Two days after Nvidia enlisted Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to underwrite more than $500B in AI compute financing, Ornn Exchange put a number on what those banks are actually pricing: a B300 forward curve running out 60 months, built from real transactions across its Nvidia cloud network rather than the spreadsheets and MOUs financiers have relied on until now. The distinction matters because a forward curve — unlike a one-off lease valuation — lets a lender mark GPU collateral to a price that updates daily, the same infrastructure that underpins financing for oil, freight, and power. Compute gets closer to a fully liquid asset class the moment enough of these curves exist to arbitrage against each other. (Ornn Exchange)
Private Companies
Vantage Data Centersfundraise
Vantage Data Centers, backed by DigitalBridge and Silver Lake, is exploring an IPO at a roughly $100B valuation or an outright sale, with preliminary discussions underway and financial advisers approached informally. A listing near that size would be the largest data center IPO on record, and the fact that a private operator can credibly float that number shows how far AI-driven demand has pulled data center valuations from anything comparable before 2023. No formal process has launched, and Vantage may still opt for a stake sale instead. (Reuters)
Public Markets
Q4 FY2026 revenue hit $11.1B, up 93% year over year. Guidance of $65B-$72B for FY2027 blew past the $54.4B Street consensus — evidence the AI-server upgrade cycle is accelerating into next year, not leveling off. (Yahoo Finance)
Q4 FY2026 revenue more than doubled to $1.01B, with non-GAAP gross margin above 50%. Guidance implies another 130% jump next quarter as 1.6T transceiver production ramps alongside continuing 200G shipments — the tightest constraint in the AI buildout right now is optical, not GPUs. (24/7 Wall St.)
Emerging
Analog inference, minus the ADC tax: Analog compute-in-memory chips have been bottlenecked by their own analog-to-digital converters, which eat most of the energy budget. A new architecture, Charge-CIM, folds ADC, multiply-accumulate, and readout into a single switched-capacitor fabric, cutting ADC energy by 91.7% and roughly doubling throughput over the best published charge-domain CIM accelerators. It's a paper-stage result, not silicon, but it targets the specific bottleneck that has kept analog inference chips from displacing digital accelerators at the edge. (arXiv)
The GPU that won't retire: On its Q2 call — where CoreWeave reported $2.58B in revenue, up 112% year over year, and a backlog that grew to $104B — CEO Mike Intrator disclosed a contract for 2020-era Nvidia A100 GPUs that runs through 2029, priced at full rate. Jensen Huang points to CUDA's backward compatibility as the reason: the same software stack that runs on Blackwell also runs unmodified on hardware that's nine years old, so an A100 rented out for inference or lower-priority training keeps earning long after a GPU with a shorter software shelf life would be scrapped. In a market gated by power, not chips, a fleet that stays revenue-generating for a decade changes the depreciation math the whole neocloud sector is built on. (Tom's Hardware)