Data center expert pulse: financing, debates, insights

Experts see compute still scarce and pricing firm, but credit has repriced and post-midterm policy risk is now the swing factor for 2027 builds.

Pulse metrics
High-yield DC funding cost
+300bp in 4-5 months
Soluna CFO, 3 Nov; LTC ~70% vs 80%  [1]
Power / capacity premium
5-10%
Ex-Meta DC finance head, 5 Nov  [2]
Rubin 5-yr contract (expected)
$6-7 / GPU-hr
Ex-GPU cloud VP, 4 Nov; ~$9 on 2-yr  [3]
Blackwell 3-yr contract
~$5 / GPU-hr
Ex-GPU cloud VP, 4 Nov  [3]
Build cost per MW
$12-14M
+$4M vs $8-10M prior
Ex-director, public mining firm, 29 Oct  [4]
AWS backlog
$578B
Q3 call; triple-digit YoY growth  [5]

As of 6 Nov 2026. Summary excludes channel checks. Expert calls 7 Oct-6 Nov 2026; earnings calls Oct-Nov 2026; news 28 Oct-6 Nov 2026. Excludes sell-side research.

Bottom line from the experts

Leading-edge capacity is still short: B200s are described as super scarce, and Rubin is expected to price above Blackwell on every contract term.  [3][3]

The constraint has moved to capital. Lenders now want an investment-grade wrapper, a long-term contract and only ~70% loan-to-cost.  [1]

Power has become a priced asset; a former Meta infrastructure-finance head says securing power and capacity can justify a 5-10% premium.  [2]

The bear view centers on overbuild: a former Lightning AI VP favors 5-10 MW distributed sites and flags risk in speculative gigawatt campuses and newer neoclouds.  [6]

What changed in the news (28 Oct-6 Nov)
  • Microsoft declined an option to more than double Nscale's Texas capacity and tightened oversight of Nebius and IREN projects.  [7]
  • Democrats flipped governorships in Texas, Ohio, Georgia, Nevada and Iowa, many of them data center hot spots.  [8]
  • 39 of 41 local data center measures produced outcomes favoring restriction.  [9]
  • ING is auctioning ~$500M of data center loans, including a Stargate-linked loan, with bids due 19 Nov.  [10]
  • Neocloud stocks fell more than 10% Thursday vs ~2-4% for hyperscalers.  [9]
What management said on Q3 calls
  • Amazon: AWS +39.3% YoY, sixth straight quarter of acceleration.  [5]
  • Microsoft: Q1 capex of $45.2B was held back by delayed third-party capacity deliveries.  [11]
  • Alphabet: capex expected to increase significantly in 2027.  [12]
  • IREN: capital markets more selective; questions now turn on delivery, financing and profitability.  [13]
  • Core Scientific: access to financing will increasingly set the pace of development.  [14]
Freshest expert reads (4-6 Nov)
  • 6 Nov, Eaton senior engineer: in backup power, who delivers fastest now matters more than product performance.  [15]
  • 6 Nov, Semtech principal engineer: 1.6T optics reach mass volume in 2H next year; CPO adoption stays limited.  [16]
  • 5 Nov, ex-Meta DC finance head: building ahead of demand is necessary; behind-the-meter power bridges grid delays.  [2]
  • 5 Nov, ex-Lightning AI VP: favors 5-10 MW distributed sites; flags overbuild risk at speculative GW campuses.  [6]
  • 5 Nov, AWS product director: Oracle's backlog is meaningful but not guaranteed cash flow, given customer concentration.  [17]
  • 4 Nov, ex-GPU cloud VP: financing needs 4-5 year contracts; customers want 3 years or less.  [3]
Expert call log, newest first
Expert call log, newest first
Published Expert perspective Theme Key view / datapoint
6 Nov Senior engineer, Eaton Power equipment Backup-power winners are decided by delivery speed and support on aggressive schedules; market big enough for multiple winners  [15]
6 Nov Manager, US hyperscaler Memory DRAM +15-20% in 3Q26; NAND ~+30% on enterprise SSD for AI data centers  [18]
6 Nov Principal engineer, Semtech Optics 800G stays key near term; 1.6T mass volume 2H next year; NPO from 2028; CPO limited  [16]
5 Nov Ex-head DC infrastructure finance, Meta Power / capex Securing power and capacity can justify a 5-10% premium; behind-the-meter bridges grid delays  [2]
5 Nov Ex-VP, Lightning AI Demand / overbuild Inference to reach 70/30 vs training within two years; overbuild risk in speculative GW campuses and newer neoclouds  [6]
5 Nov Director of AI, IT services provider Workload mix Inference and fine-tuning 60-70% of deployments; regulated enterprise half to two-thirds of the market  [19]
5 Nov Director PM, AWS Oracle backlog Backlog meaningful but not guaranteed cash flow; utilization, idle-capacity and chip-cost terms decide returns  [17]
5 Nov Account executive, IT reseller Neocloud share CoreWeave gaining share in specialized GPU cloud; related Q3 business +17.5-20% YoY  [20]
5 Nov Advisor, AI cloud provider Supply chain Critical electrical lead times up to five years vs 24-30 months pre-2020  [21]
5 Nov Ex-EMEA senior manager, Amazon Cooling / power Gas-engine chillers show ~35% better $/MW economics vs electric chillers  [22]
4 Nov Ex-VP GPU cloud, PaleBlueDot AI GPU pricing Blackwell ~$5/GPU-hr 3-yr; 20-30% down payment typical  [3]
4 Nov Ex-VP GPU cloud, PaleBlueDot AI Financing tenor Rubin $6-9/GPU-hr by term; 4-5 yr contracts needed vs customer preference of 3 yrs or less  [3]
4 Nov Director, KT Corporation Colocation pricing Korea metro colocation ~$100/kW (2019) to $170/kW (2025) per CBRE, higher in 2026  [23]
4 Nov Founder, CentraEdge Build cost Major rebuilds cost ~10-30% more per MW than greenfield  [24]
3 Nov CFO, Soluna Holdings Financing HY costs +300bp in 4-5 months; LTC ~70%; IG wrapper and long contract required  [1]
3 Nov Director energy policy, CyrusOne Power scarcity Developers turning to behind-the-meter, colocated generation and grid flexibility  [25]
3 Nov Executive, power generation company Behind-the-meter Data center BTM backlog nearly doubled vs last quarter  [26]
3 Nov Senior electrical engineer, DC company On-site generation Engine procurement +10-20% YoY; BESS demand +25-35%  [27]
2 Nov Head of energy procurement, Colt Cash-flow timing Operators struggle to match cash-flow timing against upfront investment  [28]
1 Nov Field CTO, Dell Rack density ~130 kW Blackwell to 300-400 kW Rubin Ultra; content per MW can be flat  [29]
30 Oct Senior director, ERock Bridge power Recip engines gaining share; upfront equipment payments to lock slots  [30]
29 Oct Ex-director, public mining firm Build cost Per-MW build cost up from $8-10M to $12-14M  [4]
29 Oct Ex-VP, Highrise AI Financing mix Leaders blend corporate equity, credit facilities and GPU financing; GPUs need different underwriting from real estate  [31]
28 Oct Ex-partnerships lead, Nebius Neocloud model Pure bare metal at risk from commoditization; scale and services decisive  [32]
26 Oct Director site development, Nightfall Lease pricing CleanSpark-Meta Sandersville ~$1.89M per IT MW per year, below some peers  [33]
14 Oct CRO, digital lending firm Credit tiers Hyperscalers strongest, leveraged AI cloud medium, frontier labs higher risk  [34]
Expert-quoted GPU contract pricing by term ($ per GPU-hour)

Range midpoints plotted where a range was quoted. Missing bars = not quoted.

Pricing read

Blackwell trades at ~$5/GPU-hr for three years, close to $6 for two and mid-$6-7 for one.  [3]

Rubin is expected above $6-7 on five-year terms, ~$8 on three-year and ~$9 on two-year.  [3]

GB300 on-demand runs double digits per GPU-hour at CoreWeave or Nebius.  [35]

Financing: what experts are seeing
  • Lenders want an IG wrapper, a long contract and ~70% loan-to-cost.  [1]
  • Many neocloud customers are not investment grade; idle capacity and customer failures create risk.  [3]
  • Applied Digital's ~5.5x leverage compares with ~3x bank-facility norms but should fall as leases stabilize.  [36]
Power: the binding constraint
  • Utility delays are pushing data centers to behind-the-meter generation, with gas engines and grid-forming BESS.  [27]
  • Power barriers are slowing projects, partly offset by behind-the-meter generation.  [37]
  • Hyperscalers drive most BTM demand, but AI labs and neoclouds are increasingly sourcing their own power.  [30]
Where experts agree
  • Leading-edge GPUs remain scarce; customers often want capacity sooner than it exists.  [3]
  • Workloads are shifting toward inference.  [19]
Where experts split
  • Bear: an ex-Lightning AI VP flags overbuild risk in speculative gigawatt campuses and newer neoclouds.  [6]
  • Bull: 2026 capacity coming online is up at least 15-20%, held back only by the power barrier.  [37]
  • Bear: hyperscaler and US OEM customers are trimming memory demand forecasts.  [38]
Debate scorecard: where the freshest expert evidence leans
Debate scorecard
Debate Expert lean Freshest supporting evidence
2027 demand vs overbuild Split, tilting cautious AI infra executive (4 Nov) base case -25% for 2027 deployments  [37]
Neocloud financing Cautious Soluna CFO (3 Nov): HY +300bp, LTC ~70%, IG wrapper required  [1]
GPU pricing durability Constructive Ex-GPU cloud VP (4 Nov): Rubin priced above Blackwell on every term  [3]
Power and politics Constructive for powered sites Ex-Meta DC finance head (5 Nov): 5-10% premium justified to secure power  [2]
Hyperscaler vs neocloud Favors scale players Ex-Nebius lead (28 Oct): bare metal faces commoditization; scale decisive  [32]
Inference shift Constructive IT services AI director (5 Nov): inference and fine-tuning 60-70% of deployments  [19]
1. Demand durability vs 2027 overbuild

Bull evidence

  • Ex-Meta DC finance head: building ahead of demand is necessary to avoid shortages.  [2]
  • AWS backlog of $578B is growing triple digits YoY.  [5]

Bear evidence

  • AI infra executive: 2027 deployments -25% base, ~-50% downside.  [37]
  • Ex-Lightning AI VP: meaningful overbuild risk in speculative GW campuses.  [6]
2. Can neoclouds fund the build?

Bull evidence

  • IREN: GPU financing plus prepayments fund more than 100% of Mackenzie GPU capex.  [13]
  • TeraWulf: Google credit support for financing its $19B Anthropic lease.  [39]

Bear evidence

  • Soluna CFO: HY costs +300bp in 4-5 months, LTC down to ~70%.  [1]
  • CoreWeave has $10.6B of debt due by end-2027.  [40]
3. GPU economics and useful life

Bull evidence

  • IREN re-contracted H100s at double the original price.  [13]
  • Ex-Lightning AI VP: long live the H100.  [6]

Bear evidence

  • Operators borrow against GPUs that lose value with each chip generation.  [41]
  • Ex-GPU cloud VP: providers need 4-5 year contracts to recover hardware; customers prefer 3 or less.  [3]
4. Politics and power access

Bull evidence

  • Core Scientific: Texas and Georgia uncertainty raises the value of operating, powered sites.  [14]
  • Power producer: behind-the-meter backlog nearly doubled vs last quarter.  [26]

Bear evidence

  • 39 of 41 local measures produced outcomes favoring restriction.  [9]
  • AI infra executive: European regulation is making it harder to build or energize data centers.  [37]
5. Hyperscaler self-build vs neocloud reliance

Bull evidence

  • Existing contracted capacity gains scarcity value as new builds get harder.  [9]
  • IT reseller: CoreWeave gaining share in specialized GPU cloud.  [20]

Bear evidence

  • Microsoft passed on doubling Nscale Texas capacity and tightened oversight of Nebius and IREN.  [7]
  • Ex-Nebius lead: bare-metal neocloud models face commoditization.  [32]
6. Inference shift and silicon mix

Bull evidence

  • Inference and fine-tuning already 60-70% at one integrator.  [19]
  • Inference expected to reach 70/30 vs training within two years.  [6]

Bear evidence

  • PCB vendor: ASICs likely to drive most inference on cost per token.  [42]
  • Inference neocloud General Compute is built on non-GPU silicon.  [41]
Disclosed contracted revenue / backlog by operator ($B)

Definitions differ: IREN = revenue backlog; TeraWulf = single Anthropic lease; Cipher = contracted revenue over ~20 years; DLR and IRM = signed-not-commenced backlog.

Contract book

IREN's backlog exceeds $20B, two-thirds with investment-grade customers.  [13]

TeraWulf's Justified campus carries ~$19B of contracted Anthropic revenue.  [39]

Cipher's three campuses represent $16.6B of contracted revenue.  [43]

Digital Realty's backlog reached a record $2.5B.  [44]

Iron Mountain lifted data center backlog to $575M.  [45]

Landscape by segment (latest calls and news)
Landscape by segment
Company Segment Latest datapoint
Amazon (AWS) Hyperscaler AWS +39.3% YoY; backlog $578B  [5]
Microsoft Hyperscaler Capex $45.2B, ~3/4 short-lived assets; finance leases $7.4B  [11]
Alphabet Hyperscaler Q3 capex $51.4B; FCF negative $9.1B  [12]
Meta Hyperscaler Capex incl. finance-lease principal $34.9B; FCF negative $2.1B  [46]
IREN Neocloud $22M/MW 5-yr contract; H100s re-priced at 2x  [13]
CoreWeave Neocloud $35.6B debt at 30 Jun; $10.6B due by end-2027  [40]
Core Scientific Converted miner / colocation Billing 485 MW for CoreWeave; >1.1 GW contracted incl. AMD  [14]
TeraWulf Converted miner / colocation Justified project at ~$10-12M per critical IT MW; RFS late 2027-2Q28  [39]
Cipher Digital Converted miner / colocation 700 gross MW nearing completion; three project financings  [43]
Applied Digital AI data center developer $1.59B 7% senior secured notes due 2031 at par  [47]
Galaxy (Helios) AI data center developer Exit-2027 leasing run-rate guided to $800-850M  [48]
Digital Realty Colocation REIT Occupancy 93.4%, churn 0.7%, re-leasing spreads ~12%  [44]
Equinix Colocation REIT Record annualized gross bookings $533M (+35%)  [49]
Vertiv Power and cooling Expects above-market organic growth in 2027 despite power and permitting constraints  [50]
Competitive shift toward scale

Microsoft is prioritizing its own data centers over new neocloud deals; Nebius and IREN fell >13% and CoreWeave 11%.  [7]

Gartner warned 80% of neoclouds could disappear by 2030.  [41]

Policy and tax overhang

Draft legislation would levy a gross-receipts excise tax on data center operators.  [51]

Ameren is allocating costs of new large loads to those customers under Senate Bill 4.  [52]

Credit market plumbing

An estimated $800B of data center financing sits outside public markets.  [53]

One ~$17.4B five-year GPU capacity contract only takes effect once the provider secures financing.  [54]

References