Marathon vs Riot: Bitcoin Mining’s Two Giants Bet on AI

For most of the past decade, Marathon and Riot competed on a single axis: who could aim more computing power at the Bitcoin network. In July 2026 that contest looks almost like a sideshow. Both companies now trade less on how many blocks they win and more on whether they can turn cheap, contracted electricity into artificial intelligence capacity before the math of pure Bitcoin mining squeezes them out.

The backdrop is unforgiving. Bitcoin changed hands near $63,000 on July 7, well below its October 2025 record above $126,000, and hashprice (the daily revenue a miner earns per unit of computing power) sits near $29 per petahash per day, close to the lowest level the industry has ever recorded, according to Hashrate Index. Shares of both MARA Holdings and Riot Platforms have been punished for it, and both have answered with the same two-letter word: AI.

This is a head-to-head look at the two largest US-listed Bitcoin miners: how they mine, what they hold, how badly the first quarter hurt, and which AI strategy looks more credible now that real capital is on the line.

The state of play in July 2026

Bitcoin’s total network hashrate has pushed to record highs near 963 exahashes per second (roughly 0.96 zettahashes), which means every miner is fighting for a thinner slice of a fixed block subsidy. Combine that with a Bitcoin price down more than 15% in June and you get the squeeze that CoinShares and other trackers flagged through the spring: miner profitability at or near all-time lows, with the least efficient machines switching off.

The equity market has drawn the obvious conclusion. By early July, Riot and MARA sat roughly 27% and 25% below their highs for the year, with MARA trading near $12 and Riot near $28. Benzinga reported the miners sliding in unison as the AI pivot that was supposed to rescue them ran into fresh questions about timing and demand. In short, the story that lifted these stocks is now the story that has to be proven — and by August, Riot had produced the clearest proof yet with its Anthropic contract.

How the two miners stack up

On raw scale, Marathon is the bigger machine. On discipline and early non-Bitcoin revenue, Riot has a case. The table below pulls the headline first-quarter numbers straight from MARA’s and Riot’s filings with the SEC.

Two caveats keep this from being apples to apples. MARA reports energized hashrate, which counts machines that are plugged in even if some sit idle, so its real output edge over Riot is narrower than 72 versus 42 suggests. And Riot’s higher share price reflects a smaller Bitcoin treasury and, until recently, less dilution, not obviously better operations.

Marathon’s playbook: scale and vertical integration

MARA finished the first quarter with 72.2 EH/s of energized hashrate, up 33% year over year, and about 1.9 GW of power across 18 data centers. CEO Fred Thiel has spent the past year recasting the company as a “vertically integrated digital infrastructure” business rather than a pure miner, and its balance-sheet moves back that up.

The centerpiece is energy ownership. On April 30, MARA agreed to buy Long Ridge Energy, owner of a 505 MW combined-cycle gas plant in Hannibal, Ohio, for about $1.5 billion, a deal CoinDesk framed as the company’s clearest AI data center push yet. MARA says the purchase lifts owned and operated capacity by roughly 65%, to about 2.2 GW, and hands it a site it can build into a flagship AI campus. Alongside that, a joint venture with Starwood Digital Ventures aims to deliver around 1 GW of near-term IT capacity.

In early August 2026, MARA widened the plan considerably, acquiring roughly 1,200 acres of powered land in Matagorda County, Texas, that it aims to build out to as much as 1 GW of grid capacity by October 2027 and 2 GW by April 2028. Taken together with the pending Long Ridge acquisition, which is still expected to close in the second half of 2026, MARA now frames its total power footprint as reaching about 4.8 GW — a step-change in scale, but one that still hinges on filling those gigawatts with paying tenants.

Thiel has been careful to add a discipline clause. MARA is following a build-to-suit model and will not pour capital into construction before a tenant signs, he told investors in May, noting that Google, Microsoft, AWS and Anthropic would each want different specifications. That caution matters, because it is also the crux of the bear case: unlike Riot, MARA has not yet announced an anchor hyperscaler contract at scale for its AI build-out.

Riot’s playbook: power first, activists pushing

Riot came into 2026 with a different weapon: land and power in Texas. Its Rockdale and Corsicana sites hold roughly 1.7 GW of capacity, and management has moved to monetize part of it outside Bitcoin. Deployed hashrate reached 42.5 EH/s in the first quarter, up 26%, while a new data center segment already contributed $33.2 million to revenue, its first material non-Bitcoin line.

At Corsicana, Riot filed a roughly $400 million permit for a new 335,430 square foot data center building, with construction running from April 2026 through 2028, as DataCenterDynamics documented. AMD has doubled its contracted capacity at the campus from 25 MW to 50 MW, and Riot has signed a memorandum of understanding with Terrestrial Energy to co-locate molten salt nuclear reactors, a framework that contemplates scaling to several gigawatts over time.

The loudest voice, though, belongs to an outsider. On February 18, activist investor Starboard Value published a letter arguing that Riot’s 1.7 GW could support premier data centers and throw off more than $1.6 billion in annual EBITDA, unlocking up to $21 billion in value, per The Block. Riot shares jumped about 7% on the news, CoinDesk reported, and the company later added three directors with data center conversion experience, among them Jaime Leverton.

By late summer the thesis had teeth. On Aug. 11, 2026, Riot disclosed its largest non-Bitcoin commitment yet: a 20-year, 191-megawatt AI compute agreement with Anthropic, reported at roughly $9.0 billion to $9.1 billion in base value and running through June 2048, with extension options that some coverage said could push the total to as much as $16.1 billion. It is the anchor hyperscaler contract the AI-pivot story had been missing, and it recasts a slice of Riot’s Texas power base as contracted, long-dated AI revenue rather than a hypothetical — the sort of signature Starboard had been demanding.

The balance-sheet reckoning

Both companies bled red ink in the first quarter, and both did it for the same accounting reason: fair-value marks on the Bitcoin they hold. MARA’s revenue fell 18% to $174.6 million, and it posted a net loss of $1.26 billion, or $3.31 per share, driven by roughly $1.0 billion in negative fair-value changes on its crypto, as Blockspace detailed. Riot’s revenue rose 2% to $167 million, yet it still lost about $500 million, or $1.44 per share, after a $326.7 million markdown on its own coins.

The pressure turned both into sellers of the asset they mine. MARA sold 15,133 BTC in March for roughly $1.1 billion, using the proceeds to buy back convertible notes at a discount and cut debt by about 30%, from $3.3 billion to $2.3 billion; across the first half of 2026 its sales reached about 23,093 BTC for roughly $1.6 billion at an average price of $70,631. Riot, which sold 3,778 BTC at an average of $76,626 in the first quarter, sold 9,665 BTC in total over H1. The much-loved HODL model, in which miners hoard every coin, is quietly bending under the weight of the cycle.

The AI bet: real pivot or narrative rescue?

Here is the uncomfortable truth for both. The pitch, that a power-rich mining shell can be reborn as an AI landlord, is not unique to them. Core Scientific, IREN and Applied Digital moved earlier, and Core Scientific’s colocation arrangement with a major AI cloud set the template that every miner now cites. Being late to a crowded trade is not fatal, but it does raise the bar for proof.

The moat, if there is one, is power. Contracted electricity at gigawatt scale is genuinely scarce, and grid interconnection queues run for years. That is the strongest part of both stories. The weakness is everything downstream of the substation: mining sheds are not AI-grade data centers. Liquid cooling, redundancy, low-latency networking and tenant-specific fit-outs cost real money and take real time, which is precisely why MARA’s build-to-suit rule and Riot’s phased Corsicana schedule exist. Riot has now converted a marquee hyperscaler lease — its 20-year Anthropic compute deal, disclosed in August — into contracted revenue at meaningful scale, the clearest proof point either miner has produced. Marathon, for all its land and generation, still has to sign an anchor tenant for the capacity it is building, and the market’s summer wobble shows investors want those signatures turned into cash flow.

Where the SEC fits in

Both miners are SEC registrants, which is why every figure above traces back to a 10-Q or 8-K rather than a press release. Two regulatory threads matter here. First, the securities status of mining itself has cleared up: in a March 20, 2025 staff statement, the SEC’s Division of Corporation Finance said that proof-of-work mining on public, permissionless networks does not, in its view, involve the offer or sale of securities under the Howey test, as summarized by law firm Fenwick. That removes an overhang, though the statement carries no legal force.

Second, the AI pivot reshapes what these companies must disclose. Riot now breaks out a separate data center segment and, after the Anthropic agreement, carries a multi-year contracted revenue backlog that investors will track closely; MARA’s forward-looking claims of roughly 4.8 GW and beyond are exactly the kind of statements the SEC watches closely when capital raises follow. And because both firms carry Bitcoin at fair value, every swing in the coin’s price runs straight through the income statement, turning quarterly results into a partial bet on BTC regardless of how the mining or AI lines perform.

What to watch next

• Tenants, and execution. Riot has cleared the highest bar with its 20-year, 191 MW Anthropic compute deal; the question now shifts to delivering that capacity on schedule. MARA’s Matagorda build-out and Starwood venture still need an anchor customer of their own to convert powered land into cash flow.
• The Long Ridge close. MARA’s Long Ridge Energy deal is still expected to complete in the second half of 2026, subject to regulatory approval; owning generation changes the cost math and underpins the roughly 4.8 GW footprint target.
• Starboard’s next move. The Anthropic deal validates the activist’s thesis; watch whether Starboard presses for more, from further conversions to capital returns, or declares victory.
• Hashprice and the 2028 halving. With the block subsidy at 3.125 BTC and the next halving due in 2028, mining margins only get tighter; every month of low hashprice strengthens the case for diversifying away from it.

The bottom line

Marathon brings scale, vertical integration and, after the Long Ridge deal, its own power generation, but it also carries the heavier Bitcoin exposure and the larger quarterly loss. Riot is smaller and slower, yet it has an earlier non-Bitcoin revenue line, a cleaner treasury and an activist lighting a fire under management. Strip away the branding and both are becoming power companies that happen to mine Bitcoin. Riot has now signed the first bankable AI tenant of the two — a 20-year Anthropic deal worth roughly $9 billion in base value that turns its Texas power into contracted revenue through 2048 — so the open question is whether it executes on time and whether Marathon can land an anchor of its own before its build-out gets ahead of its lease book. The summer selloff was the market’s way of demanding signatures, not slide decks; Riot has one, and the burden of proof now shifts to delivery.

This article is for information only and is not investment advice.

By the HOGE Wire Mining and Staking Desk.