Smile while the liquidity drains.
But this time, the liquidity isn’t crypto. It’s power. Pure, raw, gigawatt-scale electricity—the kind that used to keep Bitcoin’s ASICs humming. Now, it’s being rerouted to fuel the next generation of AI reasoning. On Tuesday, TeraWulf—a mid-tier Bitcoin mining outfit with a knack for securing cheap nuclear and hydro power—dropped a bomb that sent shockwaves through both the crypto and AI worlds: a 20-year, $19 billion infrastructure contract with Anthropic, the $60 billion AI safety darling behind Claude.
The chart lies. The crowd feels.
I’ve been in this game since the ICO sprint of 2017, back when I was a junior dev in Nairobi chasing EtherDelta rumors. I’ve seen miners pivot to DeFi, pivot to NFTs, pivot to HPC. But this? This is different. This isn’t a pivot. It’s a mutation. TeraWulf isn’t just renting out spare rack space. They’re selling the entire skeleton of their operation—the power, the cooling, the land, the grid interconnection—to an AI giant for two decades. In return, they get a steady revenue stream that makes Bitcoin price volatility look like a minor nuisance. But the real story isn’t the contract. It’s the capital structure behind it.
Context: Why now?
Let’s rewind. The bear market of 2022-2023 nearly killed the mining sector. Core Scientific filed for bankruptcy. Marathon’s stock halved. TeraWulf survived by cutting costs and selling power back to the grid during peak hours. But the narrative was clear: miners were commodity producers, at the mercy of Bitcoin’s hash price and energy markets. Then AI happened. Suddenly, every hyperscaler—Microsoft, Google, Amazon—was scrambling for data center capacity. And who had the power? Miners. Not just any power, but pre-permitted, already connected, often green energy. TeraWulf’s Lake Mariner facility in upstate New York sits on a massive hydroelectric grid. It’s the kind of asset that AI companies can’t build fast enough. So when Anthropic needed guaranteed compute for the next decade, TeraWulf was the phone call.
The Core: Breaking down the deal
Here’s what we know. TeraWulf signed a 20-year agreement to host Anthropic’s AI training clusters. The headline number: $19 billion. That’s total revenue over the contract life. Annualized, about $950 million. For context, TeraWulf’s 2024 mining revenue was roughly $80 million. This deal would 12x their top line. But here’s where my experience as a 7x24 market surveillance analyst kicks in. A 20-year contract in a hardware-intensive business is rare. The typical hyperscaler lease is 5-10 years. 20 years implies massive upfront capital commitment. TeraWulf will need to deploy thousands of NVIDIA H100 or B200 GPUs, plus the supporting networking, cooling, and power infrastructure. At current H100 prices (~$30K each), a 10,000-GPU cluster costs $300 million in hardware alone. Scale that to meet Anthropic’s needs—likely 50,000+ GPUs—and you’re looking at $1.5 billion in just chips.
But here’s the kicker: TeraWulf is funding this by selling assets.
Simultaneously with the Anthropic deal, TeraWulf announced the sale of its majority stake in a joint venture—rumored to be a digital infrastructure company—for an undisclosed amount. I’ve seen this playbook before. In 2021, during the NFT art heist I broke, a studio sold a minority stake to a Hollywood fund to finance marketing. TeraWulf is doing the same: selling non-core assets to raise cash for the AI buildout. Smart, but risky. The JV sale likely brings in $200-500 million, which covers initial GPU procurement. The rest will come from debt or equity dilution. The market is cheering the revenue. I’m watching the balance sheet.
Let’s talk margins. $950 million annual revenue sounds huge, but what’s the cost? Power: TeraWulf’s blended PPA rate is around $0.04/kWh. For a 100MW facility running 24/7, that’s ~$35 million per year. GPUs depreciation: assume 5-year straight-line on $1.5B hardware = $300M/year. Cooling and maintenance: another $50M. Labor, network, insurance: $30M. Total operating cost: ~$415M/year, leaving $535M gross profit. That’s a 56% gross margin—healthy for a data center operator but not the 80-90% margins AI software companies enjoy. Net profit after interest and taxes? Maybe $300-400M. On $19B total revenue, that’s a 15-20% net margin. The chart lies. The crowd feels a gold rush. The numbers tell a story of a capital-intensive utility.
Contrarian angle: The real blind spot
Everyone is focused on TeraWulf’s revenue. No one is asking: What happens if Anthropic can’t pay? Or if the US government restricts GPU exports? Or if a cheaper competitor like CoreWeave builds next door? I’ve been in this industry long enough to know that 20-year contracts have fine print. There will be minimum capacity guarantees, force majeure clauses, and—most importantly—termination rights. If TeraWulf fails to deliver a specified uptime (say 99.999%), Anthropic can walk. And TeraWulf is relying on a single customer for 90% of future revenue. That’s a concentration risk that would make any risk manager wince.
But the contrarian angle that keeps me up at night isn’t about TeraWulf. It’s about the entire mining sector. This deal is a massive signal that energy is the new crypto. Not Bitcoin, not ETH. Energy. The value of a miner’s stock is now more correlated to their power contracts than their hash rate. Every mining CEO is going to try to copy TeraWulf. But most don’t have the grid interconnection or the liquidity to land an Anthropic. The ones that do—like Marathon’s partnership with a renewable energy provider—will get bought out by hyperscalers at a premium. The ones that don’t will be left holding ASICs that lose value every halving. The industry is bifurcating: AI-ready miners vs. pure play miners. TeraWulf just jumped to the former category.
Smile while the liquidity drains. The liquidity is the old mining business model—fragile, cyclical, dependent on Bitcoin. TeraWulf is draining that liquidity and pouring it into a more stable, long-term revenue moat. But draining liquidity is painful. They had to sell assets. They might dilute shareholders. They’re betting the farm on hardware that could be obsolete in 3 years (NVIDIA Blackwell vs. Hopper).
Takeaway: The next watch
The market will price TeraWulf like a growth stock for now. But the real test comes in six months. Two data points: (1) Did TeraWulf secure GPU allocation? If they announce a partnership with NVIDIA or Supermicro, the stock flies. If they go quiet, short it. (2) What’s the capex per MW? If their buildout cost exceeds $10M/MW, the returns will be mediocre. I’ll be watching the quarterly earnings for the “AI Infrastructure Revenue” line item. And I’ll be asking the same question I asked during the Terra meltdown: Who is the last person left holding the bag? In this case, it’s not retail. It’s the shareholders who bought the hype without reading the fine print.
Wake up. The 24/7 clock never blinks. TeraWulf just reset the clock for the entire mining industry. The next handshake? It’s already being arranged in another boardroom. I’ll be there, listening.