Written by: Xiaobing

The mining companies have just wrapped up their earnings season. MARA's quarterly revenue decreased by 27% year-on-year, CleanSpark's decreased by 30.5%, and both recorded significant net losses. Meanwhile, both have been frequently mentioning their "digital infrastructure platform" and continue to invest in AI/HPC.

The narrative has been spoken for nearly two years, and the market is starting to develop antibodies.

MARA: 4.8 GW Power Pipeline, Near Zero AI Revenue

MARA released its Q2 report on August 6. Revenue was $174.9 million, a year-on-year decrease of 26.7%, lower than analysts' expectations of $209.4 million, missing by 16.1%. The net loss was $611.3 million, compared to a profit of $808.2 million the year before. The loss per share was $1.60, while Wall Street expected a profit of $0.35.

The main reason for the loss was the impairment of BTC fair value. During the quarter, 2,422 BTC were mined, and 2,213 BTC were sold at an average price of about $73,078, holding about 35,577 BTC. Total assets shrank from $7.3 billion at the end of 2025 to $4.3 billion, and digital assets dropped from $4.7 billion to $2.1 billion. Operationally, the activated computing power was 70.3 EH/s, a year-on-year increase of 22%, with a network share rising to 5.9%.

In terms of AI/HPC transformation, MARA is simultaneously advancing in three directions: co-building data centers with Starwood Capital (recent target of 1 GW, long-term target of 2.5 GW, currently in the permitting and tenant negotiation phase, with no signed customers yet); acquiring the French HPC operator Exaion (completed, but expected annual revenue is only "low eight-digit," which translates to a few million to ten million dollars); acquiring Long Ridge Energy & Power (enterprise value of $1.5 billion, including a 505 MW gas power plant, annualized EBITDA of $144 million, initial target of 200 MW AI load, planned delivery in mid-2028, acquisition awaiting approval).

Together, these three paths result in a claimed MARA power pipeline of 4.8 GW. However, as of now, AI/HPC's contribution to revenue is nearly zero. Management's statements on the conference call were candid: the first half was about "expansion and transformation," while the second half focuses on "signing customers, launching assets, and demonstrating profitability."

No money has been made yet; the framework has just been set up.

CleanSpark: 100% Mining Revenue, But $6.6 Billion Lease Changes Valuation Logic

CleanSpark also released its Q3 report for fiscal year 2026 (ending June 30). Revenue was $138 million, a year-on-year decrease of 30.5%. The net loss was $239.8 million, compared to a profit of $257.4 million the year before. Adjusted EBITDA switched from a positive $377.7 million to a negative $113 million.

A key fact: 100% of this quarter's revenue came from Bitcoin mining, with zero AI/HPC revenue.

However, on July 14, CleanSpark signed a contract that could rewrite the company's valuation: a 20-year triple net lease with a confidential "high investment-grade global technology company" to deploy 175 MW of critical IT load at the Sandersville site in Georgia, expected to begin delivery in Q4 2027. The total contract revenue is about $6.6 billion, with extension options potentially reaching $11.6 billion. Management stated that the NOI contribution rate is nearly 100%, with an average annual NOI of about $330 million.

The $6.6 billion contract revenue is a significant leap for a company with annual revenue of less than $600 million. But the construction cost is also high, at $10 to $12 million per MW, making the total investment for 175 MW about $1.75 billion to $2.1 billion, while the company currently has long-term liabilities of $1.8 billion. CleanSpark has completed equity financing and made advance payments for long-cycle equipment, moving forward with real cash, which is qualitatively different from transformation statements based on PPT.

The tenant has also signed an exclusive letter of intent for a Texas portfolio (718 acres, 885 MW), suggesting that if Sandersville is a prologue, Texas will be the main text.

Control Group: Core Scientific Has Passed the Inflection Point

In the same earnings season, Core Scientific provided a model of completed transformation.

Q2 revenue was $164.2 million, doubling year-on-year. Among this, AI/HPC co-location revenue was $136.7 million, a year-on-year increase of over 1190%, accounting for 83% of total revenue. Self-mining shrank to $21.5 million, accounting for only 17%. By mid-July, it had started billing for a 437 MW customer capacity, with annualized co-location revenue of about $635 million. In addition to securing a large contract of 529 MW with AMD, the total leasing portfolio reached 1.1 GW, with potential contract revenue exceeding $24 billion.

Core Scientific had a net loss of $1.16 billion in Q2, but the source of the loss is completely different from MARA; it was driven mainly by changes in the fair value of warrants (a non-cash valuation change of $1.05 billion), with adjusted EBITDA positive at $41.1 million, and operating cash flow positive at $230.9 million in the first half of the year.

From signing the first 200 MW contract with CoreWeave in June 2024 to achieving co-location revenue of $136.7 million in Q2, Core Scientific's conversion cycle is about two years.

Three Stages, Three Valuations

When placing the three companies together, the progress of mining companies' transformation to AI is very clear.

Core Scientific is in the "revenue realization phase."

83% of its revenue comes from AI/HPC, operating cash flow is positive, and the market has valued it as a data center operator.

CleanSpark is in the "contract lock-in phase." The $6.6 billion lease has been signed, funding is in place, but AI revenue is zero. Investors are betting on future cash flow beginning in Q4 2027, with valuation relying on discounting contract value.

MARA is in the "platform building phase." It has the largest pipeline of 4.8 GW, yet has no signed AI clients and no billable capacity, with valuation still primarily driven by BTC holdings.

To assess the real progress of mining companies' AI transformation, one should not look at the number of GW in the power pipeline, but rather at the number of billable MW. The disparity in value between MARA's 4.8 GW and Core Scientific's currently billable 437 MW is stark.

The underlying logic of mining companies' transformation is solid. The demand for AI computing power is growing much faster than the speed of electricity infrastructure construction; new substations in the U.S. have waiting times of 5 to 7 years. Mining companies hold large-scale power capacity connected to the grid, along with land and cooling infrastructure, which are precisely the most scarce resources for super-scale cloud providers. This is why Starwood is willing to cooperate with MARA and why tech giants are willing to sign 20-year contracts with CleanSpark.

Once the same kilowatt-hour is integrated with AI workloads, the revenue generated can reach several times that of mining; electricity assets are transitioning from being "priced by BTC" to "priced by computing power rent."

However, the total revenue of MARA and CleanSpark still relies on mining. BTC prices have been stagnant in the $64,000-$65,000 range, with production costs continuing to rise after the halving. MARA sold 2,213 BTC in Q2 to maintain operations, and Long Ridge used BTC to collateralize debt financing, which is equivalent to consuming Bitcoin assets to purchase computing power assets.

For investors, there is only one thing to care about: "Which company can truly convert electricity into AI revenue before cash flow runs out."