How much of their construction plans can large financing deals cover?

  • META plans to issue a total of 95 billion dollars in common stock, Class C shares and preferred stock, of which it is currently raising 35 billion dollars. SPCX says it has raised 40 billion dollars in debt led by Apollo; under the proposed finance lease arrangement, leases would commence separately as batches of equipment are delivered, with terms of approximately 5 years to 5 years and 3 months.
  • Existing financing still falls short of subsequent construction needs. In a scenario where SPCX builds 10 GW in 2027, Deutsche Bank estimates a funding gap of approximately 175 billion dollars and expects it to be financed through investment-grade and GPU-backed debt. Its assessment that financing is manageable depends on those subsequent channels.
  • Nvidia’s support comes from cash flow. Barclays estimates its customer-related off-balance-sheet commitments at approximately 250 billion dollars, mainly capped residual-value guarantees. Under the bank’s model, existing cash plus free cash flow over the next 12 months, after buybacks and dividends, can cover 125% of those commitments.

Why might high capital expenditure not depress cloud business margins?

  • UBS expects Amazon’s capital expenditure in 2027 could reach 250 billion–300 billion dollars, but expects AWS revenue to increase by approximately 80 billion dollars, with depreciation and amortization rising by only about 20 billion dollars. Operating leverage could offset lower-margin AI workloads. The bank therefore expects an AWS operating margin of approximately 46% that year, above the market consensus of 38%.
  • For Microsoft, the hurdle is delivering growth. Oppenheimer expects Azure’s constant-currency growth in the first fiscal quarter to be 46%, below the buy-side expectation of approximately 47%, and to approach 50% in the second fiscal quarter as new capacity comes online. Truist separately expects Microsoft to be the only company among the “Big Six” it covers to maintain positive free cash flow in calendar 2027, leaving room for slower AI monetization.

Why can an increase in orders coincide with a target-price cut?

  • Wells Fargo cut its Oracle price target from 280 dollars to 250 dollars: remaining performance obligations rose from 455 billion dollars to 664 billion dollars, but approximately 105 billion dollars of that involves customer-provided hardware or prepayments, with revenue economics different from traditional cloud contracts; higher GPU and component prices also affect returns on older contracts.
  • Demand has likewise not removed DigitalOcean’s funding hurdle. UBS says its pre-leasing lead time has exceeded one year, but estimates that by mid-2029 it will still need an additional 4 billion dollars in capital for an expansion of approximately 300 MW. It therefore maintained its neutral rating and cut its price target from 125 dollars to 105 dollars.

Why can a reduction in memory specifications still support a bullish view?

  • Rubin’s published HBM4 bandwidth was reduced from 22 TB/s to 19.2 TB/s. After meeting Micron management, Bank of America concluded that the specification cut reflects supply constraints and that memory still determines token output and energy efficiency at a given unit cost. New 3–5-year take-or-pay agreements have also improved demand visibility.
  • The technical constraints have concrete implications: according to ZDNet Korea, 12-layer stacks have encountered heat dissipation and yield problems. Samsung and SK Hynix plan mass-production shipments of 8-layer HBM4 in the second quarter of 2027; separate certification is still required, and the timing is later than the previously expected second half of 2026.
  • The shortage also has a time horizon. The Chosun Business Daily, citing industry forecasts, says current investment is going first into cleanrooms and HBM capacity, with significant new supply potentially appearing in 2028; capacity expansion could drive price normalization from the end of 2027 to early 2028.

Can growing model usage translate directly into customer retention?

  • OpenRouter data show that after Opus 5.5 was released, the 7-day average token usage of Anthropic models rose to 3.6 times its previous level. But the platform leans toward price-sensitive developers and small teams, giving it a different user mix from large enterprises.
  • Price-driven migration is occurring among enterprise customers at the same time. According to The Information, OpenAI offers 30%–40% discounts and free credits; Anthropic/OpenAI’s shares of coding usage at Rippling changed from 57%/16% in May to 18%/52%. Anthropic requires higher spending commitments to extend discounts and is losing some customers, so product popularity and customer retention are moving differently.