Cango’s latest AI milestone covers just 6% of its Georgia mining site. That small footprint may be the most useful part of the story. The company has completed infrastructure modifications supporting up to 3 MW of AI capacity within its 50 MW facility. For miners considering AI, this offers a practical starting point: qualify a portion of the site, test customer demand and operating economics, and let the results determine whether to expand. Cango’s Georgia rollout shows how miners can test AI in stages. But a small deployment is not the same as a small financial risk.
The AI Story Worth Watching is Only 3 MW
Cango’s Route to AI
A Pilot is a Different Decision from a Full-Site Conversion
Build the Business in Layers, not Just the Infrastructure
Reuse the Asset, not its Assumptions
Make the Next Megawatt Earn its Place – Premium Insights
When Staying with Mining is the Stronger Decision – Premium Insights
FBOX’s 1.25MW and 10MW AI data center solutions combine direct-to-chip liquid cooling with hybrid air-side cooling systems to meet the thermal demands of modern AI and HPC workloads. Designed for rack densities exceeding 150kW, FBOX delivers efficient, reliable cooling infrastructure built for high-density compute environments.
The AI Story Worth Watching is Only 3 MW
For a Bitcoin miner evaluating AI, the most useful case study may not be another billion-dollar hyperscale lease. It may be a few containers beside an existing mine. On September 3, Cango said its EcoHash subsidiary had completed infrastructure modifications supporting up to 3 MW at its 50 MW Georgia facility. An initial batch of GPU servers was operating, with some already providing commercial compute services.
That is infrastructure capacity equivalent to 6% of the site, not confirmation that 3 MW of GPUs is fully installed, occupied or profitable. A distinction that matters. Cango has demonstrated an operating deployment, not yet a publicly proven investment return. Nevertheless, it offers a useful starting point for miners: AI integration need not begin with abandoning mining or underwriting a full-campus transformation. It can begin with a defined project that earns the right to expand.
40% MORE HASHRATE. SAME HARDWARE 🤯
GDA tested BraiinsOS against stock firmware at their 40MW facility. Same machines, same conditions. The only difference was software.
Braiins OS averaged ~99 TH/s per miner vs ~70 TH/s on stock. No new ASICs, no extra capex. The full case study breaks down exactly how 👇
Cango’s Route to AI
Cango’s route was not a seamless progression from mining success to AI expansion. It combined greater infrastructure control with a difficult financial reset.
In August 2025, the company acquired the operating 50 MW Georgia mine for $19.5 million, moving toward owning facilities rather than relying entirely on third-party hosting. Site ownership created a platform for subsequent development. The acquisition price, however, was for a mining facility not an AI-ready campus.
Mining economics then deteriorated. Cango reported a $452.8 million net loss for 2025; impairment and Bitcoin valuation effects contributed materially, so this was not simply an equivalent cash operating loss. In February 2026, average operating hashrate fell to 34.55 EH/s against 50 EH/s deployed as Cango optimized equipment, relocated operations and renegotiated hosting.
The balance-sheet response was substantial. On February 9, Cango disclosed selling 4,451 BTC for approximately $305 million. All proceeds repaid part of a Bitcoin-collateralized loan. This reduced leverage and supported its ability to pursue AI, it was not a $305 million GPU purchase. Three days later, it announced a closed $10.5 million equity investment and agreements for another $65 million, then still subject to closing conditions.
April brought EcoHash’s public launch and a plan for Georgia to demonstrate modular inference infrastructure. Cango’s AI conversion had finished in early July, GPUs were arriving in batches, and in September the company announced operating GPU services.
The lesson is twofold: phased development can limit a project’s initial exposure, but Cango’s wider transition also involved losses, deleveraging and new equity. Its mining contraction should not be mistaken for megawatts directly converted into AI.
A Pilot is a Different Decision from a Full-Site Conversion
A contained deployment lets an operator test customer demand, technical performance and operating capability before committing additional land and power. The objective is not merely to switch GPUs on. It is to learn whether a repeatable business exists.
Cango’s September release describes modules installed alongside mining infrastructure, supported by transformers, precision cooling and fiber. It does not disclose project CAPEX, GPU specifications, contracted utilization, customer pricing or site-level returns. Those omissions prevent a credible payback calculation.
It is important to understand that modular does not always mean inexpensive. GPU procurement, resilient infrastructure and specialist staffing can create substantial exposure even when the footprint occupies a small share of a mining facility.
For other operators, the right pilot size is therefore not automatically 3 MW. It is the smallest commercially credible deployment that tests the intended customer offering without jeopardizing the existing business.
Build the Business in Layers, Not Just the Infrastructure
Our framework for applying this case has four stages, this is not a claim that Cango has already passed every test.
Qualification: Identify the customer and workload first. Establish whether power conditions, connectivity, cooling and location can meet that customer’s requirements. A broad forecast for inference growth is not evidence of demand for your particular site.
Pilot: Bound the budget, customer offering and technical scope. Set an acceptance test and a maximum loss before ordering equipment. Protect the mining operation from construction disruption and shared-system failures.
Commercial validation: Measure paid utilization, realized pricing, service performance, support workload and cash collections. An expression of interest is not a contract, and a powered-on server is not a billable server.
Expansion: Release the next tranche of capital only when observed performance and credible customer commitments justify it. Standardize what worked, correct what failed, and preserve the option to stop.
Cango plans to offer both bare-metal GPU services and colocation. Its disclosures suggest that Cango will operate GPUs under the bare-metal model, while colocation customers may provide their own hardware. The distinction matters because GPU ownership determines who supplies the capital and carries the utilization and obsolescence risk. Cango has not yet disclosed the exact ownership and financing structure of its GPU fleet.
Reuse the Asset, not its Assumptions
Start the engineering review with what might retain value: land, access roads, civil works, existing grid infrastructure and site operations. Then verify suitability component by component. Ownership alone does not make a transformer, permit or power agreement suitable for the new use.
Budget explicitly for workload-specific distribution, backup and redundancy, cooling, network routes, security and operational support. Specify these against the chosen hardware and customer service commitments rather than a generic “AI-ready” checklist.
Existing staff bring valuable electrical and site-management experience. Customer-facing compute adds responsibilities around provisioning, network performance, access controls, incident response and service reporting. Decide which capabilities to build and which to obtain through a partner.
Premium: In the next section, we examine the investment tests that should govern expansion and the sites where even a phased AI deployment is the wrong move.
