Hong Kong-based Tianfeng Securities’ overseas technology research team said its review of the latest quarterly earnings across the AI industry points to the return of an AI bull market, backed by what it described as strong underlying fundamentals.
The report said cloud growth has steepened again, demand for compute is running well ahead of supply, and customers are not only accepting price increases but, in some cases, prepaying build-out costs. Tianfeng said unit economics for AI infrastructure are improving across the board, and that the market narrative has shifted away from worries over excessive capital expenditure, or CapEx, and back toward confidence that demand and profitability can be delivered.
Cloud earnings seen as the clearest signal
Tianfeng said the most important signal this quarter was that new cloud contract bookings reached a 20-quarter high. Google, Amazon Web Services, or AWS, and Microsoft Azure each showed different strengths, but the report said they all support the same conclusion: the long-term growth case for cloud remains intact.
The firm described Google as an “all-in or out” survival bet. Even with negative free cash flow, or FCF, each additional $1 of cloud revenue can still translate into about $0.54 of operating profit, with a payback period of roughly two years, according to the report.
AWS, Tianfeng said, answered long-running questions around return on investment with a 39.4% margin. The report argued that AWS still has more than 5x long-term growth potential, could eventually reach $1 trillion in annual revenue, and has a CapEx payback period of less than three years.
For Microsoft Azure, Tianfeng said the reacceleration in growth has eased market concerns over whether demand can be sustained.
Neocloud providers show pricing power and prepayment demand
The report also focused on newer compute suppliers, or Neocloud players, especially CoreWeave (CRWV) and Nebius (NBIS).
Tianfeng said CoreWeave’s compute capacity is close to sold out, with multiple customers competing for GPU allocations. The company added about 500MW of effective capacity in a single quarter and raised its year-end target to more than 1.85GW. Its customer base has expanded beyond major AI model developers and hyperscalers to include industrial, financial, government and traditional enterprise clients, which the report said shows AI moving beyond a training-led frenzy and into wider industry deployment.
The report added that CoreWeave raised prices across its SKU lineup in July, lifting gross margin contribution on new contracts by 5 to 10 percentage points.
One of Tianfeng’s main arguments centered on older GPU assets. It said NVIDIA A100 chips launched in 2020 are still being signed into high-priced long-term contracts extending to 2029, directly challenging the bearish claim that GPU value would depreciate sharply or even fall to zero within three to five years.
The report also mentioned that CoreWeave issued strong earnings guidance, had a backlog of more than $100 billion, and saw its stock rise 15% in after-hours trading.
Nebius posted another set of figures Tianfeng described as strong. The company’s total value of new contracts in the second quarter grew by about 4x from the previous quarter, while the number of contracts from new customers rose by more than 9x. The report said annualized revenue per MW on new contracts reached $20 million to $25 million. About 70% of new contracts include prepayment terms, covering 50% to 60% of related CapEx. That, Tianfeng said, cuts the expected payback period from two to three years to one year and 10 months.
On that basis, the firm argued that compute infrastructure is no longer a business model defined by expansion and cash burn at the same time. It said the sector is now showing solid and mature unit economics.
The report also cited Nebius second-quarter revenue of $580 million, up 454% year over year, and a 34% rise in its share price.
Optical, memory and storage names seen nearing a bottom
In optical communications, Tianfeng said the impact of U.S. Federal Communications Commission, or FCC, policy appears limited. Lumentum (LITE) reported results above market expectations and raised its outlook. Demand for EML was about 30% above supply, while shortages of ultra-high-power lasers continued to widen.
For Innolight, the report said laser products are nearly sold out and the company plans to expand capacity by 4x over the coming quarters. The bottleneck, it added, is delivery capability rather than demand.
Tianfeng also addressed the debate over network architecture. The report said NPO is not a competing technology meant to replace CPO, but an additional intermediate architecture. Major customers have not changed their CPO deployment plans, and demand signals continue to strengthen, with AI system design moving toward deeper optical integration.
In memory and storage, Tianfeng said a bottom has almost certainly formed. Its core reasoning is “trading time for space”: the gap between falling share prices and delayed price increases has already been fully reflected in market expectations. Visibility on long-term agreements, or LTA, and shareholder return policies are also improving.
On the compute side, the firm said the market narrative is returning to AI training and open-source themes, and it kept a positive view on NVIDIA (NVDA) retesting its previous high.
Report says earnings season has chipped away at major bearish arguments
Tianfeng concluded that this earnings season has systematically dismantled the bearish arguments that weighed on sentiment over the past several quarters, including concerns about excessive CapEx, rapid GPU depreciation, unsustainable demand and difficulty quantifying returns.
The report said that when compute providers start receiving prepaid construction funding from customers, and when 2020-vintage GPUs can still secure premium contracts running through 2029, returns on AI infrastructure investment are not only real but arriving faster and at a larger scale than expected.