This week, Anthropic reportedly agreed to spend about $45 billion over six years renting 460MW of computing capacity from Nscale, a British data-centre company. The capacity will come from a site in West Virginia equipped with Nvidia’s new Vera Rubin systems. On the same day, Nvidia reported $89 billion of quarterly data-centre revenue, up 117% from a year earlier.
Both announcements depend on the increasingly urgent answer to one question... How much useful and saleable work can the AI industry produce from a given amount of electrical consumption?
The AI industry typically measures this amount in megawatts (MW). For context, a megawatt is enough electricity to power four or five racks of Vera Rubin systems, about 300 GPUs in a single row of cabinets. The same electricity would supply more than 3,000 British homes. Nvidia now sells about $40 million of equipment into each megawatt, up from $18 million with the old Hopper architecture. Under the Nscale terms, Anthropic would pay about $16 million a year to rent one. Dylan Patel of SemiAnalysis discusses this on this week’s Dwarkesh Patel podcast. He says Anthropic’s revenue has reached as much as $50 million per megawatt, against a base cost of $10 million to $15 million. That $50 million is roughly what 20,000 people would pay for a $200-a-month subscription.
Early data-centre spending after ChatGPT was defensive and paid from cash. Today it is backed by revenue from coding agents, subscriptions and business products. These are compelling returns that are drawing ever more capital into the sector. They are also changing how data centres are financed. A developer with a long contract from an AI company can borrow against the promised payments before construction starts. Nscale relies on Anthropic paying. Anthropic relies on customers buying enough AI. The lenders rely on both.
That favours labs and hyperscalers that already have the users. A contract from a company with proven revenue is easier to borrow against than one from a challenger nobody has heard of. The leaders’ capacity gets built first and their lead compounds. Patel expects Anthropic and OpenAI to hold most of the world’s compute by 2028.
Nvidia sells the equipment to every party in this chain, from the labs to the hyperscalers to new operators like Nscale. It is also arranging the capital. In August it agreed with six large investment firms to raise more than $500 billion for AI infrastructure, and it has the option to backstop up to a quarter of that itself. Customers who could not borrow on their own can now buy Nvidia systems with Nvidia-arranged credit. That pulls demand forward. It also ties Nvidia’s future sales to the loop continuing to turn.
Activity in the UK shows how overheated this is becoming. Applications to connect new demand to the grid tripled in seven months, from 41GW to 125GW, against peak national demand of about 45GW. Data centres account for at least 80GW. Of about 140 data-centre projects in the queue, only 71 had reached a final investment decision. Ofgem’s answer is a commitment fee of £237,500 to £712,500 per megawatt, refunded when the site is energised and forfeited if the project drops out.
What happens next depends on how much value each megawatt can generate, how much of it customers use and what they pay for it. In the best case, Rubin, efficient models (see the new GLM and Qwen Flash models from this week) and always-on agents multiply the work per megawatt. Prices fall, usage grows faster and data centres stay full. In the middle case, open models and custom chips make AI cheap faster than usage grows. The infrastructure stays useful, but labs lose pricing power and contracts get cheaper on renewal. In the worst case, prices fall much faster than demand. A lab expecting $50 million a year from a megawatt still owes $16 million of rent after revenue drops to $20 million. Labs renegotiate. Developers lose the payments behind their debts. Lenders stop lending and unfinished sites are cancelled. Speculative grid applications go first. Familiar bubble mechanics take over, with every party trying to get out before the others.
Data from OpenRouter shows how fast usage can respond to price. After it halved the price of two OpenAI models on 27 July, their daily token volumes rose to six and fourteen times their earlier average. Models without a discount stayed flat. At half price, six times the volume is three times the revenue.
All of these potential future paths reach well beyond AI. The buildout needs trillions of dollars while governments are already borrowing heavily. The best case keeps demand for capital intense. Long-term borrowing rates stay higher, and governments and homeowners pay more to refinance. The worst case brings losses on AI debt and tighter credit for our debt-fuelled world. AI can help indebted economies carry their debts if it raises productivity. If investment keeps growing faster than AI’s economic impact, the financing becomes a gigantic burden, even if the long-term benefits are real.
Takeaways: A boom like this has its own momentum. No single organisation can steer it, even as compute concentrates in a few hands. What an organisation can control is whether it understands the value it is buying. Much of today’s demand rests on individual impressions of productivity rather than sustained evidence of value. That evidence exists, but only for those who take the time to measure. Benchmarks and evaluations tied to your own operations show which systems earn their cost. Without that understanding, prices, revenue and demand can fall suddenly because nobody checked whether the value was real. The worst outcome would be losing faith before the proof is in, and stalling a machine that is still delivering.