Semiconductor vendors are underwriting AI infrastructure leasing, and the value of miners' power assets continues to be re-rated
AI summary card
Semiconductor vendors are underwriting AI infrastructure leasing, and the value of miners' power assets continues to be re-rated
Bernstein believes that miners such as CORZ, HUT, CLSK, and WULF are becoming key "landlords" in the expansion of AI compute capacity by leveraging approved grid interconnections and deliverable data center resources, while the credit and ecosystem support of semiconductor vendors such as AMD and NVDA is improving deal structures and financing conditions.
- CORZ's roughly 530 IT MW partnership related to AMD is expected to correspond to about $14 billion of revenue over 15 years and includes an option for AMD to expand future leasable capacity to a cumulative 2.5 GW.
- Over the past two years, miners have contracted about 8 GW of power capacity with hyperscalers, neocloud providers, and AI chip vendors, with contract revenue exceeding $160 billion.
- Signing master leases directly with investment-grade tenants is viewed as superior to older structures that relied only on neocloud providers or credit guarantees, helping to reduce financing costs and counterparty risk for long-term leases.
- HUT, CLSK, and WULF have recently disclosed large AI hosting leases respectively, showing that projects in the 350 to 500 MW range are becoming the industry's mainstream deal size.
- The report remains positive on Emerging AI Infra companies with power pipelines and delivery capabilities, while also believing that semiconductor companies such as NVDA and AMD can strengthen their ecosystems by supporting customers' deployment of compute capacity.
Report interpretation
Overview
This report discusses new deal structures at the intersection of global digital assets and AI infrastructure, focusing on hosting transactions involving CORZ and AMD, HUT and a suspected NVDA-related tenant, CLSK and a global technology company, and WULF and Anthropic. The core judgment is that AI compute demand is extremely strong, while assets with grid interconnection, land, and deliverable "powered shells" are scarce, driving up the position of former bitcoin miners in the AI data center supply chain.
Core views
The report's core views include: first, miners are no longer merely crypto asset producers but are transforming into AI infrastructure developers and data center landlords; second, long-term master leases directly facing investment-grade tenants are superior to older structures using neocloud providers as intermediaries, improving financing costs and long-term revenue visibility; third, semiconductor vendors such as AMD and NVDA are using credit, balance sheet, and ecosystem resources to help customers deploy compute capacity, which is not equivalent to the most aggressive "circular transactions" because behind it are real compute deployments and scarce data center constraints; fourth, miners with large power portfolios, execution records, and financing channels are more likely to benefit.
Analysis framework
The report uses a cross-sectional comparison of transaction cases, sorting through the recently signed capacity, contract amount, lease term, lease structure, capital expenditure, delivery pace, financing support, and implied returns of companies such as CORZ, HUT, CLSK, and WULF, and evaluates these transactions within the ecosystem structure among AI infrastructure, neocloud providers, hyperscalers, and semiconductor vendors.
Methodology notes
Direct master lease and credit support structures
The report compares long-term leases directly with investment-grade tenants, structures leased by neocloud providers and guaranteed by large technology or semiconductor companies, and new models in which semiconductor vendors participate directly, arguing that direct master leases are generally more favorable for reducing financing costs and counterparty risk.
Revenue yield, capital expenditure, NOI margin, and unlevered IRR
The report compares the economics of different hosting transactions using annual revenue per IT MW, capex per IT MW, NOI margins approaching 100% under triple-net leases, and estimated unlevered IRRs of 8% to 13%.
Approved power, grid connection resources, and delivery of powered shells
The report emphasizes that the constraints on AI compute expansion are not only chips, but also land, grid interconnection, construction progress, and the ability to deliver powered data center space on time.
Outperform, Market-Perform, Underperform
Bernstein's branded equity ratings are typically based on expected performance relative to the market index over the next 12 months, with Outperform meaning expected to beat the index by more than 15 percentage points, Market-Perform meaning roughly in line, and Underperform meaning expected to lag by more than 15 percentage points.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- CORZA core beneficiary AI infrastructure developer, rated Outperform by the report with a $32 target price.
- Strengths
- Has signed large-scale contracts with AMD and CoreWeave, with about 1.1 GW of IT load already contracted, and has about 2 GW of new site opportunities plus AMD's potential expansion rights.
- Weaknesses
- Project capex is high; total capex for the initial 530 IT MW project is about $6 billion, and future project-level financing support is still needed.
- Comparison
- Compared with structures relying only on neocloud providers, CORZ's combination of direct leasing and credit support with AMD improves counterparty quality.
- Risks
- Delivery delays, rising financing costs, AMD's expansion option not being fully exercised, and actual construction cost overruns under long-term leases.
- HUTNot a covered company, but an important recent AI hosting transaction case.
- Strengths
- The Beacon Point site has been fully commercialized, and its 704 IT MW of contracted capacity demonstrates improved credibility in large-scale AI hosting.
- Weaknesses
- The report does not provide a formal rating, and revenue per IT MW for the second-phase lease is lower than in WULF's recent transaction.
- Comparison
- HUT's repeat contracting proves its execution ability, but its revenue yield is lower than that of the WULF-Anthropic transaction.
- Risks
- Project delivery is phased from 2027 to 2028 and still faces construction, equipment, and financing execution risks.
- CLSKIts first AI hosting contract has been signed, and the report assigns an Outperform rating and a $24 target price.
- Strengths
- A 20-year triple-net lease, about $6.6 billion of contract revenue, and an NOI margin close to 100% improve long-term revenue visibility.
- Weaknesses
- As its first AI contract, pricing may reflect that its execution track record is still being established.
- Comparison
- Annual revenue of about $1.9 million per IT MW is lower than WULF's roughly $2.4 million, but the report believes there is room for improvement in economics as an execution track record is built.
- Risks
- First project delivery, project financing terms, subsequent conversion of the Texas portfolio, and construction cost control.
- WULFA beneficiary of large-scale AI hosting orders, rated Outperform by the report with a $36 target price.
- Strengths
- Signed an about $19 billion, 401 IT MW, 20-year contract with Anthropic; AI order book reaches about $27 billion; power assets cover Kentucky, Maryland, and New York.
- Weaknesses
- Specific margin and capex details of the Anthropic transaction have not yet been fully disclosed.
- Comparison
- WULF's deal term and revenue yield are among the best in the peer group, with a richer blue-chip customer mix.
- Risks
- Long-term construction pace, changes in regulation and grid policy, and project capex and financing execution.
- NVDAA beneficiary of the semiconductor ecosystem, rated Outperform by the report with a $315 target price.
- Strengths
- The data center opportunity is enormous and still in its early stages; semiconductor vendors can help customers find places to deploy compute capacity through credit, investment, or tenant relationships.
- Weaknesses
- Investors are concerned about "circular transactions" when semiconductor vendors participate in customer financing or lease support.
- Comparison
- Compared with neocloud providers bearing long-term leases alone, NVDA-related support can improve tenant credit quality and facilitate deployment of its chip ecosystem.
- Risks
- Market concerns over transaction structure transparency, customer concentration, and the sustainability of future demand.
- AMDA beneficiary of the semiconductor ecosystem, rated Outperform by the report with a $600 target price.
- Strengths
- Locks in large-scale AI compute deployment locations through the CORZ partnership and can support growth in CPU and GPU demand.
- Weaknesses
- The cooperation includes credit support and warrant arrangements, which may raise market concerns about structural complexity and capital usage.
- Comparison
- In the CORZ transaction, AMD is participating more deeply than traditional chip suppliers in the issue of infrastructure availability.
- Risks
- AI demand expectations are high; if customer deployments slow or project expansions fall short of expectations, the related ecosystem benefits may be lower than expected.
- CRWVThe report assigns an Underperform rating and a $67 target price.
- Strengths
- As a participant in GPUaaS and AI compute leasing, it remains on the growth track of AI infrastructure demand.
- Weaknesses
- The report believes its fundamentals are weaker and that it may struggle to remain competitive as the GPUaaS space evolves and traditional power assets are leased out.
- Comparison
- Compared with miners that own power and data center assets, the report assesses CRWV's competitive position more cautiously.
- Risks
- Intensifying GPUaaS competition, cost structure pressure, asset lock-in, and demand volatility.
Key data
- AI power capacity contracted by minersAbout 8 GWOver the past two years, the power capacity contracted by miners with hyperscalers, neocloud providers, and AI chip vendors.
- AI contract revenue related to minersMore than $160 billionThe scale of contract revenue corresponding to more than 20 transactions.
- Planned power portfolio of emerging AI infrastructureAbout 30 GWThe report believes this planned portfolio supports solving the problem of "time to deliver compute capacity."
- Initial CORZ-AMD cooperation capacityAbout 530 IT MWDistributed across 5 Core Scientific sites, of which 377 IT MW is leased directly to AMD, with another 152 IT MW leased to an undisclosed neocloud provider and credit-supported by AMD.
- Initial CORZ-AMD contract revenueAbout $14 billion, over 15 yearsAverage annual revenue of about $900 million, equivalent to about $1.8 million per IT MW per year.
- Potential CORZ-AMD expansion capacityAn additional about 1.9 IT GW, for a cumulative total of about 2.5 GWAMD has an option to contract additional expansion capacity.
- CORZ project capex guidanceAbout $11 million to $12 million per IT MWTotal capex for the initial 530 IT MW project is about $6 billion, of which about $1 billion has already been invested.
- CORZ capacity delivered to CoreWeave437 IT MWAs of mid-July, delivered capacity corresponds to about $635 million in average annualized hosting revenue; the full 590 IT MW is expected to be completed by early 2027.
- HUT Beacon Point contract size704 IT MW; second lease about $9.8 billionThe second-phase 352 IT MW lease doubles the site's contracted capacity to 704 IT MW, with an initial base term of 15 years.
- CLSK first AI hosting contract175 IT MW; about $6.6 billion; 20 yearsAt the Sandersville site in Georgia, with the tenant being an undisclosed high-investment-grade global technology company.
- WULF-Anthropic contract401 IT MW; about $19 billion; 20 yearsAt the Justified Data campus in Kentucky, with two additional 5-year renewal options.
- Estimated unlevered IRR range for hosting dealsAbout 8% to 13%The report believes returns on large-scale deals are moderate, but triple-net leasing and lower retrofit capex can improve project IRR.
Impact & implications
For investors, the implication of the report is that the AI infrastructure value chain is expanding from pure chip supply to power, grid connection, land, and on-time delivery capability. If miners can convert existing power and data center assets into long-term investment-grade leases, their valuation logic may shift from crypto-cycle exposure to infrastructure-like cash flows; semiconductor vendors, meanwhile, are reinforcing chip demand and ecosystem stickiness by supporting customers in deploying compute capacity. However, this thesis depends on projects being delivered on time, financing costs remaining controllable, tenant credit being reliable, and long-term AI compute demand continuing.
Risks
- Large data center projects may experience construction delays, equipment delivery delays, or grid-connection progress falling short of expectations.
- Capex per IT MW is relatively high; if project financing costs rise or loan-to-value ratios decline, equity returns may be compressed.
- Long-term leases of 15 to 20 years improve revenue visibility, but also increase dependence on tenant credit, contract terms, and long-term AI compute demand.
- Semiconductor vendors' participation in credit support, warrants, or customer deployment arrangements may be interpreted by the market as circular transactions or opaque demand quality.
- There is uncertainty around power regulation, state-level grid policy, natural gas supply, and behind-the-meter development.
- If AI compute demand slows or competition in the GPUaaS business model worsens, pricing and renewal capacity for related hosting capacity may come under pressure.
- Some transaction details in the report come from company disclosures and media reports, and certain tenant names or economic terms have not yet been fully disclosed.
What to watch
- CORZ's actual delivery progress for AMD and CoreWeave capacity in 2027 to 2028.
- Whether AMD exercises its additional expansion option of about 1.9 IT GW, and the related financing and lease terms.
- The initial energization, equipment procurement, and delivery pace of HUT Beacon Point Phase 1 and Phase 2.
- Project financing details for CLSK's first AI hosting contract and whether the Texas portfolio converts into formal contracts.
- Disclosure of capex, margin, and financing support details for the WULF-Anthropic contract.
- Whether more semiconductor vendors or hyperscalers shift to direct master leases rather than participating only through neocloud providers or guarantee structures.
- Whether rent per IT MW, NOI margins, unlevered IRR, and project financing costs for emerging AI infrastructure companies continue to improve.
- The market's acceptance of semiconductor companies such as NVDA and AMD supporting customers' compute deployment, and whether stricter disclosure requirements emerge.