bitcoin mining ai hpc halving infrastructure

Bitcoin's hash rate hit a historic high. Its miners have never earned less.

bitcoinindex.net · · 7 min read
Bitcoin's hash rate hit a historic high. Its miners have never earned less.

On September 2, 2025, Bitcoin’s network hashrate crossed 1 zettahash per second on a sustained seven-day average. For context, 1 ZH/s is one sextillion hashes per second. Bitcoin first crossed 1 exahash in 2016. It took nine years to grow one thousandfold.

Then, five months later, hashprice hit an all-time low.

On February 2, 2026, the expected daily revenue per petahash of computing power fell to $33.31, per Luxor’s Hashrate Index. The network is computing more than ever. The miners running that network are earning less per unit than at any point in Bitcoin’s history.

This is not a contradiction. It’s the halving, playing out on schedule.

The math that broke mining

The April 2024 halving cut block rewards from 6.25 BTC to 3.125 BTC. Daily new Bitcoin issuance dropped from 900 coins to 450. Revenue per unit of computing power fell in half, roughly.

What made it worse: the hashrate kept growing anyway. Each individual operator has to keep mining to recoup their hardware investment. So more machines came online, difficulty rose, and every existing machine’s share of the reward shrank further. By Q2 2025, the weighted average cash cost to produce one bitcoin among publicly listed miners had risen to approximately $74,600. The all-in cost (including depreciation and stock-based compensation) averaged $137,800 per coin. Transaction fees contributed an average of 0.7% of block rewards in September 2025, effectively nothing.

Sector-wide debt tells the story: Bitcoin miners collectively owed $2.1 billion at the end of 2024. By mid-October 2025, that had grown to $12.7 billion.

As MARA CEO Fred Thiel put it: “Bitcoin mining 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 2028 prediction is stark: “By 2028, you’ll either be a power generator, be owned by one, or be partnered with one. The days of being a miner plugged into the grid are numbered.”

Why AI makes sense

A Bitcoin mining facility built in the last four years shares most of what an AI data centre needs: power infrastructure at scale (often 100+ megawatts connected directly to grid substations), cooling systems designed for dense heat loads, approved sites with existing permits, and 24/7 operations teams.

The economics are dramatically different. CoinShares estimates that building and operating a Bitcoin mining facility costs around $700,000 to $1 million per megawatt. AI-capable data centre infrastructure costs approximately $20 million per megawatt, reflecting the redundancy and uptime requirements (99.99%+) that enterprise AI clients demand. But the revenues follow a completely different model too: long-term contracts with hyperscalers, predictable multi-year cash flows, no halving exposure, no difficulty adjustment eating into margins.

Meltem Demirors, general partner at Crucible Capital, summarised the dynamic in Wired: “Bitcoin mining created the blueprint for the AI compute boom and the modern data center. They have found that their cost of capital is much lower if they go into the AI narrative. They have the powered shell, they’re ripping out the [mining machines], and their tenant is bringing the GPUs.”

Needham analyst John Todaro was direct about what capital markets are signalling: “Investors are almost exclusively valuing Bitcoin miners for their HPC/AI opportunities at this point. The revenue per megawatt and EBITDA margins are far higher for HPC and AI colocation than for mining.”

Who’s actually doing it vs. who’s announcing it

This is where the story gets important. Across the sector, over $43 billion in AI and HPC contracts have been announced. That number covers everything from signed 10-year contracts to non-binding frameworks to vague strategic statements. The distinction matters enormously.

As of Q3 2025 earnings, only two companies reported meaningful AI or HPC revenue.

IREN (formerly Iris Energy) signed a $9.7 billion, five-year GPU cloud services contract with Microsoft in November 2025, providing access to NVIDIA GB300 GPUs. The GPUs are being acquired from Dell at a cost of $5.7 billion; annualised revenue once fully deployed will be approximately $1.94 billion. In Q3 2025, IREN reported $7.3 million in AI cloud services revenue, a small number but scaling fast. They had 10,900 NVIDIA GPUs deployed.

TeraWulf has Google as a shareholder at 14% and a $3.2 billion financial backstop. It has a 10-year, 360 MW agreement with Fluidstack at its Lake Mariner campus in Western New York. In Q3 2025, it reported $7.2 million in HPC lease revenue. The company runs on nuclear and hydroelectric power, which gives it both a cost advantage and an ESG story that attracts institutional infrastructure capital.

Core Scientific sits in a complicated category. It has $10.2 billion in contracted HPC infrastructure revenue with CoreWeave across 590 megawatts at six sites. That contract is real and signed. But in July 2025, CoreWeave announced a $9 billion all-stock acquisition of Core Scientific. Shareholders rejected the deal in October 2025, arguing the exchange ratio undervalued them. The contract survives the failed merger, but Core Scientific’s strategic direction is now uncertain.

Cipher Mining signed a 168 MW, 10-year AI hosting agreement with Fluidstack at its Barber Lake site in Texas in September 2025, a deal worth approximately $3 billion over the term. It raised $1.4 billion to finance construction. No AI revenue yet, but the contract is signed and funded.

Riot Platforms has earmarked 600 megawatts of capacity at its Corsicana, Texas facility (the world’s largest Bitcoin mine) for AI and HPC. It commissioned an independent feasibility study from Altman Solon, which confirmed the site’s suitability. No signed contracts yet and no AI revenue as of Q3 2025.

CleanSpark announced its AI pivot in October 2025, hired an SVP of AI Data Centers, and won a bid to operate a 100 MW AI data centre in Cheyenne, Wyoming (beating Microsoft for the contract, per CEO Matt Schultz). As of late October 2025, the Wyoming deal was a non-binding framework pending definitive agreements. No AI revenue reported.

MARA, the largest publicly traded Bitcoin miner by installed capacity at 57.4 EH/s, has framed its strategy around energy vertical integration. It announced a nearly two-thirds stake in French utility Exaion and a new subsidiary focused on natural gas and renewable energy. Primarily still a Bitcoin miner. No AI revenue yet.

Bitfarms is the most committed to the exit. CEO Ben Gagnon announced in November 2025 that the company would “wind down our Bitcoin mining business in 2026 and 2027.” He explained the logic: “Bitcoin mining is still profitable. It’s that HPC creates so much more value per unit of energy and does so predictably for years into the future that the company can’t justify further investment into bitcoin mining.” The Washington State facility conversion is underway. Still 100% mining revenue.

What this means for Bitcoin

The obvious concern: if publicly listed miners pivot to AI, who runs Bitcoin’s hashrate?

The answer may actually be fine. CoinShares notes that publicly traded miners added roughly 80 EH/s to the network in 2025, while the overall network grew by around 300 EH/s. The majority of new hashrate came from private operators, sovereign-backed entities, and smaller unlisted companies. Public miners are not the network’s backbone.

The more structural shift: as institutional facilities near major cities convert to AI (better connectivity, higher margins), Bitcoin mining migrates toward stranded energy sources, flared natural gas, remote hydro, wind curtailment. Portable containerised mining units are increasingly deployed to capture this energy. The result could be more geographic decentralisation of mining, not less.

The long-term risk is a different one. Thiel’s observation about fees is blunt: “If Bitcoin’s [fiat value] doesn’t grow at 50% or more annually, the math gets very tough after 2028 and even tougher in 2032,” when block rewards fall below 1 BTC per block. The fee market that was supposed to replace block subsidies hasn’t materialised at scale. Transaction fees averaged 0.7% of block rewards in September 2025.

If mining profitability stays structurally depressed and the fee market doesn’t develop, securing the network long-term becomes a harder problem. The AI pivot is rational for individual companies. Whether it’s good for the network at the margin is a different question.

Charles Chong, VP of Strategy at BlockSpaceForce and former director of strategy at Foundry, said it plainly in Wired: “The economics are terrible today. If I buy a bitcoin mining machine today, I don’t know if I can make the money back.”

For now, the hashrate is at historic highs, the network is secure, and the math is pushing capital toward AI rather than ASICs. That might be exactly what Bitcoin needs, the industrial machines leaving for more profitable applications while mining disperses to wherever energy is cheapest. Or it might not be. We’ll find out by 2028.

Sources

  1. America’s Biggest Bitcoin Miners Are Pivoting to AI — Wired, December 9, 2025
  2. Bitcoin Mining Report Q4 2025 — CoinShares (James Butterfill/Luke Nolan)
  3. Bitcoin Hashprice Falls to Five-Year Low — CoinDesk, November 18, 2025
  4. Bitcoin difficulty projected to drop 15% as hashprice hits all-time low — Yahoo Finance/Blockspace (Luxor’s Ben Harper), February 2026
  5. BTC’s 7 Day Average Hash Rate Hits 1 ZettaHash for First Time — CoinDesk, September 2, 2025
  6. Block reward miners can’t pivot to AI fast enough — CoinGeek, November 2025
  7. IREN Secures $9.7bn AI Cloud Contract with Microsoft — GlobeNewswire/IREN, November 3, 2025
  8. Crypto Miner TeraWulf to Raise $3B in Google-Backed Debt Deal — CoinDesk, September 27, 2025
  9. CoreWeave to Acquire Core Scientific — CoreWeave, July 7, 2025
  10. Core Scientific shareholders reject $9 billion CoreWeave offer — Reuters, October 30, 2025
  11. Why bitcoin miner CleanSpark beat Microsoft for Wyoming AI data center deal — CNBC, October 28, 2025
  12. Bitfarms Plans to Wind Down Bitcoin Mining by 2027 — TheMinerMag, November 13, 2025
  13. Cipher Mining Signs 168 MW, 10-Year AI Hosting Agreement with Fluidstack — Cipher Mining, September 25, 2025
  14. Post Halving: Bitcoin Miners Landscape — AMINA Bank, November 19, 2025
  15. TeraWulf stock gains more than 4% as Google boosts stake in data center operator — CNBC, August 18, 2025 (Google’s 14% stake in TeraWulf)

Data and pipeline status as of February 18, 2026.