases. This approach ensures income generation throughout the project’s lifecycle.

Blockware Solutions

Blockware Solutions has recently expanded its operations at Soluna’s Project Dorothy 2 data center in Texas, increasing its hosting capacity from an initial 5 MW to 8 MW. This expansion follows a two-year hosting agreement signed in April 2025, under which Blockware deployed approximately 1,500 of its next-generation S21+ mining units at the site. The deployment began on April 21, 2025, with operations ramping on June 2025.

This partnership is a colocation arrangement, where Blockware hosts its mining equipment within Soluna’s facility. Soluna provides full-stack infrastructure support, including power management, internet connectivity, and on-site security, creating a sustainable and efficient environment for high-performance Bitcoin mining

Navitas Global

Soluna Holdings and Navitas Global have formed a strategic partnership to jointly develop and operate Project Dorothy 1B, a 25-megawatt (MW) Bitcoin mining facility located in Texas. This collaboration was formalized in May 2023 through a $14 million investment agreement, comprising a $2 million loan and a $12 million equity investment from Navitas Global. In exchange, Navitas acquired a 49% equity stake in the project, while Soluna retained a 51% ownership interest and is responsible for operations and maintenance.

Dilution: Prospectus $87M ATM

Soluna Holdings maintains an at-the-market (ATM) equity program with a prospectus ceiling of $87 million, giving the company flexibility to raise capital by selling shares directly into the open market. The ATM structure allows opportunistic capital raises to fund project-level growth, working capital, or strategic initiatives such as AI-capable infrastructure, without relying heavily on debt or large, one-time equity offerings.

Recent $4 Million Sale

On Friday, Soluna executed a $4 million ATM sale, a modest draw against the $87 million capacity. Approximately 1.5 million shares were sold for $2.75, based on the closing stock price on that date.

Takeaway: The sale provides the company with immediate resources to execute on its upcoming pipeline while taking advantage of the recent rally in the stock price. The modest ATM issuance is immaterial if Soluna delivers on its 2.8 GW pipeline—successful project execution should grow revenue and cash flow faster than the increase in share count.”

Closing Thoughts

1) “Megawatts matter” — AI is driving structural demand for power

• Independent energy agencies and sector reports project data-center electricity demand to grow sharply over the next decade, with AI the largest incremental driver; the IEA projects data-center demand to more than double by 2030 and identifies AI as a major contributor. High-density AI racks and new GPU architectures push per-rack power toward MW-class needs. This creates a sustained market for behind-the-meter low-cost MW capacity.

• Credible ESG angle: Soluna supplying MW to anchor customers (Galaxy, BitMine, Compass, Blockware) validate the model: Galaxy’s 48 MW commitment to Project Kati and BitMine/Compass deployments show commercial demand for renewably powered hosting. Anchor tenants reduce offtake risk and make project financing easier.

2) The plausible probability of an AI colocation deal

• Demand profile: AI requires contiguous blocks of high-density MW and low-cost, flexible power — the exact product Soluna builds. Industry players and chip vendors are placing very large bets on new data-center capacity

• Financing & strategy: Generate Capital’s $100M scalable facility explicitly includes language supporting AI-related long-lead equipment procurement for Soluna projects — that’s capital earmarked to make AI deals possible.

• Execution precedent: Soluna already signs anchor offtakes (Galaxy 48 MW) and hosts multiple mining/hosting customers; that shows the commercial team can close and operate partnerships. An AI anchor is a plausible next step for a shovel-ready phase that can be configured for GPU racks.

3) Colocation is scalable, flexible, and capital-efficient

• Capital-light, low-risk model — Unlike GPUaaS providers who must purchase, maintain, and refresh expensive GPU hardware, Soluna avoids inventory and obsolescence risk. Tenants fund the GPUs themselves, freeing Soluna’s balance sheet for additional project development.

• Proven commercial precedent — Galaxy Digital’s partnership at Helios with Coreweave illustrates the colocation model’s advantages in practice. Galaxy is converting its Helios campus to support to AI/HPC, while Coreweave supplies the GPUs.

• “Flip” Joint-Venture Model: The flip structure gives Soluna a repeatable template for scaling. Because each project sits in its own special-purpose LLC, the parent company avoids cross-collateralizing debt and can recycle its limited corporate capital across many builds. The model also creates flexibility in exit or monetization: once a project has flipped, Soluna’s developer-profit share rises from roughly 7.5 % to a full 50 % of all project free cash flow, while it still owns only about 25 % of the equity. Soluna can choose to retain its 25 % equity stake for ongoing income or sell that stake to recycle capital into the next site.

Disclaimer: This article is for entertainment and informational purposes only and does not constitute financial advice. It is not a recommendation to buy, sell, or hold any securities. Readers should conduct their own research or consult a licensed financial advisor before making investment decisions.