Marathon vs Riot: Two Bets on Bitcoin Mining’s AI Future

Twelve months ago, an investor comparing MARA Holdings and Riot Platforms would have struggled to separate them on anything but size. Both were large, US-listed Bitcoin miners running warehouses of application-specific chips across Texas and beyond. Both carried their freshly mined coins on the balance sheet, and both lived or died by the same two numbers: the price of Bitcoin and the cost of the power used to make it. In 2026 that symmetry is gone. The two companies have placed deliberately opposite bets on what a Bitcoin miner should become, and the market has begun pricing them under two different rulebooks.

The two miners that stopped being the same company

The backdrop is a brutal one for anyone whose only product is hashes. Bitcoin trades around $78,000, well below its October 2025 record near $126,000, according to Fortune’s daily price tracker. The reward for finding a block is still 3.125 BTC after the 2024 halving, and the network’s difficulty keeps climbing as more machines chase the same fixed issuance. That combination has squeezed mining margins to the point where the least efficient rigs run at a loss; by CoinShares’ math, roughly 15 to 20 percent of the global fleet was underwater even before this year, per its 2026 mining report. Both MARA and Riot reached for the same escape hatch: selling their electricity and their buildable real estate to the artificial-intelligence boom instead of only to the Bitcoin network.

How each is doing that, and how different the two look as a result, is the whole story. One is trying to own the power and stay long Bitcoin. The other is turning itself into a landlord with a two-decade lease and a shrinking coin pile. The Q2 2026 numbers came in near-identical at the top line and wildly different underneath, and an August selloff that hit both stocks even as Bitcoin climbed showed that the market has quietly re-classified what these companies are.

One note on names before going further. Marathon Digital Holdings rebranded to MARA Holdings in 2024, but plenty of traders still call it Marathon, and the ticker is still MARA. Riot Blockchain became Riot Platforms back in 2023. This article uses the current names and treats both as what they are today: capital-intensive infrastructure companies that happen to have started as miners.

From twins to opposites: how the split happened

The April 2024 halving cut the block subsidy from 6.25 to 3.125 BTC, and it did what every halving does: it took a knife to revenue per unit of hashrate. Hashprice, the industry shorthand for daily mining revenue per petahash, fell to multi-year lows in the low thirties of dollars per petahash per day, and the six-month forward market has been pricing it near $30.86, according to the Hashrate Index roundup. When the reward halves but the electricity bill does not, only two things save a miner: cheaper power or a second product to sell.

Both companies were sitting on two assets that the AI wave suddenly wanted: gigawatts of contracted power near transmission, and land already zoned and interconnected. Training and serving large models needs exactly that. So both pivoted, but they chose different sides of the same trade. MARA doubled down on owning generation and kept a huge Bitcoin treasury, betting that vertical integration plus coin upside would win. Riot went the other way, converting itself into a data-center developer that leases capacity to a single anchor tenant and deliberately runs light on Bitcoin. MARA chief executive Fred Thiel framed the stakes bluntly to CoinGeek: mining, he said, “is a zero-sum game. As more people add capacity, it gets harder for everybody else. Margins compress, and the floor is your energy cost.” His conclusion: “By 2028, you’ll either be a power generator, be owned by one, or be partnered with one.”

The pivot did not arrive as a single decision so much as a year-long change of identity. Through 2025 both companies still described themselves primarily as miners that might, someday, host other workloads. By early 2026 the language had flipped: data centers were the headline and mining was the legacy business paying the bills while the transition happened. The AI capital-spending boom did the rest, as hyperscalers and frontier labs raced to secure any site with power and a grid connection, and suddenly the megawatt, not the terahash, became the unit that mattered. The two companies read that shift the same way and drew opposite conclusions about how to capture it.

Reading the Q2 2026 scorecard

Start with the headline coincidence: for the June quarter both companies reported almost exactly the same revenue, near $174 million. Underneath, they could hardly be more different. MARA’s top line was essentially all mining. Riot’s included roughly $60.5 million of non-mining revenue, split between a data-center segment and an engineering business, the first time a material slice of its income came from something other than Bitcoin.

The losses tell the same story from a different angle. MARA’s $611.3 million net loss, or $1.60 a share, was driven mostly by an unrealized mark-to-market hit on its Bitcoin, filed with the SEC. Riot’s $237.2 million loss was smaller in absolute terms and far smaller relative to its coin pile, because there is far less coin to mark. Adjusted EBITDA, which strips out those paper swings, was negative $360.9 million at MARA versus negative $69.7 million at Riot. Same revenue, very different quality of earnings.

Marathon’s bet: own the power, keep the Bitcoin

MARA describes itself as vertically integrated digital infrastructure, and it means it literally. In late April it agreed to buy the Long Ridge Energy generation complex in Ohio, a 505-megawatt combined-cycle gas plant, for roughly $1.5 billion, giving it power it owns rather than power it rents. It has since raised its power portfolio target to about 4.8 gigawatts, and it runs a joint venture with Starwood Digital Ventures aimed at more than a gigawatt of near-term IT capacity. Thiel’s operating rule is build-to-suit: no heavy capital spending on a site until a tenant signs, which keeps the balance sheet flexible but also means the AI revenue is still mostly a promise.

The other half of the MARA bet is the Bitcoin itself. The company held 35,577 BTC at quarter end, worth roughly $2.1 billion, one of the largest corporate stacks in the world. That makes MARA part miner, part leveraged Bitcoin holder, which is exactly how some analysts value it. The missing piece is a marquee anchor tenant at scale, and its absence is the core of the bear case. The stock changed hands near $12.06 in late August, up about 32 percent on the week but still down around 23 percent on the year, per market coverage. Analysts are split: Cantor Fitzgerald trimmed its target to $12 while keeping an Overweight, Clear Street cut to $10 and a Hold, and Morgan Stanley lifted its target to $6 on the assumption that at least one HPC lease and two Starwood JV site leases land by year end.

The Starwood tie-up is central to how MARA hopes to close the gap. The joint venture with Starwood Digital Ventures is meant to supply the development expertise and capital a former miner lacks, with a near-term target of roughly a gigawatt of IT capacity and a longer path beyond 2.5 gigawatts, per its quarterly update. Thiel has said the venture is seeing HPC demand greater than first expected, but demand for space is not the same as a signed, funded lease, and until one lands the market is left valuing an option rather than a contract. That is the sharpest contrast with Riot, which turned the same kind of raw power into a binding twenty-year agreement.

Riot’s bet: become the landlord

Riot took the opposite path. Chief executive Jason Les began calling the company an active, revenue-generating data-center operator earlier in the year, and Q2 backed the label up: the data-center and engineering lines together produced that $60.5 million of non-mining revenue. Rather than hoard coins, Riot deliberately shrank its treasury to 11,380 BTC, some of it pledged as collateral, using coin sales to help fund construction. It is doing to its balance sheet the reverse of what MARA is doing: going Bitcoin-light on purpose.

The pivot did not happen in a vacuum. In February the activist investor Starboard Value pushed Riot publicly, arguing in a letter reported by CoinDesk that redirecting its roughly 1.7 gigawatts toward AI could generate more than $1.6 billion in annual EBITDA and be worth up to $21 billion, and it seeded the board with directors who had converted power sites into data centers before. Riot also signed AMD as a smaller tenant for 50 megawatts. The market has rewarded the shift: the shares traded near $20.88 in late August, up roughly 58 percent on the year even after some profit-taking, per Yahoo Finance, and post-deal price targets from Bernstein ($35) and Citi ($32) sit far above where MARA’s cluster.

The Anthropic lease, decoded

The centerpiece of Riot’s transformation, and the single event that most cleanly separates the two companies, is a 20-year lease at its Rockdale, Texas campus. The terms are large: 191 megawatts of critical IT capacity, running through June 2048, with roughly $9.1 billion of base contract revenue and up to about $16.1 billion if two five-year extension options are exercised. Riot has guided to cumulative net operating income of $7.3 to $8.2 billion over the life of the deal, an average of roughly $365 to $411 million a year, with 96 megawatts due to come online by December 2027 and the full 191 megawatts by June 2028. Morgan Stanley provided a $573 million interim facility to get construction moving.

There is one wrinkle worth stating carefully. Riot’s own filings name the tenant only as one of the world’s leading frontier AI labs, and its SEC exhibit keeps that language. The identity was reported separately: CNBC, Bloomberg and The Block all identified the counterparty as Anthropic. Combined with AMD, Riot now has about 241 megawatts contracted and close to $9.8 billion of committed revenue. Les called it “a defining moment in our evolution into a leading developer of large-scale data centers.” What those halls will actually run is AI training and inference at scale, the compute layer that increasingly needs to prove what it did, a question explored in our guide to verifiable compute and the trust layer for AI agents.

To see why that contract reset the story, weigh the lease against the day job. Riot’s entire Bitcoin-mining line brought in $113.7 million in Q2, a business exposed to every twitch in the coin price and the difficulty adjustment, per its results. The Anthropic lease, by contrast, is guided to average $365 million to $411 million of net operating income a year for two decades, income that arrives whether Bitcoin sits at $40,000 or $120,000. One signed tenant, in other words, promises to out-earn the mining operation that defined the company for a decade, and to do it with a fraction of the volatility. That is the arithmetic that turned Riot from a miner into an infrastructure name in the market’s eyes.

Power is the real battleground

Underneath the accounting and the leases sits the asset both companies are really fighting over: electricity. Thiel’s warning that by 2028 a miner will need to generate power, be owned by a generator or partner with one is not rhetoric, it is MARA’s capital plan. Buying the Long Ridge gas plant outright is a bet that owning the electrons, and the interconnection that comes with them, is worth more than any hosting contract. Owned generation hedges against power-price spikes and the multi-year interconnection queues that now gate every new data center, though it also loads the balance sheet with merchant-power risk and the operational burden of running a power station.

Riot took the capital-light route and became a virtuoso of the grid instead. Rather than own generation, it turned its flexible mining load into a revenue line by curtailing during periods of grid stress, collecting $10.1 million of power-curtailment credits in Q2 and $31.1 million across the first half of 2026, per its filings. Those credits, earned through the Texas grid operator’s demand-response programs, effectively lower Riot’s power cost, because the only thing it gives up when it powers down is foregone hashrate, with no spoiled product or restart penalty. The catch is that an AI tenant on a 20-year lease wants firm, always-on power, the opposite of interruptible mining load, so the data-center pivot forces Riot to lock in firmer supply than it ever needed as a pure miner. Reconciling cheap interruptible power for hashing with firm power for AI on the same campus is the operational puzzle that will define the next two years for both firms.

Why the market now prices them by different rulebooks

The clearest evidence that these are no longer two versions of the same stock came on August 18. Both fell that day, MARA about 5 percent and RIOT about 4 percent, even though Bitcoin was rising, as the 10-year Treasury yield pushed toward 4.7 percent, per 24/7 Wall St. A pure Bitcoin proxy should not care much about the long bond on a day when Bitcoin is up. An infrastructure developer with two decades of contracted cashflow should, because those distant dollars get discounted harder when rates rise. The tape was telling you that the market now values both as capital-intensive infrastructure names, not as leveraged coin.

That reframing helps explain why RIOT is up strongly on the year while MARA is roughly flat to down: contracted, bond-like AI cashflow is being valued on a net-operating-income basis, closer to how the market prices a real-estate developer, while MARA is still valued as a Bitcoin option with an AI call attached. It also means the September macro calendar matters as much as any block reward. The path of Federal Reserve policy under a more hawkish chair, and the rate expectations that come with it, now feed directly into both valuations, a dynamic we unpack in the hawkish reset rewiring crypto’s September countdown.

The cost-to-mine trap, and why you cannot compare them directly

Investors love a single cost-per-Bitcoin number, and both companies give you one. Do not put them side by side. MARA reported a cost of about $38,690 per coin, but that is a narrow figure covering purchased energy on owned power near four cents a kilowatt-hour, not a full cash cost. Riot reported $49,912 per coin excluding depreciation, and $90,631 including it, which worked out to roughly 126.5 percent of the value of the Bitcoin it produced. In other words, Riot was cash-profitable on mining and GAAP-unprofitable in the same quarter, purely because of how depreciation is charged. The two disclosures measure different things.

The hashrate figures carry the same trap. MARA quotes 70.3 exahashes per second energized, meaning plugged in, while Riot quotes 44.4 exahashes deployed. Energized capacity flatters the comparison because idle-but-connected rigs still count. What actually matters is coins produced against the whole network, and there the self-correcting mechanism that keeps miners honest is difficulty, which sat near 125.81T at the latest retarget, per CoinWarz. JPMorgan’s Nikolaos Panigirtzoglou, who has pegged all-in industry production cost near $78,000, described the feedback loop to TFTC: “When bitcoin trades below its production cost, higher-cost miners power down, the hashrate declines, and difficulty adjusts lower.” Where those machines physically sit, and how that geography is shifting, is mapped in our look at the new map of hashrate growth.

The hashrate gap rewards a second look for the same reason. MARA’s 70.3 exahashes energized translated into 2,422 coins for the quarter, a little over 26 a day, which is only a low-single-digit percentage of everything the network mined; energized capacity counts machines that are plugged in, not necessarily the ones producing at full tilt. Riot’s 44.4 exahashes deployed produced 1,587 coins. The lesson is the one the cost figures already teach: a single big number, whether exahashes or dollars per coin, tells you almost nothing until you know exactly what the company chose to count. Any honest comparison has to go line by line rather than lean on the headline metric each investor-relations team prefers.

The Bitcoin treasury question

The 35,577-versus-11,380 gap in coins held is not a rounding difference in strategy, it is the strategy. Under current accounting, both companies mark their Bitcoin to fair value through the income statement, so every swing in the coin price runs straight through reported earnings. For MARA, with a $2.1 billion stack, that turns each quarter into partly a Bitcoin bet: a rally pads the numbers, a selloff produces losses like the $611 million print in Q2 even when the mining business is fine. Some investors treat MARA explicitly as a Bitcoin-treasury company with a mining and AI option bolted on.

Riot chose to remove most of that volatility, running lean and pledging part of its remaining coins as collateral to fund the buildout. The trade-off is symmetrical. If Bitcoin reclaims six figures, MARA’s treasury delivers outsized upside that Riot has deliberately given up, a scenario we weigh in our year-end outlook on whether Bitcoin can reclaim $100K. If it slides, MARA’s book value bleeds while Riot’s contracted lease income keeps flowing regardless. Holding coins at this scale is also an operational security problem in its own right, since a miner self-custodying billions of dollars faces the same cold-storage risks as anyone, a topic we cover in our 2026 hardware wallet reviews after the Coldcard hack.

The accounting itself amplifies the divergence. Under the fair-value rule that now applies to corporate crypto holdings, companies run both gains and losses on their coins through the income statement every quarter, where the old regime only ever let them write the value down. For a business holding 35,577 BTC, that turns the balance sheet into a live Bitcoin position capable of swinging reported profit by hundreds of millions of dollars in either direction, independent of how many machines were humming. Riot’s leaner stack means the same rule barely moves its numbers, which is precisely the point of running Bitcoin-light.

Everyone is doing this: the peer set

MARA and Riot are not alone, and the scale of the sector-wide pivot is the context that makes their choices legible. Cumulative announced AI and HPC contracts across public miners now exceed $70 billion, according to CoinShares, and several names have moved faster than either of these two. James Butterfill, head of research at CoinShares, told The Block that some listed miners could draw “as much as 70% of their revenues from AI by the end of this year, up from roughly 30% today.”

Read against that table, the two protagonists sit at opposite ends of the same spectrum. Riot is now among the leaders on contracted AI revenue, having converted a power site into a two-decade lease. MARA is a laggard on anchor tenants but arguably ahead on owned generation, having bought a power plant outright rather than leasing capacity to someone else. Whether owning the electrons or owning the contract proves smarter is the open question the market is trying to price.

Now comes the hard part: execution and financing

Signing a contract is not the same as banking the cash, and this is where the next chapter of the story will be written. Riot has to actually build: 96 megawatts live by December 2027 and the full 191 by June 2028, which means construction management, grid interconnection, equipment procurement and cost control on a project the size of Corsicana’s expansion. The financing follows the real-estate playbook: the $573 million Morgan Stanley interim facility now, then most likely longer-dated project debt raised against 20 years of contracted net operating income. That is cheaper capital than a miner usually gets, but it commits Riot to a large, fixed spending schedule whether or not the AI cycle stays hot.

MARA’s execution risk is the mirror image. Its build-to-suit rule means little capital goes out the door before a tenant signs, so the risk is not construction, it is commercial: landing a hyperscale customer at all. The company funds itself with convertible notes and with its Bitcoin treasury as a war chest and collateral source, which keeps optionality high but ties its cost of capital to the coin price. Morgan Stanley’s $6 target is explicitly conditional on MARA converting that optionality into signed leases by year end. John Todaro of Needham captured why the prize is worth chasing, telling CoinGeek that “the revenue per megawatt and EBITDA margins are far higher for HPC and AI colocation than for mining.” The catch is that Riot has already booked those margins on paper, and MARA has not.

Both paths carry a risk the bull cases tend to skip: concentration. Riot’s two decades of contracted income lean heavily on a single frontier AI lab, so that net operating income is ultimately only as durable as the tenant’s own compute demand and balance sheet; a shift in the economics of AI training, or in the tenant’s fortunes, would be felt at Rockdale. MARA faces the opposite timing risk. The build-to-suit discipline that protects its cash also means it is racing a closing window, since every quarter without a signed anchor is a quarter in which rivals lock up the hyperscale customers and the cheap interconnection capacity first. Neither risk shows up cleanly in a single quarter’s numbers, which is exactly why they are easy to underprice.

What each stock is actually a bet on

Strip away the shared history and the two names resolve into two clean, almost opposite theses. Owning MARA is a bet on Bitcoin first and AI second: you get one of the largest corporate coin treasuries as leveraged upside, plus a cheap call option on an eventual AI lease, in exchange for tolerating earnings that lurch with the Bitcoin price and a commercial pivot that is still unproven. Owning Riot is a bet on contracted infrastructure first and Bitcoin a distant second: you get bond-like AI cashflow and a de-risked coin position, in exchange for construction and delivery risk on a committed capital program and sensitivity to interest rates.

The analyst spread mirrors the divergence. MARA targets cluster in a wide $6 to $18 band that reflects genuine disagreement about whether its option pays off, while post-deal RIOT targets sit up at $32 to $35 on the strength of a signed anchor. For readers trying to form their own view, a short watch list helps.

• Whether MARA announces a genuine hyperscale anchor tenant, not just a joint-venture site.
• Whether Riot hits its 96-megawatt December 2027 milestone on time and on budget.
• The 10-year Treasury yield, which now moves both stocks more than a good week for Bitcoin does.
• The Bitcoin price itself, which flows straight through MARA’s earnings and barely touches Riot’s.
• Hashprice and difficulty, the twin gauges that decide whether the legacy mining business is a cash cow or a cash drain.

The regulatory and macro backdrop

Both companies are SEC registrants that file 10-Qs and 8-Ks, and the regulatory mood toward the mining side is, for now, permissive. In March 2025 the SEC’s Division of Corporation Finance issued a staff statement that proof-of-work mining on public, permissionless networks is not itself a securities offering under the Howey test. It carries no formal legal force, but it signals a posture, and it removes one tail risk that hung over the sector. The AI data-center business, by contrast, is regulated as ordinary commercial real estate and power procurement, not as crypto at all, which is part of its appeal to a company like Riot.

The nearer-term swing factor is macro. With the block subsidy set to halve again in 2028 and difficulty grinding higher, the pure-mining margin only gets tighter from here, which is precisely why both companies are diversifying. Rates, the dollar and the Fed’s path will keep pushing these stocks around, and the Bitcoin price will keep deciding how much MARA’s treasury is worth on any given morning. Neither company’s fate is settled. What has changed in 2026 is that they are no longer running the same race: one is trying to become a power company that still mines, the other a data-center landlord that used to. A year from now, the winner will not be whoever mined more coins, but whoever turned megawatts into the most durable cashflow.

Frequently Asked Questions

Is Marathon (MARA) or Riot the bigger Bitcoin miner in 2026?

MARA is the bigger pure miner. In Q2 2026 it ran 70.3 EH/s of energized hashrate, produced 2,422 BTC and held 35,577 BTC, against Riot’s 44.4 EH/s deployed, 1,587 BTC produced and 11,380 BTC held. Riot, however, now earns far more of its revenue outside mining, roughly $60.5 million in the quarter from data-center and engineering work.

What is Riot’s Anthropic data center deal?

It is a 20-year lease of 191 megawatts of critical IT capacity at Riot’s Rockdale, Texas campus, running through June 2048 and worth about $9.1 billion in base revenue, up to $16.1 billion if two five-year options are taken. Riot’s filings name the tenant only as a leading frontier AI lab, while Bloomberg, CNBC and The Block reported the counterparty is Anthropic.

Why did MARA and Riot stock fall when Bitcoin rose in August 2026?

On August 18 both stocks fell, MARA about 5 percent and Riot about 4 percent, even as Bitcoin climbed, because the 10-year Treasury yield pushed toward 4.7 percent. The market now values both as capital-intensive infrastructure developers whose long-dated cashflows are discounted more heavily when rates rise, rather than as simple proxies for the Bitcoin price.

Does MARA or Riot hold more Bitcoin?

MARA holds far more, about 35,577 BTC worth roughly $2.1 billion, and marks it to fair value through earnings, so its quarterly results swing with the coin price. Riot deliberately runs Bitcoin-light at about 11,380 BTC, some pledged as collateral, and sells coins to help fund its data-center construction.

Which Bitcoin miner is the better investment, MARA or Riot?

It depends on your thesis. MARA is a leveraged Bitcoin bet with a large coin treasury and an unproven AI option, while Riot is a contracted AI-infrastructure developer with lighter Bitcoin exposure and construction risk. As of late August 2026, analyst price targets ranged from about $6 to $18 for MARA and $32 to $35 for Riot, reflecting how differently the market values the two strategies.

By Yuki Tanaka, senior mining and infrastructure correspondent at HOGE Wire.