Stop #314 - How mining paves the way for AI

“A watt used for intelligence can yield 15 to 25 times more than the same watt used for hashrate.”

Erik Hersman, co-founder and CEO of Gridless, summed up his company’s future this way. Gridless has just turned four, and to celebrate it announced its first steps toward an African network of small data centers powered by renewables, where Bitcoin mining and artificial intelligence compete for the same electricity.

Taken in isolation, the sentence sounds like a miner surrendering to AI. In the rest of the post, artificial intelligence captures the profitable watts and Bitcoin confirms its usefulness when it is no longer the best customer.

Hersman writes from Nairobi. The city exists, he says, because 130 years ago the British Empire decided to build the Uganda Railway through a region lacking economic activity capable of justifying its cost. Critics dubbed it the “Lunatic Line”: impossible costs to predict, undefined demand, uncertain destination.

Nairobi began as a depot along that railway and became the answer to all the questions its detractors considered unanswerable.

For Hersman, today’s tracks are data centers, chips, fiber, and energy. The cargo they carry is intelligence. This infrastructure, too, is being built before all the demand needed to pay for it exists, in the hope that available capacity will give rise to new uses.

A railway, however, must survive long enough to see passengers arrive.

More than 560 million people in sub-Saharan Africa still live without electricity, according to the latest United Nations report on the Sustainable Development Goal. Economically sustainable demand, especially in rural areas, grows much more slowly.

The Benchmarking Africa’s Minigrids 2024 report, built on data from 27 operators, measures average consumption of 16.6 kWh per user per month. Households remain below 7 kWh and generate average monthly revenues of $2 or $2.50, depending on the region. The overall average, including commercial activity, barely reaches $3.16 per customer.

A hydroelectric, solar, or geothermal plant requires capital upfront. Consumption from homes, shops, schools, and clinics arrives over the years. In between are underutilized machines, expensive batteries, currency risk, and tariffs the population struggles to afford. The river flows and the sun shines while the project remains a poor investment.

Gridless, as covered in stop #126, had started out to produce bitcoin using unused renewable energy. Less than a year later, Hersman realized mining could fill that gap. A container of ASICs can be installed next to the energy source, buy every unused watt, and shut down when the community needs it. Geographic location matters little, the required connection is modest, and the end customer is a global market for hashrate.

Bitcoin thus becomes an energy buyer of last resort. It guarantees a revenue base from day one, while local customers connect and begin consuming.

According to figures disclosed by the company, after four years Gridless operates six sites in three countries, has helped bring electricity to 30,000 more people than before its arrival, and has seen hundreds of commercial businesses emerge. In the villages it serves, energy prices are said to have fallen by one-third. These are company figures and should be read as such, but they describe a model that has moved beyond the PowerPoint presentation stage.

The move toward AI had been brewing for some time. In November 2025, Hersman had already introduced the “Emergent Grid”: renewable plants directly connected to modular containers with miners, storage systems, and GPUs. In that design, Bitcoin made it possible to build new energy capacity, while AI inference promised higher margins.

Gridless has now added figures, sites, and an order of priority.

The community comes first: homes, businesses, clinics, and schools. The second buyer is contracted inference, meaning the real-time execution of models for local customers or entities required to keep data in-country. Next come batch workloads, less urgent jobs that can be scheduled when energy is abundant. Mining absorbs everything left over.

Gridless claims that an AI container operating at roughly 4% utilization generates the same revenue as an equally powerful Bitcoin container running continuously. Each kWh sold for AI computing would therefore be worth 15 to 25 times as much as one allocated to hashrate.

The multiplier comes from the company’s internal model, and no public methodology is available to verify it. It also compares revenue generated by electricity rather than returns on invested capital. The cost of a GPU is, according to Hersman, around 100 times that of an ASIC. It also requires better cooling and connectivity and depreciates quickly. Without a customer, it sits idle in the middle of nowhere.

Gridless wants to separate the risks. GPUs remain with financial partners and plants with producers. Communities and AI customers buy energy and computing, respectively. Gridless provides the sites, integrates the energy, and operates the system that continuously shifts loads. ASICs monetize spare capacity without forcing the company to bet its balance sheet on thousands of GPUs.

Africa hosts nearly one-fifth of the world’s population and only 0.6% of global data center capacity, according to the Africa Data Centres Association. Gridless envisions a constellation of small and medium-sized nodes: regional hubs connected to fiber and edge sites powered by hydroelectric, solar, or geothermal energy, with terrestrial and satellite links.

Training the largest models remains the domain of campuses with large GPU clusters tied together by high-speed internal networks. Edge nodes, by contrast, can handle local inference, public workloads subject to data sovereignty requirements, and batch jobs coming from abroad. It is the same principle behind distributed cloud systems: bringing some computing closer to the data, users, or available energy.

In Kenya, Microsoft and G42 announced in 2024 an initiative with an initial $1 billion investment, including a data center powered by geothermal energy from Olkaria. The project has remained stuck in the planning stage. Bloomberg attributes the delay to the lack of agreement on capacity purchase guarantees requested from the government; Al Jazeera also reports community concerns over the potential impact on water resources. Gridless can proceed with separate installations, each financed and assessed independently.

It will start with a small solar site in the Democratic Republic of the Congo and a hydroelectric one in Kenya. Demonstration racks in Nairobi and a geothermal hub should follow. Each installation will cost between $500,000 and $6 million, be financed separately, and have to prove in the field the connectivity, cooling, reliability, AI demand, and economics of the mixed system.

Last July, CleanSpark, a company focused in recent years on Bitcoin mining, signed a 20-year agreement to provide 175 MW of capacity to a “global technology company” at its Sandersville, Georgia site. The stated value is $6.6 billion over the initial term, with possible extensions up to $11.6 billion.

In its quarterly filing with the SEC, CleanSpark writes that it will continue mining until the agreement takes effect. Initial deliveries are expected in the fourth quarter of 2027: at that point, the energy will be transferred to the AI campus and mining will cease. The physical decommissioning of existing infrastructure is expected to be completed during fiscal year 2028.

IREN made an even sharper choice. In November 2025, it signed an agreement with Microsoft worth around $9.7 billion for 200 MW of computing load and Nvidia GB300 GPUs, involving roughly $5.8 billion in investment in machinery and equipment. Last May, it added a five-year $3.4 billion agreement with Nvidia for AI capacity hosted in 60 MW of existing data centers.

To prepare for the conversion, IREN retired part of its mining hardware and recorded $140.4 million in impairments, largely related to those machines.

In both cases, capital is migrating from ASICs to GPUs.

Beyond the multiplier cited by Gridless, these contracts show the value of already-contracted energy. When a site also has land, grid connection, and operational expertise, the AI customer offers revenues large enough to justify multibillion-dollar investments.

AI pays more where it can reach. An inference service subject to a service-level agreement demands continuity, low latency, and predictable response times.

Mining can stop within moments, restart when energy returns, operate far from population centers, and sell its work into a global market without finding a customer for every site. Revenue per kWh is lower, but the hashrate market is always open.

At the sites described by Gridless, clinics, schools, and businesses have priority. AI takes the energy when it offers a more profitable use. Mining joins the back of the line and turns electricity that would otherwise be curtailed or remain without a buyer into revenue.

The thesis applies where energy is isolated, intermittent, or produced ahead of demand. On a congested grid, during hours when households and industry compete for the same capacity, mining must step aside. That is the feature that makes it useful: it can do so without losing customers, data, or an ongoing computing session.

AI’s arrival therefore scales back mining as the final destination for energy and strengthens its role as an economic foundation. Bitcoin makes the first watt financeable, absorbs fluctuations, and keeps buying as local and computing demand grow. When a better use appears, ASICs shut down.

The best customer arrives after fiber, cooling, and contracts.

Mining, on the other hand, is already there.