Oct 13, 2026 4:43 PM
The market is starting to price execution, not announcements.
The financing model has crossed a threshold. The five major U.S. hyperscalers are projected to spend more than their combined operating cash flow in 2026. Yet rising investment is no longer translating automatically into richer valuations or easier project debt. The more useful question is not whether the buildout continues, but which contracts, counterparties and assets deserve financing on yesterday's terms.
Backlog is becoming a diligence exercise. Nscale's $103 billion of contract value is dominated by two customers; only $2.6 billion was active at the end of August. Oracle's Project Jupiter notice raises a different issue: even where the tenant remains committed, the date on which cash starts flowing can become contested. The emerging discount is to conversion risk, not necessarily to long-run AI demand.
Opposition can redistribute value before it removes supply. A moratorium count is not a megawatt count. SemiAnalysis identifies much less directly delayed capacity than nominal exposure, while Pew finds a sharp deterioration in public attitudes. Both can be true. Existing approvals may become more valuable even as future phases become harder and more expensive to secure.
Agentic AI is broadening the infrastructure market. Akamai's seven-year, $11.6 billion Anthropic agreement explicitly supports CPU workloads. General Compute is assembling an inference service that routes work across more than one chip architecture. The opportunity is expanding beyond the largest GPU cluster toward the combination of accelerators, ordinary compute, networking and software that completes useful work.
Related stories
Reliability is moving ahead of the ribbon-cutting. SpaceX's reported shift toward earlier redundancy and testing, Nvidia's DSX Ready program and factory-certified power systems point toward the same commercial standard: capacity must pass acceptance and remain usable. Fast construction and fast revenue are not the same metric.
The cheapest electricity need not create the best energy investment. Falling ERCOT battery revenue alongside stronger grid flexibility exposes a split between system value and asset returns. Google's proposed nuclear-uprate subscription offers a different model: pay explicitly for incremental attributes and allocate benefits through a tariff. The investment edge lies in capturing value, not merely creating it.
Bottom line. The next phase favors platforms that can convert commitments into accepted capacity, distinguish customer diversification from contract proliferation, and retain economics after financing, reliability and community obligations. This is a differentiation cycle within the boom, not a simple verdict for or against it.
NUMBERS OF THE WEEK | CHARTS & KEY INDICATORS
CAPITAL: THE SELF-FUNDING MODEL BREAKS
A growing capital program meets a widening gap between investment, internal cash generation and ultimate end-market revenue.
NUMBERS OF THE WEEK | CHARTS & KEY INDICATORS
CONTRACTS: SIZE IS NOT THE SAME AS QUALITY
Headline commitments conceal concentration, conditions precedent and a long interval between signing and cash generation.
NUMBERS OF THE WEEK | CHARTS & KEY INDICATORS
PERMISSION: MEASURE THE PROJECT, NOT THE HEADLINE
The public backlash is real. Its immediate capacity impact and its effect on the next development cycle are different questions.
NUMBERS OF THE WEEK | CHARTS & KEY INDICATORS
ECONOMICS: SYSTEM VALUE IS NOT ASSET RETURN
Better infrastructure can lower customer costs while squeezing the merchant owner. The contract determines who captures the benefit.
WEEKLY SPOTLIGHT | 01
BACKLOG IS NOT COLLATERAL. DELIVERY IS.
The credit market is distinguishing the promise of future compute from the enforceable right to receive cash.
The critical distinction in this cycle is between a large customer and a financeable obligation. A recognizable tenant can support a strong credit profile, but it does not remove construction risk, acceptance conditions, termination rights or the possibility that rent starts later than the debt model assumes.
Nscale makes the conversion problem visible. Its September filing, as reported by Bloomberg, showed $103B in total contract value, 85% attributable to Microsoft and Anthropic, and $2.6B active at August-end. The Anthropic agreement carries delivery and service conditions, and binding financing commitments had not been obtained at the filing date. The later $3.36B corporate financing announcement does not, by itself, demonstrate full funding of that individual project.
TIMING RISK CAN SURVIVE A STRONG TENANT
Oracle's Project Jupiter notice is not evidence that the tenant has defaulted or abandoned the campus. Bloomberg reports an effort to preserve rights around potential timing of rent. Oracle said the project remains on its planned schedule; Blue Owl said the notice does not change its financial commitments. The analytical issue is nevertheless material: who finances the interval between scheduled completion and an accepted, revenue-producing facility?
THE MARKET IS STILL OPEN, BUT AT A PRICE
The Information reports wider discounts and tighter protections on recent project debt, while also noting that no major active syndications had been pulled. This is repricing, not a funding shutdown. Hut 8's parent revolver offers a useful contrast: committed liquidity and a letter-of-credit sublimit can fund deposits and development before long-term project financing is optimized.
WHAT DESERVES THE PREMIUM
Underwrite the exact obligor, the acceptance test, the funded completion plan and the path through a delay. Look through separate project companies to common end demand. A diversified lender roster cannot diversify a cash-flow base ultimately dependent on the same two model developers.
BOTTOM LINE: The relevant multiple is on durable, converted cash flow. Headline backlog deserves a schedule- and contract-specific discount, not a uniform sector multiple.
WEEKLY SPOTLIGHT | 02
THE MORATORIUM PREMIUM
Restrictions may enrich already-entitled assets before they materially reduce near-term national supply.
The apparently contradictory evidence is the investment signal. Data Center Watch reports 45 projects worth $68B blocked or delayed in the second quarter. SemiAnalysis maps roughly 20 GW inside local restricted boundaries but identifies only 1.525 GW directly delayed by those restrictions. The studies measure different things: broad opposition in project dollars versus identified moratorium effects in megawatts. They should not be forced into a single cancellation statistic.
That distinction creates a potential scarcity premium for existing approvals. A parcel that can proceed while competing applications pause may acquire additional negotiating value. But the premium belongs to the specific entitlement, phase and schedule, not automatically to every acre inside an established campus. Expansion rights, air permits, water and fuel corridors can remain exposed.
THE BACKLASH IS ABOUT MORE THAN POWER BILLS
Pew's largest deterioration is nearby quality of life: a 19-point increase in negative assessments. Electricity-cost concern increased 12 points. A tariff response alone therefore cannot settle issues involving noise, visual impact, diesel or gas operation, water and local control. Amazon's community spending and transparency commitments recognize that broader bargain.
A HIGHER BAR IS NOT A UNIVERSAL BAN
Prince William County reduced its by-right overlay rather than prohibiting all development. Nevada's binding new conditions apply to applicants for state tax breaks. New Zealand's announced approach is a party policy proposal. These distinctions determine actual exposure. The investor's task is to price jurisdiction-specific permissions, not simply rank states by political label.
THE CAUSALITY CHECK
NAR finds higher home values in data-center-heavy counties but explicitly warns that correlation is not causation. County wealth is not proof that the house next to a campus benefits. Equally, opposition is not proof that every project damages housing. Credible local evidence has become part of the development budget.
BOTTOM LINE: Near-term delivery can remain strong while the price of the next permit rises. That combination favors execution-ready assets, but only when the new scarcity rent is not fully surrendered in costs or concessions.
WEEKLY SPOTLIGHT | 03
AGENTIC AI IS UNBUNDLING THE FACTORY
The next unit of infrastructure demand is a completed workflow, not a GPU in isolation.
Akamai's Anthropic agreement is the clearest example of an AI growth market hiding in plain sight. The company explicitly describes CPU workload demand, not a purchase of GPU capacity. The seven-year base commitment is $11.6B, supported by an estimated $5.5B of capex. Anthropic also receives expansion-linked warrants. It is a large opportunity for conventional compute, but neither the contract multiple nor the warrant package establishes the project's return.
The economic logic is broader than this one transaction. An agent reasons with a model, calls tools, runs code, reads and writes data, and coordinates additional tasks. That workflow needs accelerators, but it also needs CPU instances, memory, storage, security and network transport. A campus or cloud designed around one dominant chip need not capture every dollar of the resulting demand.
MULTI-SILICON IS A SERVICE MODEL, NOT JUST A CHIP CONTEST
General Compute plans to finance Cerebras systems and sell inference, while using Nvidia for prefill and other architectures across its platform. The provider absorbs procurement complexity so the customer can buy a usable service. This may increase the value of integration and workload routing even where individual processors become more substitutable.
CLOUD DIVERSIFICATION MAY STILL BE CUSTOMER CONCENTRATION
Anthropic's reported discussions with Apollo-backed Stream would give the model developer more direct control over infrastructure. SpaceXAI's shift toward external cloud customers is another change in the division of labor. Yet the same few labs can sit behind several apparently separate landlords, neoclouds and financing vehicles. Count independent end demand, not just logos.
CAPACITY IS NOT AUTOMATICALLY PAID UTILIZATION
Epoch estimates that hardware shipped through 2027 could support 30-170 million concurrent frontier-model agents once deployed and allocated to those workloads. Its working-hour comparisons are not equivalent to human productivity or jobs replaced. At 20% utilization of the central capacity case, modeled API-equivalent spending reaches $2.6T-$5.3T annually: a demand hurdle, not contracted revenue.
BOTTOM LINE: The investable opportunity is a broader infrastructure mix. The underwriting error would be to treat every potential agent-hour as a paying customer, or every new provider as a new source of demand.
WEEKLY SPOTLIGHT | 04
STANDARDIZATION IS BECOMING A RISK PRODUCT
Factory certification, reference designs and integrated delivery matter because they can reduce the variance around acceptance, not merely shorten an advertised schedule.
The industry's definition of speed is changing. The Information reports that SpaceX's new data-center leadership wants redundancy and more thorough testing in place before operation, rather than completing backup systems afterward. That choice can slow first energization while improving the date at which capacity is consistently useful to third-party customers. The two milestones should never have been treated as interchangeable.
Several supplier announcements address the same problem from the other direction. Nvidia's DSX Ready program starts with qualifying battery storage and cooling distribution products against applicable reference-design requirements. Compass describes UL 2755-certified prefabricated power centers. Wartsila, Schneider and Stanley are coordinating generation, electrical systems, controls and execution. These are attempts to move uncertainty upstream, where repeated designs and tests can resolve it.
THE MARGIN OPPORTUNITY IS IN THE INTERFACES
A repeatable module is valuable when it removes engineering hours, rework or commissioning uncertainty at each deployment. The more attractive commercial position may sit with the party accountable for integration rather than the cheapest component supplier. An alliance announcement, however, is not evidence of a single wrapped performance guarantee.
KEEP CLAIMED SAVINGS ON THEIR OWN BASIS
Schneider's announced 2-2.5 MW prefabricated power products can have deployment lead times as short as six months in the stated European offering. That is not a six-month U.S. campus schedule. Bloom's DC-power report claims 27% lower non-compute capex and 9% lower five-year total cost in its scenario; those are vendor-model results, not independently proven universal savings.
AN ORDER BOOK CAN BE REAL AND STILL CHANGE
Crusoe and Boom ended their $1.25B launch partnership. Boom said the turbines are no longer part of Crusoe's near-term primary power mix at Abilene, and Crusoe said it remains flexible on sourcing; the cancellation does not prove that the turbines are technically unsuitable. It does show why manufacturing announcements, binding orders, delivered equipment and an accepted power system must remain separate milestones.
BOTTOM LINE: Time to dependable revenue is the better KPI. Standardization can earn a premium by shrinking the range of possible delivery outcomes, provided accountability extends beyond the factory gate.
WEEKLY SPOTLIGHT | 05
THE BATTERY PARADOX: BETTER SYSTEM, TOUGHER ASSET
Grid resilience and merchant returns can move in opposite directions. Contract design decides which side of that divergence an investor owns.
Amperical's ERCOT work puts the tension in numbers: modeled 2026 year-to-date battery revenue averages about $1.9/kW-month, compared with $4.63 in 2024. Meanwhile, Bloomberg documents much faster battery deployment in Texas and California than in PJM and sharply different summer wholesale-price outcomes. Those observations are consistent with flexibility reducing scarcity rents; they are not a controlled estimate of batteries' causal contribution.
The practical consequence is that a technology can become more important to the power system while its standalone merchant economics deteriorate. More storage can reduce the volatility from which storage earns money. For a data-center customer, that may be a benefit. For the merchant asset owner, it raises the importance of location, dispatch, operating cost and contracted compensation.
THE NEW PRODUCT IS A DEFINED SERVICE
A battery supplying ride-through, a battery earning ancillary-service revenue and a long-duration store firming renewable output are not interchangeable investments. Each has a different duty cycle, availability obligation and revenue stack. Avoid underwriting data-center resilience with a merchant-price forecast that assumes the same capacity is always free to trade.
GOOGLE'S UPRATE DEAL POINTS TO A DIFFERENT BARGAIN
Georgia Power proposes 96 MW of additions to its ownership share at Vogtle and Hatch. Google would subscribe to NU-1 and receive the associated zero-emission credits. The utility estimates $900M in customer benefits over the units' lives. These are projected benefits, not a $900M payment by Google. The Vogtle uprates were approved in the 2025 IRP; the Hatch uprates and the NU-1 tariff still require PSC approval.
VALUE CREATION AND VALUE CAPTURE MUST BE SEPARATED
The uprate approach pays for a specific incremental product within a regulated allocation of costs and benefits. It is analytically different from diverting existing output to a new load or buying energy certificates without additional production. The same discipline applies to onsite generation: energy, dependable capacity, schedule insurance and environmental attributes should not be collapsed into one price.
BOTTOM LINE: The best system solution is not automatically the best merchant asset. Prefer underwriting that identifies the service sold, the counterparty paying for it and the operational rights surrendered in return.
WEEKLY SPOTLIGHT | 06
CHINA'S WARNING: EMPTY SPACE IS NOT SPARE AI CAPACITY
National capacity totals obscure the mismatch between legacy buildings and the workloads willing to pay for new ones.
SemiAnalysis's new China model describes a market with substantial legacy vacancy and strong wholesale AI demand at the same time. Those are not contradictory conditions. Low-density retail stock can be poorly suited to the power, cooling, network and physical requirements of a modern AI deployment. The unit of supply is a qualified facility for a specific workload, not an undifferentiated megawatt.
The order mix is striking: GDS and VNET together signed about 1.3 GW of wholesale orders in the first half of 2026, against less than 10 MW of retail orders. The report also estimates that private ByteDance occupies roughly one-fifth of delivered Chinese capacity. Public-company disclosures therefore reveal only part of both the landlord market and the underlying tenant concentration.
THE U.S. READ-THROUGH IS A RESIDUAL-VALUE QUESTION
Long-life land and buildings do not guarantee a long economic life for the electrical and cooling configuration inside them. Van Nieuwerburgh's Brookings financing paper estimates that IT equipment represents roughly two-thirds of a new AI campus's initial capital. When lenders assign a residual value, the replacement tenant must be able to use the asset economically, not merely occupy it.
GEOGRAPHIC EXPANSION DOES NOT MAKE COMPUTE FUNGIBLE
Alibaba is adding new regions in Turkey, Finland and the Netherlands and expanding existing capacity in the Gulf, while Palantir and Armada emphasize customer-controlled sovereign infrastructure. Jurisdiction, chip access, latency, data control and physical resilience can outweigh a lower power price. Capacity in one country is not an automatic substitute for capacity in another.
SECURITY IS NOW AN INFRASTRUCTURE CHARACTERISTIC
CNBC's September report that AWS still could not restore affected Gulf facilities after the March attacks is an unusually severe reminder that redundancy within a campus does not eliminate regional risk. Distributed sovereign systems may address some requirements, but they introduce their own network, operational and scale trade-offs. Neither centralization nor modularity is a universal answer.
BOTTOM LINE: The next oversupply debate should distinguish obsolete capacity from qualified capacity. Hardware-refresh flexibility, tenant portability and jurisdictional resilience deserve explicit value in the asset model.
ABOUT DIGITAL POWER DIGEST
Digital Power Digest is Keel Infrastructure’s institutional intelligence briefing on the intersection of power markets and digital infrastructure - interconnection, generation, gas supply, capital markets, equipment, regulation and the physical AI stack - edited by Chris Ruppel, SVP - Power. Analysis reflects public information and the editor’s judgment as of the issue date. For informational purposes only; not investment advice.
Comments
Powered by RoundtableAn AI/DeFi-powered Enterprise Media Operating System. Learn more at Roundtable.io.© Roundtable 2026. By using this site you agree to the Terms of Use and Privacy Policy


