The Constraint Is Electrons, Not Chips

If the financing story is about who fronts the money and the inference story is about who serves the cheapest token, this one is about the thing underneath both that nobody can optimize away: electricity, and the multi-year, multi-billion-dollar commitments required to secure it. Two in-window announcements make the point with unusual clarity, and both come from a corner of the market that didn't exist as an AI story a year ago.

A Bitcoin miner becomes a frontier lab's landlord

On July 6, TeraWulf signed a 20-year lease with Anthropic for its Justified Data campus in Hawesville, Kentucky: approximately 401 MW of critical IT load, expected to generate roughly $19 billion in contracted lease revenue over the initial term. Per the company's SEC 8-K, initial capacity comes online in the second half of 2027, with the full 401 MW in place by early 2028, and Anthropic holds two successive five-year renewal options. WULF shares jumped roughly 17% on the day.

TeraWulf is a Bitcoin miner. Or was. The Hawesville site is a former Century Aluminum smelter TeraWulf bought in February for $200 million, and that lineage is the whole point. Aluminum smelting and Bitcoin mining have exactly one thing in common with frontier AI training: they are enormous, dependable consumers of electricity sited next to cheap, abundant power. The asset was never really the miners or the smelter. It was the interconnect, the substation, the megawatts already wired and permitted. TeraWulf is monetizing that, and it's not alone, IREN, Hut 8, and Cipher all rose on the read-across, because nearly every miner is now pivoting the same way.

The stoic in me appreciates the economy of it. TeraWulf didn't build a new capability; it recognized that the capability it already had, the ability to source power, develop infrastructure, and secure a long-term customer commitment, was the scarce thing all along. More with less: same substation, radically different tenant.

Two honest caveats. The $19 billion is expected contracted revenue over twenty years with capacity ramping through 2028, not money in the bank, and the release carries the usual forward-looking-statement disclaimers. Separately, TeraWulf is selling its 50.1% stake in the Abernathy, Texas JV to a Fluidstack-led group for about $530 million, monetizing a ~$450 million investment at a premium and redeploying into sites it wholly owns. Read that as a company deciding that owning the power and the customer relationship outright beats a minority stake in someone else's.

Brookfield 5x's its bet on skipping the grid entirely

On June 30, Brookfield expanded its financing framework for AI-infrastructure power projects using Bloom Energy's solid-oxide fuel cells from $5 billion to $25 billion, a fivefold expansion since October 2025, under Brookfield's $100 billion AI Infrastructure Fund.

The interesting word is "islanded." Grid interconnection is the invisible bottleneck of the entire buildout, the queue to connect a gigawatt-scale campus to the grid runs years in most of the US, which is a comically long time when the models you're trying to serve are obsolete in months. On-site fuel cells sidestep the queue: generate power at the campus, don't wait for the utility. Brookfield's Sikander Rashid framed the ambition as delivering "from electrons to tokens", which is marketing, but it's marketing pointed at the right layer of the stack. A $25 billion financing framework for on-site generation is a bet that the grid will not catch up in time, and that the winners will be whoever can manufacture their own electrons on-site.

Same caveat as always: $25 billion is a financing framework, a commitment to deploy under certain conditions, not deployed capital.

The pattern worth publishing

Step back and the two deals rhyme with everything else this week. NVIDIA's own "Build in America" framing from July 1 said the same thing from the silicon vendor's side: energy availability and financing, not chip supply, are the pacing items. The Rubin generation is being pitched on being 100% liquid-cooled and grid-flexible, because thermal and power envelopes are now the design constraints that matter.

So here's the through-line across all three stories from this week. Compute financing (NVIDIA underwriting the buyers), inference economics (squeezing more tokens from each watt-hour), and power procurement (TeraWulf, Bloom/Brookfield securing the watt-hours in the first place) are three faces of one shift. The industry has spent its chip anxiety and moved on to its power-and-capital anxiety. You can have all the Blackwell you can pay for; the questions that actually gate a gigawatt-scale plan now are where does the power come from, who finances the twenty-year commitment to secure it, and how many tokens can you wring from each megawatt once it's lit.

Chips were the story we told for three years because they were the part we could see on a spec sheet. The real constraint was always the least glamorous line item in the whole build: the electricity bill, and the willingness to sign a twenty-year lease against it.

Sources: TeraWulf press release and SEC 8-K; Bloom Energy / Brookfield release; Data Center Dynamics. Revenue and financing figures are forward-looking with standard disclaimers; $19B is expected 20-year contracted revenue and $25B is a financing framework, not deployed capital.