Capital intensity is becoming the defining moat in artificial intelligence, and two deals struck on 6 and 7 July 2026 put an exact price on it: $44 billion in two days.
What happened on 6 and 7 July 2026
TeraWulf Inc. (Nasdaq: WULF) announced the lease with Anthropic on Monday 6 July 2026. The contract covers TeraWulf's "Justified Data" campus in Hawesville, Kentucky, and runs for 20 years at a contracted value of approximately $19 billion. Anthropic will draw up to 401 megawatts of critical IT load once the site is complete, with initial capacity due online in the second half of 2027 and full capacity by early 2028.
The campus sits on roughly 790 acres that TeraWulf bought for about $200 million in February 2026 — the former site of an aluminium smelter, chosen partly because industrial-scale power transmission and fibre-optic lines were already running to the property. TeraWulf is investing a further $3-4 billion of its own capital to build the data centre out, a fraction of the $19 billion the lease is worth over its full term. TeraWulf shares rose almost 5% on the announcement.
TeraWulf used the same announcement to disclose a second transaction. The company is selling its 50.1% stake in the Abernathy joint venture — a 168-megawatt site in Abernathy, Texas, established in 2025 — to an investor group led by Fluidstack, its partner in that venture, at a premium to the roughly $450 million TeraWulf had invested. The sale narrows TeraWulf's holdings toward sites it owns outright rather than shares in a joint venture.
💡Two commitments, one week
Two numbers from one week: $19 billion committed to a single 20-year lease, and $25 billion raised in a single bond sale a day later. Neither figure describes money either company has already spent. Both describe what each has just promised to spend.
Amazon followed on Tuesday 7 July 2026, opening an eight-tranche senior unsecured bond offering with maturities running from three to forty years. Barclays, Goldman Sachs, JPMorgan and Morgan Stanley underwrote the sale. Orders peaked near $62 billion before Amazon trimmed the final book to roughly $41 billion against a minimum $25 billion raise, putting demand at roughly 1.6 times the size of the sale.
Amazon described the use of proceeds as "general corporate purposes," language a company spokesperson said can cover "future capital spending or paying down debt already on the books." The timing lines up with Amazon's 2026 capital-expenditure guidance of roughly $200 billion, up from $131 billion in 2025, weighted toward data centres, custom chips and AI infrastructure. Amazon also told underwriters it does not plan to issue further debt this year.
The new bond sale brings Amazon's total 2026 debt raised toward its AI build-out to close to $89 billion, layered on top of $54 billion raised earlier in US and European bonds and a $10 billion Canadian-dollar offering in June. The spending follows a quarter in which AWS revenue grew 28% and Amazon's own Trainium and Graviton chips were projected to generate more than $10 billion a year — the demand this capital is meant to keep pace with.
Why capital intensity is becoming the moat
Both deals point at the same underlying fact. Building frontier artificial intelligence now requires balance-sheet size that only a small number of companies have. TeraWulf itself is a useful measure of the gap: the company posted $34 million in revenue last quarter, yet is committing $3-4 billion of its own capital to build the Kentucky campus, roughly a hundred times its quarterly income, financed against a signed 20-year contract with a tenant it trusts to pay.
Anthropic can make that kind of tenant look safe because its own numbers are enormous. The company recently filed confidentially to go public at a valuation of $965 billion, and the Kentucky lease is one part of a much larger compute-buying spree: Anthropic already holds commitments for more than 10 gigawatts of capacity with cloud providers including Google, Amazon Web Services and CoreWeave. The same handful of companies — a small circle of AI labs, hyperscalers and their financiers — keep reappearing on multiple sides of these deals.
The Anthropic lease validates our strategy and establishes a long-duration revenue stream with one of the world's leading AI companies.
— Paul Prager, Chairman and CEO, TeraWulf
Prager's language is corporate, but the strategy behind it is not subtle. TeraWulf is moving away from joint ventures like Abernathy, where returns are shared, toward sites it owns outright and leases directly to a single, creditworthy tenant. That is the same logic SoftBank applied at ten times the scale in France, and the same logic behind SpaceX's move into AI compute leasing: whoever controls the physical infrastructure now holds more leverage, and more certain revenue, than whoever merely promises to use it.
AI's physical footprint: a smelter becomes a campus
The Hawesville site is not a coincidence of geography. Aluminium smelting is one of the most electricity-hungry industrial processes there is, which is precisely why a decommissioned smelter comes with the two things an AI campus needs most: heavy-duty power transmission already wired to the property, and land zoned for industrial-scale energy draw. DataCenterDynamics reported that TeraWulf is reusing that existing infrastructure rather than building new grid connections from nothing.
The scale is worth holding in view. TeraWulf's Kentucky campus will draw up to 401 megawatts once complete, more than double the 168 megawatts of the Abernathy site the company just sold. NVIDIA has argued that every country now needs its own "AI factory," but a 401-megawatt facility backed by a $19 billion contract and $3-4 billion in construction capital is not a factory a country builds by simply deciding to. Such a campus requires a landlord able to raise billions against a signed lease, and a tenant able to sign one.
⚠️A different business now
A former aluminium smelter and a fresh $25 billion bond sale describe the same shift in different registers. The industry that once measured itself in parameters and benchmarks now measures itself in megawatts, acres and years of committed revenue. That is a different business, with a much shorter list of who can play.
A dignity-first reading: who gets to build the infrastructure of intelligence
Emergent Intelligence (EI) — the dignity-first lens through which I read artificial intelligence — asks a question a balance sheet cannot answer: not only who can afford to build this infrastructure, but who is excluded from the decision by the scale of the number itself. A $19 billion lease and a $25 billion bond sale are not neutral facts. Both function as a gate, built from capital requirements so large that entire regions are priced out before a single negotiation begins.
I have written before that Africa holds roughly 18% of the world's population and under 1% of global data-centre capacity, and that the continent's five largest data-centre markets combined hold less capacity than France alone. Set that figure beside this week's news. Two companies committed close to $44 billion to AI infrastructure in two days, and none of it will be built on the African continent. The gap is not a coincidence of timing; it is the same capital-intensity moat, viewed from the side that cannot clear it.
This is not an argument against the Kentucky lease or the Amazon bond sale, both of which are ordinary, defensible corporate finance. It is an argument that deciding who gets to build the infrastructure of intelligence is now a global-equity question, not only a Wall Street one. When the entry price for a single campus runs into the billions and the underwriting depends on decades of committed revenue, the list of eligible builders is set long before anyone sits down to negotiate, and that list does not yet include most of the world.
⚠️The number worth sitting with
Under 1% of global data-centre capacity for 18% of the world's people, against nearly $44 billion committed by two companies in two days. The dignity-first question is not whether this capital should be spent. It is whether the capacity to spend it should keep deciding, on its own, who gets to build the infrastructure everyone will eventually depend on.
What a broader base of builders would require
Closing a gap this size will not happen through goodwill, and it will not happen by wishing capital intensity away. The physics of a 401-megawatt data centre are real, and so is the credit risk a $19 billion lease represents. What can change is who counts as a plausible builder in the first place: financing structures that let sovereign and regional investors co-own capacity rather than only rent it, transmission and power investment that serves whole regions rather than a single campus, and deal terms that price in trained local operators rather than only imported ones.
The precedent already sits inside the deals themselves. TeraWulf financed a $19 billion commitment against $3-4 billion of its own capital by anchoring it to a single, creditworthy tenant. Nothing about that structure requires the tenant to be a US artificial-intelligence lab or the landlord to be a US infrastructure company; the obstacle is not the financial instrument, which is well understood, but the assumption about who is allowed to be a tenant, a landlord or an underwriter in deals like this one.
Two days, forty-four billion dollars, and a question of who
By the end of the week, a former aluminium smelter in rural Kentucky had a signed 20-year, $19 billion tenant, and Amazon had more than $25 billion in fresh capital lined up for the same buildout. Both are ordinary corporate decisions, backed by real demand for artificial intelligence compute. Neither company did anything wrong.
In short, the story is not the deals themselves but what the deals reveal. Artificial intelligence's frontier is now gated by balance-sheet size as much as by algorithmic skill, and that gate has a geography. It runs through Kentucky and through Wall Street, and it does not yet run through most of the African continent, or most of the Global South.
The infrastructure of intelligence is being built at a pace and a price that only a handful of companies can match. The dignity-first task is not to slow that building down. It is to ask, deal by deal, whether the list of who gets to build is widening or simply repeating itself, and to keep asking until the answer changes.
Frequently Asked Questions
The questions below address the most common queries about the Anthropic-TeraWulf lease and Amazon's July 2026 bond sale, drawn from the primary announcements and the reporting around them.
What is the TeraWulf-Anthropic lease at the Justified Data campus?
It is a 20-year lease, announced on 6 July 2026, under which Anthropic will occupy TeraWulf's "Justified Data" campus in Hawesville, Kentucky, on the site of a former aluminium smelter. The contract is worth approximately $19 billion over its term, covers up to 401 megawatts of critical IT load, and TeraWulf is investing $3-4 billion of its own capital to build the facility, with initial capacity due online in the second half of 2027.
Why is Amazon raising $25 billion in bonds for artificial intelligence infrastructure?
On 7 July 2026 Amazon opened an eight-tranche senior unsecured bond offering, underwritten by Barclays, Goldman Sachs, JPMorgan and Morgan Stanley, raising more than $25 billion after demand peaked near $62 billion. Amazon described the proceeds as being for "general corporate purposes," language reporting has tied to the company's roughly $200 billion 2026 capital-expenditure guidance for data centres, chips and AI infrastructure, up from $131 billion in 2025.
How does the Kentucky data-centre deal reuse a former aluminium smelter?
TeraWulf bought the roughly 790-acre former aluminium-smelting site in Hawesville, Kentucky for about $200 million in February 2026. Aluminium smelting requires very large amounts of electricity, so the site already had heavy-duty power transmission and fibre-optic infrastructure in place, exactly what a data centre needs, which let TeraWulf build an AI campus without starting the grid connection from scratch.
Which companies can compete at this scale of AI infrastructure spending?
Very few. The Kentucky lease and the Amazon bond sale both depend on either a multi-decade contracted revenue stream from a company able to promise it, such as Anthropic, which recently filed confidentially to go public at a $965 billion valuation, or direct access to public debt markets at investment-grade rates, such as Amazon, which has raised close to $89 billion in bonds during 2026 alone. That combination of scale and creditworthiness is held by a small number of companies worldwide.
What is the Emergent Intelligence view of who builds AI infrastructure?
Emergent Intelligence (EI) is a dignity-first reading of artificial intelligence. Applied to infrastructure spending, it holds that a capital-intensity threshold this high functions as a gate on global participation, not simply a market outcome, and that a continent holding under 1% of global data-centre capacity while two companies commit nearly $44 billion in two days is a global-equity question, not only a financial one. The EI argument is for financing structures that widen who can build, not merely for building faster.
Sources and Further Reading
Cover image: power transmission lines silhouetted against an industrial plant at dusk. Photo by Plato Terentev via Pexels.
